'UUC[UQP'FWECVKQPCN%JQKEGU CPF+PVGITCVKQP individuals...

Essays on Educational Choicesand Integration Elisabet Olme

Elisabet Olm

e Essays on Edu

cational C

hoices an

d Integration

Dissertations in Economics 2019:1

Doctoral Thesis in Economics at Stockholm University, Sweden 2019

Department of Economics

ISBN 978-91-7797-628-8ISSN 1404-3491

Elisabet OlmeElisabet holds a B.Sc. and M.Sc. inEconomics from StockholmUniversity.

This thesis consists of four self-contained essays exploring theimplications of educational policy for school segregation andindividuals' responses to integration and migration policy.

Essays on Educational Choices and IntegrationElisabet Olme

Academic dissertation for the Degree of Doctor of Philosophy in Economics at StockholmUniversity to be publicly defended on Wednesday 29 May 2019 at 13.00 in Nordenskiöldsalen,Geovetenskapens hus, Svante Arrhenius väg 12.

AbstractAre Parents Uninformed? The Impact of School Performance Information on School Choices and SchoolAssignments. We study the impact of providing information about schools' performances on households' choice of school.A randomly selected subset of households with children about to start middle school in a Swedish municipality wereprovided with information about the schools’ performances on standardized tests. We find that this information made themmore likely to apply to the top-performing schools, compared to households in the control group. The effect is driven bynative children and children to high-skilled parents. Next, we simulate how this would affect the allocation of students toschools, under the assumption that all households would have access to this information. As expected, enrollment in top-performing schools increase, but the effect is muted by the schools' capacity constraints. Again, native and high-skilleddrive the effect, by shifting their applications from mid- to top-performing schools. This leads to reduced school segregationby foreign background as children with a foreign background are overrepresented at the top-performing schools to beginwith. Furthermore, school segregation by parental education increases slightly as children with highly educated parentscongregate at the top-performing schools.School Choice, Admission Rules and Segregation in Primary School. We study the impact on primary schoolsegregation of three different admission criteria, in a school choice program using deferred acceptance to allocate studentsto schools. Using Swedish administrative register data, we simulate the allocation of students using proximity to the school,a lottery and affirmative action to determine students' priorities to oversubscribed schools. To predict the application listsunder counterfactual admission schemes, households’ preferences for schools are estimated using administrative schoolchoice data from a Swedish municipality. The results suggest that school segregation by family background decreaseswhen lottery-based priorities or affirmative action is used, compared to proximity-based priorities. When proximity-basedpriorities are used, the allocation of students to schools widely resembles the allocation that results from using schoolcatchment zones (ignoring parents’ preferences and base admission solely on students’ residential locations). The cost ofabandoning proximity-based priorities in favor of lotteries or affirmative action is modest; children are assigned to schoolsthat are highly ranked on their application lists under each admission scheme.Should I Stay or Must I Go? Temporary Refugee Protection and Labor-Market Outcomes. We study the impactof a prolonged period of temporary protection before being eligible to apply for permanent residency on human capitalinvestments and labor-market outcomes of refugees. In 2002, Denmark prolonged the period with temporary protectionfrom three to seven years, for refugees. The reform was implemented retroactively, allowing the effects to be estimatedusing a regression discontinuity design. Furthermore, we set up a theoretical search and matching model in order tounderstand the mechanisms at work. The empirical results show that enrollment in education increased, following thereform. Females and low-skilled individuals drive the effect. This is in line with the theoretical predictions from the model.We do not estimate any significant effects on labor-market outcomes for the full sample.The Effects of Performance Based Bonuses in the Swedish Language-Training Program. I study the impact of theintroduction of performance-based monetary bonuses in the Swedish language-training program for immigrants. Under thebonus system, a passing grade in certain courses was rewarded with up to 12,000 SEK (approximately 1,666 USD). Thebonus system was introduced in 2010, and eligibility was completely determined by the date of immigration to Swedenand the type of residence permit held. Using a sharp regression discontinuity design, the effects on enrollment and coursecompletion are estimated. The results do not indicate any impact on the enrollment rate, which might be explained bythe initially high enrollment rate of 80 percent. Furthermore, the prospect of receiving a bonus does not seem to increaseoverall course completion. The effects are also estimated for the bonus-qualifying courses separately, but the results areinconclusive. A positive effect of a few percentage points cannot be excluded, but the results are sensitive to the regressionspecification.

Keywords: School Choice, School Segregation, Information, Refugees, Human Capital, Immigration, LanguageTraining, Performance Bonuses.

Stockholm 2019http://urn.kb.se/resolve?urn=urn:nbn:se:su:diva-167838

ISBN 978-91-7797-628-8ISBN 978-91-7797-629-5ISSN 1404-3491

Department of Economics

Stockholm University, 106 91 Stockholm

ESSAYS ON EDUCATIONAL CHOICES AND INTEGRATION

Elisabet Olme

Essays on Educational Choicesand Integration

Elisabet Olme

©Elisabet Olme, Stockholm University 2019 ISBN print 978-91-7797-628-8ISBN PDF 978-91-7797-629-5ISSN 1404-3491 Cover picture: Tallbackaskolan in Solna. Photo: Gustav Trodin. Back-cover photo by Gustav Trodin. Printed in Sweden by Universitetsservice US-AB, Stockholm 2019

Till Mariann.

Acknowledgments

First and foremost, I would like to thank my two supervisors, JonasVlachos and David Seim, for continuously supporting me along thewinding path towards the completion of this thesis. Jonas gave methe courage to explore my own ideas, while providing firm guidancealong the way. His reading and commenting of my work has steeredmy research projects into more productive directions. In addition tobeing an excellent supervisor, he has also formed my views on how toconduct research and the role of academia in a broader context. FromDavid, I have learned a lot about being a researcher. He has helpedme improve my empirical skills and taught me how to navigate theacademic world. I am also very grateful for his support during thedemanding job market period.

I could not write an acknowledgement without mentioning DanyKessel. In addition to being a close friend, he is also the co-authorof the first two chapters of this thesis. I cannot imagine what theseyears as a PhD student would have been like, had he not been a partof it. Working alongside Dany, with his sharp and creative mind, hascertainly been intense, but never boring. We have shared coursework,research projects, travels, and many memorable moments outside ofacademia. I am very happy that I got to do all of this with you!

Another important person during these years is Matilda Kilström,one of the co-authors of the third chapter of this thesis. We got toknow each other during the master’s program, and since then, I amover and over again struck by her ambition and competence. Anyonewould find it hard to compete with her level of productivity, unlessthere is a baby goat to visit nearby; in which case she will put workaside immediately. I consider myself lucky to have gotten to work ona joint project with Matilda. But most of all, I am grateful that myinterest in economics also brought with it a dear new friend.

Several members of faculty at the Department of Economics atStockholm University, and elsewhere, have also contributed to this

thesis. Peter Fredriksson, Hans Grönqvist, and Eskil Wadensjö wereall important for my decision to apply to the PhD program in thefirst place. They have also continuously offered their support, duringthe course of the program, whenever I needed it. The first two yearsof coursework were intense, but immensely educative. Especially in-vigorating were the classes taught by Per Krusell and Arash Nekoei.In addition to being excellent teachers, they have also continued toencourage me throughout the program. A piece of advise from mysupervisor, Jonas, that turned out to be very significant was to reachout to Tommy Andersson. With a theorist’s perspective, Tommy hascontributed with valuable feedback on the second chapter of this the-sis. That same chapter was also improved by several conversationswith David Strömberg. Birthe Larsen has been an appreciated co-author of the third chapter of this thesis, a chapter where I have alsobenefited from great comments by Anna Seim.

I am also grateful to Lena Edlund, for inviting me to spend twosemesters at Columbia University. This proved to be a very rewardingtime period and allowed me to expand my network to scholars fromother institutions in the US as well. The conversations with ChrisNeilson and Peter Bergman turned out to be especially useful. Backin Stockholm, it was time to finalize this thesis. Karin Edmark did anexcellent job as the opponent on my final seminar, providing usefulcomments on all chapters of this thesis. Anna Tompsett, Ines Helm,Sergio de Ferra, Ferenc Szucs, and Diego Battiston deserves a specialthanks for helping me prepare for the job market. Finally, I would liketo thank the administrative staff at the department. Especially AnitaKarlsson, Audrone Mozuraitiene, Anne Jensen, and Marit Fahlén hasprovided outstanding administrative support during these years.

Next, I would like to thank my fellow PhD students at StockholmUniversity and Stockholm School of Economics. Elin Molin standsout. Not only is she smart and supportive, she is also exceptionallyfunny and has the ability to bring out the best in those around here.

Regardless of whether you find yourself in front of a crowded seminarroom or at a late-night karaoke bar, you are better off with her byyour side. Another person that distinguishes himself from the others isErik Lindgren. While his impressive coding-skills has benefited me onmultiple occasions, he proved his friendship when helping me repairthe leaking roof on my summer house in time before my departureto the US. I have also appreciated our spontaneous after works, of-ten joined by Niklas Blomqvist and Jonas Cederlöf. Despite his earlylunch-habits, Niklas’ impeccable sense of humor has made days at theoffice much more fun than they otherwise would have been. He alsoproofread parts of this thesis for which I am very grateful. Jonas alsobrightened up the days at the office, with his contagious laughter,except for the day I lost a bet against him and had to wear an uglysoccer scarf all day.

There are many other graduate students that I have been for-tunate to get to know during these years. My classmates from the2013 cohort deserves to be mentioned; Dany Kessel, Matilda Kilström,Erik Lindgren, Jonna Olsson, Magnus Åhl, Matti Mitrunen, JaakkoMeriläinen, Serena Cocciolo, Selene Ghisolfi, and Jósef Sigurdsson.I am very happy that I ended up in the same year as you! Severalother graduate students in the Stockholm area also enriched my timeas a PhD student. There is not enough space to mention them all,but among them are Christine Alamaa, Malin Tallås Ahlzén, DivyaDev, Emma Heikensten, Siri Isaksson, Adam Altmejd, Agneta Berge,Fredrik Paues, Louise Lorentzon, and Karin Kinnerud. In addition,I am happy that I got to be part of the organizing committe of theFENSU network, through which I got to extend my network to otheryoung female economists in the Stockholm-Uppsala area.

I am also grateful for my friends outside of academia. Life is aboutso much more than work. Thank you for reminding me about this!Two persons in particular has made sure I kept my feet on the ground;Eva and Diana. In addition to being a supportive friend, Diana has

proofread parts of this thesis. Eva has, by repeatedly offering me aseat at her dinner table, made sure I took my mind off work (at leastfor some hours). Spending time with them has given me the breaksI needed in order to get back to work, motivated instead of tired.The two of them, therefore, contributed more to this thesis than theyprobably understand. I look forward to more Friday night’s pizzaparties with you in the future!

Last, but certainly not least, I want to express my gratitude to myfamily. My father taught me early on that anything can be achieved byhard work, and spurred my interest for politics and questions of socialjustice. These are likely two reasons for why I later chose to pursue aPhD in economics. My brother and sister, who bring so much joy tomy life and always have my back. My mother and grandmother, whoare no longer with us, but continue to inspire me every day.

Elisabet OlmeStockholm, Sweden

April 2019

Contents

Introduction 1

1 Are Parents Uninformed? The Impact of School Per-formance Information on School Choices and SchoolAssignments 71.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 81.2 Institutional Setting and Experimental Design . . . . . 181.3 Empirical Strategy . . . . . . . . . . . . . . . . . . . . 251.4 Results from the Experiment . . . . . . . . . . . . . . 381.5 General Equilibrium Effects . . . . . . . . . . . . . . . 461.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . 61References . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631.A Complete Letter . . . . . . . . . . . . . . . . . . . . . 681.B Additional Figures . . . . . . . . . . . . . . . . . . . . 70

2 School Choice, Admission Rules and Segregation inPrimary Schools 732.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 742.2 Theoretical Framework . . . . . . . . . . . . . . . . . . 842.3 Institutional Setting . . . . . . . . . . . . . . . . . . . 902.4 Estimating School Preferences . . . . . . . . . . . . . . 1042.5 Simulation Strategy and Results . . . . . . . . . . . . 1182.6 An Information Experiment . . . . . . . . . . . . . . . 132

2.7 Concluding Remarks . . . . . . . . . . . . . . . . . . . 137References . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1402.A Proof of Claims Regarding Truth-Telling . . . . . . . . 1452.B Descriptive Figures and Tables . . . . . . . . . . . . . 1472.C Regularized Logistic Regression . . . . . . . . . . . . . 1532.D Robustness of Estimated Preference Parameters . . . . 1562.E Implementing DA with Reserved Seats . . . . . . . . . 1592.F Measures of School Segregation . . . . . . . . . . . . . 1602.G Additional Simulation Results . . . . . . . . . . . . . . 1632.H Test of Randomization . . . . . . . . . . . . . . . . . . 1722.I The Letter . . . . . . . . . . . . . . . . . . . . . . . . . 173

3 Should I Stay or Must I Go? Temporary Refugee Pro-tection and Labor-Market Outcomes 1753.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 1763.2 Institutional Setting . . . . . . . . . . . . . . . . . . . 1863.3 Theoretical Framework . . . . . . . . . . . . . . . . . . 1903.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . 2023.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 2133.6 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . 2183.7 Other Outcomes . . . . . . . . . . . . . . . . . . . . . 2233.8 Concluding Remarks . . . . . . . . . . . . . . . . . . . 2273.9 Figures and Tables . . . . . . . . . . . . . . . . . . . . 230References . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2533.A The Danish Asylum Process . . . . . . . . . . . . . . . 2583.B Other Reform Components . . . . . . . . . . . . . . . 2603.C Model Details . . . . . . . . . . . . . . . . . . . . . . . 2623.D Details on the Data . . . . . . . . . . . . . . . . . . . 265

4 The Effects of Performance Based Bonuses in theSwedish Language-Training Program for Immigrants 2674.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2684.2 Institutional Setting . . . . . . . . . . . . . . . . . . . 274

4.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2794.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . 2864.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 2914.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . 298References . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3004.A Additional Descriptive Evidence . . . . . . . . . . . . 3034.B Test for Manipulation . . . . . . . . . . . . . . . . . . 3054.C Regression Discontinuity Graphs of Covariates . . . . 3064.D Varying the Choice of Bandwidth . . . . . . . . . . . . 309

Sammanfattning (Swedish Summary) 311

Introduction

This thesis consists of four self-contained chapters about educationand integration. Chapter 1 and 2 study the determinants ofhouseholds’ choice of primary school and its implications for studentsorting across schools. Chapter 3 and 4 study the integrationprocess of immigrants in their new host country, by evalu-ating their response to changes in integration and immigration policy.

The opportunity to exercise choice regarding the school placement ofone’s children has become an increasingly common feature of manyeducational systems. While affluent households have always had theability to influence their childrens’ school placement, e.g. by movingto a specific school’s catchment zone, other families have had less in-fluence over which school their children have been assigned to. Duringrecent decades, school districts worldwide have implemented policiesto encourage parents to be more engaged in this decision. The mostprominent example is the introduction of centralized school choiceprograms, where parents are asked to express their preferences forschools by submitting rank-ordered application lists. A growing sci-entific literature studies the consequences of this development andaddresses the question of how to optimally design school choice pro-grams.

In the first chapter of this thesis, Are Parents Uninformed?The Impact of School Performance Information on School

1

2 INTRODUCTION

Choices and School Assignments, jointly written with DanyKessel, we study the impact of having access to information aboutschools’ performances when choosing which schools to apply to.We use data from a randomized experiment conducted in theSwedish municipality of Linköping in 2016. A set of randomlyselected households, with children about to start middle school, wereinformed about the schools’ performances on standardized tests aswell as whether the schools performed better or worse than expectedgiven the composition of students at the school.

While we find no effects on the share or composition of house-holds opting-out of their default schools, demand for top-performingschools increase by about five percentage points. This is explainedby a shift in applications from mid- to top-performing schools amongnative and highly educated households. We simulate how this wouldimpact the match between students and schools under the assumptionthat all households would have had access to the same set of infor-mation. As expected, enrollment in top-performing schools increasesbecause of the larger number of applicants. The magnitude representsan increase by about two percentage points in the enrollment rateat these schools, implying that capacity constraints prevent schoolsfrom fully accommodating the increased demand. Furthermore, thenewly admitted students to some extent displaces students with for-eign backgrounds at the top-performing schools. This reduces schoolsegregation by migration background, as students with foreign back-grounds are over-represented at the top-performing schools.

In the second chapter, School Choice, Admission Rules andSegregation in Primary Schools, also authored together withDany Kessel, we instead focus on the admission criteria used in schoolchoice programs. Using data from another primary school choice pro-gram in Sweden, we analyze how segregation across schools dependson how students’ priorities to schools are determined. We have ac-cess to the application lists, including three rank-ordered schools,

3

submitted by parents to four cohorts of primary school starters inBotkyrka municipality. To study counterfactual student allocations,we first model parents’ preferences for schools in order to determineeach household’s complete ranking of schools. Next, using simula-tions, we match students to schools using three alternative admissioncriteria and compare the outcome to the allocation of students whenpreferences are ignored and catchment zones are used to determineschool assignments.

We find that when admission is based on proximity to the school,such that students residing closer to the school are given higher pri-ority, student sorting is approximately the same as when catchmentzones are used. When priorities are instead determined using a lottery,schools become less segregated by students’ family background. Thethird admission criteria implements affirmative action, by giving pri-ority to students that would contribute to the diversity of the studentbody at the school. This reduces segregation even further. The costof replacing proximity-based admission with a lottery or affirmativeaction policies seems modest. The effects on the average rank of theassigned school and indirect utility of the households are small. Forpolicy-makers, reconsidering how priorities to schools are determinedcould therefore be worthwhile. Still, much of the initial segregationpersists due to residential segregation in combination with preferencesfor proximity to the school. Achieving a fully integrated school systemis therefore a task that will likely require changes beyond the schoolsystem.

In the next two chapters, the topic of school choice is abandonedand focus is turned to the integration process of newly arrived im-migrants. The inflow of migrants to Europe is large. In recent years,the development in some areas has also caused many people to leavetheir home countries to seek protection elsewhere, including Europe.The integration of immigrants in their new home countries is there-fore a highly policy relevant topic today. Several European countries

4 INTRODUCTION

have responded to the large inflow of asylum seekers by implementingstricter immigration policies. One such policy is the shift from per-manent to temporary residence permits, which introduces uncertaintyabout being allowed to stay in the host country in the long run.

The third chapter, Should I Stay or Must I Go? TemporaryRefugee Protection and Labor-Market Outcomes, co-authoredby Birthe Larsen and Matilda Kilström, analyzes this question in theDanish context, using a reform that was implemented in 2002. Thereform prolonged the time period with temporary residence permitsfor refugees from three to seven years. Once these years have passed,the refugee becomes eligible to apply for permanent residency. Weinterpret this change as lowering the ex-ante probability of asylumholders being granted permanent residency, and study the effects onhuman capital investments and labor-market outcomes.

In addition to having grounds for asylum, residency can also besecured by demonstrating a stable attachment to the labor market.While the reform may discourage investments in country-specific hu-man capital due to a lower expected return to such skills, it could alsoincrease the importance of such skills as entering the labor market be-comes an alternative way to secure permanent residency. In that case,the effects would go in the opposite direction. In the end, which effectdominates is an empirical question. The retroactive implementationof the reform enables us to study this question using a regression dis-continuity design. As more investments are likely required by somegroups, compared to others, in other to establish themselves on thelabor market, a heterogeneity analysis is conducted. In addition, tounderstand the mechanisms at work, we complement the empiricalanalysis by setting up a theoretical search and matching model withheterogeneity in skill-levels.

We estimate a significant positive effect on enrolling in educa-tion, driven by females and low-skilled individuals. This is in linewith the predictions from the model, suggesting that individuals fur-

5

ther away from the labor market will invest in education in orderto improve their position on the labor market. For the full sample,we find no statistically significant effects on labor-market outcomessuch as being employed (including self-employment), or on income.Exploring the impact of the reform further, we also study the effectson criminal activity, utilization of health care and the propensity tohave children. We find a negative impact on the crime rate, with thestrongest response among males. Although anecdotal evidence sug-gest that the reform had a negative health impact among refugees,we cannot confirm this by looking at the number of times individualsseek formal health care. We do find some evidence of a negative effecton childbearing, suggesting that the increased uncertainty about thetime horizon in Denmark might have deterred some individuals from,or delayed the timing, of having children.

In the fourth and final chapter of this thesis, The Effects ofPerformance Based Bonuses in the Swedish Language Train-ing Program for Immigrants, I turn to analyze the effects ofperformance-based bonuses in language training. Learning the hostcountry language is a key determinant behind successful integration.Still, many immigrants never come to master their new language. In2010, performance-based bonuses was introduced in the Swedish lan-guage training program to address this problem. A passing grade incertain courses of the program qualified the immigrant for a tax-freebonus of up to 12,000 SEK (approximately 1,666 USD).

Implicit in the introduction of the bonus-system is the notion thatimmigrants exert too little effort in their language studies. This maybe because they underestimate the returns to language skills, or be-cause they neglect potential positive externalities of language pro-ficiency. The prospect of receiving a bonus could therefore improvematters, as incentives to pass courses are strengthened. On the otherhand, if immigrants use there time more productively in other ac-tivities, the bonus could actually lead to a sub-optimal allocation of

6 INTRODUCTION

their time. Furthermore, introducing monetary incentives in an edu-cational context could also have reverse effects by crowding out theintrinsic motivation of students. In the end, understanding how mon-etary bonuses impact students’ performances in language training istherefore an empirical question.

The effects of the bonus-system are analyzed using a regressiondiscontinuity design, exploiting the fact that eligibility to the bonusprogram was completely determined by the date of arrival to Sweden.I find no impact on the enrollment rate, suggesting the the bonus-system did not induce more immigrants to start language training.This may be explained by the fact that the enrollment rate was al-ready high to begin with, leaving relatively little room for improve-ment. Next, the effect on the pass rate is examined. While there isno evidence for an increase in the overall pass rate, the results look-ing specifically at bonus-entitling courses are inconclusive. A positiveeffect of up to four percentage points cannot be excluded, but theresults are sensitive to the empirical specification.

Chapter 1

Are Parents Uninformed?The Impact of School Performance Information onSchool Choices and School Assignments∗

∗This chapter was co-authored by Dany Kessel. We thank Niklas Blomqvist,Matz Dahlberg, Karin Edmark, Jonas Vlachos, and seminar participants at Stock-holm University and Stockholm School of Economics for valuable comments andfeedback. We also gratefully acknowledge financial support from the Swedish TradeUnion Confederation (LO), the National Union of Teachers in Sweden (LR), andthe Swedish Teachers’ Union. This research also benefited from financial supportfrom Handelsbanken’s Research Foundations. Finally, we thank Linköping munici-pality for allowing us to study school choice in their municipality. The experimentaldesign and data used in this study have passed ethical vetting by the Stockholmethical review board (DNR 2015/1252-32). All errors are our own.

8 CHAPTER 1

1.1 Introduction

Privileged households have always been able to exercise school choice,for example by choosing their residential location. During the lastdecades, many countries have put policies in place in order to ex-pand this opportunity to all.1 The introduction of centralized schoolchoice programs means that an increasing number of households arenow faced with the decision of what school to send their children to.The motivation for this development is twofold. First, school choiceis argued to increase school quality as schools will have to competein order to attract students. The empirical evidence to support thisclaim is however mixed. While some studies estimate positive effectson students’ academic outcomes, others find no or even negative ef-fects.2 Second, school choice is also argued to reduce sorting of stu-dents across schools based on their migration background and socioe-conomic status, by allowing those in disadvantaged neighborhoods toopt out of nearby low-quality schools. Contrary to this, the literatureshows that schools become more segregated by students’ backgroundwhen school choice is introduced.3

One possible explanation for why the gains envisioned by propo-nents of school choice has not been realized is lack of easily accessible

1For example, Musset (2012) shows that two thirds of all OECD countrieshave introduced or expanded households’ opportunities to impact what schooltheir children are assigned to during the last decades.

2See e.g. Böhlmark and Lindahl (2015), Hoxby (2000), and Lavy (2010, 2015)for studies that show positive effects. Cullen et al. (2006) and Hsieh and Urquiola(2006) find no effects while Abdulkadiroğlu et al. (2018) estimate a negative impacton student achievement. This list is not exhaustive, and the robustness of theresults in Hoxby (2000) is discussed in a comment by Rothstein (2007) to whichHoxby (2007) replied.

3See Böhlmark et al. (2016) and Söderström and Uusitalo (2010) for studiesfinding increased segregation due to school choice using Swedish data. Bifulcoand Ladd (2007) show that the introduction of charter schools in North Carolinaincreased racial sorting and Hsieh and Urquiola (2006) find that the nationwideintroduction of school choice in Chile made schools more segregated by parentaleducation and income.

1.1. INTRODUCTION 9

information about schools’ performances. For the market-mechanismto work, households need to base their decision about what school toapply to (at least in part) on how the schools perform. As there isusually no price mechanism to guide households on the school mar-ket, this may be challenging. If households are uninformed about theperformances of schools in their choice set, providing such informa-tion could strengthen the aforementioned mechanism and increasethe competitive pressure on schools. Furthermore, the ability to dis-criminate between schools based on their performances may not beequally distributed among households. There is evidence that familybackground matters for what schools households choose to apply to.For example, disadvantaged households are more likely to apply tolow-performing schools compared to other households.4 It is not clearwhether this is explained by heterogeneous preferences for schools,differences in access to high-performing schools, or variation in house-holds’ knowledge about schools’ performances.5 In the last case, low-ering the cost of obtaining information about schools’ performancescould reduce the segregating effects of school choice.

To shed further light on this question, a large-scale randomizedcontrolled trial was conducted in the Swedish municipality ofLinköping. With about 1,600 students per cohort, they operateone of the largest school choice programs in Sweden. In 2016,households with children about to start middle school (7th grade)were randomly assigned to a control or treatment group. Prior tosubmitting their school applications to the municipality, households

4See Abdulkadiroğlu et al. (2017), Abdulkadiroğlu et al. (2018), and Hastingset al. (2009).

5Hastings and Weinstein (2008) show that lack of information about schools’performances is one explanation behind the applications to low-performing schoolsby disadvantaged families in the Charlotte-Mecklenburg Public School District,North Carolina, but proximity to better schools determines the impact of providingsuch information. Burgess et al. (2015) show that a significant proportion of thedifferences in what schools advantaged and disadvantaged households apply to inthe UK is due to variation in the quality of available schools.

10 CHAPTER 1

in the treatment group received a letter with information about theschools’ performances on the standardized tests taken in 9th grade.The letter included the schools’ average score as well as an adjustedperformance measure considering the composition of students at theschools. Using this variation in access to information about theschools’ performances, the effect on what schools are are chosen isstudied. Moreover, this random variation in access to information iscombined with knowledge about the assignment mechanism usedin this school choice program to simulate the impact on schoolassignments, had all households received the information.

This is not the first study on the topic. A number of studiesuse school rankings published in newspapers, school accountabilityschemes, or government inspections in order to explore the effects ofschool performance information. One identification approach is to ex-ploit variation in the timing of when such information is made public.Using this strategy, Hart and Figalo (2015) study the introduction ofa school accountability scheme in Florida and find a six percent in-crease in the enrollment rate of schools receiving an A compared toschools receiving a B or C. Hussain (2013) exploits the timing ofschool inspections in the UK and documents a three percent increasein enrollment the year after a school is rated as outstanding and afour percent decline in enrollment the year after a fail rating. Koningand van der Wiel (2013) find even larger effects when analyzing theimpact of school quality indicators published in a national newspaperin the Netherlands. They document a 20 percent increase in enroll-ment in academic tracks in the year following publication for schoolswith good quality scores. There are also similar studies using housingprices as the outcome in an attempt to estimate how much parents arewilling to pay for school quality, but the evidence from these studiesis less easy to interpret.6

6Figalo and Lucas (2004) analyze the introduction of the school accountabilityscheme in Florida and document that house prices increase in schooling zones

1.1. INTRODUCTION 11

Another approach to identify the effects of school performance in-formation is to exploit cutoffs that arise when information is madepublic. Using a sharp regression discontinuity design, Mizala andUrquiola (2013) studies the SNED program in Chile which identi-fies high-performing schools among a set of similar schools in terms oftheir student composition. They find no effects on enrollment of beingidentified as an effective school. In a similar study in Brazil, whereresults on standardized tests are published only for schools with aminimum number of test takers, Lépine (2015) finds no effect on en-rollment of published test scores for neither low- nor high-performingschools.

The fact that these two approaches yield different results raisesthe question if one, or both, fail to reproduce the results that a ran-domized controlled trial would give. The limited number of studiesmakes it possible that those exploiting cutoffs for identification wereconducted in settings where information had no effect whereas thoserelying on the timing of information were conducted in settings whereinformation had an impact. Another potential explanation lies in thechoice of the outcome variable. All these studies focus on enrollment.The (absence of) effects on enrollment is not necessarily informativeabout whether households reacted to the information when choos-ing which school(s) to apply to. Institutional factors, such as theschools’ capacity constraints or the assignment mechanism used toallocate students to schools, may suppress the impact on enrollmenteven though demand for certain schools increases. Moreover, whilean effect on enrollment implies that some households adjusted their

where the local school received a good score. Fiva and Kirkeboen (2011) studythe release of previously unpublished information on school quality in Norway andfind that housing prices are affected by this information. Imberman and Lovenheim(2016) exploit the release of value-added measures in Los Angles and find no effecton enrollment. The effects in both Figalo and Lucas (2004) and Fiva and Kirkeboen(2011) are short-lived reverting back to the pre-information prices in just a fewmonths. This is not easily explained and could indicate that the results are, atleast in part, driven by some sort of speculative behavior from house buyers.

12 CHAPTER 1

school applications, it is not informative about the magnitude of thiseffect. Even if only a small group of households change what schoolsthey apply to, this could have an impact on the school assignmentsof a much larger group as the effects propagate through the system.The pitfall of using enrollment as a proxy for which school(s) house-holds apply to is pointed out by Ruijs and Oosterbeek (2017), whofind no evidence that the school quality indicators studied by Koningand van der Wiel (2013) predict households’ applications to schools.7

This means that in order to understand the role of information inschool choice programs, it is necessary to study its impact on bothhouseholds’ applications to schools and what schools their children areassigned. Due to data limitations there are not many studies on theformer, but two exceptions should be mentioned. Hastings and Wein-stein (2008) use two experiments (one field experiment and one natu-ral experiment which arose because of the design of the No Child LeftBehind Act) that provided direct information about school test scoresin lower-income areas in the Charlotte-Mecklenburg Public SchoolDistrict. Analyzing data on school choices, they document a five per-centage point increase in the probability of opting out from one’sneighborhood school and instead applying to a higher-performingschool when information on school performance became available.Corcoran et al. (2018) randomly selects a number of high-poverty NewYork City middle schools and provide the students in these schoolswith a list of high schools with graduation rates above the city median(70 percent). They do not find that the treated students, on average,apply to higher-performing schools but they become more likely to

7Koning and van der Wiel (2013) use enrollment in the third year of secondaryschool as a proxy for applications in the first year. Ruijs and Oosterbeek (2017)argue that doing so, in addition to capturing effects on initial enrollment, they mayalso pick up effects on the proportion of students held back and/or transferringfrom/to other tracks. With access to data on households’ school applications fromAmsterdam, Ruijs and Oosterbeek (2017) estimate a conditional logit model whereit cannot be confirmed that the published school quality indicators predict schoolapplications.


apply to schools where their admission probability is higher.8This study of a information intervention in a Swedish municipal-

ity makes three main contributions to the literature. Firstly, becausethe treated population is not restricted to households in disadvan-taged neighborhoods, it is possible to say more about the impact onthe whole distribution of households. The universe of school choicesis observed for an entire cohort of middle school starters in a ratherwell-off Swedish university town and treatment is not restricted tolower-income households. This is important as it speaks to the poten-tial effects of making school performance information available on thecompetitive pressure in the educational system as a whole. Secondly,it can be studied whether households react mainly to raw or adjustedtest scores, as the letter included information on both. A school’s av-erage test scores are, in large part, determined by the school’s studentcomposition.9 Therefore, any changes in school choices driven by in-formation on raw test scores could be due to parents looking for aschool with a stronger peer composition rather than a more produc-tive school. This could however not explain a reaction to adjustedscores as these are, by construction, orthogonal to the peer composi-tion. Thirdly, because the assignment mechanism used in this schoolchoice program is known and high-quality Swedish register data areavailable, it is possible to simulate the impact on general equilibriumoutcomes. The question in mind is how any changes in school choicestranslates to an impact on school assignment, had all households hadaccess to same information on school performances.10

8There has also been some semi-structural work using data on households’applications to schools, including previously mentioned studies by Ruijs and Oost-erbeek (2017) and Hussain (2013).

9A regression of raw test scores on student/family characteristics (gender,parental education, migration background) using all schools in Sweden gives avalue of adjusted R2 equal to 0.653.

10Again, due to data limitations, only a few previous studies have been able toanalyze the effects of school performance information on other equilibrium out-comes than enrollment. A notable exception is a recent study by Abadie et al.

14 CHAPTER 1

The first set of results concern the impact of providing informationabout school performance on school choices. If households have pref-erences for high-performing schools but are unable to identify theseamong the set of schools to choose from, it would be expected toobserve a higher performance of schools applied to by households inthe treatment group compared to the control group. The results arein line with this hypothesis, with a 5.2 percentage point (12 percent)increase in the probability of listing a top-performing school as one’stop choice in the treatment group.11 The schools losing applicantsare the second-tier schools while the low-performing schools are un-affected. To understand whether the effects are heterogeneous, twosample splits are made. First, the sample is divided by migrationbackground to look separately at native households and householdswith a foreign background. Second, the sample is divided by the par-ents’ educational level, looking at low- and high-skilled householdsone at a time. The results show that the shift in applications to top-performing schools is driven by native and high-skilled households.Interestingly, these households react mainly to the adjusted perfor-mance measure indicating that they are looking for productive schoolsrather than a favorable peer composition. This is, to our knowledge,the first documented effect of this kind.

All in all, these findings indicate that providing school perfor-mance information could increase the competitive pressure on schoolsto improve their performance. This is in line with the predictions fromWalters (2018) model of charter school choice. Nonetheless, the effectsmay be local in the sense that low-performing schools will not be af-fected. Hastings et al. (2009) suggest that the competitive pressure in-duced by school choice is more prominent for already high-performing

(2017) in which report cards are distributed to a randomly selected set of vil-lages in Pakistan. They find effects on enrollment, average test scores, and pricescharged by private schools.

11A top-performing school is defined as being one of the five highest-performingschools among the 17 schools included in the analysis.


schools that are competing for students whose demand for quality ismore elastic. This compares well with the finding in this study thatmainly advantaged households respond to the information providedand the failure to document any significant effects of test score in-formation on school choices made by disadvantaged households (asfound in Hastings and Weinstein, 2008 and Gallego et al., 2012).

Next, simulations of how the intervention would have impactedthe allocation of students to schools are performed, under the assump-tion that all households had been provided with the information. Theresults show that the increased demand for top-performing schoolsincreases the share of students assigned to these schools by about 2percentage points. This finding is in line with the positive effect onenrollment found by, for example, Hart and Figalo (2015). However,the increase corresponds to only 40 percent of the documented in-crease in applications for top-performing schools. This confirms theconcerns expressed by Ruijs and Oosterbeek (2017) about using en-rollment as a proxy for school choice in a setting where schools havecapacity constraints. Furthermore, it lines up well with the finding byHussain (2013) that the effects of providing information about schoolquality is larger in areas where capacity constraints do not bind.

In the setting analyzed in this study, the increased enrollmentrate in top-performing schools is mainly driven by native childrenand children to high-skilled parents, from whom a shift in applica-tions was observed. These children, to some extent, displace studentswith a foreign background previously admitted to these schools. Aschildren to high-skilled parents were already to begin with enrolledin higher-performing schools than children to low-skilled parents, theperformance gap between the assigned schools of these two groupsincrease in consequence. It also reverses the performance gap of theassigned schools of children with native versus foreign background. Onaverage, children with a foreign background initially attended higher-performing schools than their native counterparts (this is generally

16 CHAPTER 1

true also for other school districts in Sweden). With the increased de-mand for top-performing schools by native households, their childrencongregate in the top-performing schools and reverses this relation-ship as children with foreign backgrounds are pushed out of theseschools.

Finally, the impact on school segregation is studied. This speaksdirectly to the second overarching questions posed in the beginning ofthis section: can school segregation be reduced by informing parentsabout schools’ performances? There are some insights from previousliterature, but the evidence is inconclusive. For example, Hart andFigalo (2015) find that the share of high-SES students increases athigh-performing schools when information on school quality is re-leased, while Mizala and Urquiola (2013), when conducting a similarexercise, find no significant effects. Corcoran et al. (2018) conductsubgroup analysis and find that disadvantaged households, in gen-eral, do not seem to react more to school performance informationcompared to advantaged households.12 However, Hussain (2013) findsthat poorer households seem to react more to school-performance in-formation than other households when it is presented in a simplifiedway.

In the simulations presented in this study, schools become moresegregated by parental skill-level as students from high-skilled house-holds congregate in the top-performing schools. At the same time, a6 percentage points significant decrease in segregation by migrationbackground (measured by the Duncan Dissimilarity Index) is docu-mented. The reason for this is however not the one usually envisioned;disadvantaged households with foreign backgrounds becoming moreprone to apply to schools with more native students. Instead, thisreduction is explained by a larger share of native households applyingto and being assigned to the top-performing schools which in general

12An exception is non-english speaking households who were much less likelyto be assigned a low-performing school.


have a high concentration of students with foreign backgrounds. Thiseffect accounts for about half of the observed drop in school segre-gation. The other half is explained by an increase of children withforeign background at the mid-performing schools, assigned the seatsthat native students left behind when switching to the top-performingschools. As foreign students were initially underrepresented at theseschools, segregation decreases.

The remainder of this chapter is organized as follows. Section 1.2outlines the institutional setting and the design of the experiment,Section 3.4 describes the data and the empirical strategy, Section 1.4presents the results from the experiment, Section 1.5 discusses thesimulation strategy and presents the results on general equilibriumoutcomes and Section 3.8 concludes.

18 CHAPTER 1

1.2 Institutional Setting and ExperimentalDesign

The Swedish compulsory school system consists of nine grades withchildren, in general, starting 1st grade the year they turn seven.13 Thesystem is highly decentralized with the municipalities being responsi-ble for providing and financing compulsory education. School choicehas been part of the Swedish school system since 1992, in the sensethat parents have had a legal right to express a preference for whichschool they would like their child to attend. However, it took manyyears before municipalities started running centralized school choiceprograms allowing parents to submit a rank-ordered list of preferredschools. Today, these types of programs are the norm. The experimentwas conducted in the municipality of Linköping in 2016.14 In this sec-tion, the school choice program in Linköping is described togetherwith the design of the experiment.

1.2.1 Linköping’s School Choice Program

Linköping is the fifth largest municipality in Sweden with a totalof 158,000 inhabitants. Most of them (109,000) live in the town ofLinköping, located in the middle of the municipality and the rest livein smaller villages or rural areas surrounding the town. The averagesize of a student cohort is about 1,600. The focus of this study is onthe choice of middle school (7th to 9th grade).15 There are 20 middle

13From the school year 2018-2019, grade K is mandatory for all children follow-ing the adoption of Proposition 2017/18:9, implying changes to Skollag 2010:800.Grade K precedes the 1st grade in compulsory school.

14The reason for studying the school choice program in Linköping is that it wasone of the few municipalities that was both interested in letting participating inan experiment and able to provide access to high-quality data on school choices.

15The reason for focusing on the choice of middle school, and not elemen-tary school, is that that school performance is likely to be relatively more im-portant when choosing middle school compared to choosing elementary school. InLinköping, proximity to the school seems to completely dominate the elementary

1.2. INSTITUTIONAL SETTING AND EXPERIMENTALDESIGN 19

schools in Linköping of which 13 are public schools. The other 7 arevoucher schools which are independently run but publicly funded.16The middle schools in Linköping are concentrated in the town ofLinköping with the highest concentration just northwest of the citycenter, as can be seen in Figure 1.1.

In the beginning of January of each year the municipality sendsout a letter to all households with children about to start 7th grade.The letter informs the parents that it is time to apply for middleschool. The letter also includes information on the household’s de-fault school, where the student will be placed if no school choice issubmitted or if the student is not accepted to any of the schoolsapplied to. The default school is usually the school closest to thehousehold’s residence. In the letter there is also a user name and apassword for an online tool, where the households can log in and sub-mit their school choice(s) in the form of a rank-ordered list of desiredschools.17 Further, the letter contains information on how students’priorities to oversubscribed schools are determined, where one canfind more information about the middle schools and information inother languages than Swedish.

The online tool is open for about three weeks at the end of Jan-uary/beginning of February, during which parents can log in andsubmit their application list. They can list up to three rank-orderedchoices, and both public and voucher schools can be ranked. There aresome middle schools in Linköping with special tracks such as drama,science and soccer. If one wants to apply for one of these it has tobe included as one of the three possible choices on the application

school choice with 75 percent of all households opting for the closest school.16The public schools in Linköping either covers grade 1 to 6 (elementary school)

or grade 7 to 9 (middle school). Two of the seven voucher schools follow thispattern and start in the 7th grade, while the other five starts earlier. 4 percent ofthe students attend one of these five voucher schools in 6th grade.

17They also have the option to accept the seat at their default school. Thisis equivalent to not submitting anything as, in both cases, the child would beassigned to the default school.

20 CHAPTER 1

Figure 1.1: Location of Middle Schools in Linköping

Notes: This map shows the locations of the middle schools in Linköping. Of the 20 middle schools inthe district, 16 are displayed on the map. Two schools are excluded because they will not be part ofthe empirical analysis in this study (for reasons discussed in Section 1.3.1). The other two schools areexcluded because their geographical location is missing in the data. Both of these schools opened in2014 and geographical data on school locations are available only up until 2013. The map also showsthe division of the municipality into SAMS-areas in order to give a sense of the population density.SAMS stands for Small Areas for Market Statistics and divides Sweden into about 9,500 small areasbased on municipal partitioning and electoral districts (depending on the size of the municipality).SAMS-areas are designed to have roughly the same population. c©Statistics Sweden


list. When the school choice period ends the municipality sends thelists of applicants to the voucher schools, so they can rank the appli-cants in order of priority. All voucher schools in Linköping operatesa first-come, first-served queue system, where parents can place theirchildren in the queue from a young age. Sometimes the queue systemis combined with priority for those who have older siblings alreadyenrolled in the school. The municipality determines the priority eachstudent has at the public schools based on the distance from the stu-dent’s home to the school, with higher priority to those living closerto the school. Special tracks are an exception, where tests and tryoutsare used instead.

When each student’s priority has been established the municipal-ity assigns students to schools using a modified version of the Bostonmechanism.18 In short, the municipality starts by giving everyone aseat at their default school. All students are then allowed to applyto their highest-ranked school and students are accepted in order ofpriority until the school is at full capacity. This is done in severalrounds as once a student is accepted at a new school, a seat becomesempty at the old school which another student might have ranked ontop. When as many students as possible have been placed at theirtop-ranked school, the process is repeated with second-ranked choicefor those who could not be assigned their top-ranked school. Finally,the process is repeated again using the the third choice. Once all stu-dents are assigned to a school the municipality sends out a secondletter informing the households about which school they have beenassigned. This usually happens in the middle/end of March. Figure1.2 visualizes this process.

The fact that the municipality uses the Boston mechanism meansthat households have incentives to act strategically when decidingwhich schools to apply to and how to rank them. The probability of

18The Boston mechanism is described in detail in Abdulkadiroğlu and Sönmez(2003).

22 CHAPTER 1

being accepted to a school is reduced with the rank of the school onthe application list. This means that careful consideration should betaken when deciding which school should be ranked on top, as a highweight is given to that choice. In Linköping, the schools applied toby students opting out from the default school are in general over-subscribed. Ranking one of these schools second or third would thusimply a very low probability of admission. Therefore, second and thirdchoices have little impact on which schools students are assigned to.This can be inferred from the fact that 97 percent of all students areassigned either their top-ranked school or their default school.19

Figure 1.2: Time-line of the School Choice Program in Linköping

Jan

2011

Feb

2011

Mar

2011

Apr

2011

Schoolchoices

subm

itted

Inform

ationsent

out

Students

assigned

toschools

Decisions

communicated

When deciding where to apply, households have access to a num-ber of different sources of information. The main source is the munic-ipality’s own online school-comparison tool, available on the munici-pality’s website and referred to in the first letter to the households.The tool covers both public and voucher schools in Linköping. It con-tains information on the schools’ location, the number of studentsin different grades, a short text written by the schools themselves,and results from a municipality-administered survey conducted eachyear where students and parents rate their school in terms of learn-ing environment, facilities, food etc. The tool does not contain any

19Of those rejected at their top-ranked school, 28 percent are admitted to asecond- or third-ranked school. However, this corresponds to only 3.2 percent ofthe total population.


information on grades, test scores, graduation rates, student-teacherratios or proportion of certified teachers. Information on these met-rics can be found online in the Swedish National Agency for Educa-tion’s database. The existence of this database is however not com-mon knowledge and the database is not easy to navigate. In a surveyconducted in a number of Swedish municipalities in 2012, parents re-port that they have difficulties making an informed choice becausethey lack information about the schools (Malmberg et al., 2014). Af-ter that, two websites targeting parents and students about to applyto schools were launched.20 The first of these websites was launchedin 2013, but is no longer active and no data on the number of vis-itors at the time are available. The second website was launched in2014 by the Swedish National Agency for Education, but was nevermarketed. Furthermore, the use was not widespread outside of theStockholm metropolitan area.21 The schools also offer an open houseforum, giving students and parents a chance to visit the school beforesubmitting their applications. While this might be an opportunity tolearn more about a single or a few schools, it is unlikely to be anefficient way to get an overview of how the 20 schools compare interms of their performance. Taken together, information on schoolperformance appears limited for parents in Linköping.

1.2.2 The Experiment

The experiment took place in Linköping at the beginning of 2016.With the toss of a coin all households with a child expected to start7th grade were sorted into a treatment group and a control group. Si-multaneously to receiving the letter from the municipality, households

20https://www.grundskolekvalitet.se/ and http://valjaskola.se/.21During 2016, according to statistics from Google Analytics provided by the

National Agency for Education, the site had about 550 visitors from Linköpingduring the whole of 2016. Considering that the website provides information aboutschools from grade K to 9 in the compulsory school system as well as high schools,this confirms limited use of the site at the time.

24 CHAPTER 1

in the treatment group also received a letter containing school perfor-mance information. For each middle school in Linköping, the averageperformance on the national standardized tests was presented. Morespecifically, for each school, the average score over the two last yearson the standardized tests in 9th grade in Swedish, mathematics andEnglish was calculated.22 The households were also provided with ameasure of how each school performs relative to how it is expectedto perform given its student composition. A detailed description onhow this measure was constructed will be given in Section 1.3.1. Theletter was translated into nine different languages to ensure that ev-eryone would be able to understand it and included contact details,should they have any questions. The letter in its entirety can be foundin Appendix 1.A. Households in the control group received only theinformation supplied by the municipality.

22In Sweden standardized tests are taken by all students in 3rd, 6th and 9thgrade.

1.3. EMPIRICAL STRATEGY 25

1.3 Empirical Strategy

This section describes the data sources, outlines the empirical strategyand discusses the identifying assumptions.

1.3.1 Data

There are three main sources of data used in this study. The first isadministrative data from the municipality of Linköping, the second isindividual registries from Statistics Sweden, and the third is school-level data from the Swedish National Agency of Education. The datafrom Linköping contains the full set of application lists submitted tothe Linköping school choice program in 2016.23 These applications arelinked to individual-level register data from Statistics Sweden, usingindividual identifiers and the Swedish multi-generational register.24From register data, demographic and geographical information on allmiddle school starters is extracted as well as students already attend-ing schools in Linköping. Further, for each household with at least onechild about to start 7th grade, there is an indicator for whether thehousehold were assigned to the treatment or control group.25 From theSwedish National Agency of Education, data on school performances(test scores and grades) and measures of the student composition ofall middle schools in Sweden are available. A more detailed descriptionof the data will follow.

23Data are also available on the application lists from the 2014 school choiceprogram. These are not used in the main analysis, but to test one of the identifyingassumptions. This is discussed in further detail in Section 1.3.3.

24If the student is adopted, the characteristics of the adopting parents are usedinstead of the biological parents as the adopting parents are more likely to beinvolved in the schooling decision.

25In addition, there is an indicator for households that could not be located bythe postal services in which case the letter was returned. This happened in threecases and in all cases, another parent on a different address received the letter(when parents were not registered on the same address, letters were sent to bothaddresses).

26 CHAPTER 1

The population consists of all students that were invited to par-ticipate in the school choice program in Linköping, i.e. all children inthe municipality expected to start middle school by fall 2016. Withinthis population, children were randomly assigned to a treatment orcontrol group. Some sample restrictions are made. First, two childrenthat never participated in the school choice program are dropped.These children were not assigned and did not enroll in a middle schoolin Linköping, suggesting that the households likely moved to an-other municipality. Second, seven students enrolled in a small religiousschool are dropped for reasons discussed later in this section. Fur-thermore, applications are observed for 41 households not expectedto participate in the school choice program and therefore never as-signed to the control or treatment group. They are also dropped fromthe analysis.26 In the end, the population consists of 1,608 householdswith children about to start 7th grade, where school choices, relevantbackground characteristics, and treatment status is observed.

The municipality of Linköping provided data on the full applica-tion lists of those participating in the school choice program, contain-ing up to three rank-ordered school choices for each child. In additionto the application lists, each student’s default school is observed aswell as which school they were assigned to. In order to keep the empir-ical analysis clean and simple, focus is on the households’ top-rankedchoices. As mentioned, the design of the system makes the secondand third choice more or less inconsequential and a large fraction ofhouseholds do not even bother to submit second or third choices.27If a household did not apply to any school(s), the default school istreated as their top-ranked choice.

26These individuals either moved to Linköping before the start of the schoolyear, but after January 2016 when the treatment and control group was definedor they were students from other municipalities that participated in Linköping’sschool choice program.

27Of 783 households in the control group, 22.5 percent report a second-rankedschool that is not their default school or the same school as their top-ranked school(which can happen if the top- or second-choice is a special track).


From Statistics Sweden’s LISA register, information on the regionof birth and level of education of the middle school starters’ parentsas well as the parents of all children already enrolled is available.Indicators for students with a foreign background (defined as beingborn abroad or having two parents born abroad) and students withhigh-skilled parents (defined as having at least one parent with morethan upper secondary education) are created. Combining this with thecompulsory school register enables measures of the student composi-tion of all middle schools in Linköping to be created.28 Statistics Swe-den also provide data on the geographical location of both studentsand schools. Data on the residential location of students are basedon their registered address on the first of January 2016 while schoollocation is based on addresses from 2013. As no school has changedlocation between 2013 and 2016, this is not a problem. Location isindicated by the midpoint of a 250m×250m grid (1000m×1000m forthe rural areas). The distance between each student and each schoolis calculated using these midpoints.29

Lastly, the Swedish National Agency for Education provide dataon school performance on the standardized tests in 9th grade for allschools in Sweden. This was used to construct the measures of schoolperformance included in the letter sent out to households in the treat-ment group. Each school’s average score on the tests in Swedish, math-ematics, and English is observed. The scale ranges from zero to 20,with ten points usually considered a passing score and 15 points con-sidered a pass with distinction. For each school, the average scoreover all these subjects is calculated. Further, the score is averaged

28The latest compulsory school register available in this study is from 2014,which means that this is the latest year for which the student composition of theschools can be observed. As there has been little variation over time in the studentcomposition at the middle schools in Linköping, this is likely to be a good measureof the peer composition considered by those choosing schools in 2016.

29The distance is calculated as√

(xstudent − xschool)2 + (ystudent − yschool)2,where x and y represent the longitude and latitude of the location of the stu-dent/school.

28 CHAPTER 1

over the years 2013-2014, in order to reduce the noise that arisesfrom the fact that some schools have few students per cohort.30 Thisis the first measure included in the letter, henceforth referred to asthe schools’ raw test scores. In addition, the letter included a measureon how the schools perform relative to what would be expected giventheir student composition. This measure is calculated by estimatingthe following regression using data on all middle schools in Sweden,separately for 2013 and 2014:

Yi = α+X ′iβx + εi, (1.1)

where Yi is school i:s score on the standardized tests in 9th gradeand Xi is a vector of variables describing the student compositionat school i.31 The adjusted performance measure of a school is thendefined as the average of their residuals over the two years.

For four of the 20 middle schools in Linköping, student compo-sition, location, and/or performance on the standardized tests aremissing. In two cases, this is explained by the schools opening in 2015implying that no students in these schools had taken the standardizedtests by the time the letter was sent out. Further, location is not ob-served since the last year in the geographical data set is 2013. One ofthese schools is an subsidiary of an already existing school, expectedto eventually take over the middle school students from that school.The old and the new school are located close to each other and stu-dents with either of these two as their default school have them bothas their default school. When the analysis demands it, they are there-fore treated as the same school. The other new school is not linked

30See Kane and Staiger (2002) for a discussion on why this is important. In-cluding additional years would have been preferred but was not possible due todata limitations.

31The vector includes the share of boys, the share of students with high-skilledparents and the share of newly arrived students. The specification follows theSwedish National Agency for Education’s method used in their so-called SALSA-measure.


to an existing school in the same way. Throughout the analysis, itis assumed that parents (in absence of information) treat it as anaverage-performing school. The other two schools for which data aremissing are dropped from the analysis. One is a small religious schoolwith about five to ten students per grade. In 2016, there were sevenstudents with this school as their default school (because they werealready enrolled in that school). None of these households applied toany other school and no other household applied to this school. Itcan therefore be dropped, without any loss for the analysis. The sec-ond is a small school aimed at helping children from extremely roughconditions. None of the households had this as their default schooland there were no applications to this school. It can therefore also bedropped without consequence for the analysis.

1.3.2 Summary Statistics

Table 1.1 reports the attributes of the middle schools in Linköping.The average score on the standardized tests is 13.32 (on a 20-gradescale), 0.42 points less than expected given their student composi-tion. On average, the schools have 22 percent of students with a for-eign background and 62 percent of students with high-skilled parents.There is considerable variation between the schools both in terms ofperformance and student composition. For example, the average scoreat the lowest-performing school is 9.75, which is not considered a pass-ing grade. The highest-performing school scores 16.10 which is wellover the threshold for a pass with distinction.

Table 1.2 reports descriptive statistics for the population of school-choosing households. It is clear that the population in this study is notas disadvantaged as in previous studies on this topic (such as Hastingsand Weinstein, 2008 and Corcoran et al., 2018). The average defaultschool is similar to the average school in Table 1.1, despite defaultschools being a subset of all schools.32 Further, households have on

32Specifically, voucher schools are never default schools. In addition, the special

30 CHAPTER 1

Table 1.1: Summary Statistics for Schools

Mean SD Min Max

PerformanceRaw test scores 13.32 1.62 9.75 16.10Adjusted test scores -0.42 0.79 -2.03 1.03

Student compositionForeign background (share) 0.22 0.21 0.03 0.86High-skilled parents (share) 0.62 0.16 0.15 0.86

Observations 16Notes: Raw test scores is the average score (on a scale from 0 to 20)on the standardized tests in Swedish, mathematics and English in 9th

grade during 2013-2014. Adjusted test scores is the difference betweenthe actual and expected score, where the expected score is the scorepredicted by the regression in equation (2.11). A student is consideredto have a foreign background if the student is born abroad or if bothparents are born abroad. A student is considered as having high-skilledparents if at least one of the parents have more than upper-secondaryeducation.

average six schools within a reasonable distance (5 kilometers) fromtheir home. The variance in the number of schools close to home islarge, with some households having no schools within 5 kilometerswhile others have more than ten. This reflects the fact that somehouseholds reside in the town of Linköping where the schools are con-centrated while others reside in the smaller villages or the rural areasurrounding the town. Even within the set of schools close to home,households face considerable variation in terms of the schools’ per-formance and student composition. For example, the average differ-ence in test scores between the highest- and lowest-performing schoolwithin 5 kilometer is over 4 points. This variation in school charac-teristics, in combination with the fact that most schools are locatedin the city center (see Figure 1.1) where the municipality’s bus linesconnect, implies that most households could respond to the treatment

tracks are never a default option.


by applying to a higher-performing school without too much of a costin terms of increased travel distance.

Table 1.2: Summary Statistics for Households

Mean SD Min Max

Predetermined variablesForeign background 0.22 0.41 0.00 1.00High-skilled parents 0.66 0.47 0.00 1.00

Characteristics of default schoolsRaw test scores 13.16 1.43 9.75 16.10Adjusted test scores -0.41 0.65 -2.03 1.03Foreign background (share) 0.23 0.19 0.03 0.86High-skilled parents (share) 0.60 0.14 0.15 0.86

Within 5 km from homeNo. of schools 6.00 4.74 0.00 13.00Diff (max-min), raw test scores 4.05 2.30 0.00 6.35Diff (max-min), adj. test scores 2.12 1.09 0.00 3.06Diff (max-min), share foreign background 0.46 0.32 0.00 0.81Diff (max-min), share high-skilled parents 0.43 0.25 0.00 0.71

Observations 1608Notes: Raw test scores is the average score (on a scale from 0 to 20) on the standardizedtests in Swedish, mathematics and English in 9th grade during 2013-2014. Adjustedtest scores is the difference between the actual and expected score, where the expectedscore is the score predicted by the regression in equation (2.11). A student is consideredto have a foreign background if the student is born abroad or if both parents areborn abroad. A student is considered as having high-skilled parents if at least one ofthe parents have more than upper-secondary education. Diff indicates the differencesbetween the maximum and minimum value of the mentioned variables for all schoolslocated within five kilometers of the students’ home.

Finally, in Table 1.3, the school choice of households in the controlgroup is described. This illustrates the choices made by householdsnot provided with information about schools’ performances and isthe baseline to which the treated households’ choices will be com-pared. Panel a) presents the characteristics of the top-ranked schoolof all households in the control group. About half of them tries toopt out of their default school by listing another school as their top

32 CHAPTER 1

choice. In general, native and high-skilled households are more proneto opt out of their default schools compared to households with aforeign background and low-skilled households. They also apply tohigher-performing schools, schools with relatively more children ofhigh-skilled parents, and schools with relatively fewer children fromimmigrant households. When it comes to adjusted test scores, thepattern is not as clear. High-skilled households choose schools withbetter adjusted scores than low-skilled households, but householdswith foreign backgrounds choose schools with better adjusted scoresthan native households.

In Panel b), the difference between the top-ranked schools andthe default schools is reported for the subset of households trying toopt out of their default schools.33 Conditional on applying to anotherschool, all groups choose higher-performing schools (both raw and ad-justed) compared to their default school. For high-skilled and nativehouseholds, the difference is larger compared to low-skilled householdsand households with a foreign background. This is in line with the pre-vious findings that disadvantaged households are less likely to choosehigh-performing schools (Abdulkadiroğlu et al., 2018, Abdulkadiroğluet al., 2017, and Hastings et al., 2009). All groups do also, on aver-age, choose schools with a larger fraction of students with high-skilledparents and a lower fraction of students with a foreign background.Especially low-skilled and foreign households apply to schools withsmaller shares of immigrants than in their default school.

1.3.3 Empirical Specification

Estimating treatment effects in a randomized controlled trial isstraightforward. Let Ti be an indicator equal to one if household i

is assigned to the treatment group and zero otherwise. The effect

33The number if observations is lower than what would be implied by 50 percentof the 783-individual sample in panel a). This is explained by the lack of data ona few schools mentioned previously.


Table 1.3: Summary Statistics of School Choices for Households in theControl Group

All Native Foreign High- Low-skilled skilled

a) Characteristics of top choiceChoosing non-default 0.50 0.56 0.24 0.56 0.39Distance 4345 4840 2690 4439 4240Raw test scores 13.81 13.91 13.47 14.05 13.38Adj. test scores -0.21 -0.23 -0.13 -0.17 -0.29Share foreign 0.21 0.16 0.36 0.18 0.24Share high-skilled 0.67 0.69 0.61 0.70 0.62

Observations 783 615 166 510 265

b). Differences between top choice and default school)Distance 1956 2047 1296 1931 2037Raw test scores 0.90 0.92 0.78 0.89 0.92Adj. test scores 0.49 0.53 0.22 0.52 0.40Share foreign -0.07 -0.06 -0.13 -0.06 -0.11Share high-skilled 0.17 0.17 0.15 0.16 0.18

Observations 318 280 37 244 74Notes: Panel a) reports the characteristics of the top-ranked school for all stu-dents in the control group while panel b) reports the differences between thetop-ranked school and the default school for those listing a school other than thedefault as their top choice. Distance is the distance in meters measured as thecrow flies from the student’s home to the top-ranked school. Raw test scores isthe average score (on a scale from 0 to 20) on the standardized tests in Swedish,mathematics and English in 9th grade during 2013-2014. Adjusted test scores isthe difference between the actual and expected score, where the expected score isthe score predicted by the regression in equation (2.11). A student is consideredto have a foreign background if the student is born abroad or if both parents areborn abroad. A student is considered as having high-skilled parents if at leastone of the parents have more than upper-secondary education.

34 CHAPTER 1

of the treatment on an outcome Yi can then be estimated using alinear regression model:

Yi = α+ βTi + εi, (1.2)

where εi is a (heteroskedasticity robust) error term. β is the coef-ficient of interest, indicating the effect of the treatment on outcomeY .34 For β to have a causal interpretation, randomization of house-holds into the control and the treatment group needs to have beensuccessful. To confirm random assignment, the treatment indicator(Ti) is regressed on all variables included in Table 1.2. If random-ization was successful, these variables should not predict treatmentstatus. Table 1.4 shows the results from this regression, as well as thep-value from an F-test of joint significance. Only one out of eleven ofthe included predetermined variables is significant at the 10 percentlevel. The F-test confirms the null hypothesis, allowing the conclusionthat randomization was successful.

Another potential threat to identification is that information couldhave spread from treated households to households in the controlgroup. The time window for school choice is relatively short (lessthan three weeks), limiting the scope for information to spread, andthere was no mentioning of the experiment in local media. Anecdotalevidence from households contacting us also suggest that the informa-tion sent out was not subject to discussion between parents.35 Thereis no perfect test of information dispersion, but data from the 2014school choice program can be used to assess this further.

34As neither sampling nor assignment of treatment was clustered, clusteredstandard errors are not reported. See Abadie et al. (2017) for a discussion aboutwhen to use clustered standard errors. All statistical tests in this study have alsobeen run using randomization inference (see Young, 2018). Doing this does notaffect the results.

35Of 30 individuals contacting us regarding the information sent out, one haddiscussed it with another parent and one had heard another parent mentioningthe letter.


Table 1.4: Test of Randomization

Treated

Foreign background 0.0317(0.0381)

High-skilled parents -0.0101(0.0319)

Raw test scores (default school) 0.0440(0.0452)

Adj. test scores (default school) -0.0714(0.0513)

Share foreign background (default school) 0.0529(0.1370)

Share high-skilled parents (default school) -0.1870(0.3830)

No. of schools within 5 km -0.0085(0.0113)

Diff (max-min), raw test scores (schools within 5km) -0.0307(0.0437)

Diff (max-min), adj. test scores (schools within 5km) 0.0777∗(0.0471)

Diff (min-max), share foreign (school within 5km) -0.1050(0.2410)

Diff (min-max), share high-skilled (schools within 5km) 0.2860(0.2270)

Observations 1608P-value for F-test of joint significance 0.601

Notes: Standard errors in parenthesis. Significance levels indicated by *p < 0.1, ** p < 0.05, *** p < 0.01. Dependent variable is the treatmentindicator (Ti). Raw test scores is the average score (on a scale from 0 to 20)on the standardized tests in Swedish, mathematics and English in 9th gradeduring 2013-2014. Adjusted test scores is the difference between the actualand expected score, where the expected score is the score predicted by theregression in equation (2.11). A student is considered to have a foreign back-ground if the student is born abroad or if both parents are born abroad. Astudent is considered as having high-skilled parents if at least one of the par-ents have more than upper-secondary education. Diff indicates the differencesbetween the maximum and minimum value of the indicated variable for allschools located within five kilometers of the students’ home.

36 CHAPTER 1

Data on the school choices made by households with childrenabout to start 7th grade in Linköping in 2014 are available for compar-ison. If information did not spread from the treatment group to thecontrol group then the control group should have access to the sameinformation as those with children starting in 2014. Similar choicesbetween the two groups would therefore be expected. Nevertheless,time trends in the schools’ popularity over time could give rise to dif-ferences that have nothing to do with the presence of information onschool performances. The test is therefore not perfect, but observingsimilar choices for the two groups would be in line with the absence ofinformation dispersion. For each of the 16 schools available to choosein both 2014 and 2016, the change in the probability of being chosenis estimated by running the following regression:

Si = α+ β2016i + εi, (1.3)

where Si is an indicator variable taking the value of one if theschool in question was household i:s top-ranked school and zero oth-erwise. 2016i is an indicator variable equal to one if household i ispart of the control group in 2016 and zero otherwise and εi is a (het-eroskedasticity robust) error term. β is the difference in the shareof students ranking that school at the top of their application listbetween the 2016 control group and 2014 population. In Figure 1.3,the β for each school is plotted together with the confidence interval.The schools are ordered by their raw test scores (subfigure a) andtheir adjusted score (subfigure b). If information spread, it would beexpected that the β:s increase with school performance.

There is only one significant coefficient. One of the low-performingschools is considerably less chosen in the control group of the exper-iment compared to the general population of households 2014. Thecoefficient is very large, implying a reduction in the share of studentsapplying to this school by almost 80 percentage points. The schoolwent from over a hundred applicants in 2014 to only eleven in 2016.


One middle school in Linköping was evacuated in 2015 due to ex-tremely poor working conditions. The school was temporarily housedin the cellar of a nearby high school and local media reported thatstudents and teachers alike were fleeing the school. It is a reasonableto assume that this is the school in question, and that the evacuationrather than the treatment explains this extreme outlier. Except forthis single school, there is no sign of an downward slope (or indeed anyother pattern) that would suggest that information spread betweentreated and untreated households.

Figure 1.3: Change in the Schools’ Probability of Being Top-Ranked on theHouseholds’ Application Lists Between 2014 and 2016

(a) Schools ordered by raw testscores

-.1

-.05

0

.05

Cha

nge

in p

roba

bilty

1 3 5 7 9 11 13 15School performance rank

Beta 95% Confidence interval

(b) Schools ordered by adjustedscores

-.1

-.05

0

.05

Cha

nge

in p

roba

bilty

1 3 5 7 9 11 13 15School performance rank

Beta 95% Confidence interval

Notes: This figure presents the differences in the probability of being a household’s top-ranked schoolfor each of the 16 schools available in both the 2014 and 2016 school choice program. The differences inprobabilities are obtained from the estimated β:s in regression (1.3). Both subfigures present the samecoefficients but the schools are first ordered by their raw test scores (subfigure a) and then by theiradjusted score (subfigure b). Rank one indicates the highest-performing school.

38 CHAPTER 1

1.4 Results from the Experiment

In Table 1.5, effects on the extensive margin are reported in order toexamine whether households became more likely to try to opt out oftheir default school by applying to another school. The first columnshows that the coefficient on the share of households applying to otherschools is small and insignificant. Moreover, effects on the compositionof households applying to a school other than their default school arealso reported. Column two and three show no significant effects on theshare of high-skilled or foreign households among those trying to optout of their default school. Overall, Table 1.5 suggests that providingschool performance information does not impact participation in theschool choice program.

Table 1.5: Treatment Effects on the Extensive Margin

Total Households opting out

Opting out of Share Sharedefault school high-skilled foreign

Treated 0.0109 0.0251 0.0314(0.0250) (0.0307) (0.0227)

Constant 0.499∗∗∗ 0.732∗∗∗ 0.103∗∗∗(0.0179) (0.0225) (0.0155)

Observations 1608 812 810Notes: Significance levels indicated by * p < 0.1, ** p < 0.05, ***p < 0.01. The first column shows the effect of treatment on theshare of households ranking a school other than their default schoolas their top-choice. The second and third column shows the effectof treatment on the share of high-skilled/foreign households amongthose trying to opt out of their default school. A student is consid-ered to have a foreign background if the student is born abroad orif both parents are born abroad. A student is considered as havinghigh-skilled parents if at least one of the parents have more thanupper-secondary education.

Next, whether the treatment had any effect on the performance

1.4. RESULTS FROM THE EXPERIMENT 39

of the schools that households ranked on top of their application listsis studied. Figure 1.4 shows the cumulative distribution of the top-ranked schools for the control and treatment group separately, or-dered by the schools’ performances (with rank one indicating thehighest-performing school). In subfigure a), schools are ordered bytheir raw test scores and in subfigure b) by their adjusted score. Inboth, the gap between the groups are initially widening up until thefifth highest-performing schools with treated households being morelikely to apply to the top-performing schools. The gap between thedistributions then starts to diminish and closes at about the tenthschool in the performance rank. In other words, the treatment causea shift in applications from mid- to top-performing schools. The sizeof the gap is considerably smaller when ordering schools by their rawtest scores compared to the adjusted scores. Subfigure c) shows thesame, but with schools ordered by their average of the two previousperformance measures.36 The correlation between raw and adjustedtest scores is 0.75, and the relationship is displayed graphically insubfigure d).

Table 1.6 confirms and quantifies these results. In panel a),schools are ranked by their raw test scores and in panel b) bytheir adjusted scores. The schools are grouped into top-, mid- andlow-performing. Top-performing schools are the five schools withhighest (raw or adjusted) scores, mid-performing are the sixth totenth highest-performing schools, while the remaining schools arecategorized as low-performing. Applications to top-performingschools according to their adjusted score increases by about 5.2percentage points (corresponding to 12 percent). Applications tomid-tier schools decrease by 5.5 percentage points (corresponding to15 percent). The coefficient becomes smaller and insignificant whenschools are instead ranked according to their raw test scores. Panelc) corresponds to subfigure c) in Figure 1.4 with the schools ranked

36(Raw test scores rank+Adjusted test scores rank)/2.

40 CHAPTER 1

Figure 1.4: Cumulative Distribution of Top-choices, by School Perfor-mance Rank

(a) Raw test scores

0

.2

.4

.6

.8

1

Cum

ulat

ive

prob

.

1 5 9 13 17School

Control Treatment

(b) Adjusted test scores

0

.2

.4

.6

.8

1

Cum

ulat

ive

prob

.

1 5 9 13 17School

Control Treatment

(c) Average score

0

.2

.4

.6

.8

1

Cum

ulat

ive

prob

.

1 5 9 13 17School

Control Treatment

(d) Correlation

10

12

14

16

Raw

test

sco

res

-2 -1 0 1Adjusted test scores

Notes: Subfigure a) to c) show the cumulative distribution of the households’ top-ranked schools overthe schools’ performance rank, for the control and treatment group separately. In subfigure a), schoolsare ordered by their raw test scores, in subfigure b) by their adjusted scores and in subfigure c) by theaverage of these two performance measures, with rank one indicating the highest-performing school.Subfigure d) shows the correlation between raw test scores and adjusted test scores.


by their average performance and gives a similar picture as that ofpanel a).

Table 1.6: Treatment Effects on Performance of Top-Ranked Schools

Top- Middle- Low-performing performing performing

a) Schools ranked by raw test scoresTreated 0.0311 -0.0306 0.0043

(0.0247) (0.0205) (0.0216)Constant 0.559∗∗∗ 0.230∗∗∗ 0.248∗∗∗

(0.0178) (0.0151) (0.0155)

b) Schools ranked by adjusted test scoresTreated 0.0516∗∗ -0.0551∗∗ 0.0036

(0.0239) (0.0247) (0.0202)Constant 0.336∗∗∗ 0.459∗∗∗ 0.205∗∗∗

(0.0170) (0.0179) (0.0145)

c) Schools ranked by average performanceTreated 0.0522∗∗ -0.0557∗∗ 0.0036

(0.0248) (0.0238) (0.0202)Constant 0.420∗∗∗ 0.376∗∗∗ 0.205∗∗∗

(0.0177) (0.0174) (0.0145)Observations 1608 1608 1608

Notes: Standard errors in parenthesis. Significance levels indicatedby * p < 0.1, ** p < 0.05, *** p < 0.01. The dependent variablesTop-performing, Middle-performing and Low-performing are indi-cator variables that take the value one if a household’s top-rankedschool is categorized as top-performing (top 5), mid-performing (6-10) or low-performing (11+) and zero otherwise. The categorizationof schools into three performance groups is done using their averageperformance of raw and adjusted scores.

All in all, looking at the distributions in Figure 1.4 it seems likethe average rank of the schools is the best metric to use when evaluat-ing the effects of the treatment. The results from subfigure a) and b)combined with the results in Table 1.6 indicate that if one dimensionof the information provided was more important it was the informa-tion on adjusted performance of the schools. This could be because

42 CHAPTER 1

households have already taken the raw test scores into account whenmaking their choices, either because they actually have access to theinformation on raw test scores directly or because they have a goodenough proxy (such as student composition or reputation). It couldalso be because they are aware of the fact that the test scores ofa school is not necessarily a good measure of school quality. Theseresults indicate that the observed changes in school choices are notdriven by households looking for a stronger (in terms of observablecharacteristics) peer group, but rather something else.

To further understand this shift in applications, it is investigatedwhether the effect is driven by specific types of households. Table 1.7reports results for the subgroups. In panel a), focus is on high-skilledhouseholds. In this group, applications to top-performing schoolsincrease by about 8.1 percentage points and applications tomiddle-performing schools decrease by 6.7 percentage points. Inpanel b), where results for low-skilled households are reported,there are no significant effects. A test for the different treatmenteffects between the two groups is close to significant (p-value =0.116). In panel c) and d), the sample is instead cut by theirmigration background. Among native households applicationsto top-performing schools increase by about 7.6 percentagepoints. Applications to middle-performing schools decrease by 6.8percentage points. There are no significant effects for householdswith a foreign background and the difference in treatment effectsbetween these two groups is significant (p-value = 0.073).

The shift in applications from second-tier to top-performingschools could potentially increase the competitive pressure onschools to improve their performance. Note also that the presenceof strategic incentives in this school choice program may suppressthe impact of providing information about school performance onschool choices. Households may update their preferences for schoolswhen informed about school performance, but refrain from updating


Table 1.7: Treatment Effects on Performance of Top-ranked Schools, byHousehold Type

Top- Middle- Low-performing performing performing

a) High-skilledTreated 0.0808∗∗∗ -0.0669∗∗ -0.0139

(0.0308) (0.0289) (0.0225)Constant 0.476∗∗∗ 0.360∗∗∗ 0.163∗∗∗

(0.0222) (0.0213) (0.0164)

Observations 1050 1050 1050

b) Low-skilledTreated 0.0014 -0.0447 0.0433

(0.0400) (0.0419) (0.0393)Constant 0.314∗∗∗ 0.410∗∗∗ 0.276∗∗∗

(0.0288) (0.0305) (0.0277)


c) NativeTreated 0.0758∗∗∗ -0.0685∗∗ -0.0073

(0.0279) (0.0272) (0.0220)Constant 0.403∗∗∗ 0.407∗∗∗ 0.190∗∗∗

(0.0198) (0.0198) (0.0158)


d) ForeignTreated -0.0332 -0.0057 0.0388

(0.0541) (0.0476) (0.0484)Constant 0.481∗∗∗ 0.263∗∗∗ 0.256∗∗∗

(0.0396) (0.0349) (0.0346)

Observations 343 343 343Notes: Standard errors in parenthesis. Significance levels indicatedby * p < 0.1, ** p < 0.05, *** p < 0.01. See Table 1.6 for definitionsof the dependent variables and Table 1.5 for definitions of foreignbackground and high-skilled.

44 CHAPTER 1

their school choices if they think that the admission probability attheir now preferred school is low. In that sense, the effects shouldbe interpreted as a lower bound for the information-effect in astrategy-proof school choice program. Finally, the effects suggestthat the substitution from middle- to top-performing schools ismainly driven by native and high-skilled households. This meansthat it is mainly these schools that will face an increased competitivepressure to improve their performance, which may not benefit allstudents. Rather, disadvantaged students in low-performing schoolsmight get left behind.

The finding that well-off households respond to more sophisti-cated performance measures by switching their applications to top-performing schools has to our knowledge not been documented previ-ously. Still, it is consistent with the findings in Hart and Figalo (2015)and the theoretical predictions of Walters (2018). One question thatremains, however, is how to reconcile these results with the signif-icant response by disadvantaged households found by Hastings andWeinstein (2008) and Gallego et al. (2012). A first hypothesis focuson the differences between the contexts in which these studies wereconducted. Sweden is a more equal society than both Chile and theUS. The disadvantaged households in Linköping are relatively well-offcompared to the disadvantaged households in their studies, suggestingthat they might not be as uninformed. In other words, the treatmentmight have added less new information in this study compared totheirs.37 There is also a difference in the type of information that wasprovided. Unlike previous studies, the letter sent out in this studyincluded not only raw test scores but also a performance measuretaking into account the peer composition of the schools. It is possi-ble that this made the information harder to understand, especiallyfor disadvantaged households, and therefore had less of an impact on

37For example in Hastings and Weinstein (2008) only 11 percent of the studentsopt out of their default school before receiving information. In the sample analyzedin this study, the corresponding number is 24 percent.


school choices (as discussed in Imberman and Lovenheim, 2016).

46 CHAPTER 1

1.5 General Equilibrium Effects

This section first describes the simulation strategy and then describesthe general equilibrium effects under the assumption that householdswould have had access to the same information about school perfor-mances. The focus is on how the observed change in school choicesimpacts the assigned schools of all students and what this implies forthe level of school segregation.

1.5.1 Simulation Strategy

While understanding whether and how households react to schoolperformance information is important, an equally important questionis how the households’ responses to such information impacts whichschools children are assigned to. As schools usually have a fixed num-ber of students they can admit, which school a student is assignedto depends on both the own school choice(s) as well as the choices ofothers. The increased demand for top-performing schools among na-tive and high-skilled households may therefore very well impact theassigned schools of children in other households as well and capac-ity constraints means that increased demand for certain schools maynot be fully accommodated. To explore this, a simulation exerciseis conducted. A simpler strategy would be to compare the assignedschools of students in the control group to those of students in thetreatment group. This would however not represent the general equi-librium effects as only half of the population were informed about theschools’ performances. A comparison of the assigned schools betweenthese two groups therefore runs the risk of overestimating the effectsof providing school performance information on school assignments.

Instead, knowledge of the assignment mechanism and priorityrules used in this school choice program is used to simulate the as-signment of students to schools under the assumptions that i) nohousehold had access to school performance information and ii) all

1.5. GENERAL EQUILIBRIUM EFFECTS 47

households had access to school performance information. Focus isstill on the households’ top-ranked choice in order to keep the anal-ysis consistent throughout the study. Furthermore, the analysis isrestricted to the direct (or short-term) effect on school assignments.Any potential effects that the initial reallocation of students may haveon the school choices of households in years to come is thus left forfuture work.

In order to simulate the allocations of students to schools in thesetwo scenarios, each school’s capacity is approximated by the numberof students with the school as their default or the number of studentsthat were assigned to the school in 2016 (whichever is higher).38 Next,a simulation sample the size of the observed cohort is drawn withreplacement from either the control or treatment group. Students inthe simulation sample are placed at their default school, as observedin the data. All students are then allowed to apply to their top-rankedschool, again, as observed in the data. Students are accepted to theirtop-ranked school in order of priority if that school is not alreadyat full capacity. A student’s priority is determined by the distancebetween the students home and the school, unless the school is avoucher school for which a random number is used. Students notaccepted at their top-choice is referred back to their default school,at which they always have top priority.

As some students are accepted to their top-ranked school, theyleave empty seats behind at their default school. Other students mighthave ranked those on top of their application lists. The capacity of allschools is therefore adjusted and students who were rejected at theirtop choice are allowed to apply again. This process is repeated until nostudents are placed in new schools. At this point, the final allocationof students to schools is reached and the outcome measures of interestare documented. In the next section, the distribution of these outcome

38This means that the capacity may be underestimated for undersubscribedschools. As these are the least popular schools, it does not cause concern.

48 CHAPTER 1

measures using 1,000 independently drawn simulation samples willbe presented. The differences in outcomes between the simulationsamples drawn from the control group and the treatment group canbe interpreted as the effect on school assignments, had all householdsbeen given the same information on school performance. The first setof outcomes are related to how the students are distributed acrosstop-, middle- and low-performing schools. The next set of outcomesmeasures the level of school segregation by migration background andparental skill-level.

Segregation between groups divided into organizational units(such as schools) can be studied using either measures of evennessor measures of exposure (Massey and Denton, 1988). Exposuremeasures are sensitive to the share of minority students in thepopulation, something that evenness measures are generally not. AsAllen and Vignoles (2007) point out, the share of minority studentsin the population is not affected by educational policy. Therefore,a measure of evenness will be used to evaluate how segregationis impacted. Following Massey and Denton (1988), the impact ismeasured using the Duncan Dissimilarity Index (DDI), defined as:

DDI = 12

N∑j=1| aja− bj

b|, (1.4)

where N is the number of schools, aj is the number of studentsfrom group A in school j, a is the total number of students fromgroup A, bj is the number of students from group B in school j andb is the total number of students from group B. The DDI rangesfrom zero, no segregation, to one, total segregation. It has a clearinterpretation as the percentage of one of the two groups that wouldhave to move to a different organizational unit (school) in order toproduce a distribution in each organizational unit that matches thedistribution of the entire population (Duncan and Duncan, 1955).

The DDI is however not without it’s drawbacks. Most notably,


it fails the transfer principle (James and Taeuber, 1985 and White,1986). This means that it is insensitive to the redistribution of minor-ity group members among organizational units with minority propor-tions above or below the overall minority proportion. Only transfersof minority members from units where they are overrepresented tounits where they are underrepresented (or vice versa) affect the valueof the index. Therefore, effects are also reported using Theil’s EntropyIndex. This index was originally proposed by Theil and Finizza (1971)and is another measure of evenness for which the transfer principleholds. Theil’s Entropy Index (TEI) is given by:

TEI = 1N

N∑i=1

xjµln(xj

µ), (1.5)

where N is the number of schools, xj is a characteristic of schoolj (such as the share of students with a foreign background assignedto school j) and µ is the mean of x for all schools in N . The TEI canbe interpreted as the difference between the diversity (entropy) of thesystem and the weighted average diversity of individual organizationalunits, expressed as a fraction of the total diversity of the system(Reardon and Firebaugh, 2002).

1.5.2 Effects on School Assignments

Figure 1.5 shows the distribution of simulation results for the shareof students admitted to top-, middle- and low-performing schools forthe control and treatment group separately. In subfigure a), the dis-tribution shifts to the right for the treatment group, indicating that alarger share assigned to top-performing schools. The magnitude rep-resents an average increase in the enrollment rate in these schools by2.1 percentage points (from 36.6 percent when sampling from the con-trol group to 38.7 percent when sampling from the treatment group).This increase corresponds to about 40 percent of the increased de-

50 CHAPTER 1

mand for the top-performing schools documented in Table 1.6. Thisimplies empty seats in the top-performing schools initially, but notenough to fully accommodate the increased demand due to provi-sion of school-performance information. Instead, these schools becomeoversubscribed and some applicants have to be rejected. In subfig-ure b), the leftward shift of the distribution implies a 1.9 percentagepoint drop in the share of students assigned to the middle-performingschools (from 36.1 percent to 34.2 percent). In subfigure c) it can beseen that assignment to the low-performing schools is not affectedmuch. Not surprisingly, as students are reallocated from mid- to top-performing schools, the rank by average performance of the assignedschool is improved by 0.1 points (subfigure d).

In Figure 1.6, the effects on being assigned to a top-performingschool are reported for each subgroup separately. Subfigure a) showsthat the share of students with high-skilled parents assigned top-performing schools increases by 3.3 percentage points (from 41.8 per-cent to 45.1 percent). In subfigure c), a similar increase is seen forstudents with a native background (3.4 percentage points increasefrom 34.2 percent to 37.6 percent). This is expected as it is for thesehouseholds that a significant effect on school choices is documented.Subfigure b) shows that the share of students with low-skilled parentsin the top-performing schools is unaffected. Subfigure d) shows thatthe share of students with foreign background in the top-performingschools decreases by 2.3 percentage points (from 46.1 percent to 43.8percent). All students with foreign background with a top-performingschool as their default school in the treatment group are assigned toa top-performing school. This implies that the decrease in the shareof students with a foreign background assigned to a top-performingschool is due to them being displaced by the increased pressure fromother students applying to these schools.

Figure 1.7 shows the simulation results on the effect on the aver-age performance-rank of the assigned school for different subgroups.


Figure 1.5: Distribution of Simulated Shares Assigned to Top-, Middle-and High-Performing Schools

(a) Top-performing

0

2

4

6

8

Perc

ent

0.34 0.36 0.38 0.40 0.42Share

Control Treatment

(b) Mid-performing

0

2

4

6

8

Perc

ent

0.30 0.32 0.34 0.36 0.38 0.40Share

Control Treatment

(c) Low-performing

0

2

4

6

Perc

ent

0.24 0.26 0.28 0.30 0.32Share

Control Treatment

(d) Rank of assigned school

0

2

4

6

Perc

ent

7.20 7.40 7.60 7.80 8.00Rank

Control Treatment

Notes: Subfigure a) to c) presents the simulation results on the share of students in the control andtreatment group assigned to top-, middle-, and low-performing schools respectively. Subfigure d) presentthe average rank of the assigned school, where rank one indicates the highest-performing school asmeasured by the school’s average of the raw and adjusted test score. The simulation strategy is describedin detail in Section 1.5.1.

52 CHAPTER 1

Figure 1.6: Distribution of Simulated Shares Assigned to a Top-PerformingSchool, by Household Type

(a) High-skilled

0

1

2

3

4

Perc

ent

0.38 0.40 0.42 0.44 0.46 0.48Share

Control Treatment

(b) Low-skilled

0

1

2

3

4

Perc

ent

0.20 0.25 0.30 0.35Share

Control Treatment

(c) Native

0

2

4

6

Perc

ent

0.32 0.34 0.36 0.38 0.40Share

Control Treatment

(d) Foreign

0

1

2

3

4

5

Perc

ent

0.35 0.40 0.45 0.50 0.55Share

Control Treatment

Notes: This figure presents the simulation results on the share of students in the control and treatmentgroup assigned to top-performing schools, by household type. A student is considered to have a foreignbackground if the student is born abroad or if both parents are born abroad. A student is consideredas having high-skilled parents if at least one of the parents have more than upper-secondary education.The simulation strategy is described in detail in Section 1.5.1.


In subfigure a) and c), it can be seen that students of high-skilled par-ents and native students improve in terms of the average performancerank of their assigned school. Subfigure b) reports the results for stu-dents with low-skilled parents, showing that the average performance-rank of their assigned school is not affected. This is expected sinceno change was seen in their school choices. Neither was there anyevidence of them being displaced from the top-performing schools byhigh-skilled households. Subfigure d) reports the results for studentswith a foreign background.They are, on average, assigned to lower-ranked schools. All in all, the additional performance information inthe letter increases the difference in performance between the assignedschool of students from high-skilled and low-skilled households. Fornative and foreign households, the performance gap is reversed so thatstudents with a foreign background are assigned to schools that arelower performing (according to the average of their raw and adjustedscores).

1.5.3 Effects on School Segregation

Figure 1.8 shows the simulation results on school segregation. Subfig-ure a) suggests that the level of segregation by parental skill-level isabout 0.26-0.27 as measured by the DDI, implying that about 26-27percent of the students in either group (high- or low-skilled) wouldhave to switch school to reach a perfectly desegregated school system.The fact that the two distributions are almost completely overlap-ping means that providing information on school performances didnot impact how segregated students are in terms of their parents’educational level.39 However, when measuring segregation using TEIinstead, a slight increase in student sorting by parental skill-level isobserved (see subfigure b).

These results can be reconciled by considering the definition of39The difference in average DDI between the control and treatment group is

0.006.

54 CHAPTER 1

Figure 1.7: Distribution of Simulated Average Performance-Rank of As-signed School, by Household Type

(a) High-skilled

0

2

4

6

8

Perc

ent

6.50 7.00 7.50Rank

Control Treatment

(b) Low-skilled

0

2

4

6

Perc

ent

8.00 8.50 9.00 9.50Rank

Control Treatment

(c) Native

0

2

4

6

8

10

Perc

ent

7.00 7.20 7.40 7.60 7.80 8.00Rank

Control Treatment

(d) Foreign

0

1

2

3

4

5

Perc

ent

6.50 7.00 7.50 8.00 8.50Rank

Control Treatment

Notes: This figure presents the simulation results on the average performance-rank of the assignedschool, by household type. Rank one indicates the highest-performing school as measured by the school’saverage of the raw and adjusted test score. A student is considered to have a foreign background ifthe student is born abroad or if both parents are born abroad. A student is considered as having high-skilled parents if at least one of the parents have more than upper-secondary education. The simulationstrategy is described in detail in Section 1.5.1.


these segregation indexes and the composition of students already en-rolled in the schools. The top-performing schools have above-averageshares of students with high-skilled parents (74.8 percent comparedto 56.3 percent in the other schools). When children to high-skilledparents switch from other schools to these, they become even moreclustered at the top-performing schools. This generates an increase inthe TEI which considers the entropy of the whole system. The DDIchanges only when individuals switch from a unit where they are over-represented to a unit where they are underrepresented (or vice versa).The absence of a change in the DDI therefore tells us that the childrenof high-skilled parents switching to the top-performing schools wouldotherwise have ended up in schools where many of their peers wouldalso have had highly educated parents.

Subfigure c) and d) display similar results for segregation by for-eign background. A considerable drop in the DDI is noted when sam-pling from the treatment group instead of the control group (subfig-ure c). In other words, the simulations indicate that providing schoolperformance information reduces sorting by foreign background. Themagnitude is about 6 percentage points, a drop from 0.44 to 0.38,corresponding to a 14 percent decrease. Looking at the results sub-figure d), when measuring segregation by the TEI instead, confirmsthis reduction in school segregation. In fact, this drop is large enoughto push down school segregation below the level that would prevail ifall students been assigned their default school (0.41 using the DDI).

To understand the mechanism behind the reduction in school seg-regation by foreign background, an important piece of information isthe student composition at the top-performing schools. These schoolshave relatively many students with foreign backgrounds, comparedto the other schools in the district.40 The information on school per-formances provided increased demand for these schools among native

40The top-performing schools have, on average, 23 percent children with foreignbackgrounds. The corresponding number for the mid-performing schools is 13.

56 CHAPTER 1

Figure 1.8: Distribution of Simulated Effects on School Segregation

(a) Parental skill-level (DDI)

0

1

2

3

4

Perc

ent

0.15 0.20 0.25 0.30 0.35Duncan Dissimilarity Index

Control Treatment

(b) Parental skill-level (TEI)

0

1

2

3

4

Perc

ent

0.04 0.06 0.08 0.10 0.12 0.14Theils Entropy Index

Control Treatment

(c) Foreign background (DDI)

0

1

2

3

4

Perc

ent

0.30 0.35 0.40 0.45 0.50 0.55Duncan Dissimilarity Index

Control Treatment

(d) Foreign background (TEI)

0

1

2

3

4

5

Perc

ent

0.10 0.15 0.20 0.25 0.30Theils Entropy Index

Control Treatment

Notes: This figure displays the simulated effect on school segregation by parental skill-level and foreignbackground. Subfigure a) and c) measures school segregation using the Duncan Dissimilarity Indexand subfigure b) and d) uses Theil’s Entropy Index. The simulation strategy is described in detail inSection 1.5.1.


households and the simulations presented in Figure 1.6 suggested anincreased enrollment rate of native children in these schools. This hasa desegregating effect on the schools in this district, but the increasein the share of native children enrolled in the top-performing schoolsof 3.4 percentage points can only explain slightly more than half ofthe simulated reduction by 6 percentage points in school segregation(as measured by the DDI).

Given that the school choices of households with foreign back-grounds were not significantly impacted by the information providedon school performances, the remaining reduction in school segregationis instead explained by how the initial effect propagates through thesystem as children are assigned to schools. The children that switchinto the top-performing schools may displace other students and alsoleave empty seats behind them in the schools they would otherwisehave attended. Subfigure a) of Figure 1.9 shows whether the schoolsare over- or undersubscribed. Two of the mid-performing schools (per-formance rank 6-10) have more applicants than seats. The studentcomposition of these schools may thus be affected if the compositionof those waiting in line to be admitted is different from those leavingfor the top-performing schools. Further, two of the top-performingschools were also oversubscribed implying that if new applicants areadmitted to these schools, they necessarily push other students out.This is in line with subfigure d) of Figure 1.6, with fewer foreignstudents assigned to the top-performing schools. To explain the re-maining reduction in school segregation, subfigure b) of Figure 1.9shows how that the share of students with foreign background in-creased by 3.8 percentage points (from 7.4 to 11.2 percent) at the twooversubscribed mid-performing schools. This effect combined with theincreased share of native students in the top-performing schools cor-responds well to the total observed decrease in the DDI.

The mechanism behind this reduction in school segregation is notthe one usually envisioned with disadvantaged students taking seats at

58 CHAPTER 1

Figure 1.9: School Subscription Rate and Change in Student Composition

(a) Subscription rate

0

.5

1

1.5

Subs

crip

tion

rate

1 5 9 13 17Average score

(b) Share foreign in oversubscribedmid-performing schools

0

1

2

3

4

Perc

ent

0.00 0.05 0.10 0.15Share

Notes: Subfigure a) shows the subscription rate (number of applicants divided by the number of seats)for all schools ordered by their performance rank. Rank one indicates the highest-performing school asmeasured by the school’s average of the raw and adjusted test score. Subfigure b) shows the simulatedshare of students with foreign background in the two mid-performing schools that are oversubscribedaccording to subfigure a). A student is considered to have a foreign background if the student is bornabroad or if both parents are born abroad. The simulation strategy is described in detail in Section1.5.1.


high-performing schools with favorable student compositions. Rather,it is driven by a shift in applications from native students to high-performing schools attended by many students with foreign back-ground. The native students leave empty seats at schools that wereattended mainly by native students, which are then taken by studentswith a foreign background who previously applied to, but were notaccepted to, these schools or by foreign students that are no longeradmitted to the top-performing schools. It should therefore be notedthat these results on segregation do not necessarily generalize to othersettings. In a case where the top-performing schools have a low shareof students with a foreign background the effects could be smalleror, depending on the pre-existing application pattern and the insti-tutional setting, potentially even reversed.

To examine the extent to which these results are generalizable, therelationship between school performance and the share of studentswith foreign backgrounds is examined. Using data from the SwedishNational Agency of Education, schools are assigned a percentile rankwithin their municipality based on their adjusted performance mea-sure.41 Municipalities with less than ten middle schools are excluded(leaving 32 municipalities for the analysis). Figure 1.10 in Appendix1.B displays how the percentile rank of their adjusted performancescore covaries with the share of foreign-born students as well as theshare of students with foreign backgrounds. The pattern observed inthis school district, with top-performing schools having more foreignstudents than mid-performing schools, seems to be general. This im-plies that results in this section may generalize to other school districtsin Sweden and school districts elsewhere, where foreign students areoverrepresented at the best-performing schools.

The impact on school segregation was presented using two dif-ferent indexes, both belonging to the group of evenness-measures.

41Adjusted performance is defined as in the letter provided to households in thetreatment group but, due to data limitations, using grades instead of test scores.

60 CHAPTER 1

As Carrington and Troske (1997) show, these types of segregation-measures are sensitive to randomness when the average size of theorganizational unit is small. Because the size of the organizationalunits, or schools, in this study is the same in the two scenarios com-pared, this should not pose a problem. Still, Figure 1.11 in Appendix1.B presents simulation results when allocating students to schoolsrandomly, with sampling from the control and treatment group re-spectively. The absence of a shift in the distribution confirms thatthe impact mentioned above does not arise due to randomness. Fur-ther, to confirm that the results are not driven by the properties ofthe evenness-measures, the results are replicated using the Isolation-index. The properties of this index is described in detail in Lieberson(1981), and belongs to the group of exposure-measures. The outcomeof this exercise is also presented in Figure 1.12 Appendix 1.B andconfirms the conclusions regarding how school segregation would beaffected.

1.6. CONCLUDING REMARKS 61

1.6 Concluding Remarks

This study reports results from a randomized controlled trial whereinformation on raw and adjusted (for student composition) test scoreswas distributed to households about to choose a middle school in aSwedish municipality. The analysis show that providing informationon school performances impact which schools households apply to.The effects are heterogeneous and mainly driven by native and high-skilled households. Furthermore, the effect is local in the sense thatit shifts applications from mid- to top-performing schools withouthaving an impact on applications to low-performing schools. The in-formation provided to households included both raw test scores andscores adjusted for the student composition at the school. Householdsseem to react mainly to the latter, suggesting that they are not ad-justing their applications because they are looking for schools with astronger (observable) peer composition. All in all, this indicates thatschool performance information can have an effect on the competitivepressure on schools to improve performance. However, given that theeffects are restricted to certain groups and certain schools, everyonemight not benefit from this increased competition, as disadvantagedstudents in low performing schools risk being left behind.

Using simulations, general equilibrium effects are studied. Thedocumented changes in school choices translate into effects on theassignment of the households that did alter their school choices. Ca-pacity constraints do however seem to mute this effect. This suggeststhat it is important to couple information interventions with a readi-ness to expand capacity at the schools for which demand increasesto the extent possible. The shift in applications for high-skilled andnative households also seems to displace students with a foreign back-ground from the top-performing schools. Therefore, the interventionis not without drawbacks, at least as long as equality is striven for, asit increases the school performance gap between advantaged and dis-advantaged students. Finally, the initial effects on assignment propa-

62 CHAPTER 1

gates through the system as capacity frees up, thereby also potentiallybenefiting students that did not react to the information. Throughthis mechanism school segregation in terms of migration backgroundis decreased.

REFERENCES 63

References

Abadie, A., Athey, S., Imbens, G. W., andWoolridge, J. (2017). Whenshould you adjust standard errors for clustering? NBER WorkingPaper No 24003. National Bureau of Economic Research, Inc.

Abdulkadiroğlu, A., Agarwal, N., and Pathak, P. A. (2017). TheWelfare Effects of Coordinated Assignment: Evidence from theNew York City High School Match. American Economic Review,107(12):3635–3689.

Abdulkadiroğlu, A., Pathak, P. A., and Walters, C. R. (2018). Free toChoose: Can School Choice Reduce Student Achievement? Ameri-can Economic Journal: Applied Economics, 10(1):175–206.

Abdulkadiroğlu, A. and Sönmez, T. (2003). School Choice: A Mech-anism Design Approach. American Economic Review, 93(3):729–747.

Allen, R. and Vignoles, A. (2007). What should an index of schoolsegregation measure? Oxford Review of Education, 33(5):643–668.

Bifulco, R. and Ladd, H. F. (2007). School Choice, Racial Segrega-tion, and Test-Score Gaps: Evidence from North Carolina’s Char-ter School Program. Journal of Policy Analysis and Management,26(1):31–56.

Böhlmark, A., Holmlund, H., and Lindahl, M. (2016). ParentalChoice, Neighbourhood Segregation or Cream Skimming? An Anal-ysis of School Segregation after a Generalized Choice Reform. Jour-nal of Population Economics, 29(4):1155–1190.

Böhlmark, A. and Lindahl, M. (2015). Independent Schools and Long-run Educational Outcomes - Evidence from Sweden’s Large-scaleVoucher Reform. Economica, 82(327):508–551.

64 CHAPTER 1

Burgess, S., Greaves, E., Vignoles, A., and Wilson, D. (2015). WhatParents Want: School Preferences and School Choice. The Eco-nomic Journal, 125(587):1262–1289.

Carrington, W. J. and Troske, K. R. (1997). On Measuring Segrega-tion in Samples with Small Units. Journal of Business & EconomicStatistics, 15(4):402–409.

Corcoran, S. P., Jennings, J. L., Cohodes, S. R., and Sattin-Bajaj, C.(2018). Leveling the Playing Field for High School Choice: Resultsfrom a Field Experiment of Informational Interventions. NBERWorking Paper No 24471. National Bureau of Economic Research,Inc.

Cullen, J. B., Jacob, B. A., and Levitt, S. (2006). The Effect ofSchool Choice on Participants: Evidence from Randomized Lotter-ies. Econometrica, 74(5):1191–1230.

Duncan, O. D. and Duncan, B. (1955). A Methodological Analysis ofSegregation Indexes. American Sociological Review, 20(2):210–217.

Figalo, D. N. and Lucas, M. E. (2004). What’s in a Grade? School Re-port Cards and the Housing Market. American Economic Review,94(3):591–604.

Fiva, J. H. and Kirkeboen, L. J. (2011). Information Shocks and theDynamics of the Housing Market. The Scandinavian Journal ofEconomics, 113(3):525–552.

Gallego, F. A., Lagos, F., and Stekel, Y. (2012). Types of Informationand School Choice: An Experimental Study in the Chilean VoucherSystem. url: http://www.webmeets.com/files/papers/LACEA-LAMES/2012/735/preliminary%20120312.pdf (accessed 1 April2018).

REFERENCES 65

Hart, C. M. D. and Figalo, D. N. (2015). School Accountability andSchool Choice: Effects on Student Selection Across Schools . Na-tional Tax Journal, 68(3S):875–900.

Hastings, J. S., Kane, T. J., and Staiger, D. O. (2009). Heteroge-neous Preferences and the Efficacy of Public School Choice. url:http://people.virginia.edu/ sns5r/microwkshp/hastings.pdf (ac-cessed 1 April 2018).

Hastings, J. S. and Weinstein, J. M. (2008). Information, SchoolChoice, and Academic Achievement: Evidence from Two Experi-ments. Quarterly Journal of Economics, 123(4):1373–1414.

Hoxby, C. (2000). Does Competition Among Public Schools Ben-efit Students and Taxpayers? The American Economic Review,90(5):1209–1238.

Hoxby, C. (2007). Does Competition Among Public Schools BenefitStudents and Taxpayers? Reply. The American Economic Review,97(5):2038–2055.

Hsieh, C.-T. and Urquiola, M. (2006). The Effects of GeneralizedSchool Choice on Achievement and Stratification: Evidence fromChile’s school Voucher Program. Journal of Public Economics,90(8-9):1477–1503.

Hussain, I. (2013). Not Just Test Scores: Parents’ Demand Re-sponse to School Quality Information. url: http://www.sole-jole.org/13363.pdf (accessed 1 April 2018).

Imberman, S. A. and Lovenheim, M. F. (2016). Does the Market ValueValue-added? Evidence from Housing Prices after a Public Releaseof School and Teacher Value-added. Journal of Urban Economics,91(C):104–121.

66 CHAPTER 1

James, D. R. and Taeuber, K. E. (1985). Measures of Segregation.Sociological Methodology, 15:1–32.

Kane, T. J. and Staiger, D. O. (2002). The Promise and Pitfalls ofusing Imprecise School Accountability Measures. Journal of Eco-nomic Perspectives, 16(4):91–114.

Koning, P. and van der Wiel, K. (2013). Ranking the Schools: HowSchool-Quality Information Affects School Choice in the Nether-lands. Journal of the European Economic Association, 11(2):466–493.

Lavy, V. (2010). Effects of Free Choice Among Public Schools. TheReview of Economic Studies, 77(3):1164–1191.

Lavy, V. (2015). The Long-Term Consequences of Free School Choice.NBER Working Paper No. 20843. National Bureau of EconomicResearch, Inc.

Lépine, A. (2015). School Reputation and School Choice in Brazil: aRegression Discontinuity Design. Working Papers, Department ofEconomics 2015-38, University of Sï£¡o Paulo (FEA-USP).

Lieberson, S. (1981). A Piece of the Pie: Blacks and White Immigrantssince 1880. University of California Press, Oakland, California.

Malmberg, B., Andersson, E. K., and Bergsten, Z. (2014). CompositeGeographical Context and school Choice Attitudes in Sweden: AStudy Based on Individualls Defined, Scalable Neighborhoods. An-nals of the Association of American Geographers, 104(4):869–888.

Massey, D. S. and Denton, N. A. (1988). The Dimensions of Residen-tial Segregation. Social Forces, 67(2):281–315.

Mizala, A. and Urquiola, M. (2013). School Markets: The Impactof Information Approximating Schools’ Effectiveness. Journal ofDevelopment Economics, 103(C):313–335.

REFERENCES 67

Musset, P. (2012). School Choice and Equity: Current Poli-cies in OECD Countries and a Literature Review. OECD Ed-ucation Working Papers, No. 66. OECD Publishing, Paris,url: http://dx.doi.org/10.1787/5k9fq23507vc-en (accessed 1 April2018).

Reardon, S. F. and Firebaugh, G. (2002). Measures of MultiGroupSegregation. Sociological Methodology, 31(1):33–67.

Rothstein, J. (2007). Does Competition Among Public Schools Ben-efit Students and Taxpayers? Comment. The American EconomicReview, 97(5):2026–2037.

Ruijs, N. M. and Oosterbeek, H. (2017). School Choice in Amsterdam:Which Schools are Chosen when School Choice is Free? EducationFinance and Policy, 14(1):1–64.

Söderström, M. and Uusitalo, R. (2010). School Choice and Segre-gation: Evidence from an Admission Reform. The ScandinavianJournal of Economics, 112(1):55–76.

Theil, H. and Finizza, A. J. (1971). A Note on the Measurement ofRacial Integration of Schools by Means of Informational Concepts.The Journal of Mathematical Sociology, 1(2):187–193.

Walters, C. R. (2018). The Demand for Effective Charter Schools.Journal of Political Economy, 126(6):2179–2223.

White, M. J. (1986). Segregation and Diversity Measures in Popula-tion Distribution. Population Index, 52(2):198–221.

Young, A. (2018). Channeling Fisher: Randomization Tests and theStatistical Insignificance of Seemingly Significant Experimental Re-sults. The Quarterly Journal of Economics. Forthcoming. url:https://doi.org/10.1093/qje/qjy029.

68 CHAPTER 1

Appendices

1.A Complete Letter

2 (3)

Further Information Regarding the School Choice

You have recently, or will soon, receive a letter from Linköping municipality

informing you that it is time to choose a school for your child. As we are studying

school choice, we would like to give you some additional information that we hope

can be of use.

In the attached table you can see how the schools in Linköping municipality have

performed on the standardised tests in grade nine in recent years. The first column

shows the average score on the tests. In the second column we have calculated how

the students of the school have performed on these standardised tests compared to

schools that are similar in terms of the students' educational and migration

background.

If you have any questions feel free to contact us.

Dany Kessel Elisabet Olme

[email protected] [email protected]

0707173803 0735442220

Notes: Contact details blacked out.

1.A. COMPLETE LETTER 69

1 (1)

Skola

School

Genomsnittspoäng

Average Score

(max=20)

Över/underprestation

Over/underperformance

(poäng/points)

Arenaskolan - -

Berzeliusskolan 14,4 +0,2

Björkö Friskola 13,2 -1,2

Dar al Uloum - -

Ekholmsskolan 12,6 -1,0

Elsa Brändströms Skola - -

Folkungaskolan 15,4 0

Frösunda Pandionskolan - -

Internationella Engelska

Skolan

15,7 +0,6

Klara Privata Grundskola 13,5 +0,2

Kungsbergsskolan 11,7 -1,4

Kunskapsskolan Linköping 11,7 -2,0

Linghemsskolan 12,8 -0,8

Ljungsbro Skola 12,5 -0,8

Malmslättskolan Tokarp 13,7 -0,3

Nya Munken 14,2 -0,5

Skäggetorpsskolan 9,8 -0,8

Tornhagsskolan 13,1 -0,1

Vittra Lambohov 16,1 +1,0

Ånestadsskolan 13,0 +0,1

Såhär läser du tabellen: Om det står 15 i kolumnen ”Genomsnittspoäng” och +0,2

i kolumnen ”Över/underprestation” betyder det att genomsnittspoängen för skolans

elever var 15 och att detta var 0,2 poäng mer än genomsnittet på skolor med

liknande elever.

How to read the table: If it says 15 in the column “Average performance” and

+0,2 in the column “Over/underperformance”, it means that the students average

score at the school was 15, and that this was 0,2 points higher than the average

score at schools with similar students.

Källa/Source: Skolverket/The National Agency for Education.

70 CHAPTER 1

1.B Additional Figures

Figure 1.10: Adjusted Performance and Share of Foreign Students inSwedish Compulsory Schools

(a) Foreign born

5

10

15

20

25

30

Shar

e

0 .2 .4 .6 .8 1Percentile performance in municipality

(b) Foreign background

5

10

15

20

25

30

Shar

e

0 .2 .4 .6 .8 1Percentile performance in municipality

Notes: This figure shows the relationship between the schools adjusted performance and share of foreignborn students (subfigure a) or students with a foreign background (subfigure b). The adjusted perfor-mance measure is calculated using grades from year 9, adjusted for the student composition. Schoolsare then assigned a percentile rank within municipality.

1.B. ADDITIONAL FIGURES 71

Figure 1.11: Distribution of Simulated Effects on School Segregation UnderRandom Allocation

(a) Parental skill-level

0

1

2

3

4

Perc

ent

0.04 0.06 0.08 0.10 0.12 0.14Duncan Dissimilarity Index

Control Treatment


0

1

2

3

4

Perc

ent

0.05 0.10 0.15 0.20Duncan Dissimilarity Index

Control Treatment

Notes: This figure shows the simulated effect on school segregation measured using the Duncan Dissim-ilarity Index, when students are randomly allocated to schools. The simulation strategy is described indetail in Section 1.5.1.

Figure 1.12: Distribution of Simulated Effects on School Segregation, Usingthe Isolation Index

(a) Parental skill-level

0

2

4

6

8

Perc

ent

0.66 0.68 0.70 0.72 0.74Isolation Index

Control Treatment


0

2

4

6

Perc

ent

0.30 0.35 0.40 0.45Isolation Index

Control Treatment

Notes: This figure shows the simulated effect on school segregation using the Isolation Index. Thesimulation strategy is described in detail in Section 1.5.1.

72 CHAPTER 1

Chapter 2

School Choice, Admission Rulesand Segregation in PrimarySchools∗

∗This chapter was co-authored by Dany Kessel. We thank Niklas Blomqvist,Chris Neilson, David Seim, David Strömberg, Jonas Vlachos, and seminar par-ticipants at Stockholm University, the Stockholm-Uppsala Education EconomicsWorkshop, Columbia Business School, Columbia University, Princeton University,the INAS conference in Oslo, the CEN workshop in Copenhagen, SUDSWEc,IAS Norrköping, Lund University, Stockholm School of Economics, the AASLE inMelbourne, the II Workshop on Empirical Research in Economics of Educationin Reus, Tilburg University, and the ENTER conference in Toulouse for valuablecomments and feedback. We also gratefully acknowledge financial support fromthe Swedish Trade Union Confederation (LO), the National Union of Teachers inSweden (LR), and the Swedish Teachers’ Union. This research also benefited fromfinancial support from Handelsbanken’s Research Foundations. Finally, we thankBotkyrka municipality for allowing us to study school choice in their municipality.The experimental design and data used in this study have passed ethical vettingby the Stockholm ethical review board (DNR 2015/1252-32). All errors are ourown.

74 CHAPTER 2

2.1 Introduction

School choice is nowadays an integral part of many educational sys-tems. Instead of assigning children to their neighborhood school, par-ents are encouraged to apply to the school(s) of their choice.1 Pro-ponents of school choice has argued that this will broaden accessto quality schools by allowing disadvantaged households to opt outof their neighborhood schools. The empirical evidence tells anotherstory, suggesting that school choice has rather led to increased sort-ing across schools by students’ ethnic and socioeconomic background(Hsieh and Urquiola, 2006, Bifulco and Ladd, 2007). Yet, school dis-tricts often strive for integrated schools in terms of students’ familybackgrounds. This ambition is not uncalled-for, as school segrega-tion has been shown to have adverse effects on (especially disadvan-taged) students’ academic performances.2 An even greater concernis that school segregation may also impact students’ long-term out-comes.3 In order to counteract the segregating tendencies of schoolchoice, understanding the mechanisms through which school choice

1A neighborhood-based school assignment implies that parents can choose aschool indirectly by moving to an address within that school’s catchment area(given the financial resources to do so). In this study, the term school choice isused to refer to situations in which parents are asked to submit their preferencesfor school(s) through a centralized application system.

2See e.g. Hanushek et al. (2009), Billings and Rockoff (2013), Johnson (2011),Gamoran and An (2016), and Billings et al. (2016). In addition, Burgess andPlatt (2018) show that inter-ethnic relations are improved when students meetindividuals from other ethnic groups in school, which may be important from awider perspective.

3Chetty and Hendren (2018a,b) establish that the childhood environment hassubstantial long-term effects on college attendance and earnings. Furthermore,they show that high-quality schools is one important characteristic of neighbor-hoods facilitating upward mobility. If the school system becomes increasingly seg-regated, maintaining high quality schools in disadvantaged neighborhoods couldprove challenging. For example, teachers respond to the peer composition and aremore likely to stay on the job when facing high ability students (Karbownik, 2016).In a similar study, Jackson (2009) shows that teacher quality declines in schoolsin Charlotte-Mecklenburg, US, when the share of black students increases.


leads to increased sorting is important. Heterogeneous preferencesamong households may be part of the explanation, as school choiceallow them to sort accordingly. For example, Hastings et al. (2009),Borghans et al. (2015) and Burgess et al. (2015) find that advantagedhouseholds are more likely to apply to high-performing schools, whichmay lead to a segmented school market.

Another, less well-studied, determinant of the distribution of stu-dents across schools is the criteria used to determine admission tooversubscribed schools. Today, many school districts run centralizedschool choice programs where an algorithm produces a match betweenstudents and schools, taking application lists submitted by parentsinto account. This requires a decision about which algorithm to useand according to what rules students should be given priority whenthere are more applicants than seats at a school. For example, in mostcountries, students applying to high school are given priority accord-ing to their grade point average (GPA). This is not an option forchildren starting primary school, as there is no record of their previ-ous academic achievements. Instead, neighborhood-based assignmenthas been the prevailing method of allocation. With the introductionof centralized school choice, allowing parents to apply to a broaderset of schools, a need for more sophisticated priority structures hasarisen. However, there is little evidence to lean against when decidinghow students’ priorities to schools should be determined.

In this study, the impact on school segregation of three differentpriority structures under school choice is examined. First, students aregiven priority based on distance, with higher priority to students re-siding closer to the school. Proximity-based admission (either by usinga continuous distance measure or by giving extra priority points forneighborhood schools) is common at the elementary school level andis used in e.g. Sweden, US, UK, Estonia and Spain.4 Next, priorities

4See http://www.matching-in-practice.eu/elementary-schools/ for anoverview of elementary school admission in EU. Several countries, such as the

76 CHAPTER 2

are randomly determined by running a lottery before the allocation-process starts. Randomly assigned priorities are used in several schoolchoice programs, for example in Amsterdam, Beijing, and New YorkCity (see De Haan et al., 2015, He, 2017, and Abdulkadiroğlu et al.,2005). Finally, a version of controlled school choice is implemented byreserving seats for students contributing to the diversity of the school.Affirmative action in school choice exists for example in high schooladmissions in Paris where students from low income families are givenextra priority points (Fack et al., 2018). Moreover, affirmative actionhas been implemented as the result of court ordered desegregationguidelines in the US (Abdulkadiroğlu and Sönmez, 2003).

The impact of these different admission criteria is studied in thecontext of a centralized school choice program where students areassigned to primary schools using the deferred acceptance (DA) al-gorithm. A simulation strategy is used to determine the allocationof students to schools under each priority structure. Moreover, theoutcomes are contrasted with the allocation of students using schoolcatchment zones. This is a common way to assign students to schoolsin absence of school choice. Instead of considering parents’ prefer-ences for schools, each school admits all students living in a predefinedcatchment area surrounding the school. In order to perform the simu-lations, parents’ preferences for different school attributes are modeledusing a semi-structural approach.5 The estimated parameters are usedto create complete application lists for each child, where all schools intheir choice set are ranked in order of the parents’ preferences. Theseapplication lists, together with the set of primary schools and theircapacity constraints are sufficient to determine the allocation under

UK, lack national guidelines, giving rise to local variation in admission procedures.Still, proximity seems to be a common feature of elementary school admission inmany European countries. See for example Burgess et al. (2015) for a descriptionof admission to public elementary schools in the UK, where proximity is the mainpriority determinant.

5Pathak and Shi (2017) suggest that the use of structural demand models inthe school choice context can be effective when studying counterfactual outcomes.


various priority structures.The data used to simulate the allocation of students to schools

are collected from the school choice program in Botkyrka municipal-ity, a suburban school district located in the Stockholm metropolitanarea in Sweden. The school choice program is centralized, includingall elementary schools in the district. Each household with a childexpected to start primary school are invited to participate in the pro-gram, and asked to submit a rank-ordered application list includingthree schools of their choice.

Students’ priorities to public schools are determined based on howfar from the school they live and whether an older sibling is alreadyenrolled in the school. In addition, there are five voucher schools thatoperate a first-come, first-served queue system in addition to givingpriority to siblings of current students.6 The allocation of students toschools is done in two rounds. Students are first assigned to voucherschools and thereafter, using the DA algorithm, to public schools.Botkyrka provided access to the application lists submitted by parentsto the four cohorts of children starting primary school between 2011and 2014; a population of almost 4,000 children.7 Combining this datawith register data from Statistics Sweden allows for detailed analysisof the impact of different admission rules.

Botkyrka constitutes an interesting setting to study how admis-sion rules impact school segregation, as it has one of the most seg-regated elementary school systems in Sweden (see Figure 2.9 in Ap-pendix 2.B). Due to the level of residential segregation, neighborhood-based assignment is likely to transmit the residential segregation intothe school system. The prospect of being able to impact the degreeof student sorting across schools by modifying the admission scheme

6Voucher schools are publicly funded but privately run schools.7The focus on primary school starters is motivated by the fact that many

children in Botkyrka attend the same school for their entire elementary education,making the choice of primary school the important one for how segregated theschool system becomes.

78 CHAPTER 2

is therefore of great interest for this school district, as well as othersimilarly segregated school districts aiming to integrate their schools.The main measure of student sorting used in this study is the shareof the variance in students’ socioeconomic status (SES) that can beexplained by what school they are assigned to.

The simulation results show that when students are assigned toschools using proximity-based priorities, 18.5 percent of the variationin SES can be accounted for by their assigned school. This is aboutthe same as when students are assigned their neighborhood school,without consideration to the preferences for schools expressed by par-ents. In other words, the hypothesis that neighborhood-based alloca-tion consolidates school segregation when households are residentiallysegregated is confirmed. Moreover, this sorting pattern persists underschool choice when priorities are determined based on where studentslive relative to the school their parents apply to. When students’ pri-orities are determined using lotteries or affirmative action policiesinstead, students become relatively less sorted. In the first case, theshare of the variance in students’ SES-status that is explained by theschool they are assigned drops to 16.9 percent and, in the latter case,to 14.8 percent. This implies a reduction in school segregation by 8.5and 20 percent relative to sorting under proximity-based priorities.

The model of parents’ preferences for school attributes shows thatproximity to the school is an important determinant when choosingwhich school to apply to. The proximity-based priority structure istherefore well aligned with the preferences of the average household.Still, the fact that school segregation decreases when proximity-basedpriorities are abandoned implies that some households are willing tolet their children travel further in order to get a more preferred school.The probability of admission when applying for a school that is notlocated in the proximity of their home increases with lottery-basedpriorities and the affirmative action policy prioritizes specifically thosestudents that contributes to the diversity of the school. The fact that


households are residentially segregated therefore have less impact onthe level of segregation across schools under these two priority struc-tures.

The greater reduction in school segregation under affirmative ac-tion is also expected. There are two reasons for why this effect isstronger compared to when lottery-based priorities are used. Firstly,children that would contribute to the diversity of the school are nowgiven highest priority. Some of these might have been unlucky in thelottery, and would therefore not have been admitted to the school ap-plied to under that system. Secondly, lottery-based priorities may alsoenable households to escape neighborhood schools where they wouldbelong to a minority group in favor of schools where the students comefrom more similar backgrounds. This would have the opposite effecton segregation. Under the affirmative action policy, this would becomeharder as they would have very low priority to a school with studentssimilar to themselves. Hence, while both lottery-based priorities andthe affirmative action policy relaxes households from the constraintthat proximity-based priorities imply, the former allows desegregatingand segregating reallocations to take place, side by side.

The magnitude of the reduction in school segregation is sizable,given the ease with which admission schemes can be modified. Imple-menting a new priority structure in an existing school choice programis straightforward and the monetary cost of such a reform is low.However, the real challenge when modifying the admission scheme islikely how well it will be received by those participating in the pro-gram. Allowing parents to apply to schools is likely to cause discontentif the school choice program does not produce a matching betweenstudents and schools that, at least to a reasonable extent, satisfiestheir expressed preferences. If the majority of children are assignedto schools that are very low-ranked by their parents, the justificationfor introducing school choice could be questioned. To examine this di-mension, the rank of the assigned school is compared under the three

80 CHAPTER 2

different priority structures examined in this study. This exercise sug-gests that parents preferences are accommodated, with almost nineout of ten children assigned a school that was ranked top-three bytheir parents, under all priority structures.

Although the reduction in school segregation is large enough tomotivate school districts to consider what admission rules to use, re-member that 14.8 percent of the variation in student SES can beexplained by what school the students are assigned, even when af-firmative action policies are implemented. Part of the remaining ex-planation is the residential segregation in combination with strongpreferences for proximity of the school. That is, no matter the admis-sion scheme, most households apply to a school near their home. Aset of additional simulations confirms that residential segregation isa more important determinant of the remaining segregation, ratherthan heterogeneity in preferences between different types of house-holds.

A number of robustness tests are performed, to confirm the con-clusions in the main analysis. First, the analysis is complemented us-ing another measure of segregation. While the preferred measure de-scribed above captures several dimensions of a students’ background,additional evidence is presented using the Duncan Dissimilarity Index(DDI), studying sorting by foreign background and parental educationseparately. In line with the main results, affirmative action policies de-creases sorting along both these dimensions. Somewhat surprisingly,no reduction in sorting by foreign background or parental educationis observed when using lottery-based priorities. How to align theseresults with the reduced segregation by student SES is discussed inSection 2.5.3.

Another important consideration concerns the assumption under-lying the simulation exercises. These rely on the ability to predictwhich school household’s would apply to. The main worry is thathouseholds may not report their truthful ranking of schools on the


submitted application lists, due to the presence of strategic incentives.Consistently estimating the preference parameters requires house-holds to be truthful when submitting their application lists. AlthoughDA is a strategy-proof mechanism, the restriction in Botkyrka to rankonly three schools on the application list could reintroduce strategicincentives (Haeringer and Klijn, 2009). As noted by Calsamiglia et al.(2010), the extent to which this imposes a restriction on the choos-ing agent depends on the true ranking of schools with guaranteedadmission. With knowledge about how priorities are determined inthe Botkyrka School Choice Program, schools where admission is asgood as guaranteed can be identified for each household. Assumingthat households are rational utility-maximizing agents, the rankingof these schools on their application lists can then be used to catego-rize some choices as truthful. Applying this method to the sample ofprimary school starters, at least 79 percent of the sample is found tosubmit (partially) truthful application lists. This allows for credibleestimation of parents’ preferences for schools.

Finally, results from a randomized experiment where informa-tion about school performances was provided to households about tochoose primary school in Botkyrka in 2016 are presented. It has beenshown that, in some school choice programs, lack of information ex-plains part of the differences in which schools certain households applyto. If parents are uninformed, the application lists may not representtheir preferences because they are unable to identify the schools thatmatches their preferences. As previous studies shows that disadvan-taged households apply to higher-performing schools when providedwith information on school performances (see e.g. Hussain, 2013 andHastings andWeinstein, 2008), a randomized experiment is conductedin which treated households received information about schools’ passrates on the standardized tests in Swedish and mathematics. Analysisof the application lists reveals that this information had no impact onthe ranking of schools on the submitted application lists. This sug-

82 CHAPTER 2

gests that the differences between the application lists submitted byhouseholds of different types are not easily diminished by improvingthe information provided about the schools in this particular setting.

This study contributes to the small but growing number of empir-ical studies examining the importance of how school choice programsare designed. Even though the admission criteria are important deter-minants of students’ admission probabilities at different schools, thereis little evidence on how they impact the sorting of students acrossschools. The previous literature is mainly focused on the theoreticalproperties of different assignment mechanisms, starting with the sem-inal study by Abdulkadiroğlu and Sönmez (2003). This was the firsttime that the school choice problem was approached from a mecha-nism design perspective. Pathak (2011) provides a review of the the-oretical literature that followed. While these studies provide impor-tant insights about the theoretical properties of different algorithms,they offer little guidance to school districts regarding how to deter-mine students’ priorities. For this, quantitative studies are needed,acknowledging that the importance of the theoretical insights mightdepend on the environment in which the system is implemented. Acouple of studies have used administrative data to look at other as-pects of the design of a school choice program, but none focus on thepriority structure.8 The absence of studies on the topic is to someextent explained by the lack of data on parents’ applications and/orinstitutional features preventing such data to be used for counterfac-tual analysis. This study overcomes these problems by having accessto high-quality data from a centralized school choice program in asetting with limited strategic considerations by parents.

The remainder of this chapter is outlined as follows. Section 2.2presents the theoretical framework and is followed by Section 2.3 de-

8For example, Luflade (2017) studies the welfare effects of sequential imple-mentation of DA when application lists are truncated (allowing students to up-date their expected admission probability) in the context of university admissionin Tunisia.


scribing the institutional features, the data and some summary statis-tics of the Botkyrka School Choice Program. Next, in Section 2.4, theempirical strategy for estimating school preferences is discussed to-gether with the determinants of school choice. Section 2.5 discussesthe simulation strategy and results from counterfactual analysis ofalternative priority structures. In Section 2.6, the randomized ex-periment giving additional information about school performance torandomly selected households and the results are presented. Finally,Section 2.7 concludes.

84 CHAPTER 2

2.2 Theoretical Framework

From a theoretical perspective, allocating students to schools is amany-to-one matching problem. A finite set of students should beallocated to a finite set of schools. While all students should be allo-cated to a school, no one should be allocated to more than one school.Furthermore, schools have a cap on the number of students they canadmit. These are the constraints that any solution to the problemhave to respect. If all students can be allocated to their most pre-ferred school without any school admitting more students than theyhave capacity for, the solution is trivial. Everyone gets their mostpreferred school. In reality, some schools are often more popular thanothers and cannot admit all applicants. In this case, the school choiceproblem becomes more complex and demands a set of rules in orderto select a matching. This set of rules is referred to as a matchingmechanism.

The first to address the school choice problem in this way wasAbdulkadiroğlu and Sönmez (2003). Concluding that many, at thetime commonly used, school choice mechanisms were flawed, theyproposed two alternatives; the DA algorithm and the Top TradingCycles (TTC) mechanism. The DA algorithm has since then becomeincreasingly popular and is now implemented in school choice pro-grams worldwide. While an extensive theoretical literature is avail-able to guide this choice, school districts also need to decide how todetermine students’ priorities to schools. Most mechanisms require apriority ordering of students for each school as an input in order toselect a matching between students and schools. The set of rules usedto determine these orderings of students is referred to as a prioritystructure.

This study focuses on the impact of different priority structureswhen students are allocated to schools using the DA algorithm, andthe data are collected from a school choice program using DA. Un-derstanding how this algorithm works will therefore be important for

2.2. THEORETICAL FRAMEWORK 85

how to conduct the empirical analysis.

2.2.1 The Deferred Acceptance Algorithm

The DA algorithm was first proposed by Gale and Shapley (1962)in the context of college admission and adapted to the school choicecontext by Abdulkadiroğlu and Sönmez (2003).9 Suppose that all stu-dents can rank all schools in the order of preference and that there isa priority ordering of all students for each school, stipulating the or-der in which students should be admitted.10 The DA algorithm thenworks as follows:

Step 1 Each student applies to his/her top choice. Eachschool tentatively accepts one student at a time accord-ing to their priority until their capacity is reached or noapplicants remain.

Step k Each student rejected in step k−1 applies totheir next preferred school. Each school considers all ten-tatively accepted and new applicants together and tenta-tively accepts one student at a time, according to theirpriority, until their capacity is reached or no applicantsremain.

The algorithm terminates when no student is rejected, at which pointall students are placed at their final assignment.

Different algorithms have different properties. In the mechanismdesign literature, matching algorithms are often assessed by three

9The two situations differ in that college admission is a two-sided matchingproblem while school choice is a one-sided matching problem.

10This could be thought of as the schools’ preferences over students. The rea-son for using another terminology is that schools are often not allowed to admitstudents in the order that they would prefer themselves. Rather, there are rulesfor in which order students should be admitted at, for example, the school districtlevel.

86 CHAPTER 2

desirable properties; strategy-proofness, stability, and efficiency. TheDA algorithm satisfies the first two, but not the last of these proper-ties. Strategy-proofness implies that it is a weakly dominant strategyfor everyone to truthfully report their preferences on the applicationlist. A mechanism that is not strategy-proof is open for manipulationin the sense that a student can get a more preferred school by misrep-resenting his/her preferences when submitting the application list. Inthe DA mechanism, priorities are independent of how a student ranksa school. A student who was tentatively assigned to one school cantherefore be rejected from that school in a later round. This featureof the algorithm ensures that it cannot be manipulated.11

The DA is also stable, implying that it always produces a match-ing between students and schools that is individually rational, non-wasteful and eliminates justified envy. Individual rationality meansthat no student assigned a school would prefer not to be assignedat all. Non-wastefulness means that no student prefers a school withempty slots over the assigned school. Eliminating justified envy meansthat no student prefers another school where a student with lowerpriority was admitted, over the assigned school. Stability is an im-portant feature, as unstable mechanisms tend to be abandoned in thereal world (Roth, 2007).

Efficiency implies that a mechanism always selects an allocationthat pareto dominates the allocation selected by any other mecha-nism. An allocation pareto dominates another allocation if every stu-dent weakly prefers the assignment given by the first allocation andat least one student strictly prefers the assignment given by the firstallocation. In other words, it is not possible to improve the outcomefor one student without harming another if the mechanism is effi-cient. This is not true in the case of DA, where a situation can arisein which two students would like to switch seats with each other,making both better off and no one worse off. Letting these students

11Abdulkadiroğlu and Sönmez (2003) provide a proof.


switch seats would however imply that stability is not guaranteed. Inpractice, school officials therefore face a trade-off between efficiencyand stability.

One of the reasons behind the popularity of the DA mechanismis the fact that it is strategy-proof. However, when implemented inpractice, the application lists are often truncated by putting a restric-tion on the number of schools that can be included. In Botkyrka, forexample, parents cannot list more than three out of the 25 schoolsin the municipality. Haeringer and Klijn (2009) show that this rein-troduces strategic incentives, as households try to avoid a situationwhere their child is rejected by all schools applied to. This can bedone by including a safe school, i.e. an acceptable school where theexpected admission probability is high.12 Luflade (2017) studies uni-versity applications in Tunisia and find that truncated applicationlists substantially reduce welfare because some students are rejectedby all schools on their application list. However, the extent to whichthis causes the choosing agents to act strategically is likely to dependon the specific setting. Abdulkadiroğlu et al. (2005) and Abdulka-diroğlu et al. (2009) report that in NYC, high school applicants wereallowed to list five programs before the redesign of the system in2003-2004 and twelve programs after, with more than 500 programsto choose from. With the five-school restriction, about 50 percent ofhouseholds ranked less than the maximum allowed number of schools,after the redesign the corresponding number was 72-80 percent.

With knowledge about which schools can be considered safe fora given student, it is possible to assess the truthfulness of the appli-cation list. Let school j be defined as a safe school for household i ifthe probability of admission to this school is equal to 1 conditional onapplying. Suppose that application lists are truncated as in Botkyrka,with only three choices allowed. Then:

Claim 1. If the 1st choice is a safe school, the 1st choice is truthful.12Calsamiglia et al. (2010) discuss the concept of safe schools.

88 CHAPTER 2

Claim 2. If the 1st choice is not a safe school but the 2nd choice isa safe school, both the 1st and 2nd choices are truthful.

The intuition is straightforward. If you include a safe school any-where on your application list except for the lowest rank and thereis another school that you prefer above the safe school, you wouldbe better off ranking the preferred school above the safe school. InAppendix 2.A, proof of these claims are provided together with thenecessary assumptions. Taken together, they imply that the submit-ted application lists can be considered truthful up until the highestranked safe school, if a safe school is included and is not the lowest-ranked school on the application list. If a safe school is ranked laston the application list, it cannot be inferred whether the ranking istruthful or not. The reason is that there might exist a more preferredschool that was not included due to a strategic decision to insteadinclude the safe school to avoid being rejected at all schools appliedto. The inclusion of safe schools is therefore informative when rankedanything but last on the truncated application list.

Another important note to make regarding the truthfulness of sub-mitted applications when DA is used is made by Fack et al. (2018);strategy-proofness only implies that truth-telling is a weakly domi-nant strategy. In other words, there may exist another strategy (an-other ranking of schools) that will result in an equally good outcomefor the applicant. To understand why, consider the case where anapplicant has a zero probability of admission at the most preferredschool. Whether this school is included on top of the application listor excluded from the same is not going to matter for which school thestudent is assigned to. In the first case, the student will be rejectedfrom the most preferred school and the algorithm will continue andtry to place the student at the next preferred school. In the secondcase, the algorithm will start with the school that was ranked sec-ond on the truth-telling application list. As the rank of a school doesnot affect the admission probability, the student ends up in the same


school in both cases. Furthermore, the admission probability does nothave to be zero for this to occur. It is enough that the applicant per-ceives that the admission probability is zero. Fack et al. (2018) use theterm skipping the impossible to describe this applicant behavior. Thisimplies that the DA mechanism can give rise to multiple equilibria.

The question of truthfulness will be returned to, once the institu-tional setting surrounding the school choice program in Botkyrka hasbeen described.

90 CHAPTER 2

2.3 Institutional Setting

The Swedish elementary school system covers ten grades with chil-dren usually starting school the year they turn six. The first year(grade K) was voluntary at the time period studied, followed by ninemandatory years (grade 1-9).13 Despite grade K being voluntary, al-most all children (97 percent) were enrolled. Since the early 1990’smunicipalities are responsible for providing and financing the primaryeducation of all children residing in the municipality. There are twotypes of schools. Public schools are run by the municipalities andvoucher schools are operated by independent providers, but publiclyfunded. All schools follow the same curriculum, stipulated in nationallegislation. School funding is based on vouchers directly connectedto each student, but municipalities are allowed to adjust the amountbased on a student’s background or disabilities.14 Neither public norvoucher schools are allowed to charge any fees.

School choice was formally introduced in Sweden in 1992, togetherwith the reform that allowed voucher schools to operate. Prior to this,students would as a rule attend their local public school. Initially,school choice was decentralized to the school level. This meant thatparents, who did not want their children to attend the local publicschool, had to contact the school(s) of interest and ask whether aseat for their child was available. Not surprisingly, the vast majoritycontinued of children continued to attend their local public school. Atthe turn of the millennium the expansion of voucher schools took off,introducing competition for students on the educational market inSweden.15 In the mid-2000, municipalities started to run centralizedschool choice programs. This was in part a reaction to the expansion

13Since the school year 2018-2019, grade K is mandatory for all children follow-ing the adoption of Proposition 2017/18:9, implying a change in Skollag 2010:800.

14In Botkyrka, 28-44 percent of the school budget is compensatory (EY, 2014).15Böhlmark and Lindahl (2015) show that the share of the students enrolled in

voucher schools increased from 1.6 percent in 1998 to 11 percent in 2009.

2.3. INSTITUTIONAL SETTING 91

of voucher schools. By making it easier to apply other public schools,parents wouldn’t have to leave the public school system if not satisfiedwith their local public school.

These days, the majority of Swedish municipalities run their owncentralized school choice programs, at least in the urban areas. Theprograms typically include all public schools in the municipality. Insome places, voucher schools are also included in the program, mean-ing that parents can apply to all schools in their municipality throughthe same system. By law, voucher schools are not allowed to discrim-inate between students based on where they live. This means thatsome parents apply to voucher schools in other municipalities, butthis is usually done by contacting the school directly. Furthermore,parents can apply to a public school in other municipalities althoughthey have no obligation to admit students not residing within theirborders.16 In this case, the parents would have to contact the munic-ipality in which the public school of interest is located.

There is considerable variation in the design of the school choiceprograms run by Swedish municipalities. While the legislation isvague, it imposes some restrictions on the municipalities in terms ofthe admission criteria they can use. To begin with, they are obligedto take parents’ expressed preferences for schools into account. Atthe same time, they need to ensure that no child is placed in aschool too far away from their home, unless the parents applied toa school far from their home. Also, children always have the rightto stay in their current school. Voucher schools also have somerestrictions on which admission criteria they can use. Either, theycan base their admission on the distance from the students’ home tothe school or use a first-come, first-served queue system. In addition,they are also allowed to give priority to children with older siblingsalready enrolled in the school.17 Many municipalities also tend to

16Under special circumstances, a child can have the right to attend a publicschool located in another municipality.

17The regulations can be found in Swedish at http://www.skolverket.se/

92 CHAPTER 2

give priority to schools where older siblings are enrolled, but whetherthis is in line with the legislation or not is unclear.

2.3.1 The Botkyrka School Choice Program

Botkyrka is the fifth largest municipality in Stockholm county, withabout 90,000 inhabitants. It is a multicultural community, where amajority of the population are first or second generation immigrants.The level of residential segregation is high, which also shows in theeducational system. As can be seen in Figure 2.9, their schools areamong the most segregated in the country in terms of the students’migration background and their parents’ level of education.

Almost 1,000 children enters the primary school system every year,and the vast majority of families take part in the school choice pro-gram run by the municipality.18 In January, households with childrenexpected to start grade K in the upcoming academic year are informedabout the school choice program and how to participate. They are alsoprovided with information about all the public and voucher schoolsin the municipality, all of which are included in the program. Theparents are given three weeks to register the schools of their choiceonline, in the form of a rank-ordered application list including threeschools. At each rank, instead of selecting one of the schools in themunicipality, parents can also indicate that they have applied to aschool elsewhere. Although there is no gain to informing the munic-ipality about this, about three percent of parents use this option.Participation in the program is encouraged by the fact that there isno default school where a child is guaranteed a seat if the householddoes not participate in the program. This gives strong incentives tosubmit an application list.

There are 25 primary schools in Botkyrka to choose from. 14 ofthese continues up until grade nine, implying that most children stay

regelverk/mer-om-skolans-ansvar/val-av-grundskola-1.210176.18The participation rate during the time period studied is around 95 percent.


in the same school for their entire compulsory education. Althoughmost schools are public, there are five voucher schools, enrolling about8 percent of the students in Botkyrka combined. While one of thevoucher schools widely resembles the public schools, the other four arelikely to attract a very specific group of applicants.19 The location ofthe schools are shown on the map in Figure 2.1. The map also makes itclear that there are three distinct neighborhoods in the municipality,where the schools are located and where the population density ishigher compared to the surrounding areas.

Figure 2.1: Location of Primary Schools in Botkyrka, 2014

kilometers

0 10

Notes: This map displays the location of the 22 primary schools in Botkyrka in 2014 included in theanalysis. See Section 2.3.2 for a discussion about sample restrictions on schools. Public schools areindicated in orange and voucher schools in green. The map also shows the division of the municipalityinto SAMS-areas in order to display the population density. SAMS stands for Small Areas for MarketStatistics and divides Sweden into about 9,500 small areas based on municipal partitioning and electoraldistricts (depending on the size of the municipality). SAMS-areas are designed to have roughly the samepopulation. c©Statistics Sweden

In April, the allocation of students to schools is usually done and

19One is a Christian school, one is a Muslim school, one is a bilingual school(Finnish-Swedish), and one uses a specific teaching method inspired by Freinet.

94 CHAPTER 2

households are informed about their child’s school placement. In prac-tice, the allocation process takes place in two rounds. First, childrenare assigned to voucher schools. They give priority to children witholder siblings already enrolled in the school and otherwise accordingto a first-come, first-served queue system. Furthermore, they con-sider only applicants listing them as their top choice. Next, childrenwho were not assigned to a voucher school are allocated to the publicschools using the DA algorithm. As the application lists are truncated,a child might be rejected at all schools applied to. These children are,together with those whose parents did not submit an application list,placed at schools with empty seats once all other children have beenassigned a school.

The priority orderings of public schools are based on proximityand whether siblings are already enrolled in the school. Children thathave a sibling enrolled in grade K-4 at a school have priority to thatschool over children that do not. To get a strict priority ordering,children with and without already enrolled siblings are ranked usinga relative distance measure meant to minimize the walking distancefor all children. The relative distance is calculated by subtracting thedistance to the school applied to from the distance to the closest pub-lic school.20 A higher relative distance measure gives higher priority,implying that children whose distance to school would increase moreif rejected are admitted first.21

20When the school applied to is the closest public school, the next closest publicschools is used instead. The distance is measured as the walking distance using amap with only roads assessed by the municipality to be safe for a 6-year old towalk on. This is done to ensure that the distance-measure reflects a route to theschool which does not require the child to be accompanied by an adult to get toschool.

21Note that this measure does not consider whether the child, if not given aseat at the school applied to, would actually be admitted to the alternative schoolused to calculate the relative distance to the school applied to.


2.3.2 Data and Summary Statistics

This study uses administrative school choice data for all primaryschool starters in Botkyrka during 2011 to 2014, provided by Botkyrkamunicipality. This gives a population of 3,822 primary school startersfor whom applications with three ranked-ordered choices are observed.These data are linked to register data from Statistics Sweden, usingindividual identifiers. The multigenerational register allows identifi-cation of their parents and siblings. At the school level, performanceon the standardized tests in Swedish and mathematics taken in gradethree is observed. The compulsory school register provides informa-tion about all students enrolled in a primary school in Botkyrka,making the student composition observable. Furthermore, Botkyrkamunicipality shared information about the number of students eachpublic school had capacity to admit in grade K for the years 2012to 2014. Below, the characteristics of the primary school starters aswell as the schools in their choice set is described together with thesample restrictions.

Primary schools There are 25 primary schools offering grade K inBotkyrka, five of which are voucher schools. All schools are includedin the school choice program, meaning that applications to all theseschools can be observed. In order to estimate school preferences, thecharacteristics of the school needs to be observed. Due to lack of data,three schools are dropped from the analysis. Two are confessionalvoucher schools for which data on the student composition are notavailable.22 One is a public school, opening in 2013, for which thelocation is not observed. Another public school closed in 2012, keepingthe total number of primary schools constant at 25 during the time

22One of these schools is a Muslim school and one is a Christian school, togetherenrolling 34 to 49 students per year in grade K. In order to protect the religiousrights of those attending confessional schools, data on students enrolled in theseschools are not available from Statistics Sweden.

96 CHAPTER 2

period. This implies that the choice set will consist of 23 schools forthe 2011 and 2012 cohorts, and 22 schools for the 2013 and 2014cohorts.

Table 2.1 presents the characteristics of the schools, in total andby neighborhood. It is clear that the students have a stronger socioe-conomic background at the schools in the east while those in the northhave a more disadvantaged background. Immigrant students are alsoover-represented in the schools in the north, while the opposite is truein the west and especially the east. The average test score on the stan-dardized tests in Swedish and mathematics, taken in grade three, is12.9 points on a scale from 0 to 20, with the north performing belowaverage and the other two neighborhoods above average. There areno big differences in terms of the grade K capacity of the schools,with about 50 seats per school. The number of schools is more or lessproportional to the number of students living in each neighborhood.The three voucher schools in the choice set are located in the northernand western neighborhoods.

Primary school starters The school choice data consist of appli-cations for 3,903 children.23 81 of these children are excluded from theanalysis as they are registered at an address outside of Botkyrka, giv-ing a population of 3,822 primary school starters. The excluded chil-dren are likely living in a household about to move to Botkyrka andtherefore invited to participate in the school choice program. Theyare excluded as the distance to school from their future address inBotkyrka cannot be calculated, which is important for the analysis.24Furthermore, applications for the three schools excluded from the

23To be precise, there are 3,901 children participating but two of these childrenparticipate two years in a row. Likely, they did not start school the first time andwere therefore invited to participate in the program again next year.

24The alternative of calculating the distance from their current address is not anoption, since that would not accurately reflect the expected distance once settled inBotkyrka. When estimating preferences, this would suggest that these householdsput very low value on proximity to the school which is not necessarily true.


choice set above are dropped. This implies that for some households,the application list will consist of two schools only. These applicationlists are recoded so that the highest-ranked school within the definedchoice set is considered the top choice and the second-highest rankedschool is considered the second choice, leaving the third rank missing.Some households also use one of the ranks on the application list toreport that they applied to a school in another municipality. This ishandled in the same way.

Table 2.2 presents summary statistics of the population of primaryschool starters and some characteristics of their top-ranked school. 41percent of the primary school starters have a foreign background (de-fined as being born abroad or having two parents born abroad). Thelevel of residential segregation is striking; 73 percent of the children inthe north have a foreign background, but only nine percent of childrenliving in the east. There are also large differences in their parents’ levelof education, with a more well-educated population in the east andless well-educated in the north. The neighborhood in the west is, inboth dimensions, more mixed than the other two. In all three neigh-borhoods, about half of the children are boys and about half have asibling already enrolled in a primary school in the municipality.

Many families apply to schools in proximity of their home, with59 percent ranking the closest school on top of their application list.The distance from home to the top-ranked school is about one kilo-meter in the north and the east, and about 1.4 kilometers in thewest. As previous literature suggest, distance to school seems to bean important determinant also in this setting. Still, there are somedifferences between neighborhoods worth pointing out. Households inthe north seem more willing to travel further than they would haveto, had they attended their nearest school. While 70 percent in theeast list the nearest school as their top choice, this drops to 50 percentin the north. Furthermore, six percent of the households in the northapply to a school in another neighborhood and 17 percent apply to

98 CHAPTER 2

a voucher school. These shares are lower in the west and much lowerin the east, where the population is in general more well-educatedand there are fewer households with foreign backgrounds. Having allchildren attend the same school also seems important as 42 percentof the top-ranked schools are schools attended by an older sibling.


Table 2.1: Summary Statistics for Grade K Schools, 2011-2014

All By neighborhood

North West EastAverage test score 12.92 12.38 13.13 13.70

(1.25) (1.23) (1.18) (0.86)Foreign background (share) 0.47 0.77 0.30 0.12

(0.31) (0.12) (0.19) (0.03)High-educated (share) 0.46 0.34 0.49 0.67

(0.16) (0.08) (0.10) (0.09)Capacity in grade K 52 50 53 55

(20) (17) (26) (18)Average number enrolled (K-5) 260 242 267 285

(96) (61) (136) (79)Certified teachers (share) 0.76 0.74 0.74 0.81

(0.10) (0.08) (0.13) (0.08)Student-teacher ratio 13.97 12.15 14.44 16.89

(3.19) (2.29) (3.07) (2.52)Newly arrived immigrants (share) 0.05 0.09 0.03 0.01

(0.07) (0.08) (0.05) (0.01)Parental income (100 SEK) 2736 1904 3124 3817

(954) (413) (739) (388)Allowance recipient (share) 0.19 0.29 0.16 0.05

(0.14) (0.11) (0.12) (0.03)Days in unemployment 44.20 71.62 29.40 11.54

(30.66) (20.68) (17.34) (3.70)Grade 9 school 0.62 0.80 0.67 0.20

(0.49) (0.41) (0.48) (0.41)Observations 90 40 30 20

Notes: This table presents the mean and standard deviation of each variable for thesample primary schools in Botkyrka 2011-2014. Average test score is the school averagescore (max = 20) on the standardized tests in Swedish and mathematics taken in gradethree. Foreign background indicates the share born abroad or with both parents bornabroad. High-educated indicates the share with at least one parent with universitylevel education. Capacity in grade K indicates the average number of seats availablefor primary school starters. Average number enrolled indicates the average number ofstudents enrolled in grade K-5. Certified teachers is the share of certified teachers atthe school. Student-teacher ratio indicates the number of students per teacher. Newlyarrived immigrant indicates the share with less than 4 years in Sweden. Parental incomeis the average parental income measured in 100 SEK. Allowance recipient indicatesthe share of students living in households that receive social assistance or housingallowances. Days in unemployment indicates the parental yearly average of days withunemployment benefits. Grade nine school indicates that the school starts at grade Kand continues up to grade nine.

100 CHAPTER 2

In Table 2.3, the characteristics of the schools that students wereassigned to are summarized. First, note that a very high share is as-signed their top choice, around 90 percent in all neighborhoods, andonly two percent are rejected at all three schools applied to. Partof the explanation for this is the significant overcapacity in the pri-mary school system in Botkyrka during these years, which will bereturned to in Section 2.4.1 when the presence of strategic incentivesis discussed. There are some differences across neighborhoods worthnoting in terms of the assigned school. It is somewhat more commonto not be assigned a preferred school, attend a voucher school, or at-tend a school in another neighborhood for primary school starters inthe north. Students in the west travel a bit further to their schoolscompared to students in the two other neighborhoods. This is likelyexplained by the fact that this is a much larger geographical areacompared to the other two neighborhoods. Not surprisingly, as manystudents attend a local school, the peers of the attended school differ alot by neighborhood. This reflects the differences in population char-acteristics across neighborhoods. Finally, there are large differencesin performances of the school attended, with the pass rate of schoolsattended by students from the north being only 30 percent. The passrate increases to 43 percent in the schools attended by students in thewest and doubles to 61 percent for those in the east.

Finally, evidence that the households in Botkyrka face real op-portunities regarding the school choice is presented. Table 2.4 showsthat on average, households have about five schools within two kilo-meters from where they live. Increasing the distance from two to threekilometers gives, on average, two additional schools to choose from.This is important, implying that there is a choice even for house-holds with strong preferences for proximity. Table 2.4 also suggeststhat there is substantial variation in school attributes, even when re-stricting the choice set to schools near the home. If only consideringschools within two kilometers of one’s home, the difference in the share


Table 2.2: Summary Statistics for the Primary School Starters, 2011-2014

All By neighborhood

North West EastHigh-educated (share) 0.52 0.38 0.54 0.73


(0.49) (0.44) (0.44) (0.29)Newly arrived immigrants (share) 0.04 0.06 0.03 0.01

(0.18) (0.24) (0.16) (0.08)Male (share) 0.51 0.50 0.51 0.51

(0.50) (0.50) (0.50) (0.50)Sibling enrolled (share) 0.49 0.48 0.51 0.49

(0.50) (0.50) (0.50) (0.50)Characteristics of top choiceDistance (meters) 1106 959 1399 968

(1354) (1272) (1664) (912)Additional distance from nearest school (meters) 1047 1196 1113 565

(1411) (1559) (1391) (835)Nearest school (share) 0.59 0.51 0.61 0.70

(0.49) (0.50) (0.49) (0.46)Sibling enrolled (share) 0.42 0.40 0.45 0.43

(0.49) (0.49) (0.50) (0.50)Nearest school or sibling enrolled (share) 0.74 0.68 0.76 0.81

(0.44) (0.47) (0.43) (0.39)Outside own neighborhood (share) 0.04 0.06 0.04 0.01

(0.19) (0.24) (0.19) (0.09)Voucher school (share) 0.08 0.17 0.03 0.00

(0.27) (0.37) (0.16) (0.04)Observations 3822 1566 1260 996

Notes: This table presents the mean and standard deviation of each variable for the sample ofprimary school starters in Botkyrka 2011-2014, in total and by neighborhood. High-educatedindicates the share with at least one parent with university level education. Foreign backgroundindicates the share born abroad or with both parents born abroad. Newly arrived immigrantindicates the share with less than four years in Sweden. Male indicates the share of males andolder sibling enrolled indicates the share with an older sibling enrolled in a primary school inBotkyrka. Distance is the distance to the top-choice school in meters and additional distance fromnearest school indicates how much further away from their residency the top-choice school is locatedcompared to the nearest school. Nearest school, sibling enrolled and nearest school or siblingenrolled indicates the shares listing the nearest school and/or a school where a sibling is enrolledas their top choice. Outside own neighborhood indicates that the top-choice school is not locatedin the neighborhood of residence and voucher school indicates the share listing a voucher school astheir top choice.

102 CHAPTER 2

of peers with foreign background (high-educated parents) is 28 (22)percentage points and about 2 points on the average standardizedtest score.

Table 2.3: School Assignments (Grade K), 2011-2014

All By neighborhood

North West EastDistance (meters) 1029 878 1362 805

(1330) (1203) (1647) (854)Attend top choice (share) 0.92 0.88 0.94 0.94

(0.28) (0.32) (0.23) (0.24)Attend second choice (share) 0.04 0.05 0.03 0.03

(0.19) (0.21) (0.17) (0.18)Attend third choice (share) 0.01 0.02 0.01 0.01

(0.11) (0.13) (0.10) (0.09)Rejected from all three choices (share) 0.02 0.03 0.01 0.00

(0.13) (0.17) (0.10) (0.05)Attend school outside own neighborhood (share) 0.03 0.05 0.03 0.01

(0.17) (0.21) (0.17) (0.09)Attend voucher school (share) 0.08 0.18 0.03 0.01


(0.31) (0.14) (0.16) (0.04)High-educated (share) 0.49 0.34 0.51 0.69

(0.16) (0.08) (0.09) (0.08)Average pass rate 0.42 0.30 0.43 0.61

(0.21) (0.16) (0.20) (0.15)Observations 3822 1566 1260 996

Notes: This table presents the mean and standard deviation of each variable for the sample ofprimary school starters in Botkyrka 2011-2014, in total and by neighborhood. All variables refer tothe school attended in grade K in October, the year the begin school. Distance is the distance toschool measured in meters. Attend top, second and third choice indicates the share assigned theirtop, second and third choice respectively and rejected from all three choices indicates the sharenot attending any of the schools on their application list. Attend school outside own neighborhoodindicates the share not attending a school in the neighborhood of residence and attend voucherschool indicates the share attending a voucher school. Share foreign and share high-educatedindicates the average share of students with foreign background (born abroad or both parents bornabroad) and high-educated parents (at least one parent with university level education) at theschool attended. Average pass rate indicates the pass rate on the standardized tests in Swedish andmathematics taken in grade three.


Table 2.4: Variation in School Attributes by Distance from Home,2011-2014

1km 2km 3km 5km 10kmNumber of schools 2 5 7 12 22

(1) (2) (3) (4) (2)Difference (max-min)Foreign background (share) 0.13 0.28 0.33 0.68 0.85

(0.14) (0.16) (0.17) (0.16) (0.08)High-educated (share) 0.10 0.22 0.26 0.44 0.53

(0.09) (0.08) (0.07) (0.11) (0.05)Average test score 1.07 2.05 2.22 3.20 3.90

(1.05) (1.17) (1.18) (1.21) (1.07)Observations 3822 3822 3822 3822 3822

Notes: This table displays the mean and standard deviation of each variable forthe sample of primary school starters in Botkyrka 2011-2014. Number of schoolsindicates how many schools the average household has within x kilometers of theirhome. The remaining variables are expressed in terms of the difference between themaximum and minimum value of the variable for schools within x kilometers ofhome. Foreign background indicates the share born abroad or with both parents bornabroad. High-educated indicates the share with at least one parent with universitylevel education. Average test score is the school average score (max = 20) on thestandardized tests in Swedish and mathematics taken in grade three.

104 CHAPTER 2

2.4 Estimating School Preferences

In order to simulate the allocation of students to schools under differ-ent admission schemes, it is necessary to know that the applicationlists would have looked like. As households would have no incentivesto act strategically, the input needed is their full ranking of schoolsunder truth-telling. The observed application lists are not informativeof this, as they are truncated at three choices. For this reason, theobserved application lists will be used to estimate parents’ preferencesfor schools. The estimated parameters can then be used to calculatethe households’ indirect utility derived from each school. Orderingthe schools accordingly gives rise to a complete application list.

This strategy does not only solve the problem of truncated ap-plication lists. It also makes it possible to perform dynamic simu-lations, where schools are allowed to change over time, in order tostudy the long-run effects of different admission criteria. The modeldoes not need to identify the true underlying preferences for differentschool attributes. It is sufficient that the preference parameters canbe used to predict households’ ranking of schools under truth-telling.The prediction power can be assessed directly, using cross-validationtechniques. However, the application lists predicted also needs to betruthful. In other words, the model needs to be evaluated and esti-mated on truth-telling households. Fack et al. (2018) make a novelcontribution to the literature by outlining an approach for estimatingpreferences for schools without having to require truth-telling of theobserved application lists. Unfortunately, this approach is not possibleto use in this study.25 Instead, this section will describe how a large

25Their approach assumes stability, which is a weaker requirement than truth-telling. However, estimating the preference parameters under stability requiresknowledge about the cutoffs for admission to each school in order to determineindividualized feasible choice sets. Without access to the specific road map usedto calculate the relative distance measure used to determine students’ priority inBotkyrka, the cutoffs cannot be identified and the feasible choice sets can thereforenot be determined.

2.4. ESTIMATING SCHOOL PREFERENCES 105

subset of truth-telling households is identified and then go thoughthe model used to estimate their preferences together with some ro-bustness checks. However, the estimates are likely to be very similarto what they would have looked like using the weaker assumptionof stability. This is because the significant overcapacity in the schoolsystem implies that the feasible choice set for each individual will bewell approximated by including all schools in the choice set.

2.4.1 Identifying Truth-Telling Households

While the DA mechanism is strategy proof, there are two reasons whyhouseholds may not be truth-telling when submitting their applica-tion lists to the Botkyrka School Choice Program. The first is the useof truncated application lists and the second is that households mayskip the impossibles. The extent to which this causes households to actstrategically depends on the level of congestion in the school system.Only households expecting to be rejected at all of their three mostpreferred schools or households expecting a zero percent probabilityof admission at one of their three most preferred should deviate fromtruth-telling.

The substantial overcapacity in Botkyrka’s primary school sys-tem during these years diminishes the likelihood of these situationsto arise. The size of the cohorts entering the primary school systemwere relatively small, but the municipality decided not to downsizethe educational system as future cohorts were expected to increasein size again. Because of this, most schools had empty seats once allstudents had been assigned a school. Table 2.10 shows that none ofthe schools in the northern neighborhood was oversubscribed dur-ing 2012 to 2014.26 In the other two neighborhoods, the number ofoversubscribed schools vary between zero and three. Which schools

26The municipality kept data on each school’s capacity for grade K studentsfrom 2012 and onward, implying that the overcapacity cannot be calculated forthe year 2011.

106 CHAPTER 2

are oversubscribed also varies, and no single school is oversubscribedall years. It is therefore a reasonable assumption that no school is re-garded as an impossible. Further, because the vast majority of schoolshad empty seats, very few households should find reason to strategizewhen submitting their application lists. In other words, truncationshould not impose much of a restriction on households in this setting.

Nevertheless, the presence of truth-telling households can be ex-amined more stringently using the concept of safe schools, introducedin Section 2.2. First, safe schools are identified for each child in thepopulation. Empirically, a school j is defined as safe for households iif:

− student i has a sibling enrolled in school j (in grade K-4 forpublic schools)

− school j was never oversubscribed during 2012 to 2014

− all applicants at the same residential location as student i wasaccepted at school j

Sibling status is essentially a guarantee for admission.27 Schoolsthat are never oversubscribed are categorized as a safe choice for allhouseholds and for household i, a school j is also categorized as safeif everyone residing in the same geographical unit as household i whoapplied was admitted to that school.28 Only the first criterion is usedfor voucher schools, since data are lacking on capacity and residentiallocation does not impact admission probability for these schools.

Table 2.5 shows that 2,752 households rank a safe school on top oftheir application list, implying that their top-ranked choice is truthful.

27The only scenario in which an already enrolled sibling would not be a guar-antee for admission is if there are more applicants with enrolled siblings than thecapacity to admit grade-K students. When asked, officials at Botkyrka could notremember this ever being the case.

28This criterion is applied only if there are at least 10 applicants from the samegeographical unit as household i.


Of the remaining households, 299 have a safe school ranked secondimplying truth-telling up until the second rank. Hence, 80 percentof the sample submit (at least partially) truthful application lists.The other 20 percent are not necessarily being strategic, but theirapplication lists are not informative about whether they are truth-telling or not. The model presented next will be estimated using theapplication lists of the household categorized as truthful.

Table 2.5: Categorization of Truthful Choices, 2011-2014

Top-ranked school Second-ranked schoolSafe schools 2752 299Enrolled sibling 1397 31Never oversubscribed 1237 256Residential location 118 12

Non-safe schools 1070 771Observations 3822 1070

Notes: This table presents the categorization of truthful choices for the sampleof primary school starters in Botkyrka 2011-2014. The first column displayswhether the top choice is a safe school or not. The second column displayswhether the second choice is a safe school or not, for the sub-sample ofhouseholds not listing a safe school as their top choice. The Table also displayswhether a choice is categorized as a safe school because of an already enrolledsibling in the school, because the school was never oversubscribed or because theresidential location implies a relative distance measure effectively guaranteeingadmission.

So far, voucher schools have been ignored in the discussion re-garding truthfulness. The fact that they deviate from the DA mecha-nism and consider only top-applicants implies that households shouldbe strategic. Either, they should list the voucher school of interestat the top of the application list, or exclude it entirely. However,households are not informed about the differential admission proce-dure for voucher schools. The observed application lists are consistentwith parents not being aware of this, as voucher schools are rankedboth second and third (see Figure 2.2). While it drops somewhat forvoucher schools ranked third, this is likely to be explained by two ofthe voucher schools in the choice set having a rather specific profile im-plying that those applying to these schools likely have rather strong

108 CHAPTER 2

Figure 2.2: Share Applying to Voucher Schools, 2008-2014

0

.02

.04

.06

.08

Shar

e

Top choice Second choiceThird choice

Notes: This figure displays the share of primary school starters in Botkyrka 2011 to 2014 listing avoucher school as their top, second and third choice respectively.

preferences for these schools. Anyhow, the share listing a voucherschool as their third choice is substantially higher than expected hadthey known that only top choices were considered for admission.

To further substantiate this claim, the characteristics of thoseranking a voucher school on top of their application list are com-pared to those with a voucher school ranked second. Table 2.11, inAppendix 2.B, shows that there are no big differences in the demo-graphics of these two groups. If some parents had figured this out,it would be expected to observe more well-educated parents amongthose listing a voucher school on top of their application lists andless-educated parents in the other group.

Another aspect to consider regarding the voucher schools is thefact that they admit students based on a first-come first-served queuesystem. It is possible that some households would like to apply tothese schools but abstain from doing so because they did not puttheir children in the queue. In other words, some households mayconsider the voucher schools as impossibles. Without access to dataon which children are in the queue and no information about whether


the voucher schools are oversubscribed or not, it is not possible to de-termine for which households this might be the case. This will there-fore be treated as a question of correctly defining the choice set, anddiscussed further once the model has been presented.

2.4.2 A Discrete Choice Framework

Assume that household i gets the following utility of their child beingadmitted to school j:

Uij = X ′ijβi + εij , (2.1)

where Xij is a vector of school characteristics that vary at the schoolor school-individual level, βi is a vector of preference parameters forthese attributes, and εij is an iid error term. Xij includes the loga-rithm of the distance to school, an indicator for whether the school islocated in one’s neighborhood, an indicator for schools attended byan older sibling, an indicator for schools offering all grades (K-9), thenumber of students enrolled, the average teacher experience (mea-sured in years), the share of students with a foreign background, theshare of immigrant students recently arrived to Sweden and the shareof students with at least one highly-educated parent. The selection ofvariables included in this vector is discussed in Appendix 2.C.

The household can choose between the schools j ∈ Ji, where Jidenotes the choice set of household i. Let yi be an indicator equal toj if household i chooses school j. Assuming that households act asutility maximizing agents requires:

yi = j iff Uij > Uik ∀k ∈ Ji. (2.2)

The probability of household i choosing school j is then:

Pr[yi = j] = Pr[Uij > Uik ∀k ∈ Ji], (2.3)

110 CHAPTER 2

which can be rewritten, making the standard assumption that theerror term εij has a type 1 extreme value distribution, as:

Pr[yi = j] = eX′ijβi∑m

k=1 eX′ikβi

. (2.4)

To estimate the preferences parameters, a mixed logit model isused. This model is preferred to other discrete choice models becauseit relaxes the IIA assumption and the restrictions on the substitutionpatterns that it implies.29 Since the β-vector consists of random pa-rameters, the choice probability has to be integrated over all possiblevalues of β:

Lij(βi) = Pr[yi = j] =∫

ex′ijβi∑m

k=1 ex′ikβif(β)dβ. (2.5)

Equation (2.5) is the likelihood function of household i and ex-presses the likelihood of household i choosing school j as a functionof β, where f(β) is the density function of β (in this context alsoreferred to as the mixing distribution). The mixing distribution de-scribes the distribution of β in the population and the parametersare assumed to be normally distributed random variables. An excep-tion is the variable measuring the distance to school, the indicator forenrolled siblings, and the indicator for schools in one’s own neighbor-

29Independence of irrelevant alternatives (IIA) implies that the conditionalprobability of choosing alternative x over alternative y is independent of any otheralternatives. IIA is often illustrated by the red-bus/blue-bus problem. AssumingIIA, the conditional probability of commuting by car when choosing between carand a red bus is independent of whether a blue bus is introduced. If the red andblue buses are equal except for the color, the introduction of a blue bus shouldhave little effect on the probability of commuting by car. Rather, it would be ex-pected to decrease commuting by the red bus and thereby increase the conditionalprobability of commuting by car given the choice between car and red bus.


hood which are assumed to be fixed across individuals.30As there is no variation in the choice set of schools, identification of

the preference parameters comes from the rank-ordered applicationlists. They allow the substitution pattern of each household to beobserved. One can think of the rank-ordered lists as three sequentialchoices and the model is required to replicate this substitution patternby restricting the parameters to match all three choices.31 Let y1

i

be the top choice, y2i the second choice and y3

i the third choice.Theprobability that household i would apply to school j1, j2 and j3 in thatorder can be expressed as the product of the individual probabilitiesof these choices:

Pr[y1i = j1, y2

i = j2, y3i = j3] =

∫ 3∏r=1

ex′ijrβh∑

k∈Jriex′ikβif(β)dβ. (2.6)

Note that the choice set is now indexed by r, as it will changefor each choice that is made. Suppose that there are m schools tochoose from. Once a household has chosen which school to rank ontop of their application list, they will have m−1 schools left to choosefrom when picking their second-ranked school. For the third rank, thechoice set is reduced to m− 2 schools.

2.4.3 Estimation

To get the log likelihood function, equation (2.6) is summed over allhouseholds:

30A continuous mixing distribution is specified, but it is possible to have adiscrete mixing distribution in which case the mixed logit model turns into alatent class logit model.

31Berry et al. (2004) discuss how to identify preference parameters using“second-choice” data.

112 CHAPTER 2

LL =∑i

∑j∈J1

∑j∈J2

∑j∈J3

d1ijd

2ijd

3ijln(Pr[y1

i = j1, y2i = j2, y3

i = j3]), (2.7)

where drij is equal to one if household i had school j at rank r. Asequation (2.7) does not have a closed form solution for β, estimation isdone using maximum likelihood. 200 draws of β from its distributionf(β) is used and for each draw, the choice probability in equation 2.6is calculated. Inserting the simulated choice probability in equation(2.7) gives the estimated log likelihood. The value of the parametersof the distribution of β that maximizes the value of equation (2.7) isthe maximum likelihood estimator.

To gain power, the model is estimated on the pooled populationof truth-telling households from 2011 to 2014. The model is estimatedseparately for subgroups, defined by the interaction between parentaleducation and foreign background, to allow for heterogeneity in pref-erences at the group level. The school attributes are included with aone-year lag (except for distance, the dummy for whether the schoolis located in one’s own neighborhood and the dummy for having asibling enrolled in the school), assuming that parents form their ex-pectations based on what they observe by the time they make theirchoices. Given that the school attributes seem stable over time (seeFigure 2.10), this behavioral assumption seems reasonable.

The estimated preference parameters are presented in Table 2.6,with standard errors clustered at the 250 × 250 meters level.32 Thecoefficients should be interpreted as the change in utility from a oneunit change in the corresponding school attribute. The results suggestthat households are more likely to apply to schools located closer tohome and in their own neighborhood as well as schools where a sibling

32Table 2.13 provides evidence that the results are robust to varying the levelof clustering to 500× 500 and 750× 750 meters.


is already enrolled. Larger schools and schools offering all grades upto year 9 are also more likely to be applied to. Teacher experience,measured by the average number of years in the occupation, does notturn out significant for any group. The remaining variables includedin the model describe the peer composition of the schools. For the fullestimation sample, schools with a smaller (larger) share of studentswith foreign backgrounds (high-educated parents) are more likely tobe higher ranked by the parents.

However, estimations by subgroup uncover importantheterogeneity. Natives have stronger preferences against peers withforeign background. High-educated immigrant households also prefersmaller shares of students with foreign background, which is not truefor their low-educated counterparts. Further, high-educated nativesare less likely to apply to schools with higher shares of newly arrivedimmigrant-students, while the opposite is true for low-educatedimmigrant households. Finally, a school is more likely to be appliedto by native households if the peers have well-educated parents,especially by the high-educated households themselves.

114 CHAPTER 2

Tab

le2.

6:Pr

eference

ParametersEs

timated

with

aRa

nkOrdered

Mixed

LogitModel

All

Nativeba

ckgrou

ndForeign

backgrou

nd

Higheducation

Low

education

Higheducation

Low

education

Mean

SDMean

SDMean

SDMean

SDMean

SDDistance

-1.892

∗∗

∗-2.282

∗∗

∗-1.878

∗∗

∗-1.662

∗∗

∗-1.969

∗∗

∗

(0.0658)

(0.124)

(0.0839)

(0.111)

(0.0871)

Nbdscho

ol2.360∗

∗∗

2.624∗

∗∗

2.302∗

∗∗

1.622∗

∗∗

1.634∗

∗∗

(0.148)

(0.290)

(0.218)

(0.251)

(0.268)

Enrolledsibling

3.675∗

∗∗

4.506∗

∗∗

4.067∗

∗∗

3.565∗

∗∗

3.121∗

∗∗

(0.124)

(0.272)

(0.271)

(0.234)

(0.187)

Grade

90.264∗

∗∗

0.0223

0.204∗

0.116

0.279∗

∗0.613∗

∗0.416∗

∗∗

0.367∗

0.423∗

∗∗

0.00414

(0.0666)

(0.208)

(0.0898)

(0.0874)

(0.101)

(0.218)

(0.0991)

(0.179)

(0.118)

(0.216)

Nstud

ents

0.004∗

∗∗

0.000

0.002∗

∗∗

0.002

0.003∗

∗∗

-0.000

0.004∗

∗∗

-0.000

0.006∗

∗∗

0.000

(0.0004)

(0.0001)

(0.0006)

(0.0018)

(0.0004)

(0.0003)

(0.0005)

(0.0002)

(0.0007)

(0.0005)

Avg

teacherexp.

0.0289

0.225∗

∗∗

0.0615

0.292∗

∗∗

0.0374

0.202∗

∗∗

0.00889

0.211∗

∗∗

0.000854

0.164∗

∗∗

(0.0187)

(0.0172)

(0.0352)

(0.0274)

(0.0223)

(0.0349)

(0.0271)

(0.0373)

(0.0264)

(0.0296)

Foreign

backgr.

-2.006

∗∗

∗2.911∗

∗∗

-3.097

∗∗

∗-3.362

∗∗

∗-3.609

∗∗

∗3.035∗

∗∗

-1.206

∗2.606∗

∗∗

-0.237

3.180∗

∗∗

(0.351)

(0.315)

(0.549)

(0.483)

(0.446

)(0.409)

(0.572)

(0.484)

(0.651)

(0.457)

New

lyarrived

0.315

1.614

-4.277

∗∗

2.275

0.321

1.312

-0.468

2.266

1.604∗

-0.357

(0.542)

(1.809)

(1.442)

(3.416)

(1.085

)(6.618

)(0.709)

(1.790)

(0.710)

(0.334)

High-educated

1.801∗

∗∗

-2.865

∗∗

∗5.203∗

∗∗

-4.049

∗∗

∗1.233∗

1.624

-0.284

2.591∗

-1.011

3.684∗

∗

(0.410)

(0.609)

(0.666)

(0.792)

(0.551)

(2.396

)(0.723)

(1.053)

(0.833)

(1.269)

Observation

s187953

62038

42979

35444

47492

Not

es:Thistablepresents

theestimated

meanan

dstan

dard

deviationof

thepreference

parameters.

∗p<

0.05,

∗∗p<

0.01

,∗

∗∗p<

0.00

1.Stan

dard

errors

inpa

rentheses(clustered

atthe250times

250meterslevel).The

sampleinclud

esallprim

aryscho

olstarters

inBotky

rka2011-2014

categorizedas

(partially)truth-telling(see

Section2.4.1fordetailson

thiscategorization

).Distanceisthedistan

cebetweentheho

mean

dthescho

ol,

measuredin

log(meters).Nbdscho

olis

adu

mmyforthescho

olbeing

locatedin

theneighb

orho

odof

residence.

Enrolledsiblingis

adu

mmyfor

having

asiblingalread

yenrolled

inthescho

ol.Grade

9is

adu

mmyforK-9

scho

ols.

Nstud

ents

indicatesho

wman

ystud

ents

areenrolled

ingrad

eK-5.Avg

teacherexp.

istheaveragenu

mber

ofyearsin

theoccupa

tion

.Foreign

backgrou

ndindicatesthesharebornab

road

orwithbothpa

rents

bornab

road

.New

lyarrivedindicatestheshareof

children

immigrating

toSw

eden

thelast

four

years.

High-educated

indicatesthesharewithat

leaston

epa

rent

withun

iversity

leveleducation.


These results line up relatively well with previous literature, sug-gesting that the location and peer composition of schools are impor-tant choice determinants. An important difference is that this studyfails to document any strong preferences for school attributes relatedto the quality and performance of the schools. It is left for futurework to investigate this in more detail, but a possible explanationmay be that this study analyzes primary school starters of age sixwhich is considerably younger than the populations studied in mostprevious studies. Once children grow older and start middle school oreven higher levels of education, quality concerns may become moreimportant.

2.4.4 Validity of the model

In order to assess the prediction accuracy, a ten-fold cross validationof the model is performed. First, the pooled population of primaryschool starters is divided into ten equally sized folds. Next, the modelis estimated on nine folds and the resulting parameters are used topredict the ranking for households in the excluded fold. This is re-peated ten times, excluding one fold at the time. The results are thencombined and the predicted rankings are compared to the rankingsobserved in the data.

Figure 2.3 displays the difference between the predicted and ob-served rank, separately for observed top-, second-, and third-rankedchoices. The model performs best when predicting the top-rankedschool. For 73 percent of all households, the school predicted to betop-ranked is the same as that observed on top of their application list.The prediction accuracy is 42 percent for the second-ranked school,and 24 percent for the third-ranked school. Although the predictionaccuracy decreases, very few predictions are far off suggesting thatthe model is well-suited for replicating the overall substitution pat-tern in this population. For example, the predicted top-ranked schoolis within two ranks of the observed top-ranked school for 89 percent

116 CHAPTER 2

Figure 2.3: 10-Fold Cross Validation of Estimated Preference Parameters

(a) Top-Ranked

0

.2

.4

.6

.8

Shar

e

0 5 10 15 20Rank minus predicted rank (abs value)

(b) Second-Ranked

0

.2

.4

.6

.8

Shar

e0 5 10 15 20

Rank minus predicted rank (abs value)

(c) Third-Ranked

0

.2

.4

.6

.8

Shar

e

0 5 10 15 20Rank minus predicted rank (abs value)

Notes: This figure displays histograms of the deviation of the predicted rank from the observed rankon the application lists separately for the observed top, second and third choice. The predicted rankwas created using a 10-fold cross validation approach, estimating the rank ordered mixed logit modelexcluding one fold at the time and then using the estimated parameters to predict the choices of theexcluded fold. The deviation of the predicted rank is calculated by subtracting the predicted rank fromthe rank observed on the application list and taking the absolute value.

of the households. Corresponding numbers for the second-, and third-ranked schools are 82 and 78 percent.

Next, to determine whether the presence of impossible voucherschools impact the estimated parameters, the robustness of the es-timations is examined by excluding non-listed voucher schools fromthe choice set. The results are shown in Table 2.14 in Appendix 2.D.Except for the coefficient on the share of newly arrived immigrantchildren, the estimates are not impacted much from this restriction.The change in the mentioned coefficient is not surprising, giving thatthere are quite few newly arrived immigrant children in the popu-lation and many of them attend one of the schools excluded in thisrobustness check.

Finally, a number of robustness checks is performed to confirmthat the endogeneity of households’ residential location is not a bigconcern. If households choose where to live based on the geographyof the school market, specifically by moving close to the preferredschool, preferences for proximity may be overestimated. Observingrank-ordered application lists, as in this study, makes this problemless severe compared to studies with access only to the top-ranked


choice. Rank-ordered application lists makes it possible to observethe substitution pattern between schools. This makes it possible toinfer how other school attributes are valued relative to proximity tothe school.33

Nevertheless, the model is estimated on a subsample restrictedto households that have lived in the same residential location sincetheir school-starting child was born. These are less likely than theothers to have considered the location of a specific primary schoolwhen decided where to live, and the estimates should therefore notoverestimate the proximity parameter. Furthermore, the model is alsoestimated excluding the top-ranked choices on each application list.If households moved in order to increase the admission probability ata specific school, this school will almost certainly be on top of theirapplication list. Excluding these choices will therefore reveal the sub-stitution pattern without the bias introduced by moving close to aspecific school. The results are presented in Table 2.15, in Appendix2.D, suggesting no reason for concern. The estimated parameters arevery similar to the main results in both cases. An exception is thecoefficient on the indicator for schools where an older sibling is en-rolled, which drops in magnitude when top-choices are excluded. Thisis not surprising as excluding the top-ranked choice most often meansexcluding the school where sibling are already enrolled. Importantly,the distance-related coefficients are not changing much due to thisrestriction.

33One could imagine setting up a two-step model in which residential locationand schools are chosen sequentially, but there are no data on e.g. neighborhoodcharacteristics that would be required to model the residential choice. Therefore,the previous literature is followed and this is abstracted from when estimatingschool preferences.

118 CHAPTER 2

2.5 Simulation Strategy and Results

In this section, the estimated preference parameters are used to simu-late the allocation of students to schools under counterfactual admis-sion criteria to determine the impact of different priority structures.

2.5.1 Simulation Strategy

In all simulations, students are assigned to schools using DA, butthe way priorities are determined will vary. The simulation samplesare randomly drawn with replacement from the pooled populationof the 2011 to 2014 cohorts of primary school starters. The studentsin the simulation samples keep their characteristics as observed inthe data, in order to resemble the actual population of students inBotkyrka. The simulation samples faces a choice set of 22 schools, withschool attributes as observed in 2014. Using the estimated preferenceparameters, complete application lists are constructed for each childby calculating and ranking all schools according to their expectedindirect utility. A randomly drawn error term is added to the utilityof each student. The students are then assigned schools using DA andthe priority structure under evaluation and the allocation of students,and its implied level of school segregation, is observed.34

Both the short- and long-term impact of different priority struc-tures is studied. The short-term simulations should be interpreted asthe immediate effect of implementing the priority structure. This ef-fect is simulated using 250 independently drawn simulation samples,each facing the same choice set of schools. However, the effects maybe dynamic as the implementation of a new priority structure couldimpact the peer composition of the schools which in turn may changehow households rank the schools next year. To study this, the long-term impact is simulated, using 100 independently drawn simulation

34Note that the same admission rules are imposed for both public and voucherschools in the simulations.

2.5. SIMULATION STRATEGY AND RESULTS 119

samples to enter the system in year t = 0 (where schools are as ob-served in 2014). Once the students in these simulation samples havebeen allocated, the peer composition of the schools is updated anda new set of 100 independently drawn simulation samples faces theupdated choice sets. This is iterated until t = 20, to represent thelong-term impact over a 20-year period.

When operating with significant overcapacity in the schooling sys-tem, as Botkyrka did during 2011 to 2014, the priority structures be-come less binding as there is less congestion. In other words, there willbe fewer situations where two students are competing for the sameseat and the priority rules come into effect to determine which studentto accept and reject. In order to understand the impact of alternativepriority structures, the overcapacity is removed in the simulations byincreasing the size of the cohort in the simulation sample to 1,179,which is equal to the number of seats available in 2014 at the schoolsincluded in the choice set.

The priority structures imposed, described in detail below, do notintroduce any strategic incentives. Together with DA, the simula-tion sample is therefore faced with a strategy-proof school choice pro-gram.35 For each student, all schools are ranked according to theirindirect utility of each school. In the long-run simulations, as schoolattributes are updated, some schools may become more popular overtime eventually reaching a point where some households start to con-sider these school as impossible. In reality, households may then ex-clude these schools from their application list but in the simulationsthe schools will be included. This is not a problem, since the outcomesof interest in this study are related to the assigned school and not theschools applied to per se. Whether an impossible school is includedor not has no effect on the assignment, it is only problematic whentrying to infer preferences from the application lists.

35Although DA is strategy-proof, the way priorities are determined could in-troduce strategic incentives. This is for example true for priority structures wherea student’s priority is affected by the ranking of the school applied to.

120 CHAPTER 2

In addition to looking at the three different priority structures,the level of segregation using school catchment zones is presented.This was the prevailing system in Botkyrka prior to the introductionof school choice and is still used in other countries to assign studentsto schools. With catchment zones, each school has a geographicallydefined area and all students living in this area are automatically as-signed to that school. Data on exactly how the catchment zones weredefined are not available. To approximate assignment using catch-ment zones, students are assigned to schools in a way to resemblethe purpose when designing the catchment zones in Botkyrka. First,every student is tentatively assigned to their nearest school. If morestudents are assigned to a school than there are available seats, thestudent with shortest additional distance to the next nearest schoolis reassigned to that school. This is repeated until all students areassigned a school and no school admits more students than it has ca-pacity for. This allocation is used a an approximation of the allocationunder school catchment zones.

2.5.2 Definition of Alternative Priority Structures

Here, the three different priority structures under evaluation are pre-sented.

Lottery-based priorities Each student is assigned a randomlydrawn lottery number determining his or her priority at every school,i.e. single tie-breaking. Abdulkadiroğlu et al. (2009) show that thisis favorable in terms of student welfare compared to multiple tie-breaking where a separate lottery is held for each school. Note thatDA with random priorities using single tie-breaking is equivalent torandom serial dictatorship (Abdulkadiroğlu and Sönmez, 2003).

Proximity-based priorities Students residing closer to the schoolare given priority over students residing further away. Distance is


calculated as the crow flies.

Proximity-based priorities with reserved seatsProximity-based priorities are combined with affirmative action. Ateach school j, a share st of all seats are reserved for students oftype t, where st equals the share of type t students in the cohort.Priorities are determined by proximity, but type t students havepriority to the seats reserved for their type over all other students.The types are defined by the interaction of parental educationand foreign background. For example, if there are 20 percenthigh-educated immigrant households in the population, each schoolwill reserve 20 percent of its seats to children from high-educatedimmigrant households. If there are more high-educated immigrantchildren than the seats reserved, those residing nearest the schoolare given priority.

This is a version of controlled school choice. The first theoreticalapplication of controlled school choice, by Abdulkadiroğlu and Sön-mez (2003), focused on racial quotas. They use “hard” quotas, imply-ing that if a seat is not taken by a student of the type it was reservedfor, it is left empty. Kojima (2012) shows that this way of implement-ing quotas can actually hurt rather than help the students meant tobenefit from the policy. Using simulations, Hafalir et al. (2013) con-firm the the finding of Kojima (2012) is not an exception, but ratherto expected when using quotas of this type. Instead, they suggest adifferent implementation of affirmative action in school choice calledminority reserves, where minority students are given higher priorityas long as the seats reserved for the minority are not full. However, ifthere are not enough minority applicants, the seats become availableto majority students in order to avoid empty seats.

In this study, the recommendation by Hafalir et al. (2013) is fol-lowed. Appendix 2.E describes how the quotas are implemented inthe simulations.

122 CHAPTER 2

2.5.3 Impact on School Segregation

The main outcome of interest is the level of school segregation underdifferent priority structures. School segregation by student SES is de-fined as the between-school variance in SES over its total variance,yielding a measure between 0 and 1. This has an intuitive interpre-tation as the variance in student SES that is explained by the schoolassignment and can be obtained by the R2 from regressing studentSES on a full set of school dummies. A student’s SES is measuredby their predicted GPA, which should be interpreted as a compos-ite measure of the student’s socioeconomic background. In Appendix2.F.1, the details on how this measure is constructed is described.

Figure 2.4 displays the distribution of school segregation by SESover the 250 simulation samples for each priority structure and assign-ment using school catchment zones. Firstly, note that school segrega-tion is essentially the same with proximity-based priorities as whenchildren are assigned to schools using school catchment zones, withabout 18 percent of the variation in predicted GPA being explainedby the assigned school. Secondly, abandoning proximity-based priori-ties decreases school segregation. This is not surprising, given the highlevel of residential segregation which spills over into the educationalsystem to a larger extent when priorities are based on residential lo-cation. When a lottery determines priorities instead, the reduction inschool segregation is about 8.5 percent while reserved seats leads toa 20 percent reduction.

While lottery-based priorities cut the ties between residential andschool location in terms of admission, the relationship is not elim-inated as parents have strong preferences for proximity. Still, thereare households willing to let their children travel further in order toattend a more preferred school than those nearby their home. Someof these households are going to get a high enough lottery numberto attend a school further away which they would not have been ad-mitted to if their priority had been determined by their residential


Figure 2.4: Distribution of School Segregation by SES in the Short RunUnder Alternative Priority Structures

0

5

10

15

20

25

30

Den

sity

.1 .15 .2 .25SES segregation

Catchm. zones ProximityLottery Res. seats

Notes: This figure displays the distribution of the share of the variance in student SES that isexplained by the assigned school (see Appendix 2.F.1 for details about how this segregation measureis defined) under each alternative priority structure and assignment using catchment zones. Details onthe different priority structures can be found in Section 2.5.2. The distribution is computed using 250simulations of independently drawn samples. The lines are fitted using Stata’s kdensity command,using the Epanechnikov kernel and optimal bandwidth selection.

location. Reserved seats strengthen this effect by explicitly prioritiz-ing students that contribute to the diversity of the school applied to,for example immigrants applying to schools with a mainly native stu-dent population. This priority structure thus encourages reallocationsthat lead to reduced levels of segregation across schools. Another rea-son why the lottery does not lead to the same level of reduction is thatit may also be giving for example native students in immigrant denseschools the opportunity to opt out of their local school and attend aschool with fewer immigrants further away. The impact of the lottery-based priority structure is thus the net effect of both integrating andsegregating reallocations of students.

The analysis is complemented using another type of segregationmeasure; the Duncan Dissimilarity Index. This measure is used tolook at segregation between two types, typically a minority versusmajority group. The definition of the measure is presented in Ap-

124 CHAPTER 2

pendix 2.F.2. In this study, it will be used to look at segregation byparental education and migration background. Figure 2.14, in Ap-pendix 2.G, presents the results. While proximity-based assignmentagain yields essentially the same level of segregation by parental ed-ucation as using catchment zones, it is noted that students becomeslightly less sorted in terms of foreign background when allowed tochoose schools. Next, the lottery-based priorities do not substantiallyimpact school segregation compared to proximity-based assignment.Only when seats are reserved is there an impact comparable to theeffect on student SES.

This has two potential explanations. Since lottery-based prioritiesdecrease segregation by SES using a composite measure including sev-eral dimensions of the student’s background, lottery-based prioritiesimpacts the distribution of students across schools. However, the real-location of students may take place within the group of students withsimilar migration backgrounds or parents with similar educationalbackgrounds. For example, two students with foreign background mayhave very different SES, but if they switch schools the DDI will notbe impacted. It could also be that the integrating and segregatingeffects completely offsets each other when looking at these measures.Finally, another explanation lies in the definition of the DDI measure.This measure is only impacted when students switch from a schoolwhere they belong to a majority group to a school where they belongto a minority group. In summary, these measures seem not to captureimportant aspects of the students’ backgrounds and how this impactsthe composition of students at the schools.

To further understand how these priority structures affects the al-location, the share with highly educated parents and the share with aforeign background is presented at the school level. Figure 2.5 showsthe change in these shares for lottery-based priorities and reservedseats relative to priorities based on proximity. Note that the sign ofthe change is the same for most schools, but the magnitude is ampli-


Figure 2.5: Average Change in Peer Composition in the Short Run bySchool under Alternative Priority Structures

(a) High-Educated Parents

-.15 -.1 -.05 0 .05 .1Ppt change relative to proximity-based priorities

Reserved seats Lottery

(b) Foreign Background

-.2 -.1 0 .1 .2Ppt change relative to proximity-based priorities

Reserved seats Lottery

Notes: This figure displays the change in the peer composition by school when reserving seats andusing lottery-based priorities relative to the baseline of using proximity-based priorities. Details onthe different priority structures can be found in Section 2.5.2. The change is calculated by taking theaverage change of 250 simulations of independently drawn samples. High-educated is defined as havingat least one parent with university level education and foreign background is defined as being bornabroad or having two parents born abroad.

fied when seats are reserved for underrepresented groups. The peercomposition changes dramatically at some schools. For example, theshare of students with foreign background drops by 20 percentagepoints in one school and increases by about 15 percentage points inanother. At the same time, some schools are essentially unaffectedby how priorities are determined. This suggests that the reduced seg-regation is explained by large changes in the peer composition at aspecific set of schools.

The heterogeneity across the three neighborhoods in Botkyrkaand the fact that the majority of the households seem to have strongpreferences for sending their children to a school within their ownneighborhood, motivates an analysis at the neighborhood level. Fig-ure 2.6 shows that segregation is generally higher in the west andeast, where the decrease in segregation is also greater when abandon-ing proximity-based priorities. Table 2.16, in Appendix 2.G, showsthat most students stay in a school in their own neighborhood. With

126 CHAPTER 2

proximity-based priorities, the schools in the east are populated onlyby students living in the east, but with other types of priorities, someseats are taken by students from the north. Schools in the west areleast affected by the priority structures in terms of which neighbor-hood their students come from.

The preference for attending a school within one’s neighborhoodseems to restrict how much segregation can be reduced within a schoolchoice program while still respecting parents’ preferences. Burgesset al. (2015) study school choice in England and finds that a largepart of the differences in which schools are chosen by advantaged anddisadvantaged households are explained by differences in availableschools rather than differences in underlying preferences for schoolattributes.

In order to study how much the residential segregation contributesto the segregated school assignments, students are randomly reallo-cated in terms of their residential location and the simulations arerepeated.36 This corresponds to a scenario in which there are somedifferences between household types in terms of their preferences forschools but no residential segregation. Figure 2.18-2.21, in Appendix2.G, shows that segregation decreases substantially compared to Fig-ure 2.4. Without the residential segregation, proximity-based priori-ties give an allocation of students to schools where about three percentof the variation in student SES is explained by the assigned school. Asall groups have strong preferences for proximity, it is not surprisingthat the residential segregation has a large impact on school segrega-tion.

In addition, a scenario in which the residential segregation is asobserved in the data, but all households have the same preferences

36To keep the overall residential patterns, each sampled student is given a lo-cation where at least one primary school starter was observed to reside. Further,when randomizing locations, weights proportional to the number of students re-siding in each location are used in order to replicate to overall population densityin each location.


Figure 2.6: Distribution of Segregation by SES in the Short Run underAlternative Priority Structures, by Neighborhood

(a) North

0

5

10

15

20

25

30

35

Den

sity

.05 .1 .15 .2 .25SES segregation


(b) East

0

5

10

15

20

25

30

35

Den

sity



(c) West

0

5

10

15

20

25

30

35

Den

sity



Notes: This figure displays the distribution of the share of the variance in student SES that is explainedby the assigned school (see Appendix 2.F.1 for details about how this segregation measure is defined)under each alternative priority structure and assignment using catchment zones. Details on the differentpriority structures can be found in Section 2.5.2. The left panel displays the segregation among schoolslocated in the north, the middle panel for schools in the east and the right panel for schools in thewest. The distribution is computed using 250 simulations of independently drawn samples. The lines arefitted using Stata’s kdensity command, using the Epanechnikov kernel and optimal bandwidth selection.

128 CHAPTER 2

Figure 2.7: Distribution of the Share Assigned Top and Top-Three Choicesin the Short Run under Alternative Priority Structures

(a) Admitted to Top Choice

0

10

20

30

40

Den

sity

.7 .72 .74 .76 .78 .8 .82Share admitted

Proximity LotteryRes. seats

(b) Admitted to Top-Three Choices

0

10

20

30

40

Den

sity

.84 .86 .88 .9 .92Share admitted


Notes: This figure displays the distribution of the share assigned to a top (left panel) and a top-three(right panel) choice under each alternative priority structure. Details on the different priority structurescan be found in Section 2.5.2. The distribution is computed using 250 simulations of independentlydrawn samples. The lines are fitted using Stata’s kdensity command, using the Epanechnikov kerneland optimal bandwidth selection.

for schools is studied. This shifts the distribution of the DDI slightlyto the left, implying that differences in preferences explain part of thesegregation across schools but not as much as residential segregation.Given these preferences, the priority structures impact segregation inthe same direction as in the main scenario.

2.5.4 Impact on Other Outcomes

In addition to school segregation, another important outcome of aschool choice program is how well it satisfies the parents’ preferencesfor schools. As school choice is meant to give parents a say aboutthe school their children will attend, its credibility depends on itsability to assign students to highly enough ranked schools. An obviousoutcome to look at is the share assigned to a preferred school. Panela) of Figure 2.16 shows the share assigned to their top choice andpanel b) the share assigned to a top three choice under each prioritystructure.


Abandoning proximity-based priorities decreases the shareassigned to their top choice from 78.89 to 77.37 (76.40) whenusing lottery (reserved seats), implying that between 25 and 35fewer students are assigned to their most preferred school. Thefact that proximity-based priorities outperform the other prioritystructures in assigning students to their top choices, is explained bythe fact that this priority structure aligns well with the majorityof households having a strong preference for schools located neartheir residence. In other words, proximity-based priorities tendto give most students high priority to schools that they rankhighly. However, the difference decreases to the magnitude of onepercentage point when looking at the share assigned to a top threechoice, from 89.97 to 88.59 percent for both lottery-based prioritiesand when using reserved seats.

The indirect utility of the assigned school can also be calculated,under each priority structure. Figure 2.15, in Appendix 2.G, showsthat while utility decreases when switching from proximity to lot-tery and again switching from lottery to reserved seats, the magni-tude is small. The welfare loss of not assigning students to their topchoices may therefore be acceptable for school districts putting muchweight on fighting school segregation, since households seem contentwith their second and third choices as well. To see whether there isheterogeneity in how households are impacted by different prioritystructures, Figure 2.17 in Appendix 2.G shows utility by householdtype. While the changes are generally very small, utility decreases forhigh-educated native households when they lose the privilege to theschools located near their homes.

Another important dimension is how far away the assigned schoolis located. Both previous literature and the results in this study sug-gest that proximity is an important factor. Figure 2.8 shows that theaverage distance to the assigned school is about 1041 meters when us-ing proximity-based priorities and increases by up to 180 meters when

130 CHAPTER 2

using one of the other ways to determine priorities. This increase isdriven by students assigned to a top-choice school, which points tothe complexity of using distance to school to evaluate whether oneallocation is better than another.

Figure 2.8: Distribution of Distance to Assigned School in the Short Rununder Alternative Priority Structures

0

.0025

.005

Den

sity

800 1000 1200 1400Meters

Proximity LotteryRes. Seats

Notes: This figure displays the distribution of distance (in meters) to assigned school under eachalternative priority structure. Details on the different priority structures can be found in Section2.5.2. The distribution is computed using 250 simulations of independently drawn samples. The linesare fitted using Stata’s kdensity command, using the Epanechnikov kernel and optimal bandwidthselection.

2.5.5 Long Run Simulations

As the priority structures impact the allocation of students to schools,some schools experience quite large changes in peer composition (thiswas evident in Figure 2.5, in Appendix 2.G). This could imply that thedemand for certain schools changes, as parents observe how the peercomposition changes. For example, in Section 2.4 it was noted thathouseholds prefer schools with less immigrants. If a school experiencesa sharp increase in the number of children with foreign background,parents might be inclined to avoid this school the following year. Tostudy the dynamic effects of changing the priority structure, results


from simulations in the long-run are presented. Figure 2.22-2.23, inAppendix 2.G, show segregation measured by student SES, foreignbackground and parental education over a 20 year period. Schoolsegregation is remarkably stable over time, suggesting that the initialchanges to the peer composition do not offset any greater changesover time.

132 CHAPTER 2

2.6 An Information Experiment

This section will explore whether households in Botkyrka are lackinginformation about school performances, when submitting their ap-plication lists. Previously, it was concluded that school performancedoes not seem to be an important determinant when choosing schools.Throughout the simulations, households were therefore assumed notto consider this when ranking schools.

Previous literature suggest that households may not consider thisbecause they lack information about schools’ performances ratherthan not putting any weight on this. For example, Hastings and Wein-stein (2008) show an increasing share of low income families in NorthCarolina choosing a better performing school when provided withinformation on school test scores. In Malmberg et al. (2014), usingsurvey data from Swedish municipalities collected in 2012, many par-ents report that it is hard to make an informed choice due to lackof information about the schools. Motivated by this, a randomizedexperiment was conducted in Botkyrka in 2016 to explore whetherproviding information about schools’ performances impact the rank-ing of schools on the households’ application lists.

2.6.1 The Field Experiment

All of the 1,307 children expected to start primary school in 2016 wererandomly selected to a treatment or control group. The treatmentgroup was provided with information about the performance of allgrade K schools in Botkyrka, at the same time as the municipalityinformed all households about the school choice program. The samedata sets as described in Section 2.3.2 are available for this cohort. 159children, for whom no application lists were submitted are excludedfrom the analysis.37 In addition, school choices are observed for 26

37These are likely households residing in Botkyrka but about to move to anothermunicipality explaining their non-participation in the school choice program.

2.6. AN INFORMATION EXPERIMENT 133

children who were never assigned a control or treatment group. Theseare not part of the analysis either and are likely children of householdsthat had not yet moved to Botkyrka when the randomization wasdone. In the end, the sample consists of 1,148 children, 580 of whombelong to the treatment group and 568 to the control group. Summarystatistics for these primary school starters are shown in Table 2.7.

Table 2.7: Summary Statistics for the Primary School Starters,2016

Control group Treatment groupHigh-educated (share) 0.54 0.56

(0.50) (0.50)Foreign background (share) 0.43 0.46

(0.50) (0.50)Observations 568 580

Notes: This table presents the mean and standard deviation of each variable for thesample of primary school starters participating in the school choice program in 2016,separately for those assigned to the control group and the treatment group. High-educated indicates the share with at least one parent with university level education.Foreign background indicates the share born abroad or with both parents born abroad.

Households in the treatment group received a letter with schoolperformance information. More specifically, for each school, the av-erage pass rate on the standardized tests in grade three in Swedish,mathematics and English were calculated and averaged over the years2012 to 2014.38 In addition, they were provided with a measure of howeach school performed in relation to how it was expected to performgiven its student composition. This measure was constructed by firstestimating the following regression on the full population of primaryschools in Sweden separately for each year 2012 to 2014:

Pass ratej = α+ X′jβx + εj (2.8)

where X′i is a vector of variables describing the student composi-tion of school i including the share of boys, share with foreign back-

38In Sweden, standardized tests are taken for the first time in 3rd grade andlater in 6th and 9th grade of primary school.

134 CHAPTER 2

ground, share born abroad, share newly-arrived immigrants and theshare with high-educated parents.39 The average of the residuals forthe Botkyrka schools from these yearly estimations is used as themeasure of school performance relative expected performance giventhe student composition. Table 2.8 presents summary statistics forthese two measures. There were 25 schools available to choose from,but one school is dropped from the analysis as there are no data onthat school’s performance on the standardized tests.40 In the end,there are 24 schools with an average pass rate of 83 percent. The rel-ative pass rate is -2, indicating that the schools in Botkyrka in generalperform a little worse than expected given their student compositioncompared to primary schools in the rest of Sweden. In both measures,there is considerable variation across schools.

Table 2.8: Performance Measures for Grade K Schools, 2016

Mean SD Min MaxPass rate 83.22917 9.627835 62.5 96.5Relative pass rate -2.020833 4.730749 -15 6.5Observations 24

Notes: This table presents the mean, standard deviation, minimum and maximumfor the sample of primary schools with grade K in Botkyrka in 2016, excluding oneconfessional school for which school performance data are not available. Pass rate isthe share of students passing all standardized tests in Swedish and mathematics takenin grade three. The relative pass rate is the difference between the pass rate and theexpected pass rate given the composition of students at the school (for details on howthe relative pass rate is calculated, see Section 2.6).

The letter including these performance measures can be foundin its entirety in Appendix 2.I. As Botkyrka has a large immigrantpopulation, the letter was translated into nine different languages to

39The variables included in X′i follow the specification used by the SwedishNational Agency for Education to construct their student composition measure(called SALSA).

40The school dropped from the analysis is a confessional school, for which suchdata are not collected by Statistics Sweden. In 2016, there were 22 householdsapplying to this school of which 15 listed it as their top choice.

2.6. AN INFORMATION EXPERIMENT 135

ensure that everyone would be able to understand its content. Contactdetails were included should they have any questions.

To study the impact of the information intervention, the followingregression equation is estimated using OLS:

Yi = α+ βTi + εi, (2.9)

where Yi is the outcome of interest, Ti an indicator equal to one ifhousehold i was assigned to the treatment group and zero otherwiseand εi is a (heteroskedasticity robust) error term. The coefficient ofinterest is β, indicating the change (in percentage points) in the passrate/relative pass rate from receiving the letter with information onschool performance.41 For a causal interpretation of β, randomizationneeds to be successful. Table 2.17, in Appendix 2.H, presents the re-sults from running a regression of the treatment indicator (Ti) on thebackground characteristics of the sample of primary school starters.If randomization was successful, there should be no significant rela-tionship between the background characteristics of the primary schoolstarters and their assigned treatment status. This is confirmed by theresults.

Table 2.9 presents the results, showing no significant impact onthe performance of the top-ranked school or the average performanceof the three schools included on the application lists. The magnitudeof all coefficients are small (e.g. -0.616 would represent a decreaseby less than a percentage point in the average pass rate of the top-ranked schools) and the 95 percent confidence intervals exclude effectslarger than 0.4 percentage points. This has to be considered smallgiven the variation in the schools’ pass rates presented in Table 2.8.In other words, providing information about schools’ performancesdid not nudge households to apply to higher-performing schools. Al-though there results may seem counter-intuitive, they do suggest that

41As neither sampling nor assignment to treatment/control group was clustered,clustered standard errors are not reported (Abadie et al., 2017).

136 CHAPTER 2

school performance is not an important determinant of households’choice of school in this setting. A more thorough investigation of thisis left for the future. For the purpose of this study, it is concluded thatthe observed ranking of schools on the application lists would mostlikely remain even if households’ were provided with better informa-tion about the schools’ performances. This is an important note, asthe conclusions in this paper is based on simulations where house-holds’ ranking of schools is an important input.

Table 2.9: Estimated Treatment Effects of The Information Interven-tion

Top choice Average of top three choices

Pass rate Relative pass rate Pass rate Relative pass rateTreated -0.616 -0.164 -0.527 -0.0612

(0.514) (0.238) (0.443) (0.150)

Constant 85.47∗∗∗ -1.161∗∗∗ 85.69∗∗∗ -1.148∗∗∗(0.354) (0.162) (0.308) (0.105)

Observations 1118 1118 1148 1148Notes: This table presents the estimated effects of the information intervention on theperformance of the school(s) applied to. Effects are estimated using OLS regression.Standard errors are presented in parentheses. Significance levels are indicated by * p<0.1,** p>0.05, *** p<0.01. The dependent variable is the pass rate or relative pass rate ofthe top choice (the two leftwards columns) or averaged over the top three choices (the tworightwards columns).



This study examines the effects on school segregation of altering thepriority structure in a centralized school choice program using the DAmechanism to assign students to schools. Three types of priority struc-tures are evaluated: proximity-based, lottery-based and reserved seatswith proximity as a tie-breaker. Given the widespread use of the DAmechanism in school choice programs and the potential importance ofpriority structures when allocating students, this is an understudiedsubject. The findings indicate that school segregation by student SESdecreases when proximity-based assignments are abandoned, with amagnitude of 8.5 (lottery) to 20 (reserved seats) percent. As the ma-jority of the children will continue in the same school for the entirecompulsory education, one should keep in mind that it implies a rel-atively long exposure time to a less segregated school.

The magnitude is greater when explicitly reserving seats for stu-dents contributing to the diversity of the school. Some householdsapply to schools where their child would belong to an underrep-resented group. Such applications are prioritized under affirmativeaction, while some would randomly receive low priority when usinglottery-based priorities. Moreover, the lottery may also prioritize stu-dents applying to schools in order to escape schools where they arein minority. Hence, while reserved seats encourage reallocations todecrease segregation, lottery-based admission facilitates both segre-gating and integrating reallocations. It is also documented that thedecrease in segregation by student SES is partly explained by hetero-geneity and reallocations within groups, explaining why the lottery-based priorities do not decrease segregation by parental education orforeign background.

Finally, the reduction of school segregation comes at the cost ofincreased distance to the assigned school and allocating less studentsto their top choice. From a social planners perspective, it is a matterof deciding the weights on decreasing school segregation and assigning

138 CHAPTER 2

students to preferred schools. One argument for not putting all theweight on the latter is that parents may not internalize the impact onothers of their school choice(s). Still, the trade-off between decreasingschool segregation and satisfying parents’ preferences may not be verylarge. While the share assigned to a top-choice school decreases whenabandoning proximity-based priorities, the three priority structuresperform about the same in terms of the share assigned to top-threechoices. Furthermore, the increase in distance to school is driven bystudents preferring to travel further in order to get a school that theylike. It should thus not be regarded as a welfare loss.

The impact of varying the priority structures may be dependenton the setting studied. Botkyrka is an interesting case to study, withhigh residential segregation and clearly separated neighborhoods.This study shows that the mobility between neighborhoods islimited, even when admission rules become more beneficial forapplicants from other neighborhoods. The reduced segregation isto a large extent explained by within neighborhood reallocation ofstudents. The impact of changing the priority structures is thusmitigated by parents’ unwillingness to let their children commute toa different neighborhood. In a setting with shorter distance betweenthe neighborhoods, the priority structures may thus have an evengreater impact. On the other hand, this greater impact may beoffset by dynamic effects with greater response to a changing peercomposition. In addition, the mobility is likely lower for youngerchildren, suggesting that the priority structures could have an evengreater impact on for example segregation at the middle school level.

To conclude this study, priority structures may be an importantpolicy tool for school districts to fight school segregation. Relativeto the ease with which they can be modified and the low monetarycost of such a change, the impacts are not negligible. Furthermore,the welfare costs appear relatively small. At the same time, the res-idential location of households plays an important role and restricts


the efficacy at which priority structures impact the allocation of stu-dents. Additionally, one should keep in mind that different prioritystructures are beneficial to different types of households. The questionof how to determine priorities is therefore closely related to what onethinks is a fair way of allocating school seats.

140 CHAPTER 2

References

Abadie, A., Athey, S., Imbens, G. W., andWoolridge, J. (2017). WhenShould You Adjust Standard Errors for Clustering? Working Paper24003, NBER.

Abdulkadiroğlu, A., Pathak, P. A., and Roth, A. E. (2005). TheNew York City High School Match. American Economic Review,95(2):364–367.

Abdulkadiroğlu, A., Pathak, P. A., and Roth, A. E. (2009). Strategy-Proofness versus Efficiency in Matching with Indifferences: Re-designing the NYC High School Match. American Economic Re-view, 99(5):1954–1978.

Abdulkadiroğlu, A. and Sönmez, T. (2003). School Choice: A Mech-anism Design Approach. American Economic Review, 93(3):729–747.

Allen, R. and Vignoles, A. (2007). What Should an Index of SchoolSegregation Measure? Oxford Review of Education, 33(5):643–668.

Berry, S., Levinsohn, J., and Pakes, A. (2004). Differentiated ProductsDemand Systems from a Combination of Micro and Macro Data:The New Car Market. Journal of Political Economy, 112(1):68–105.

Bifulco, R. and Ladd, H. F. (2007). School Choice, Racial Segrega-tion, and Test-Score Gaps: Evidence from North Carolina’s Char-ter School Program. Journal of Policy Analysis and Management,26(1):31–56.

Billings, S., Deming, D. J., and Ross, S. L. (2016). Partners in Crime:Schools, Neighbourhoods and the Formation of Criminal Networks.Working Paper 21962, NBER.

REFERENCES 141

Billings, S. B. and Rockoff, J. (2013). School Segregation, Educa-tional Attainment, and Crime: Evidence from the End of Busingin Charlotte-Mecklenburg. The Quaterly Journal of Economics,129(1):435–476.

Böhlmark, A. and Lindahl, M. (2015). Independent Schools and Long-run Educational Outcomes: Evidence from Sweden’s Large-scaleVoucher Reform. Economica, 82(327):508–551.

Borghans, L., Golsteyn, B. H., and Zölitz, U. (2015). Parental Pref-erences for Primary School Characteristics. The B.E. Journal ofEconomic Analysis and Policy, 15(1):1–33.

Burgess, S., Greaves, E., Vignoles, A., and Wilson, D. (2015). WhatParents Want: School Preferences and School Choice. The Eco-nomic Journal, 125(587):1262–1289.

Burgess, S. and Platt, L. (2018). Inter-ethnic Relations of Teenagersin England’s Schools: The Role of School and Neighborhood EthnicComposition. Discussion Paper Series 1807, CReAM.

Calsamiglia, C., Haeringer, G., and Klijn, F. (2010). ConstrainedSchool Choice: An Experimental Study. American Economic Re-view, 100(4):1860–1874.

Chetty, R. and Hendren, N. (2018a). The Impacts of Neighborhoodson Intergenerational Mobility I: Childhood Exposure Effects. TheQuarterly Journal of Economics, 133(3):1107–1162.

Chetty, R. and Hendren, N. (2018b). The Impacts of Neighbor-hoods on Intergenerational Mobility II: County-Level Estimates.The Quarterly Journal of Economics, 133(3):1163–1228.

De Haan, M., Gautier, P. A., Oosterbeek, H., and van der Klaauw,B. (2015). The Performance of School Assignment Mechanisms inPractice. Discussion Paper 9118, IZA.

142 CHAPTER 2

Duncan, O. D. and Duncan, B. (1955). A Methodological Analysis ofSegregation Indexes. American Sociological Review, 20(2):210–217.

EY (2014). Kan Resursfördelningen Lösa Skolkrisen? En Studie omResursfördelning till Grundskolan.

Fack, G., Grenet, J., and He, Y. (2018). Beyond Truth-Telling: Pref-erence Estimation with Centralized School Choice and College Ad-missions. American Economic Review (forthcoming).

Gale, D. and Shapley, L. S. (1962). College Admissions and the Stabil-ity of Marriage. The American Mathematical Monthly, 69(1):9–15.

Gamoran, A. and An, B. P. (2016). Effects of School Segregationand School Resources in a Changing Policy Context. EducationalEvaluation and Policy Analysis, 38(1):43–64.

Haeringer, G. and Klijn, F. (2009). Constrained School Choice. Jour-nal of Economic Theory, 144(5):1921–1947.

Hafalir, I. E., Yenmez, M. B., and Yildirim, M. A. (2013). Effec-tive Affirmative Action in School Choice. Theoretical Economics,8(2):325–363.

Hanushek, E. A., Kain, J. F., and Rivkin, S. G. (2009). New EvidenceAbout Brown v. Board of Education: The Complex Effects of SchoolRacial Composition on Achievement. Jounal of Labor Economics,27(3):349–383.

Hastings, J. S., Kane, T. J., and Staiger, D. O. (2009). HeterogeneousPreferences and the Efficacy of Public School Choice. WorkingPaper 2145, NBER.

Hastings, J. S. and Weinstein, J. M. (2008). Information, SchoolChoice, and Academic Achievement: Evidence from Two Experi-ments. Quarterly Journal of Economics, 123(4):1373–1414.

REFERENCES 143

He, Y. (2017). Gaming the Boston School Choice Mechanism in Bei-jing. Working Paper 15-551, Toulouse School of Economics (TSE).

Hsieh, C.-T. and Urquiola, M. (2006). The Effects of GeneralizedSchool Choice on Achievement and Stratification: Evidence fromChile’s Voucher Program. Journal of Public Economics, 90(8-9):1477–1503.

Hussain, I. (2013). Not Just Test Scores: Parents’ Demand Re-sponse to School Quality Information. Working Paper https://www.sole-jole.org/13363.pdf (accessed October 31, 2018).

Jackson, C. K. (2009). Student Demographics, Teacher Sorting, andTeacher Quality: Evidence from the End of School Desegregation.Journal of Labor Economics, 27(2):213–256.

Johnson, R. C. (2011). Long-run Impacts of School Desegregationand School Quality on Adult Attainments. Working Paper 16664,NBER.

Karbownik, K. (2016). The Effects of Student Composition onTeacher Turnover: Evidence from an Admission Reform. WorkingPaper 6133, CESifo.

Kojima, F. (2012). School Choice: Impossibilities for Affirmative Ac-tion. Games and Economic Behavior, 75(2):685–693.

Luflade, M. (2017). The Value of Information in Cen-tralized School Choice Systems. Job Market Paperhttps://economics.sas.upenn.edu/sites/default/files/filevault/event_papers/LUFLADE_JobMarketPaper.pdf (ac-cessed October 31, 2018).

Malmberg, B., Andersson, E. K., and Bergsten, Z. (2014). CompositeGeographical Context and school Choice Attitudes in Sweden: A

144 CHAPTER 2

Study Based on Individualls Defined, Scalable Neighborhoods. An-nals of the Association of American Geographers, 104(4):869–888.

Massey, D. S. and Denton, N. A. (1988). The Dimensions of Residen-tial Segregation. Social Forces, 67(2):281–315.

Pathak, P. A. (2011). The Mechanism Design Approach to StudentAssignment. Technical report, Annual Review of Economics.

Pathak, P. A. and Shi, P. (2017). How Well Do Structural DemandModels Work? Counterfactual Predictions in School Choice. Work-ing Paper 24017, NBER.

Roth, A. E. (2007). Deferred Acceptance Algorithms: History, Theory,Practice and Open Questions. Working Paper 13225, NBER.

2.A. PROOF OF CLAIMS REGARDING TRUTH-TELLING 145

Appendices

2.A Proof of Claims Regarding Truth-Telling

Assumptions: Assume that all schools j ∈ N are acceptable andthat the households are not indifferent between any two schools∈ N . Households are risk-neutral and rank the schools in order tomaximize their utility as expressed in equation 2.1. All schools have astrict priority ordering over students, independent of the households’ranking of schools. All students have a positive probability p ofbeing accepted to any school j ∈ N , i.e. ∀h,∀j : pij > 0. pij can beinterpreted as household i’s probability of having a priority abovethe capacity cut off for school j. A school j is defined as safe forhousehold i if the probability of household h being admitted toschool j is equal to 1, i.e. school j is a safe school for household h ifphj = 1. Assume also that households are able to determine whethera school is safe or not.

Proof of claim 1: Suppose that household h lists school j astheir top choice, and that school j is a safe school. Since phj = 1,the expected utility of submitting any ranking where school j isthe top choice is Uj . Suppose now that there exists (at least) oneother school k that is preferred to school j (Uk > Uj). If householdh would modify the ranking and list school k as their top choice,their expected utility would be equal to pkUk + (1 − pk)Uj . Since0 < pk < 1 and Uk > Uj , it follows that pkUk + (1 − pk)Uj > Ujand household h would be better off with the modified ranking.Hence, if household h submits a safe school j as their top choice,there cannot exist any other school k ∈ N that is preferred to school j.

Proof of claim 2: Suppose that household h lists the non-safeschool k as their top choice and the safe school j as their second

146 CHAPTER 2

choice. The expected utility of this ranking is pkUk + (1 − pk)Uj .Suppose that school k is not preferred to school j (Uk < Uj). Ifhousehold h would modify their ranking and list school j as theirtop choice, they would get utility Uj since pj = 1. Since Uk < Ujand pk > 0, this implies that Uj > pkUk + (1 − pk)Uj , whichmeans that household h would be better off with the modifiedranking. Hence, given that school j is ranked second, it must bethat school k is preferred to school j. Next, suppose that thereis another school l that is preferred to school j (Ul > Uj). Ifhousehold h would modify their ranking and rank school l aboveschool j (as their top or second choice), they would get expectedutility pkUk + pl(1 − pk)Ul + (1 − pk − pl(1 − pk))Uj .Since Ul > Uj and pl > 0, this implies thatpkUk + pl(1 − pk)Ul + (1 − pk − pl(1 − pk))Uj > pkUk + (1 − pk)Uj ,and household h would be better off with the modified ranking.Hence, given that school j is ranked second, there cannot exist anyother school l 6= k ∈ N that is preferred to school j.

2.B. DESCRIPTIVE FIGURES AND TABLES 147

2.B Descriptive Figures and Tables

Figure 2.9: School Segregation by Municipality in Sweden, 2011

(a) Parental Education

0.00

0.10

0.20

0.30

0.40

0.50

0.60

DD

I (ed

uc)

Municipality


0.00

0.10

0.20

0.30

0.40

0.50

0.60

DD

I (fo

reig

n)

Municipality

Notes: This figure displays the level of the Duncan Dissimilarity Index (DDI) for all municipalitiesin Sweden in terms of parental education (left panel) and foreign background (right panel). Botkyrkamunicipality is indicated by the red bar. The DDI can be interpreted as the share of students of onegroup (e.q. high-educated parents) that would have to switch to another school in order to produce adistribution of students in each school that matches the distribution of the entire student population.See Appendix 2.F.2 for more details on the DDI.

148 CHAPTER 2

Figure 2.10: School Atributes Over Time, 2011-2014

(a) Number of Students (K-5)

0

100

200

300

400

500

2011 20122013 2014

(b) Average Test Score

0

5

10

15

2011 20122013 2014

(c) Share High-Educated Parents

0

.2

.4

.6

.8

1

2011 20122013 2014

(d) Share Foreign Background

0

.2

.4

.6

.8

1

2011 20122013 2014

Notes: This graph presents the number of students enrolled in grade K-5, the average test score (max= 20) on the standardized tests in Swedish and mathematics taken in grade three, the share with high-educated parents (defined as having at least one parent with university level education) and the shareof students with foreign background (defined as being born abroad or having both parents born abroad)by year (2011-2014) for each school.


Figure 2.11: Correlation Between Test Scores and GPA

100

150

200

250G

PA

10 12 14 16Average test score

Notes: This graph presents a scatter plot of the relationship between a school’s score on thestandardized tests in Swedish and mathematics in grade three (max = 20) and the grade pointaverage in grade nine (max = 320) for all grade K-9 schools in Botkyrka.

150 CHAPTER 2

Figure 2.12: Histograms of School Attributes

(a) Average test score

0

.1

.2

.3

.4

Den

sity

10 11 12 13 14 15Average test score

(b) Share foreign

0

.5

1

1.5

2

Den

sity

0 .2 .4 .6 .8Foreign background (share)

(c) Share high-educated

0

1

2

3

Den

sity

.2 .4 .6 .8High-educated (share)

(d) Share cert. teachers

0

1

2

3

4

5

Den

sity

.4 .6 .8 1Certified teachers (share)

(e) Stud-teacher ratio

0

.05

.1

.15

Den

sity

5 10 15 20 25Student-teacher ratio

(f) Unemployment

0

.005

.01

.015

.02

.025

Den

sity

0 20 40 60 80 100Days in unemployment

(g) Newly arrived

0

5

10

15

20

Den

sity

0 .1 .2 .3Newly arrived immigrants (share)

(h) Parental inc.

0

2.0e-04

4.0e-04

6.0e-04

8.0e-04

Den

sity

1000 2000 3000 4000 5000Parental income (100 SEK)

(i) Allowance receipt

0

1

2

3

4

5

Den

sity

0 .1 .2 .3 .4Allowance recipient (share)

Notes: These graphs display histograms of school attributes for the pooled sample of primary schools2011 to 2014. Average test scores is the school average test score (max = 20) on the standardized testsin Swedish and mathematics taken in grade three. Foreign background is the share of students bornabroad or with both parents born abroad. Highly educated is the share of students with at least oneparent with university level education. Certified teachers is the share of certified teachers at the school.Stud-teacher ratio indicates the number of students per teacher. Unemployment indicates the parents’yearly average of days receiving unemployment benefits. Newly arrived is the share of students thatarrived to Sweden during the last four years. Parental income is the average yearly parental incomemeasured in 100 SEK. Allowance recipient is the share of households that receive social assistance orhousing allowances.


Table 2.10: School Capacity and Enrollment in Public Schools,2011-2014

All By neighborhood

North West East2011Number of schools 20 8 7 5Number of seats available . . . .Number of students enrolled 951 365 314 272Overcapacity (percent) . . . .Number of oversubscribed schools . . . .2012Number of schools 20 8 7 5Number of seats available 1220 534 392 294Number of students enrolled 998 382 352 264Overcapacity (percent) 18 28 10 10Number of oversubscribed schools 4 0 3 12013Number of schools 19 8 6 5Number of seats available 1050 401 387 262Number of students enrolled 913 341 353 219Overcapacity (percent) 13 15 9 16Number of oversubscribed schools 3 0 3 02014Number of schools 19 8 6 5Number of seats available 1105 447 395 263Number of students enrolled 964 375 343 246Overcapacity (percent) 13 16 13 6Number of oversubscribed schools 4 0 2 2

Notes: This table displays the number of schools, number of seats available in grade K,number of students enrolled in grade K, the implied overcapacity in number of emptyseats and in percentage terms and the number of oversubscribed schools, in total and byneighborhood for the municipal schools each year. Number of students enrolled includesall students enrolled during the autumn semester (not conditional on participating in theschool choice program). Information on school capacity is missing for 2011 and voucherschools are excluded from this table as there are no data on their capacity.

152 CHAPTER 2

Table 2.11: Summary Statistics for the PrimarySchool Starters by Rank of Voucher School, 2011-2014

Voucher school rank

Top SecondDemographicsHigh-educated (share) 0.47 0.49

(0.50) (0.50)Foreign background (share) 0.53 0.56

(0.50) (0.50)Newly arrived immigrants (share) 0.03 0.03

(0.18) (0.16)Neighborhood of residenceNorth (share) 0.88 0.92

(0.33) (0.28)West (share) 0.11 0.07

(0.32) (0.26)East (share) 0.01 0.01

(0.08) (0.10)Observations 297 272

Notes: This table presents the mean and standard deviation ofeach variable for the sample of primary school starters in Botkyrka2011-2014 listing a voucher school as their top (first column) orsecond (second column) choice. High-educated indicates the sharewith at least one parent with university level education. Foreignbackground indicates the share born abroad or with both parentsborn abroad. Newly arrived immigrant indicates the share withless than four years in Sweden. North, west and east indicates theshare living in each of these neighborhoods.

2.C. REGULARIZED LOGISTIC REGRESSION 153

2.C Using Regularized Logistic Regressionfor Variable Selection

The data contains many attributes of the schools in Botkyrka. In-cluding all of them in the model may lead to over-fitting, makingthe parameters less suitable for prediction. To avoid this, regularizedlogistic regression (LASSO) is used as a data-driven approach for vari-able selection. Regularization implies that the optimization problemis modified to include a penalty term which allows shrinkage of theestimated coefficients in order to reduce the variance and thereby im-prove out-of-sample prediction. Lasso is suitable for variable selectionas it uses L1 penalty, implying that coefficients are allowed to shrinkto zero (thus being excluded from the model). Using lasso regular-ization on a logistic regression model means that the optimizationproblem is modified. The penalized version of the log likelihood to bemaximized is:

LL =N∑i=1

[yixiβ − log(1 + exiβ)

]− λ

p∑j=1|βj | (2.10)

λ is the shrinkage penalty and as λ increases the penalty becomesmore prominent. The entire lasso regularization path is fitted for alogistic model of which school is top-ranked and second-ranked sep-arately. 10-fold cross validation is used in order to find the optimalvalue of lambda for each model. The cross-validation curves are dis-played in Figure 2.13. Table 2.12 presents the variables included inthe model, and indicates which variables are active in at least one ofthe two models.42

42The regularized logistic regression is implemented in R using the glmnetpackage.

154 CHAPTER 2

Figure 2.13: Cross-Validation Curves for the LASSO-models

(a) Top-Ranked Choices

−10 −8 −6 −4 −2

0.04

0.05

0.06

0.07

0.08

log(Lambda)

Mea

n−S

quar

ed E

rror

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

15 15 14 12 9 8 8 8 7 6 6 4 3 2 2 2 2 2 1 1

Number of non−zero (active) coefficients

(b) Second-Ranked Choices

−9 −8 −7 −6 −5 −4 −3

0.06

0.07

0.08

0.09

log(Lambda)

Mea

n−S

quar

ed E

rror

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●

●●

●

●

●

●

●

●

●●

●●

●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

15 14 14 13 9 9 9 8 8 8 8 8 8 7 5 4 4 2 2 2 1

Number of non−zero (active) coefficients

Notes: This figure shows the cross-validation curve and upper and lower standard deviation curves foreach value of the λ sequence. The vertical lines show the value of λ that minimizes the MSE (λmin)and the value of λ within one standard deviation of λmin that gives the most regularized model.

2.C. REGULARIZED LOGISTIC REGRESSION 155

Table 2.12: Variable Selection Using LASSO

Variable Coefficient 6= 0Neighborhood school YESSibling enrolled YESDistance to school, log(m) YESDistance to school (m) NOTest scores, standardized tests NOPass rate, standardized tests NOStudent teach ratio NOAverage teacher experience YESShare certified teachers NOShare of students from households receiving allowances NOAverage socioeconomic status of students NOShare with at least one high-educated parent YESAverage parental unemployment duration NOAverage parental income NOShare of students with foreign background YESShare of newly arrived immigrant-students YESShare of male students NOOffers all grades up to year 9 YESNumber of students enrolled (K-5) YES

Notes: The first columns list all variables included in regularized regression modelsof the probability of a school being listed as a top choice or second choice. Thesecond column specifies whether the shrinkage induced by L1-penalty resulted inthe coefficient being set to zero or not. The model was estimated on top and secondchoices separately. NO indicates that the variable coefficient was shrunk to zero inboth models and YES indicates that the coefficient was not set to zero in at leastone of the models.

156 CHAPTER 2

2.D Robustness of Estimated Preference Pa-rameters

Table 2.13: Preference Parameters Estimated with a Rank Ordered MixedLogit Model with Varying Levels of Clustering

250m grid 500 m grid 750 m gridMean SD Mean SD Mean SD

Distance -1.892∗∗∗ -1.892∗∗∗ -1.892∗∗∗

(0.0658) (0.0799) (0.0846)

Nbd school 2.360∗∗∗ 2.360∗∗∗ 2.360∗∗∗

(0.148) (0.148) (0.189)

Enrolled sibling 3.675∗∗∗ 3.675∗∗∗ 3.675∗∗∗

(0.124) (0.142) (0.183)

Grade 9 0.264∗∗∗ 0.0223 0.264∗∗ 0.0223 0.264∗∗ 0.0223(0.0666) (0.208) (0.0838) (0.211) (0.0966) (0.196)

N students 0.004∗∗∗ 0.000 0.004∗∗∗ 0.000 0.004∗∗∗ 0.000(0.0004) (0.0001) (0.0005) (0.0001) (0.0006) (0.0001)

Avg teacher exp. 0.0289 0.225∗∗∗ 0.0289 0.225∗∗∗ 0.0289 0.225∗∗∗

(0.0187) (0.0172) (0.0262) (0.0200) (0.0360) (0.0244)

Foreign backgr. -2.006∗∗∗ 2.911∗∗∗ -2.006∗∗∗ 2.911∗∗∗ -2.006∗∗∗ 2.911∗∗∗

(0.351) (0.315) (0.461) (0.376) (0.487) (0.411)

Newly arrived 0.315 1.614 0.315 1.614 0.315 1.614(0.542) (1.809) (0.621) (1.965) (0.817) (2.690)

High-educated 1.801∗∗∗ -2.865∗∗∗ 1.801∗∗ -2.865∗∗∗ 1.801∗ -2.865∗∗∗

(0.410) (0.609) (0.598) (0.700) (0.719) (0.804)Observations 187953 187953 187953

Notes: This table presents the estimated mean and standard deviation of the preference parame-ters. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. Standard errors in parentheses, clustered at the 250times 250, 500 times 500 and 750 times 750 meters respectively. The sample includes all primaryschool starters in Botkyrka 2011-2014 categorized as (partially) truth-telling (see Section 2.4.1for details on this categorization). For variable definitions, see Table 2.6.

2.D. ROBUSTNESS OF ESTIMATED PREFERENCEPARAMETERS 157

Table 2.14: Preference Parameters Estimated with aRank Ordered Mixed Logit Model, Inclduing and Ex-cluding Voucher Schools

All Excl. voucher schoolsMean SD Mean SD

Distance -1.892∗∗∗ -2.037∗∗∗

(0.0658) (0.0743)

Nbd school 2.360∗∗∗ 2.252∗∗∗

(0.148) (0.160)

Enrolled sibling 3.675∗∗∗ 3.669∗∗∗

(0.124) (0.127)

Grade 9 0.264∗∗∗ 0.0223 0.233∗∗ -0.144(0.0666) (0.208) (0.0744) (0.205)

N students 0.004∗∗∗ 0.0001 0.005∗∗∗ -0.0002(0.0004) (0.0001) (0.0005) (0.0003)

Avg teacher exp. 0.0289 0.225∗∗∗ 0.0709∗∗∗ 0.224∗∗∗

(0.0187) (0.0172) (0.0203) (0.0200)

Foreign backgr. -2.006∗∗∗ 2.911∗∗∗ -2.035∗∗∗ 3.142∗∗∗

(0.351) (0.315) (0.416) (0.320)

Newly arrived 0.315 1.614 2.304∗∗∗ -0.0465(0.542) (1.809) (0.529) (0.0919)

High-educated 1.801∗∗∗ -2.865∗∗∗ 2.371∗∗∗ 3.240∗∗∗

(0.410) (0.609) (0.433) (0.628)Observations 187953 175431

Notes: This table presents the estimated mean and standard deviationof the preference parameters. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001.Standard errors in parentheses (clustered at the 250 times 250 me-ters level). The full sample includes all primary school starters inBotkyrka 2011-2014 categorized as (partially) truth-telling (see Sec-tion 2.4.1 for details on this categorization). Non-movers is the subsetof households which did not change their residential location for thelast years, Movers is the subset of households that did move duringthis period and Excl. top choices estimates the model on second andthird choices only. For variable definitions, see Table 2.6.

158 CHAPTER 2

Tab

le2.

15:Pr

eference

ParametersEs

timated

with

aRa

nkOrdered

Mixed

LogitModel,Ro

-bustness

Tests

All

Non

-movers

Movers

Excl.topchoices

Mean

SDMean

SDMean

SDMean

SDDistance

-1.892

∗∗

∗-1.936

∗∗

∗-1.832

∗∗

∗-1.712

∗∗

∗

(0.0658)

(0.0788)

(0.0775)

(0.0763)

Nbdscho

ol2.360∗

∗∗

2.350∗

∗∗

2.343∗

∗∗

2.472∗

∗∗

(0.148)

(0.204)

(0.234)

(0.145)

Enrolledsibling

3.675∗

∗∗

3.631∗

∗∗

3.758∗

∗∗

2.095∗

∗∗

(0.124)

(0.154)

(0.222)

(0.210)

Grade

90.264∗

∗∗

0.0223

0.239∗

∗-0.0315

0.305∗

∗∗

0.219

0.319∗

∗∗

-0.0188

(0.0666)

(0.208)

(0.0848)

(0.0714)

(0.0789)

(1.120)

(0.0916)

(0.0669)

Nstud

ents

0.004∗

∗∗

0.000

0.004∗

∗∗

0.000

0.004∗

∗∗

-0.000

0.004∗

∗∗

-0.000

(0.0004)

(0.0001)

(0.000

5)(0.0004)

(0.0004)

(0.0002)

(0.0004)

(0.0001)

Avg

teacherexp.

0.0289

0.225∗

∗∗

0.0536

∗∗

0.205∗

∗∗

-0.00370

-0.243

∗∗

∗0.0648

∗∗

∗0.221∗

∗∗

(0.0187)

(0.0172)

(0.0203)

(0.0184)

(0.0216)

(0.0221)

(0.0187)

(0.0178)

Foreign

backgr.

-2.006

∗∗

∗2.911∗

∗∗

-2.028

∗∗

∗2.936∗

∗∗

-1.924

∗∗

∗2.834∗

∗∗

-2.184

∗∗

∗-2.174

∗∗

∗

(0.351)

(0.315)

(0.481)

(0.409)

(0.396)

(0.625)

(0.355)

(0.356)

New

lyarrived

0.315

1.614

-0.0940

-1.743

0.774

1.309

1.042

-1.951

(0.542)

(1.809)

(0.631)

(1.901)

(1.062)

(12.98)

(0.810)

(1.864)

High-educated

1.801∗

∗∗

-2.865

∗∗

∗1.820∗

∗∗

-3.063

∗∗

∗1.765∗

∗∗

0.992

2.045∗

∗∗

-1.579

(0.410)

(0.609)

(0.506)

(0.777)

(0.491)

(3.504)

(0.442)

(0.912)

Observation

s187953

108091

79862

119350

Not


theestimated

meanan

dstan

dard

deviationof

thepreference

parameters.

∗p<

0.05,

∗∗p<

0.01

,∗

∗∗p<

0.00

1.Stan

dard

errors

inpa

rentheses(clustered

atthe250times

250meterslevel).The

fullsampleinclud

esall

prim

aryscho

olstarters

inBotky

rka2011-2014categorizedas

(partially)truth-telling(see

Section2.4.1fordetailson

this

categorization

).Non

-moversis

thesubset

ofho

useholds

which

didno

tchan

getheirresidentiallocation

forthelast

years,

Moversis

thesubset

ofho

useholds

that

didmovedu

ring

this

periodan

dExcl.topchoicesestimates

themod

elon

second

andthirdchoiceson

ly.For

variab

ledefinition

s,seeTab

le2.6.

2.E. IMPLEMENTING DA WITH RESERVED SEATS 159

2.E Implementing DA with Reserved Seats

To implement affirmative action, stratified assignment withproximity as a tie-breaker (within and across student types)is used. At each school, st seats are reserved for studentsof type t. The DA algorithm is modified to incorporate this as follows:

Step 1 a) Each student applies to his/her top choice. Each schoolconsiders applicants of each type separately. Students of type t

are tentatively accepted one at a time according to their priorityuntil there are no applicants left or the school runs out of reservedseats for type t, in which case all remaining applicants of type t arerejected.b) If there is capacity left at the school (which happens when thereare less applicants than reserved seats for at least one student typet), all students rejected in this round are considered together andthe school tentatively accepts one student at a time according totheir priority.

Step k a) Each student rejected in the previous round applies tohis/her next preferred school. Each school considers all tentativelyaccepted and new applicants together for each type separately andaccepts one student at a time according to their priority until thereare no applicants left or the school runs out of reserved seats for typet, in which case all remaining applications of type t are rejected.b) If there is capacity left at the school (which happens when thereare less applicants than reserved seats for at least one student typet), all students rejected in this round are considered together andthe school tentatively accepts one student at a time according totheir priority.

The algorithm terminates when no student is rejected, at whichpoint all students are placed at their final tentative assignment.

160 CHAPTER 2

2.F Measures of School Segregation

This section presents the measures of school segregation used in thisstudy.

2.F.1 Share of Variance in Student SES Explained bySchools

The preferred measure of school segregation in this study is the ratioof the between-school and the total variance in student SES. First,a socioeconomic index is constructed by estimating the relationshipbetween GPA and an ethnic/socioeconomic background using dataon all ninth-graders in Sweden during 2011 to 2014 according to thefollowing specification:

GPAit = α+ βXit + εit, (2.11)

where X is a vector consisting of the student’s gender, the log-arithm of the father’s and mother’s income level, dummies for themother’s and father’s educational level (7 levels), dummies for theregion of birth of the student, the father and the mother (6 regions)and school fixed effects.43 The coefficients in the vector β describesthe relationship between the variables in X and grade nine GPA andare then used to create predicted GPA for the sample of primaryschool starters:

ˆGPAit = α+ βXit. (2.12)

The predicted GPA can be interpreted as a measure of a student’ssocioeconomic background, where predetermined characteristics thatare important for the student’s expected performance in school are

43Estimating the same regression excluding school-fixed effects gives very simi-lar results, but to use coefficients from the regression including school-fixed effectsis preferred in the following steps.

2.F. MEASURES OF SCHOOL SEGREGATION 161

given more weight than other factors. A student who is predicted ahigh GPA-score has a strong background in the sense that he/shecomes from an advantaged household and the opposite is true for awho is predicted a low GPA-score.

One way to measure school segregation is to assess the level ofsorting across schools by predicted GPA. To capture this, the ratioof the between-school variance and the total variance in predictedGPA is studied. Intuitively, this indicates how much of the variancein predicted GPA that can be explained by which schools the studentsattend. To get at this, the following regression equation is estimated:

ˆGPAit = γ + Zit + εit, (2.13)

where Z is a set of dummies (one for each school) and record theR2. To see that the R2 from this regression gives the ratio of thebetween-school variance and total variance in predicted GPA, the R2

can be rewritten as:

R2 =∑i(yi − y)2∑i(yi − y)2 (2.14)

=∑s

∑i∈s(yi − y)2∑i(yi − y)2

=∑s ns(ys − y)2∑i(yi − y)2

=1N

∑s ns(ys − y)2

1N

∑i(yi − y)2

= σ2s

σ2 .

2.F.2 Duncan Dissimilarity Index

To complement the analysis, the sorting across schools in terms offoreign background and parental education is studied. One way to

162 CHAPTER 2

show this would be to compare the share of students with e.g. foreignbackground across schools. This is intuitive, but has the drawback ofbeing hard to compare to other settings. To use a segregation indexis therefore preferable. Segregation indexes can be divided into mea-sures of evenness or exposure.44 Exposure measures are sensitive tothe share of minority students in the population, which evenness mea-sures are generally not. As pointed out by Allen and Vignoles (2007),educational policy cannot affect the share of minority students mak-ing measures of evenness more relevant in the context of evaluatingthe design of a school choice program. The Duncan Dissimilarity In-dex (DDI) is used, a commonly used evenness measure that rangesfrom zero (no segregation) to one (perfect segregation).45 It offers astraightforward interpretation as the share of one group that wouldhave to switch to another school in order to produce a distribution ineach school that matches the distribution of the entire population. Itis defined as:

DDI = 12

n∑i=1|aiA− biB|, (2.15)

where ai is the number of individuals in group A in school i andbi is the number of individuals in group B in school i, and A and Bmeasure the total number of individuals in these two groups.

44Massey and Denton (1988) provide a discussion about different segregationindexes and their properties.

45This measure was introduced by Duncan and Duncan (1955).

2.G. ADDITIONAL SIMULATION RESULTS 163

2.G Additional Simulation Results

Figure 2.14: Distribution of School Segregation by Parental Education andForeign Background in the Short Run under Alternative Priority Structures


0

5

10

15

20

Den

sity

.15 .2 .25 .3 .35Duncan Dissimilarity Index



0

5

10

15

20

Den

sity

.4 .45 .5 .55 .6 .65Duncan Dissimilarity Index


Notes: This figure displays the distribution of the Duncan Dissimilarity Index for parental education(left panel) and foreign background (right panel) under each alternative priority structure andassignment using catchment zones. See Appendix 2.F.2 for details about how this segregation measureis defined. Details on the different priority structures can be found in Section 2.5.2. The distributionis computed using 250 simulations of independently drawn samples. The lines are fitted using Stata’skdensity command, using the Epanechnikov kernel and optimal bandwidth selection.

164 CHAPTER 2

Figure 2.16: Distribution of Distance to Assigned School in the Short Runby Assignment to Top Choice

(a) Assigned Top Choice

0

.005

.01

.015

Den

sity

500 600 700 800 900 1000Meters


(b) Not Assigned Top Choice

0

.005

.01

.015

Den

sity

1500 2000 2500 3000 3500 4000Meters


Notes: This figure displays the distribution of distance (in meters) to assigned school for those assignedto their top choice (left panel) and those not assigned to their top choice (right panel) under eachalternative priority structure. Details on the different priority structures can be found in Section 2.5.2.The distribution is computed using 250 simulations of independently drawn samples. The lines are fittedusing Stata’s kdensity command, using the Epanechnikov kernel and optimal bandwidth selection.

Figure 2.15: Distribution of Utility of Assigned School in the Short Rununder Alternative Priority Structures

0

.1

.2

.3

.4

.5

Den

sity

-12 -10 -8 -6 -4 -2Utility


Notes: This figure displays the distribution of utility of the assigned school under each alternativepriority structure and assignment using catchment zones. Utility is calculated according to equation(2.1). Details on the different priority structures can be found in Section 2.5.2. The distribution iscomputed using 250 simulations of independently drawn samples. The lines are fitted using Stata’skdensity command, using the Epanechnikov kernel and optimal bandwidth selection.


Table 2.16: Residential Location of Students, bySchool Neighborhood and Priority Structure

Proximity Lottery Reserved seatsEastern schoolsEast 1.00 0.96 0.96

(0.00) (0.01) (0.02)North 0.00 0.04 0.04

(0.00) (0.01) (0.01)West 0.00 0.00 0.00

(0.00) (0.00) (0.00)Western schoolsEast 0.08 0.08 0.08

(0.03) (0.03) (0.03)North 0.00 0.00 0.00

(0.01) (0.01) (0.01)West 0.92 0.91 0.92

(0.03) (0.03) (0.03)Northern schoolsEast 0.03 0.04 0.05

(0.02) (0.02) (0.02)North 0.95 0.93 0.93

(0.03) (0.03) (0.03)West 0.02 0.03 0.02

(0.01) (0.02) (0.01)Observations 250 250 250

Notes: This table presents the share of students in the simulationsample from each neighborhood, by school neighborhood andpriority structure. The shares are calculated as the average using250 simulations of independently drawn samples.

166 CHAPTER 2

Figure 2.17: Distribution of Utility of Assigned School in the Short Rununder Alternative Priority Structures, by Household Type

(a) Native, High-Educated

0

.1

.2

.3

.4

.5

Den

sity

-12 -10 -8 -6 -4 -2Utility


(b) Native, Low-Educated

0

.1

.2

.3

.4

.5

Den

sity

-12 -10 -8 -6 -4 -2Utility


(c) Foreign, High-Educated

0

.1

.2

.3

.4

.5

Den

sity

-10 -8 -6 -4 -2Utility


(d) Foreign, Low-Educated

0

.1

.2

.3

.4

.5

Den

sity

-12 -10 -8 -6 -4 -2Utility


Notes: This figure displays the distribution of utility of the assigned school under each alternativepriority structure and assignment using catchment zones, by household type. Utility is calculated ac-cording to equation (2.1). Details on the different priority structures can be found in Section 2.5.2.High-educated is defined as having at least one parent with university level education and foreignbackground is defined as being born abroad or having both parents born abroad. The distribution iscomputed using 250 simulations of independently drawn samples. The lines are fitted using Stata’skdensity command, using the Epanechnikov kernel and optimal bandwidth selection.


Figure 2.18: Distribution of School Segregation by SES in the Short Rununder Alternative Priority Structures, Assuming Random Residential Loca-tion

0

20

40

60

80

Den

sity

0 .02 .04 .06 .08SES segregation


Notes: This figure displays the distribution of the share of the variance in student SES that isexplained by the assigned school (see Appendix 2.F.1 for details about how this segregation measureis defined) under each alternative priority structure. Details on the different priority structures can befound in Section 2.5.2. The distribution is computed using 250 simulations of independently drawnsamples, where the residential location of each sampled household was replaced by a randomly drawnlocation in the municipality (using weights based on the observed location in order to replicate theobserved density of the population). The lines are fitted using Stata’s kdensity command, using theEpanechnikov kernel and optimal bandwidth selection.

168 CHAPTER 2

Figure 2.19: Distribution of School Segregation by Parental Education andForeign Background in the Short Run under Alternative Priority Structures,Assuming Random Residential Location


0

5

10

15

20

25

30

35

40

45

Den

sity




0

5

10

15

20

25

30

35

40

45

Den

sity



Notes: This figure displays the distribution of the Duncan Dissimilarity Index for parental education(left panel) and foreign background (right panel) under each alternative priority structure. See Ap-pendix 2.F.2 for details about how this segregation measure is defined. Details on the different prioritystructures can be found in Section 2.5.2. The distribution is computed using 250 simulations of inde-pendently drawn samples, where the residential location of each sampled household was replaced by arandomly drawn location in the municipality (using weights based on the observed location in order toreplicate the observed density of the population). The lines are fitted using Stata’s kdensity command,using the Epanechnikov kernel and optimal bandwidth selection.


Figure 2.20: Distribution of School Segregation by SES in the Short Rununder Alternative Priority Structures, Assuming Homogeneous Preferences

0

5

10

15

20

25

Den

sity

.1 .15 .2 .25SES segregation


Notes: This figure displays the distribution of the share of the variance in student SES that isexplained by the assigned school (see Appendix 2.F.1 for details about how this segregation measureis defined) under each alternative priority structure. Details on the different priority structures can befound in Section 2.5.2. The distribution is computed using 250 simulations of independently drawnsamples, where the same preference parameters are used to construct the application lists for allsampled households. The lines are fitted using Stata’s kdensity command, using the Epanechnikovkernel and optimal bandwidth selection.

170 CHAPTER 2

Figure 2.21: Distribution of School Segregation by Parental Education andForeign Background in the Short Run under Alternative Priority Structures,Assuming Homogeneous Preferences


0

5

10

15

20

Den

sity




0

5

10

15

20

Den

sity



Notes: This figure displays the distribution of the Duncan Dissimilarity Index for parental education(left panel) and foreign background (right panel) under each alternative priority structure. See Ap-pendix 2.F.2 for details about how this segregation measure is defined. Details on the different prioritystructures can be found in Section 2.5.2. The distribution is computed using 250 simulations of indepen-dently drawn samples, where the same preference parameters are used to construct the application listsfor all sampled households. The lines are fitted using Stata’s kdensity command, using the Epanechnikovkernel and optimal bandwidth selection.

Figure 2.22: Average School Segregation by SES in the Long Run UnderAlternative Priority Structures

0

.05

.1

.15

.2

.25

.3

.35

Aver

age

0 5 10 15 20Year


Notes: This figure displays the average share of the variance in student SES that is explained by theassigned school (see Appendix 2.F.1 for details about how this segregation measure is defined) undereach alternative priority structure. Details on the different priority structures can be found in Section2.5.2. The value is computed as the average of 250 simulations of independently drawn samples.


Figure 2.23: Distribution of School Segregation by Parental Education andForeign Background in the Long Run under Alternative Priority Structures


.2

.3

.4

.5

.6

Aver

age

0 5 10 15 20Year



.2

.3

.4

.5

.6

Aver

age

0 5 10 15 20Year


Notes: This figure displays the average of the Duncan Dissimilarity Index for parental education (leftpanel) and foreign background (right panel) under each alternative priority structure. See Appendix2.F.2 for details about how this segregation measure is defined. Details on the different priority struc-tures can be found in Section 2.5.2. The value is computed as the average of 250 simulations of inde-pendently drawn samples.

172 CHAPTER 2

2.H Test of Randomization

Table 2.17: Balance

TreatedForeign (share) 0.0372

(0.0277)

Highly educated (share) 0.0347(0.0278)

Constant 0.478∗∗∗(0.0250)

Observations 1306Notes: This table presents the results from anOLS regression of the treatment indicator onforeign background (born abroad or having bothparents born abroad) and high-educated (at leastone parent with university level education) ofthe sample of primary school starters in 2016.Standard errors are presented in parentheses.Significance levels are indicated by * p<0.1, **p>0.05, *** p<0.01.

2.I. THE LETTER 173

2.I The Letter

2 (7)

Further Information Regarding the School Choice

You have recently, or will soon, receive a letter from Botkyrka municipality informing

you that it is time to choose a school for your child. As we are studying school choice,

we would like to give you some additional information that we hope can be of use.

In the attached table you can see how the schools in Botkyrka municipality have

performed on the standardised tests in grade three in recent years. The first column

shows the share of the school’s students that passed the tests. In the second column we

have calculated how the students of the school have performed on the standardised

tests compared to schools that are similar in terms of the students' educational and

immigrant background.

If you have any questions feel free to contact us.

Dany Kessel Elisabet Olme

[email protected] [email protected]

0707173803 0735442220

Notes: Contact details blacked out.

174 CHAPTER 2

3 (6)

Skola School

Andel godkända (%) Pass rate (%)

Över/underprestation Over/underperformance

Banslättsskolan 96,5 +3,5 Björkhaga skola 88 -0,5 Borgskolan 73,5 -0,5 Botkyrka Friskola 85,5 +0,5 Broängsskolan 89,5 -2,5 Brunnaskolan 79 -5,5 Edessaskolan 80,5 -3,5 Eklidsskolan 96,5 +3,5 Fittjaskolan Botkyrka Norra 67 -6,5 Freinetskolan Kastanjen 80 -6,5 Grindtorpsskolan 64,5 -10 Gryningeskolan - - Hammerstaskolan 78,5 -5,5 Karsby International School 72,5 -5,5 Kassmyraskolan 93 +3,5 Kvarnhagsskolan 81,5 -0,5 Malmsjö skola 91,5 +1 Parkhemsskolan 92 -0,5 Rikstens skola 91,5 +0,5 Skogsbacksskolan (f.d. Storvretskolan)

62,5 -15

Sverigefinska skolan, Botkyrka 85 0 Tallidsskolan 85 +6,5 Trädgårdsstadsskolan 90 -2,5 Tullingebergsskolan 89 0 Tunaskolan 85 -2,5 Såhär läser du tabellen: Om det står 80 i kolumnen ”Andel godkända” och +2 i kolumnen ”Över/underprestation” betyder det att 80 procent av skolans elever klarade de nationella proven och att det var 2 procentenheter mer än på skolor med liknande elever. How to read the table: If it says 80 in the column “Pass rate” and +2 in the column “Over/underperformance”, it means that 80 percent of the students at that school passed the standardized tests and that this pass rate is 2 percentage points higher compared to schools with similar students. Källa/Source: Statistiska Centralbyrån (SCB)/Statistics Sweden.

Chapter 3

Should I Stay or Must I Go?Temporary Refugee Protection and Labor-MarketOutcomes∗

∗This chapter was co-authored by Birthe Larsen and Matilda Kilström. Wethank Niklas Blomqvist, Jonas Cederlöf, Matz Dahlberg, Peter Fredriksson, JohnHassler, David Jinkins, Per Krusell, Jaakko Meriläinen, Elin Molin, Arash Nekoei,Peter Nilsson, Gritt Ølykke, Per Pettersson-Lidbom, Miikka Rokkanen, AnnaSeim, David Seim, Jósef Sigurdsson, Björn Tyrefors Hinnerich, Jonas Vlachos,Eskil Wadensjö, and seminar participants at the Migration and Demographicsconference in Nürnberg, the Institute for Housing and Urban Research (UppsalaUniversity), the National Institute of Economic Research, Research Institute ofIndustrial Economics, Swedish Public Employment Service, and Stockholm Uni-versity for valuable comments and helpful discussions. Financial support fromStiftelsen Söderströms Donationsfond is gratefully acknowledged by Elisabet Olmeand Matilda Kilström. This research benefited from financial support from Han-delsbanken’s Research Foundations. All errors are our own.

176 CHAPTER 3

3.1 Introduction

Recent developments around the world have led to a large inflow ofasylum seekers to Europe. In response to the increased numbers ofasylum seekers, many European countries have implemented stricterimmigration policies. The motivation has been to reduce immigra-tion and/or improve the integration of immigrants granted residency.One such policy is the shift from permanent to temporary residencepermits for refugees.1 While several countries have, or are about to,implement such reforms, empirical evidence is lacking on their effectson refugees’ integration in society in general and in labor markets inparticular.

A priori, it is possible to think that a shift to temporary permitscould have both positive and negative effects on integration in soci-ety and in the labor market. The public debate has been centeredaround the relative strengths of these effects. On the one hand, theexpected return to investment in country-specific human capital fallsif the probability of receiving permanent residency falls. There canalso be a cost in the form of increased stress from a lower probabil-ity of being granted permanent residency. On the other hand, actionsthat lead to labor-market attachment during the time with tempo-rary residency are incentivized when they increase the probability forpermanent residency. This could strengthen the incentives for labor-market investments in the host country and improve integration.

The net effect of a shift from permanent to temporary residencepermits for refugees is therefore an empirical question in much needof attention. Specifically, this study addresses what the effects areof changes in the probability of being granted permanent residency.

1For example, in July 2016, Sweden introduced a temporary law shifting frompermanent to temporary residence permits and limiting possibilities of family re-unification. Among several other changes to the refugee policy, in December 2014Australia reintroduced temporary protection visas - which cannot be promoted topermanent status - for those who arrive without a valid visa.


Furthermore, while immigrants’ entrance into the labor market isrelatively well studied, less attention has been given to the specificchallenges of those given refuge in a new country. In fact, little isknown about the integration process of refugees and their labor mar-ket prospects. In a recent study, Fasani et al. (2018) show that refugeesperform worse in the labor market than other immigrants across Eu-rope, and Dustmann et al. (2017a) highlight the indecisiveness aboutthe duration and permanence of the stay in the host country as oneof the primary reasons for the poor labor market integration. Thiscalls for more research focusing on refugee immigration to Europe,with a focus on refugees’ labor-market outcomes in connection withthe expected duration in the host country.

In this study, the effects of a Danish reform, implemented in 2002as part of a reform package, are analyzed. The reform changed theeligibility requirements for permanent residency, thereby lowering theex-ante probability of being granted a permanent residence permit.This was done by increasing the length of the time period a refugeewould have had to have been a legal resident (on a temporary resi-dence permit) in Denmark before being eligible to apply for perma-nent residency. During the time with temporary status, a residencepermit could be withdrawn if the grounds for protection were nolonger valid, and if the individual did not have the right to stay onother grounds, such as having a solid labor-market attachment. Thefact that the spell under a temporary residency was implementedretroactively makes it possible to distinguish the effect of this partof the reform from other aspects of the reform package. The changeapplied to individuals who lodged their first asylum application onor after February 28, 2002. This meant that refugees who applied forasylum from February 28, 2002, onwards faced a longer time periodwith temporary status, during which they risked losing the groundsfor protection, before they could apply for permanent residency. Allelse equal, the ex-ante probability of receiving permanent residency

178 CHAPTER 3

in Denmark on the grounds of asylum was thus higher prior to thereform.2

To understand the mechanisms at work, a theoretical search andmatching model is set up, with the objective of deriving predictionsthat can be compared to the empirical findings. The model focuses onlabor-market outcomes and features heterogeneity in terms of produc-tivity and a human capital investment decision. The model is used tostudy education and labor-market outcomes under different assump-tions about a policy change similar to the reform. Empirically, a re-gression discontinuity in time framework is used to study the impactof the reform. Register data allows individuals granted asylum, anda large set of their outcomes, to be observed over time. Thus, thisstudy is able to empirically dig more deeply into the mechanisms atwork and to consider the impact of the reform on different subgroupsof refugees. The interest is in the behavioral responses to this reformcomponent, and the focus is on outcomes that are relevant for integra-tion and/or the assessment of grounds for prolonged residency thatthe individual could affect herself. The main outcomes are thereforein terms of educational investments and status in the labor market.Labor-market attachment can, in itself, be viewed as a measure ofintegration, whereas education can be considered as an investment inintegration. The full sample is studied, as well as sample splits basedon gender and skill level.

While the reform analyzed in this study is in many ways ideal for

2However, this does not mean that the reform necessarily changed whetheran individual got to stay in Denmark or not. In fact, around 90 percent of theindividuals (in both the control and the treatment group) are still in Denmarktwelve years after their first arrival. The individual could (1) have had asylumreasons throughout the time period with a temporary permit, or (2) established alabor-market attachment which could be used as grounds for prolonged temporaryresidency. Although a refugee had no control over the development in her homecountry, or the Danish authorities’ assessment of whether the grounds for asylumwere still valid, she could affect her attachment to the labor market and thus affectthe probability of staying in Denmark.


determining the effect of prolonged temporary status, the setting alsooffers some challenges. The sample size is limited because outcomesare only studied for individuals that are actually given asylum and thetime interval needs to be restricted to four months before and afterthe reform to avoid confounding policy changes. This means that theresults should be interpreted carefully and that the sample splits arerestricted to those mentioned above. Descriptive analysis is used alongwith the estimated results to substantiate the understanding of theimpact of the reform. Furthermore, the theoretical model is used asa benchmark to which the empirical findings can be compared.

The results suggest that lowering the ex-ante probability of re-ceiving permanent residency increased enrollment in education. En-rollment is measured as the share that is ever enrolled in education,excluding Danish courses, during the first twelve years of residencyin Denmark. The enrollment rates are also higher for the treatmentgroup throughout the sample period, as shown by plotting enroll-ment rates over time. The increase in enrollment is mainly driven byfemales and low-skilled individuals, defined as individuals who lacka university education.3 The positive effect on enrollment in educa-tion is interpreted as an increased investment in human capital andintegration. In terms of labor-market outcomes, focus is on the shareof individuals that are ever employed (or self-employed) during thefirst twelve years in Denmark, and on their earnings measured threeand seven years after arrival. No significant effects are estimated onlabor-market outcomes, but the coefficients are negative.4 The sameholds true when looking at earnings conditional on employment. Theempirical findings on education are in line with the implications from

3The estimated effect for the full sample is an increase of 17 percentage pointsat the cutoff point. The effect for females is an increase of around 21–27 percentagepoints at the cutoff.

4The negative impact on earnings is significant for certain subgroups. Becausethe estimated effects are not robust to inclusion/exclusion of control variables andhave large standard errors, these results should be interpreted very cautiously.

180 CHAPTER 3

the theoretical model where low-skilled individuals are, ex ante, morenegatively affected by the reform. In the model, this is true regardlessof whether they are employed or unemployed.

There are other outcomes of interest in this context, related to dif-ferent parts of the literature discussed in the subsequent paragraph.In this study, the impact of the reform on criminal activity, health andfertility is also analyzed. Engaging in criminal activity could be seenas an alternative to entering the labor market, but criminal activitiesalso make it harder to obtain permanent residency post-reform. In-creased uncertainty from a lower probability of permanent residencymay also have a direct effect on individual health which, in turn, couldaffect future labor-market outcomes. In terms of fertility, the reformand the implied increase in uncertainty may have affected the will-ingness to bring children into the world. Finally, the asylum holders’duration in Denmark is of interest, for two reasons. First, the reformcould have affected the willingness and ability to stay in Denmarkas the prerequisites to stay changed, which in itself is an interestingoutcome. Second, if the fraction staying in Denmark changed, the re-sults on other outcomes may be driven by this selection rather thanby behavioral responses among those staying in Denmark. This studyfinds a negative effect on conviction rates for property crimes duringthe first twelve years in Denmark of around ten percentage points.This decrease is concentrated among males. There are no significantdifferences, for the full sample, between the two groups in terms ofhealth, fertility, or the share that is still in Denmark twelve years aftertheir first asylum application. The latter finding suggests that any ef-fects picked up are unlikely to be due to an indirect effect - operatingthrough selection - that would occur if some group was more likelythan another to stay in Denmark.

Related literature There is, to our knowledge, no other studythat specifically analyzes the long-run effects on refugees of a pro-


longed temporary status. Changes to immigration policy are particu-larly difficult to evaluate due to difficulties in finding a valid compar-ison group. Previously, some studies have compared different types ofimmigrants to assess the importance of, for example, the time horizonin the host country. As different types of immigrants may differ in im-portant ways, it would be preferable to study the effects within onespecific group of immigrants. At the same time, even when looking atone type of immigrants (for example refugees) there may be substan-tial heterogeneity. It is well known that the characteristics of refugeesfrom a given country can change over time. All of this implies thatestimating the effects of changes to the probability of being grantedpermanent residency is challenging.

There are related studies that consider the difference between tem-porary and permanent migration spells in other contexts. For exam-ple, Chen et al. (2016) study the selection into temporary or per-manent migration. Temporary, short-term, migration is typically aresponse to fluctuations in the local labor market, while long-termmigration is more stable. They show that long-term migrants aremore strongly positively selected and relate this to higher returnsto matching. Adda et al. (2016) estimate a dynamic model of re-turn migration and human capital accumulation. They simulate theeffects of uncertainty about the permanence of an individual’s stayin the host country and find reduced investments in human capitaland decreased life time earnings because of a shortened pay-off pe-riod. These papers do not explicitly consider refugees, a group thatis fundamentally different from other types of migrants in that theyare forced to leave their home country. This implies that while othermigrants have the option to return to their home country, refugeesmay not. Cortes (2004) recognizes the importance of this distinc-tion and focuses on the heterogeneity between refugees and economicimmigrants in terms of their time horizons. Assuming that refugeescannot return to their country of origin, and thus face a longer time

182 CHAPTER 3

horizon, they have stronger incentives to invest in country-specifichuman capital. Her study is related to, and theoretically builds on,Duleep and Regets (1999) and their model of human capital accumu-lation. Furthermore, Orrenius and Zavodny (2015) study the effectson labor-market outcomes of granting specific groups of immigrantsa temporary protected status (TPS) in the United States, and showthat, in general, it appears that even having a temporary permit -compared to an illegal status - improves the labor-market opportu-nities for immigrants. In Cortes (2004) and Orrenius and Zavodny(2015), a distinction is made between immigrants with different timehorizons (refugees vs. economic immigrants) and between immigrantswith a different legal status. One benefit of the setup in this studyis that the importance of the time horizon and status in the hostcountry can be analyzed within one group of immigrants, refugees.Arguably, refugees are likely to be different in many ways comparedto, for example, economic immigrants, and, since they constitute amore marginalized group in relation to the labor market, it is partic-ularly important to understand the effects of changing the conditionsthey face.

Several papers study immigration and crime and how policy mat-ters in this context. In a recent study, Pinotti (2017) uses a regressiondiscontinuity design to show that immigrant legalization reduces thecrime rates among immigrants in Italy. The proposed mechanism isthat legalization increases the opportunity cost of crime by improvingaccess to the regular labor market. Mastrobuoni and Pinotti (2015)find a reduction in crime following the European Union enlargement.Baker (2015) also finds a negative effect on crime of legalization of un-documented immigrants in the United States. Lozano and Sørensen(2011) study the effect of legalization on earnings among Mexicanimmigrants in the United States and find an increase in occupationalwages. They interpret their findings as support of immigrants find-ing better jobs following legalization. Cascio and Lewis (2017) also


study the effect of legal status in the US context, exploiting the Im-migration Reform and Control Act of 1986, and find an increase inEITC transfers but no effects on food stamp transfers. Fasani (2018)finds small and non-persistent reductions in crime following a wave ofamnesty programs in Italy. Furthermore, Fasani (2015) highlights theimportance of policy design in shaping effects of legalization on crime.Considering a different type of outcome, Dustmann et al. (2017b)study the consumption effects of legalization. They show that undoc-umented immigrants consume less than documented immigrants, andargue that this is because of lower income. More closely related tothe outcomes analyzed in this study, Devillanova et al. (2017) ana-lyze employment effects of legalization following an amnesty programin Italy and find positive effects of prospective legal status on em-ployment probability. Legalization policies are clearly important, butthey are conceptually different from policies involving permanent ver-sus temporary residence permits. It is not necessarily the case thatfindings from the legalization literature translate into other types ofpolicy changes, such as the one analyzed in this study.

Another closely related study in terms of the type of policy an-alyzed is Blomqvist et al. (2018). In the current draft, they studyshort-run effects, over a one-year horizon, of restricting the access topermanent residency in Sweden and find no conclusive evidence onparticipation in language training programs. Finally, Mansouri et al.(2010) provide a comparative study of temporary permit regimes inDenmark, Germany and Australia. Through interviews with NGOs,they conclude that introducing temporary residence permits, or pro-longing the temporary status, increased the uncertainty for refugeesand suggest that integration has been made more difficult as a re-sult. A key advantage of this study is that the response to prolongedtemporary status can be quantified and that the mechanisms throughwhich refugees were affected can be studied.

There are several relevant papers using Danish data to study im-

184 CHAPTER 3

migrants’ outcomes. Clausen et al. (2009) analyze the effects on labor-market integration for immigrants from Danish active labor-marketprograms (ALMPs). They find mixed effects depending on the typeof program, but in general positive effects from language trainingand participation in wage subsidy programs.5 Other aspects of the2002 reform package in Denmark have also been studied. There wereseveral aspects to the general reform package; notably limiting theaccess to the welfare state and to family reunification. Huynh et al.(2007) study the employment effects of limiting access to the welfarestate, finding positive employment effects from reduced benefits. Theauthors exploit the discontinuity that arises from the fact that onlythose granted asylum after July 1, 2002 were subject to the new ben-efit rules. Similarly, Rosholm and Vejlin (2010) analyze the effects oflowering the benefits on both job finding and job separation rates.Rather than using a regression discontinuity approach, they imple-ment a mixed proportional hazard model. In line with the evidencefrom Huynh et al. (2007), the authors find small positive effects onthe job finding rate. Finally, Andersen et al. (2019) also exploits thediscontinuity in benefit levels and confirm the finding that lower ben-efits led to increased earnings and employment. However, they alsofind a strong negative effect on female labor force participation and adecrease in disposable income for most households. In addition, theyalso study other outcomes and show that the lower benefit level ledto a sharp increase in property crime and had a negative impact onthe educational outcomes of children. In this study, another part ofthe reform package is analyzed to shed some light on the effects of alower ex-ante probability of receiving a permanent residence status.

The remainder of this chapter is organized as follows. Section 3.2describes the Danish institutional setting and the reform that is stud-ied. Section 3.3 presents the theoretical framework. Section 3.4 de-

5See Sarvimäki and Hämäläinen (2016) for a study on ALMP in Finland. Theyfind positive effects on earnings following compliant participation.


scribes the empirical strategy and the data. In Section 3.5, the mainresults are presented. Robustness checks are performed in Section 3.6.Section 3.7 presents results on other outcomes. Section 3.8 concludesthe chapter.

186 CHAPTER 3


Denmark has seen the number of asylum applicants vary a great dealover the years. 2001 marked a peak in the number of asylum seekers,and between 2001 and 2002, the number of asylum seekers was cut inhalf from 12,512 to 6,068, with most of the asylum seekers arrivingfrom Afghanistan, Iraq, and the Former Republic of Yugoslavia. Thisis the time period of immediate interest to this study, and as explainedin Section 3.2.1, it is a time of substantial change in terms of asylumpolicies.

The process of applying for asylum in Denmark is governed by theAliens Act and decisions are made by the Danish Immigration Service(DIS), while appeals are handled by the Refugee Appeals Board.6 Theprocess of applying for asylum in Denmark and the different types ofpermits are described in more detail in Appendix 3.A.

3.2.1 The 2002 reform package

On November 27, 2001, a new minority center-right-wing coalitiongovernment was appointed in Denmark. This shift of government re-flected a shift in the public opinion on immigration (see, for example,Mansouri et al., 2010). The new government introduced a number oflegislative changes regarding asylum and immigration policies thatwere passed by the Danish parliament as amendments to the AliensAct and the Integration Act. This study analyzes the effects of areform component that changed the criteria for eligibility for perma-nent residency in Denmark (henceforth referred to as the reform).This change was part of a suggestion for a new Bill to amend theDanish Integration Act, presented by the new government in Febru-ary 2002 (Ersbøll and Gravesen, 2010). The Bill was passed by the

6Individuals granted asylum for humanitarian reasons are an exception andin these cases, the Ministry for Foreigners, Integration and Housing (in 2002, theMinistry for Integration) makes the decision. If an asylum seeker’s application isrejected, he/she can still be given asylum for humanitarian reasons.


Danish parliament (Folketinget) on June 6, 2002.7 The explicit aimof this reform package was to limit the number of asylum seekers inDenmark, while honoring international obligations, and to speed upthe integration process (The Danish Immigration Service, 2003).

Both before and after the reform, individuals given asylum wereinitially granted a temporary residence permit if protection wasdeemed necessary. While under temporary status, the residencepermit could be discontinued if the grounds for residency were nolonger valid. Generally, temporary protection would be sustained ifthe need for protection was intact and there were no legal reasonsto withdraw it.8 Refugees could also be allowed to sustain theirtemporary residence permits based on labor-market attachments,even if there was no longer any need for protection. After a certaintime period as a resident in Denmark, a refugee (above 18 years old)would be eligible to apply for permanent residency.

The main change implied by the reform was the change in howlong a refugee would have had to have been on a temporary residencepermit before being considered for permanent residence status. Priorto the reform, three years were sufficient, whereas after the reforma refugee would have to wait for seven years before being allowed toapply for a permanent residence permit.9 This change implied thatindividuals subject to the new rules would have to live with tempo-rary protection for a longer time period, facing the risk of having theirpermit discontinued. Once eligible to apply for permanent residency,

7Discussions began in January when a new aliens policy was introduced, andthis gave rise to the suggested Bill amending the Integration Act in February. TheBill that was eventually adopted implied changes to the Aliens Act as well. Billno. L 152 entered into force as Act no. 365 of June 6, 2002.

8Paragraph 11 in the Aliens Act.9Formally, the reform implied that if the refugee had held a legal permit on

basis of paragraphs 7–9 of the Aliens Act for at least seven years, counting fromthe date of approval of the temporary permit, he/she was eligible to apply forpermanent residency. Paragraphs 7–9 included permits for all categories of refugeesthat are consider in this study, and, in particular, paragraph 9 included specificpermits based on labor-market attachment.

188 CHAPTER 3

refugees would be granted permanent residence if the need for pro-tection remained or if they had a labor-market attachment (given thefulfillment of some supplementary conditions), unless there were legalreasons to withdraw the residence permit. Prior to the reform, theseconditions included completing an integration program and havinglimited public debt. Under the new rules, in addition to complet-ing the integration course, asylum seekers would now have to pass alanguage test and hold no overdue public debt. In addition, while acriminal record used to lead to a longer waiting time, a serious crimi-nal record would prevent permanent residency altogether post-reform(Ersbøll and Gravesen, 2010). Obtaining permanent residency wasthus made more difficult by the reform.

In addition to the changes in the requirements for being eligibleto apply for permanent residency, the 2002 reform package also en-tailed lower benefit levels, made family reunification more difficult,abolished the de facto status and the possibility to apply for asylumat Danish embassies abroad. This study is able to isolate the effectof changes to the eligibility for permanent residency from other partsof the reform package as this was the only component introducedretroactively and it applied to all individuals who lodged their asylumapplication on or after February 28, 2002 (the date when the new Billwas proposed). The other components of the reform package took ef-fect after the Bill had been passed, on July 1, 2002. For more detailsabout the other components of the reform, see Appendix 3.B.

Another potentially important reform came in 2003, allowing im-migrants that had lodged their applications on or after February 28,2002 to apply for permanent residency already after five years if theywere "well integrated", i.e., if they had a strong labor-market attach-ment and had not relied on social welfare.10 Furthermore, in case of

10This was an addition to paragraph 11 in the Aliens Act, entered into forceas Act no. 425 of June 10, 2003, and the formal requirement implies that theapplicant should have lived legally in Denmark for at least five years and havebeen self-supporting with a solid labor-market attachment for the last three years.


an exceptionally successful integration, it was possible to receive apermanent permit already after three years of legal residency (Ers-bøll and Gravesen, 2010). In terms of the analysis in this study, thisimplies that the integration motive was made stronger.11

The key takeaway from the policy change introduced by Act no.365 of June 6, 2002, for the analysis, is that it implied a lower ex-anteprobability of being granted permanent residency based on asylumreasons. At the same time, permanent residency could be obtainedthrough labor-market attachment and a potential effect of the re-form is therefore that this option became more important. In termsof such incentives, different groups of refugees may have been differ-entially affected. In particular, since not all groups are likely to facethe same labor-market prospects, the option of securing residencythrough labor-market attachment will require more investments forsome groups than others. The types of heterogeneity considered inthe empirical part of this study are in terms of gender and skill levels.

11During 2002, there were some other important changes to decision practicesfor specific refugee groups. These are unrelated to the policy changes analyzed inthis study, but they are relevant to highlight since they affected the approval ratesfor specific nationalities. In particular, changes applied to asylum seekers fromAfghanistan, Iraq and Kosovo, for whom, following a reassessment of the securitysituations, the requirements for asylum were made stricter.

190 CHAPTER 3

3.3 Theoretical Framework

Changes to the probability of being granted permanent residencyin Denmark could affect labor-market and education investments inthe host country in two opposite directions. On the one hand, anex ante lower probability of permanent residency based on asylumreasons could increase the incentives to qualify for permanent resi-dency based on labor-market attachment, and thus increase invest-ments in country-specific human capital, for example by acquiring aneducation. If so, it would be expected to see positive effects on educa-tional and/or labor-market outcomes. On the other hand, with a lowerprobability of staying in the long run, the expected payoff to country-specific investments is lower. The lower probability could, in this case,deter asylum seekers and it would be expected to see negative effectson educational and/or labor-market outcomes. A key argument infavor of temporary protection is the idea that it has positive effectson integration. This claim is, however, clearly subject to verificationbecause of these potentially counteracting effects. To shed some lighton the mechanisms at work, a search and matching model is set upusing the framework laid out in Diamond (1984) and Mortensen andPissarides (1994), modified to include a choice of whether or not toinvest in human capital.12

The theoretical model aims at providing testable predictions andto facilitate the interpretation of the empirical results in this study.The model is intentionally kept as simple as possible to focus onthe key questions of interest. In particular, focus is on steady-stateanalysis. Given the one-time policy changes undertaken in Denmark,it would be relevant to also solve for transitional dynamics, wherebyjob-finding rates would change over time until they reach a new steadystate. It is conjectured that the transition dynamics in this model arerather fast, but an examination of this conjecture is left for future

12For studies on investments in host country-specific human capital, see forexample Chiswick (1978), Cortes (2004), and Duleep and Regets (1999).


research. The steady-state model is thus solved to obtain endogenousexpressions for wages, labor-market conditions and the decision ofwhether or not to invest in education. Comparative statics are thenperformed to study the response of these variables to a policy changesimilar to the Danish reform of 2002.

3.3.1 A search and matching model with human capitalinvestments

In the model, individuals are either educated or uneducated. Educatedworkers are considered high-skill, indexed H, and uneducated workersare considered low-skill, indexed L. The different skill levels S, withS ∈ {H,L}, correspond to productivity levels yH > yL.13 A simplify-ing assumption of separate markets for high- and low-skilled workersis made. It is further assumed that refugees may be in a temporary ora permanent state, corresponding to residency R, with R ∈ {T, P}.Therefore, there are four different markets in total corresponding tothe different combinations of S and R.

The productivity in the model is host-country specific and thevalue of being in the host country is assumed to be larger than thevalue of being in the home country, as this is consistent with theasylum seeker fleeing the home country. Therefore, the home countrycan be disregarded in the model. As refugees may lose temporaryresidency, an exogenous probability of being deported from the hostcountry is included. If the individual is deported, she gets nothing.Firms supply vSR vacancies and the unemployment rates are given byuSR. The matching function is given by MS

R = (vSR)α(uSR)1−α, whereMSR is the number of matches in residency state R and for skill level

S, and α ∈ (0, 1) is the match elasticity with respect to vacancies. Thetransition rate for an unemployed refugee worker of skill level S intoemployment in residency state R is then given by fSR(θSR) = (θSR)α,

13See Bennett et al. (2015) for a model where firms supply jobs for both immi-grants and natives.

192 CHAPTER 3

where θSR = vSR/uSR is labor-market tightness. Firms’ transition rates

are given by qSR(θSR) = (θSR)α−1.14 Next, the value functions of workersand firms are defined. Let UST and EST denote the expected presentvalues of unemployment and employment in the temporary state. Thevalue functions are then given by:

rUST = fST (EST − UST ) + ρSU (USP − UST )− Γ(S)c(e)− (a+ dU )UST , (3.1)

rEST = wST + σ(UST − EST ) + ρSE(ESP − EST )− Γ(S)c(e)− (a+ dE)EST , (3.2)

where r is the exogenous discount rate, ρSU and ρSE are the prob-abilities of moving from the temporary to the permanent state asunemployed and employed, respectively, Γ(S) is an indicator variablewhich takes the value of one if the worker acquires and maintains ed-ucation and zero otherwise, c(e) is the cost of acquiring education, ais the exogenous transition rate out of the labor force, dU and dE arethe probabilities of being deported while unemployed or employed,and wST is the sectoral wage (i.e., the wage for a given skill level andresidency status). The value function for an unemployed individualin the temporary residency state, rUST , then consists of the sum ofthe expected value of transitioning into employment, into permanentunemployment minus the loss from being deported (or exiting thelabor force), and the cost of investing in education (if the individualchooses to do so). For an employed individual in the temporary state,the value function, rEST , instead consists of the wage in the currentperiod, the expected value of losing the job and transitioning intounemployment and the expected value of becoming employed in the

14The transition rates are determined as: fSR(θSR) = MSR/u

SR, and qSR(θSR) =

MSR/v

SR.


permanent sector, minus the loss from the risk of exiting the laborforce or being deported, and the cost of investing in education (if thatapplies).

It is assumed that individual workers, indexed i, have differentabilities, ei, and therefore face different costs of obtaining educa-tion, c(ei). The variable ei is assumed to be uniformly distributed,ei ∈ [0, 1], and the costs are decreasing in ability at a decreasingrate, c′(ei) < 0 and c′′(ei) > 0. Furthermore, in order to guarantee anon-trivial solution where some, but not all, individuals choose to ac-quire an education, it is assumed that the individual with the highestability faces a very low cost of education, c(1) = 0, and the individ-ual with the lowest ability faces a very high cost of education, i.e.,limei→0 c(ei) =∞. Hence, Γ(H) = 1 and Γ(L) = 0.15

For the permanent state, the values of unemployment and em-ployment are instead determined by:

rUSP = fSP (ESP − USP )− Γ(S)c(e)− aUSP , (3.3)

rESP = wSP + σ(USP − ESP )− Γ(S)c(e)− aESP , (3.4)

where wSP is the wage in the permanent state for skill group S. Thedifference as compared to the temporary residence state is that indi-viduals no longer face the risk of being deported, and they do not needto take into account the probability of transitioning to the permanentstate. From the firm’s perspective, JST and V S

T represent the expectedpresent value of an occupied job and a vacant job in the transitorystate. The value functions for a job paying the wage wST and a vacantjob are then:

15It is assumed that the educational switching does not occur in steady state.The assumption that the education cost is borne every period is a simplifyingassumption and is not important for the results.

194 CHAPTER 3

rJST = yS − wST + σ(V ST − JST ) + ρSE(JSP − JST )− (a+ dE)JST , (3.5)

rV ST = qST (JST − V S

T )− k, (3.6)

where k are hiring costs. The value of a filled position is then equal tothe productivity gain minus the wage paid, plus the expected valueof that job instead turning into a vacancy or a permanent position,minus the risk of the worker exiting the labor force or being deported.The value of a vacancy is given by the probability of that vacancyturning into a filled position minus hiring costs. The expressions forfirms in the permanent are:

rJSP = yS − wSP + σ(V SP − JSP )− aJSP , (3.7)

rV SP = qSP (JSP − V S

P )− k. (3.8)

The difference now is that firms no longer risk losing their workersbecause of deportation (and they do not need to take into accounttransition into the permanent state). Free entry implies V S

R = 0 and

equations (3.6) and (3.8) can therefore be rewritten as k

qSR= JSR.16

16To see this, simply use equations (3.6) and (3.8) together with the free entrycondition to set:

0 = qST (JST − 0)− k −→ JST = k

qST,

0 = qSP (JSP − 0)− k −→ JSP = k

qSP.


Wages

Wages are determined in Nash Bargaining where workers and firmshave equal bargaining power.17 Wages are then determined by themaximization problem:

maxwSR

(ESR − USR

)0.5 (JSR − V S

R

)0.5, (3.9)

with the first-order condition ESR −USR = JSR − V SR .18 For the perma-

nent state, equations (3.1)-(3.6) are used and free entry is assumed.Furthermore, in the baseline case, the risk of being deported is as-sumed to be the same for individuals who are employed and un-employed, i.e., dE = dU = d. Then, the impact of a change in thedeportation rate can be studied, in Section 3.3.1. The simplifying as-sumption that ρHE = ρHU is made. The rationale behind this is thatdespite the general increase in the number of years before an immi-grant could apply for permanent residency, some special conditionswere in place for workers. Hence, since uneducated workers empir-ically have a higher unemployment rate, they were worse off thaneducated workers, in particular if they happened to be unemployed.Educated workers, regardless of current employment status, would bemore likely to fulfill these conditions over the years. To capture thisin the model, equal probability of securing permanent residency is as-sumed for high-skilled workers, but a difference in the probability for

17Note that a labor market consisting only of refugees is considered. Refugeesare likely to face different labor-market conditions than natives, and this mightbe reflected in the wage setting process. To simplify the analysis, however, equalbargaining power is assumed in the labor market that refugees face.

18This follows from:

(ESR − USR

)0.5 (JSR − V SR

)0.5[

0.5 1ESR − USR

dESRdwSR

+ 0.5 1JSR − V SR

dJSRdwSR

]= 0,

(ESR − USR

)0.5 (JSR − V SR

)0.5[

0.5 1ESR − USR

· 1 + 0.5 1JSR − V SR

· (−1)]

= 0.

196 CHAPTER 3

low-skilled workers is allowed depending on their employment status.In the end, the following expression for wages in the permanent stateis arrived at:

wSP = 0.5(yS + kθSP ). (3.10)

From this expression, it can be seen that wages are increasing inlabor-market tightness and in productivity. For the transitory state,the expression for wages is slightly more complicated:

wST = 0.5(yS + r + a+ d+ ρSE

r + a+ d+ ρSUkθST

− (ρSE − ρSU )r + a

r + a+ d

r + a+ d+ ρSUkθSP

), (3.11)

which is also increasing in productivity.19

Labor-market tightness

Next, labor-market tightness, θSR, is considered. This is defined asvacancies relative to the unemployment rate. Here, expressions arederived for the transitory and permanent state in reduced form, i.e.,expressions in terms of exogenous parameters and the endogenouslabor-market tightness. For the temporary state, equations (3.5)–(3.6)are used and free entry is assumed, i.e., k

qSR= JSR, to arrive at the

following expression (in terms of θLT and exogenous parameters) foruneducated workers:

19For educated workers, it as assumed that ρHE = ρHU , and the expression canbe simplified to obtain:

wHT = 0.5(yH + kθHT ). (3.12)


(r+a+ σ + ρLE + d)2k(θLT )1−α =(r + a+ σ + ρLE)yL

r + a+ σ− r + a+ d+ ρLEr + a+ d+ ρLU

kθLT

+(ρLE − ρLUr + a

r + a+ d

r + a+ d+ ρLU− ρLEr + a+ σ

)kθLP , (3.13)

and for educated workers:

(r+a+ σ + ρHE + d)2k(θHT )1−α =(r + a+ σ + ρHE )yH

r + a+ σ− kθHT −

ρHEr + a+ σ

kθHP . (3.14)

For the permanent state, the following is obtained:

(r + a+ σ)2k(θSP )1−α = yS − kθSP . (3.15)

It can be shown that the labor-market tightness facing workers withtemporary status is lower than the labor-market tightness facingworkers with permanent status, θST < θSP .20 This is then consistentwith a higher employment rate for permanent workers thantemporary permit workers. The reason is that the firm supplyingvacancies to temporary permit workers faces a lower duration of apotential match and therefore supplies fewer vacancies for a givenpool of unemployed job seekers. In Appendix 3.C, it is shown thatlabor-market tightness, θST , increases in ρSE and decreases in ρSU .Furthermore, wages, wST , are increasing in ρSU whereas the effect of achange in ρSE on wages is indeterminate.21

20The details are available upon request.21It is also noted that labor-market tightness decreases when the deportation

rate, d, increases. The same is true for wages. The intuition behind this resultis that a higher deportation rate decreases the value of a match, and therefore

198 CHAPTER 3

Education

Next, the human capital investment decision is studied: whether ornot to invest in education. For simplicity, workers with a temporarystatus are considered, which is the main state of interest in the em-pirical part of this study. When a worker makes this decision, shecompares the value of unemployment as an educated worker, bear-ing the associated costs of education, to the value of unemploymentas an uneducated worker. The marginal worker has the ability level,e, which makes her indifferent between acquiring higher educationand remaining an uneducated worker. The condition determining theability level of the marginal worker is written as:

rUHT (e) = rULT . (3.16)

Workers proceed to higher education if the expected income gainsfrom education exceed the cost of education.22 Equation (3.1)is rewritten and subtracted from equation (3.3). The free entrycondition is then used to arrive at the following rewritten expressionfor (3.16):

{(r + a)θHT + ρHU θ

HP −

r + a+ d+ ρHUr + a+ d+ ρLU

((r + a)θLT + ρLUθ

LP

)}·

k

r + a+ ρHU= c(e). (3.17)

Condition (3.17) determines e as a function of exogenous parametersand the endogenous variables, θST and θSP for S ∈ {H,L}. Workers withability level ei below the threshold level, ei ≤ e, choose not to invest ineducation, whereas workers with ei > e choose to go to school. Hence,

worsens the labor-market conditions (and decreases the labor-market tightness).The same intuition applies to wages.

22Note that this cost can be a monetary or a time cost. Here, it is considereda monetary cost.


e and (1 − e) constitute the uneducated and educated labor forces,respectively. The right-hand side of equation (3.17) is the expectedincome gain of investing in education. This gain needs to be positivein order for at least some workers to proceed to higher education. Thefact that productivity is higher for educated workers means that thereis an educational wage premium which, in turn, provides incentivesfor higher education as well as a higher probability of getting a job.

Impact of a policy change

Finally, the impact of a policy change on employment and educationis considered. The law change can be seen as having the followingimplication. It became more difficult for all immigrants to obtain per-manent residency. As mentioned earlier, despite the general increasein the number of years before an immigrant could apply for perma-nent residency, some special cases applied for employed workers. Asemployed immigrants could apply earlier, and as the unemploymentrate for educated workers is much lower than for uneducated workers,the negative impact was more severe for uneducated workers, and theuneducated unemployed workers were the most negatively affected.Three cases are therefore considered.

In the first case, it is assumed that the likelihood of obtaining per-manent residency decreases for uneducated and unemployed individu-als only, so that dρLU < 0. In the second case, the impact of a decreasein the likelihood of obtaining a permanent permit for employed une-ducated workers is considered, dρLE < 0. Finally, to capture that alsothe educated workers were negatively affected, just to a lesser extent,the impact of an equal decrease in the likelihood of obtaining perma-nent residency for (employed and unemployed) educated workers isconsidered, dρHE = dρHU < 0.

Consistent with the implications of the law change, the order ofmagnitude is thought of in the following way,

∣∣∣dρLU ∣∣∣ > ∣∣∣dρLE∣∣∣ > ∣∣∣dρHE ∣∣∣ =

200 CHAPTER 3

∣∣∣dρHU ∣∣∣ .Case 1. If dρLU < 0, the likelihood of obtaining permanent resi-dency decreases only for the unemployed, uneducated individuals. Inthis case, education increases. There are two counteracting forces atplay. First, the relative value of being educated (and unemployed) in-creases as the value of being uneducated and unemployed decreases.This increases the value of obtaining an education. Second, there isan increase in the labor-market tightness for the uneducated workersbecause their wages fall, which increases employment (they will bemore eager to have a job when the value of being unemployed falls).This effect tends to reduce the number of individuals that acquire aneducation. The former effect dominates the latter, implying that moreindividuals acquire education.

Case 2. If dρLE < 0, the likelihood of obtaining a permanent resi-dency for employed uneducated workers diminishes. A lower probabil-ity of permanent residency reduces labor-market tightness, and thusemployment, for this group of workers. This is because the match be-tween a worker and a firm will last for a shorter period of time. Thenegative impact on labor-market tightness dominates (the effect onwages is indeterminate) and the incentives to acquire an educationbecome stronger, which increases education.

Case 3. Finally, if dρHE = dρHU < 0, the likelihood of obtaining per-manent residency falls for educated individuals. As in Case 2, a lowerprobability of permanent residency reduces labor-market tightness andthus employment for the educational group is affected. But in this case,as both employed and unemployed workers are equally affected, there isnot a direct impact on wages and hence wages fall with labor-markettightness. The reduction of labor-market tightness for the group offirms hiring educated workers will reduce incentives to acquire an ed-ucation whereby fewer workers acquire education.


As noted, the actual law change implied that the reduction in thelikelihood of obtaining permanent residency had heterogenous impactdepending on an individual’s educational level as well as employmentstatus, both of which are typically related. Educated workers usuallyhave a lower unemployment rate compared to uneducated workers andtherefore face a lower reduction in the probability of obtaining per-manent residency than uneducated workers. Among the uneducatedworkers, the employed workers were less negatively affected than theunemployed workers. Consistent with the changes considered, educa-tion is then expected to increase, the impact on wages for uneducatedworkers is expected to be ambiguous and the impact of wages for ed-ucated workers is expected to be negative.

Furthermore, as pointed out above, the analysis abstracts fromthe fact that the intensity of the effects may differ across time.23 Inthe empirical design, the interest is in changes to behavior that occurover time. Specifically, outcomes are studied over time after approvaland individuals facing different spells under temporary protection arecompared. Dynamic effects will therefore be informally discussed fur-ther in Section 3.5.

23As shown in, for example, the literature on effects of unemployment insurance(UI) on unemployment duration, where duration dependence can be expected tomatter. See for example Nekoei and Weber (2017) where extended UI benefits arefound to lengthen unemployment, but also improve matching (measured in termsof wages). Rosholm and Toomet (2005) is an example allowing for discouragement.

202 CHAPTER 3


This section describes the data sources and the empirical strategyused for the analysis.

3.4.1 Data

The main data set is register data collected by Statistics Denmark.For the purpose of this study, two sources of Danish micro data arecombined. First, register data on a broad set of outcomes for all immi-grants in Denmark 1997–2015 from Statistics Denmark are available.This data set includes all immigrants who were registered as living inDenmark on January 1 in any of the years 1997 to 2015, which meansthat individuals can be followed up until twelve years after their initialapplication for asylum was approved. Second, using unique registerdata from the Danish Immigration Service, for each individual, thegrounds for the residence permit held as well as dates of applicationand approval are observed. Using individual identifiers, these data arelinked to the main data set enabling relevant treatment and controlgroups to be defined, as discussed in more detail in Section 3.4.2.The main variables of interest include educational investments andlabor-market outcomes. Enrollment, defined as the share of individ-uals that, at some point during the twelve years of data observed,enroll in general education or in education at the university level isstudied.24 In terms of labor-market outcomes, focus is on employmentstatus and labor earnings (including earnings from self-employment).Employment status is defined as the share of individuals that areregistered as employed (or self-employed) at some point during thetwelve years observe in the data, whereas earnings are measured afterthree and seven years of residency in Denmark in the benchmark (inthe sensitivity analysis, labor market outcomes are also considered in

24The data come from the educational registers UDDA and VEUV.


each individual year in the sample).25From the register data, information on demographic character-

istics is also collected, specifically age, gender, nationality, maritalstatus, and the number of children in the household, to be used ascontrol variables in the analysis.26 In addition, from the educationalregisters, two different measures of skill level at arrival are imputed.First, the highest level of education completed before arrival in Den-mark (primary/secondary or higher) is used.27 Second, the entry levelof Danish language courses (1, 2, or 3) is used, because the entry levelis determined by the individual’s skill level.28 These measures of initialskill level are used both as control variables and to split the samplein order to study heterogeneous effects.29

Sample restrictions Individuals lacking information on applica-tion date are dropped, together with those who applied for asylumbefore November 1, 2001 or after June 30, 2002. Without informationon the application date, it is not possible to identify the relevant con-trol and treatment groups. Figure 3.1 shows a time line of the periodof interest and the way the sample is split into a control and a treat-ment group. The control group is defined as individuals applying forasylum between November 1, 2001 and February 27, 2002, while the

25The data come from the INK and RAS registers.26These variables are from the population register (BEF). To determine marital

status on arrival, it is assumed that if the date of the first change in marital statusis missing, the change must have happened before arriving in Denmark (or it wouldhave been recorded). Children at arrival is defined by considering all children bornbefore the application year and associated with the first family identifier availablein the registers after the first asylum application.

27Primary/secondary education includes early childhood education and pri-mary education as well as lower and upper secondary education. Higher educationcaptures university studies (short cycle tertiary, bachelor, master and doctoral).

28Level 1 is for students with no or limited educational background, or thosewho are considered to have limited learning abilities because of trauma, level 2 isfor students with some (normal) educational background and level 3 is for studentswith higher education (who often speak several languages).

29All information on education comes from the two registers UDDA and VEUV.

204 CHAPTER 3

treatment group includes individuals applying between February 28and June 30, 2002. The sample split is chosen to ensure that nothingelse is happening that would affect those applying prior to and postthe cutoff differentially. As described in Section 3.2.1, there were sev-eral components to the 2002 reform, apart from the prolonged waitingtime for permanent residency. To avoid confounding effects from theseother components, which mainly relied on the date of approval, thesample is also restricted to individuals whose applications were ap-proved after July 1, 2002. This restriction is made in order to ensurethat the comparison is made between asylum holders who only differin terms of which rules regarding permanent residency they are sub-ject to, and not in any other dimensions. The reform included changesto, for example, the benefit structure. Due to long processing times,this restriction on the approval date does not reduce the sample toany considerable extent. Figure 3.2 shows the fraction of individu-als in 2001 and 2002 whose applications were approved post July 1,2002, by month of application. Individuals lodging their applicationfrom abroad are also excluded.30 Throughout the study, the unit ofanalysis is the individual. Finally, because this study is interested ineducational and labor-market outcomes, focus is on individuals whoare between the age of 16–60 at the time of application.

30This means dropping three individuals that would otherwise have been in-cluded in the control group.


Figure 3.1: Overview of the time period of interest

Nov

2001

Dec

2001

Jan

2002

Feb

2002

Mar

2002

Apr

2002

May

2002

Jun

2002

Jul2002

Con

trol

grou

p

New

gove

rnm

ent

Per

man

ent

resi

denc

ych

ange

Tre

atm

ent

grou

pA

ctno

.36

5ap

prov

edA

ctno

.36

5in

effec

t

Figure 3.2: Fraction granted asylum post July 1, 2002, by month of appli-cation 2001-2002

(a)

0

.2

.4

.6

.8

1

Shar

e ap

prov

ed p

ost J

uly

1, 2

002

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Year 2001

(b)

0

.2

.4

.6

.8

1

Shar

e ap

prov

ed p

ost J

uly

1, 2

002

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Year 2002

3.4.2 Identification

The implementation of the reform implies that refugees who appliedfor asylum prior to February 28, 2002 (henceforth referred to as thecutoff) were able to apply for permanent residency three years afterapproval, whereas those who applied after the cutoff had to wait forseven years. The fact that this reform was implemented retroactively

206 CHAPTER 3

gives rise to a regression discontinuity in time setting with no possibil-ity of manipulation around the cutoff. The decision about the reformwas taken on June 6, 2002 and took effect on July 1, 2002 - but thepart of the reform that are analyzed in this study applied retroac-tively to everyone who applied from February 28, 2002 onwards. Thismeans that neither immigrants nor the decision makers at the DIScould have perfectly manipulated the date of application in order toachieve a certain treatment. Intuitively, individuals who applied justbefore the cutoff should therefore be comparable to individuals whoapplied just after the cutoff.

Looking at aggregate statistics from The Danish Immigration Ser-vice (2003) in Figure 3.3, it is concluded that there is no major changein the number of lodged asylum applications in Denmark from Febru-ary to March 2002. In the data, only individuals whose asylum ap-plications were subsequently approved are observed. Figure 3.4 showsthe number of approved applications by month of application andtype of asylum during 2002. Once more, there is no notable changein the number of approvals around the cutoff. As the actual date ofapplication is observed, a histogram of the number of granted asylumapplications is also presented using the week of application in Figure3.5.31 The absence of a spike in the density of applications made justbefore the cutoff is intuitive as the reform was implemented retroac-tively, leaving no room for manipulation.32

31Aggregation to the weekly level is necessary in order to comply with the microdata policy of Statistics Denmark.

32In Table 3.6, the results from a formal test of manipulation at the cutoff usingthe Stata package rddensity are presented. This test is implemented for a linearand a quadratic specification. For the linear specification, a p-value of 0.121 isobtained while the quadratic specification results in a p-value of 0.935.


Figure 3.3: Asylum applications lodged in Denmark, by month of applica-tion 2001–2002

0

200

400

600

800

1,000

Lodg

ed a

pplic

atio

ns

Nov Dec Jan Feb Mar Apr May Jun

Data source: The Danish Immigration Service (2003).

Figure 3.4: Number of approved applications, by normalized month of ap-plication and type of asylum 2001–2002

58

99

115

50

84

105

27

55

72

31

51

73

14

44

58

37

63

74

33

51

65

37

54

66

0

50

100

150

Appl

icat

ions

by

type

of a

ppro

ved

asyl

um

-4 -3 -2 -1 0 1 2 3

Convention De facto statusOther or humanitarian

Data source: UDLST.

In the regression discontinuity framework, treatment effects are iden-

208 CHAPTER 3

Figure 3.5: Density of running variableApproved applications lodged November 2001 to June 2002

0

.02

.04

.06

.08

Den

sity

-18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12 14 16

Notes: Week 17 excluded, due to too few observations (n < 5), to comply with the rules of StatisticsDenmark.

tified by estimating the magnitude of the discontinuity at the cutoff.While the sharp cutoff implied by the reform intuitively lends itselfto the regression discontinuity approach, ideally one would want tocompare individuals on each side close to the cutoff. As Denmark ap-proves a relatively small number of asylum seekers, a relatively broadbandwidth of four months on each side of the cutoff is used (119 daysbetween November 2001 to June 2002). This leads the attention tothe inherent trade off between precision and bias in the regressiondiscontinuity framework. Extending the bandwidth around the cut-off increases the precision, but also the risk of introducing a bias. Asthe running variable is the date of application, treatment effects areestimated parametrically in order to avoid confounding time-varyingeffects. The regression equation is specified as:

Yi = α+ βTi + h(xi) + Tih(xi) + Zi + εi, (3.18)

where Yi is the outcome of individual i, xi is the normalized date ofapplication such that February 28, 2002, is set to zero and h(·) is a


continuous function of the date of application, Ti is an indicator fortreatment status with Ti = 0 if xi < 0 and Ti = 1 otherwise, andεi is the error term. An interaction between h(xi) and the treatmentindicator Ti is included, to allow for different trends over time oneach side of the cutoff. β is the coefficient of interest measuring theeffect of being subject to the new rules on permanent residency. Inthe main specification, h(·) is specified as a linear function. In Section3.6, the order of this polynomial is varied to test the robustness ofthe results. Furthermore, all specifications are estimated with andwithout a vector of predetermined individual characteristics, Zi, toincrease efficiency and confirm that covariates do not affect the pointestimates.33

For the graphical representation (Figures 3.9–3.10), the mean ofeach outcome is plotted for evenly spaced bins of the running variable.For each plot, a global linear polynomial is fitted, h, to approximatethe population CEF, using a uniform kernel and evenly spaced bins.34For all plots, the full bandwidth of 119 days is used and covariatesare not included. Similar graphs of predetermined characteristics arepresented to substantiate the continuity assumption underlying theregression discontinuity framework; see Figures 3.6–3.8 in Section 3.9.In terms of predetermined characteristics, demographic characteris-tics and educational background is studied. For demographic charac-teristics, the fraction of males, household characteristics and averageage, as well as nationality is considered. For educational background,

33Although it has been standard practice in regression discontinuity designs tocluster on the running variable, this study follows Kolesár and Rothe (2016) andabstain from clustering using only heteroskedasticity robust standard errors. Allestimations have been repeated for the full sample with clustering on the runningvariable and it is found that not clustering is the more conservative approach forall outcomes. Results from estimations where standard errors are clustered areavailable upon request.

34These plots are produced using the Stata package rdplot; for more details,see Calonico et al. (2014). Graphs using a quadratic polynomial for the mainoutcomes are available upon request.

210 CHAPTER 3

both a self-reported measure of the highest level of education achievedand the level of Danish studies to which the individual is assigned isinvestigated. The reason to consider these characteristics is that theyare predetermined variables that may impact how individuals are af-fected by the reform. There are significant discontinuities in terms ofnationalities. There is a positive jump for the category other nation-alities (of 0.175) and a negative jump for Afghans (-0.128).35 Thereis a marginally significant discontinuity in the share of males (0.136).Labor earnings one year after approval is also estimated by regressinglabor earnings on predetermined characteristics. Then, the disconti-nuity is estimated using the predicted values for earnings. There areno significant discontinuities in this variable.

Table 3.1 compares means of predetermined characteristics for thecontrol and treatment group as well as their normalized difference.36The normalized difference gives a scale-invariant measure of the mag-nitude of the difference between groups. Differences above 0.25 areconsidered to indicate sizable differences. It is noted that the groupsare generally well balanced over the whole 8 month period that de-fines the sample of interest. Once more, the biggest differences arise interms of nationalities: there are more Iraqis in the control group andmore individuals from the category other nationalities in the treat-ment group. In addition, there are fewer males in the treatment group.Apart from these variables, the two groups seem balanced. However,as a small sample size is challenging in a regression discontinuityframework, the analysis is complemented by looking at differences in

35Other nationalities is defined as a dummy equal to one if the individual is notfrom one of the most common countries of origin: Afghanistan, Former Yugoslavia,Iraq, or Somalia.

36See Imbens and Woolridge (2009) for a motivation for using the normalizeddifference. The measure is defined as:

nd = xt − xc√(sd2

t + sd2c) /2

.


average outcomes over time, by treatment status.

Table 3.1: Comparison of means (bw 119 days)

(1) (2) (2)–(1)Control Treatment Normalized difference

Demographic characteristicsMale 0.54 0.43 -0.22

(0.50) (0.50)No. of children 1.71 1.57 -0.07

(1.90) (2.03)Partner 0.51 0.56 0.10

(0.50) (0.50)Age at application 31.00 31.47 0.05

(9.10) (9.72)

EducationDanish 1 0.21 0.24 0.07

(0.41) (0.43)Danish 2 0.40 0.36 -0.08

(0.49) (0.48)Danish 3 0.27 0.27 0.00

(0.45) (0.45)Primary or secondary 0.45 0.47 0.04

(0.50) (0.50)Higher 0.24 0.22 -0.05

(0.43) (0.42)

Country of originAfghanistan 0.34 0.30 -0.09

(0.48) (0.46)Iraq 0.19 0.10 -0.26

(0.39) (0.30)Former Yugoslavia 0.10 0.15 0.15

(0.31) (0.36)Somalia 0.17 0.13 -0.11

(0.37) (0.33)Other 0.19 0.32 0.30

(0.40) (0.47)N 372 263

Notes: Values in parenthesis are (s.d.). Demographic characteristics are mea-sured at application. Danish 1 - Danish 3 indicate the level of Danish coursesassigned at approval, whereas primary or secondary and higher education indi-cates the level of education acquired prior to applying for asylum in Denmark.The normalized difference is defined as xt−xc√(

sd2t+sd2

c

)/2

.

Heterogeneous effects To capture potential heterogeneity in re-sponse to the reform, the sample is split by (i) level of education at ar-rival (below/above university level, henceforth referred to as low/high

212 CHAPTER 3

skilled) and (ii) gender (males/females).37 The reform may have hada different impact on different groups of refugees since differences inaccess to the labor market could determine how much they were ableto affect their probability of being granted residency based on labor-market attachment. These aforementioned sample splits can captureimportant differences in terms of labor-market access. Related to thetheoretical model, these groups may also differ in their cost of ac-quiring education in the host country. Other potentially interestingsample splits would be to look at age groups and country of origin,but splitting the sample along these dimensions is not feasible dueto the sample size and the distribution of these variables. For thesubgroup analysis, the sample is split by subgroup and equation 3.18estimated.

37The overlap between these groups is investigated. Around 47 percent of thefemales are classified as low skilled, and around 22 percent as high skilled. Amongmales the division is similar with 43 percent of the men classified as low skilledand 26 percent classified as high skilled. For around 30 percent of both males andfemales, skill level at arrival is not observed. The correlation between gender andskill level is 0.07.

3.5. RESULTS 213

3.5 Results

This section presents and discusses the empirical findings. Bothgraphical evidence and estimates from the regression analysisis presented. The graphical evidence is based on the estimateddiscontinuity at the cutoff date. For these figures, a bandwidth of119 days on each side of the cutoff is always used and covariates arenot included. For the regression analysis, results are presented withand without covariates for the benchmark bandwidth of 119 days.

3.5.1 Educational outcomes

Human capital investments can be viewed as part of the integrationprocess. Table 3.2 shows the estimated results for the full sample aswell as for the subgroups based on gender and skill level. First, en-rollment in formal education is considered. This variable measuresthe share of individuals that, at some point during the twelve yearsobserved in the data, enroll in any type of formal education (primary,secondary or university). Columns (1) and (2) show the estimated co-efficients with and without covariates for the full sample. An increaseof around 17 percentage points is estimated at the cutoff. The effectis significant at the 5 percent level when controls are excluded and atthe 1 percent level in the specifications including controls. Panel (a)in Figure 3.9 confirms this picture and a clear jump is observed at thethreshold. Turning to the subgroups analyzed in columns 3–10, theeffect is driven by females and, to a lesser degree, low-skilled individ-uals. For females, the estimated effect is an increase of 22 percentagepoints at the cutoff, significant at the 5 percent level for the specifi-cation with covariates (the effect is slightly stronger when covariatesare excluded). Next, the effect on enrollment at the university level isestimated. This variable measures the share of individuals that enrollin university education at some point in time during the twelve yearsobserved in the data. The estimated coefficient is positive at around

214 CHAPTER 3

6 percentage points but not significant (for the low skilled, the es-timated coefficient is around 11 percentage points and significant atthe 10 percent level).38 Panel (a) of Figure 3.11 shows the evolutionof enrollment rates over time. Here, the share of individuals that areenrolled in education in a specific year is investigated (through a fit-ted quadratic polynomial with 95 percent confidence intervals). Thisshows that the treatment group does have overall higher enrollmentrates over time for general education. Finally, in panel (a) in Figure3.21, the estimated effect at the cutoff for enrollment in education ineach year after arrival, is presented separately. If anything, the effectappears to grow stronger over time.

The positive effect on enrollment in education is interpreted as anincreased investment in human capital and integration. However, forthe integration to be successful, it is relevant to assess what enroll-ment results in. Therefore, this study also considers different typesof education that might be particularly relevant for access to the la-bor market (adult education and labor-market training), the propen-sity to complete an education, and the number of years in education(throughout, Danish courses which are mandatory for both groupsare excluded). There are no significant effects on any of these vari-ables.39 For females, who show the largest increase in the propensityto be enrolled in education, there is support for a higher propensity tobe classified as a student in the long run (measured seven years afterapproval). This implies that women are more likely to be students ascompared to working or being unemployed and could indicate thatfemales substitute work for education in order to boost their humancapital. Alternatively, females could be having a harder time gainingaccess to the labor market and enroll in education by necessity rather

38A test for the difference in means between the control and treatment groupis also implemented. There is a marginally significant difference between the twogroups for enrollment in general education, with higher enrollment rates for thetreatment group.

39The results are available upon request.

3.5. RESULTS 215

than choice.The findings for enrollment are in line with the hypotheses in Case

1 and Case 2 from the theoretical model, i.e., that the change in theex-ante probability of obtaining a permanent residence permit wasespecially pronounced for low skill individuals or those further awayfrom the labor market. Females are, if anything, likely to be lessattached to the labor market than males. Thus, the findings, withstrong effects among females and the low skilled, are in line with thetheoretical model. Next, direct measures of labor-market outcomesare considered.

3.5.2 Labor-market outcomes

Labor-market outcomes are direct measures of attachment to the la-bor market which makes them highly important. Furthermore, in theprevious section, no significant effects of the reform on human capitalinvestments were found for males. This could be because they in-stead find jobs to a greater extent, which makes it important to lookat labor-market outcomes also for subgroups. First, whether an in-dividual was ever employed (including self-employment) in Denmarkis studied. Next, labor earnings after three and seven years of resi-dency is studied. Employment is defined as the share of individuals(in the treatment and control group) that are ever employed or self-employed during the twelve years following their initial approval ofasylum. Table 3.3 shows the estimated regression results for the fullsample and the four subgroups. In addition, panels (c)–(e) in Figure3.9 show the graphical presentation of the results. For the full sam-ple, there are no significant effects on employment or earnings. Whenlooking at earnings conditional on employment, there is a marginallysignificant negative effect when including controls (these results areavailable upon request). All coefficients are negative but impreciselyestimated. The graphical evidence also reveals a small decrease, butthere are no indications of a sizable and significant negative effect.

216 CHAPTER 3

Turning to the subgroups, the picture is very similar to the resultsfor the full sample. However, females (and to some extent the highskilled) exhibit a significant decrease in earnings. This is consistentwith more females being enrolled in education. For males, who werenot more likely to enroll in education, there is no significant effect onany of the labor-market outcomes. Figure 3.12 confirms this picturelooking graphically at employment and earnings for each year and inFigure 3.21 the coefficients are estimated for each of the different yearsafter arrival. The conclusions remain. If anything, there are signs of amore negative impact on earnings in later years. To try to understandthe potential mechanisms, Figure 3.13 looks at the fitted quadraticpolynomial over time for earnings conditional on employment in thedifferent subgroups (since the subgroups are small already before con-ditioning on employment, regressions are not estimated for this out-come variable). A difference is noted in that earnings for women whoare employed appear to be very similar in both groups, whereas thereis a divergence for the high-skilled individuals with the control groupexperiencing stronger earnings over time. Interestingly, the divergencebetween the control and the treatment group appears 3–5 years afterapproval for this subgroup. This is around the time when individualsin the control group are eligible to apply for permanent residency sta-tus and is in line with high-skilled individuals in the treatment groupaccepting jobs with lower earnings compared to the control group.This could be a sign of weaker bargaining power of the individuals inthe treatment group or that employers are more reluctant to investin individuals whose future in the country (and thus also in the firm)is more uncertain. This is consistent with Case 2 and 3 in the model,where employment falls for both educational groups.

The highest skill level ever achieved in the labor market duringthe years observed (more details on skill level in Appendix 3.D.1)is also considered together with the number of times an individualchanges workplaces. There is no difference in the number of times

3.5. RESULTS 217

individuals change workplaces during the twelve years they are ob-served in Denmark but they generally appear to do so at a decreasingrate over time (which is consistent with a more stable labor-marketattachment). For high-skilled individuals, a weakly significant nega-tive effect on the highest skill level achieved on a job is estimated.This could imply that the high skilled accept jobs for which they arepotentially over-qualified. It is concluded that the increased enroll-ment in education does not seem to have translated into improvedlabor-market outcomes at any time horizon.40

The labor-market outcomes can be related to the theoreticalmodel as follows. Cases 1 and 2 had the same implications forinvestments in education but differed in the impact on labor-markettightness and therefore employment. The empirical evidence isconsistent with the relative reduction in the probability of obtainingpermanent residency for employed versus unemployed workers beingsuch that the counteracting effects cancel out. Females, however,display reduced earnings which is more in line with Case 1, i.e.,that unemployed and uneducated individuals were most negativelyaffected, as females are likely further away from the labor market.However, there is no evidence of divergence in earnings conditionalon employment when looking at earnings dynamics graphically. Thisis, in turn, consistent with the law change implying a reduction inthe likelihood to obtain permanent residency for everyone, eventhough it was more severe for the uneducated and unemployedworkers. Hence it is consistent with the theoretical model.

40The results are available upon request.

218 CHAPTER 3

3.6 Sensitivity Analysis

In this section, sensitivity analysis is performed to assess the robust-ness of the results. In particular, standard tests for the validity ofthe regression discontinuity approach is performed. Furthermore, theimportance of calender effects is investigated. Results not presentedhere are available upon request.

3.6.1 Placebo tests

A standard test in this type of design is to test for placebo effectsby estimating the same model, but varying the location of the cutoff.Discontinuities at other cutoff points (where nothing happened) maysuggest that discontinuities at the real cutoff are not due to the re-form. The main sample is split into the control and treatment groupseparately. Then, following Imbens and Lemieux (2008), discontinu-ities in the outcome variables are tested for at the median date ofapplication in each of the two groups. The advantage of splitting thesample into the control and treatment group is that fitting a regres-sion function over a point where a discontinuity is expected to occurcan be avoided. Discontinuities could be tested for at other pointswithin each of these sub-samples, but using the median gives morepower to detect potential discontinuities. Tables 3.7–3.9 present theresults from this placebo analysis. For most variables, there are nosignificant discontinuities at the placebo cutoff. For employment, asignificant (at the 10 percent level) discontinuity is estimated whencovariates are included. In addition, for a few other variables, whenthe other outcomes discussed in more detail in Section 3.7 are in-cluded, a significant jump at the placebo cutoff is detected. This is thecase for the share registered in Denmark, the number of births, andhospital visits for the treatment group. Given the narrow bandwidthused to implement this test and the lower number of observations thisimplies, it is not surprising that a few discontinuities are detected as

3.6. SENSITIVITY ANALYSIS 219

the time trend and control function cannot be estimated as well.

3.6.2 Choice of bandwidth

Given that the placebo test detected a few discontinuities at othervalues of the running variable than the true cutoff point, the robust-ness of the results are assessed in greater detail. More specifically,the sensitivity of the results to changes in the bandwidth is carefullyinvestigated. The main results, presented in Tables 3.2-3.3, are esti-mated using a bandwidth of 119 days around the cutoff point. Thebandwidth cannot be extended further without including individualsin the treatment group that were also subject to, for example, thechange in benefit levels. For this reason, the sensitivity analysis isrestricted to analyzing the effects when decreasing the bandwidth.

For both predetermined characteristics and outcome variables, co-efficients and confidence intervals from estimating the regression dis-continuity equation using bandwidths starting at 21 days and thenincreasing the bandwidth by two days at a time until reaching 119days (the benchmark bandwidth) are presented. Figures 3.15–3.17present the results from this analysis for predetermined characteris-tics. Although confidence intervals suggest that even for smaller band-widths, the coefficients are in general not significantly different fromzero and the coefficients become much more stable at broader band-widths. This analysis corroborates the choice of using a bandwidth of119 days. Figures 3.18–3.20 present the same type of analysis for theoutcome variables, and confirm the interpretation of the results. Atbroader bandwidths, the coefficients become insensitive to bandwidthchanges.

Many studies that use the regression discontinuity approachchoose to use optimal bandwidth selection, a data-driven approachto select how many observations on each side of the cutoff shouldbe used in the estimation. In this study, the broadest bandwidthpossible is used to isolate the effect of this reform, i.e., many

220 CHAPTER 3

observations as possible is used without including individuals thatwere also subject to other components of the 2002 reform. Thisgives the bandwidth of 119 days. However, regressions for the mainoutcomes are also estimated using the optimal bandwidth. Usingthe optimal bandwidth selection, about 100 observations are usedin estimations compared to the sample size of 635 when using the119 days bandwidth. In general, coefficients estimated with theoptimal bandwidth are in line with, or larger in magnitude, thanthe preferred specification. The exception is enrollment, wherethe magnitude is smaller and non-significant when controls areexcluded (although including controls increases the magnitude andthe estimate is significant at the 10 percent level) when using theoptimal bandwidth. In light of the low number of observations usedin these estimations, the 119 days bandwidth remains the preferredchoice.

3.6.3 Assumptions on the regression specification

The main results are also replicated using a quadratic polynomial,rather than the linear function for h(·) of Section 3.5. The main rea-son to include higher-order polynomials is to capture non-linearitiesin the underlying data. In this study, using a higher order polynomialoften appears to lead to overfitting and, thus, overestimation of theeffect. Using the linear specification is therefore the more conserva-tive choice for most outcomes. However, the results for enrollment aresensitive to the inclusion of a second-order polynomial. The estimatedeffect is smaller in magnitude and imprecisely estimated. Looking atthe graphical evidence in Figure 3.9 (a), the linear fit seems moreappropriate. The linear case therefore remains the preferred specifi-cation.

The benchmark estimations employ a uniform kernel, but all theresults have been estimated for the full sample using a triangularkernel as well. The motivation for using a triangular kernel is that

3.6. SENSITIVITY ANALYSIS 221

it gives more weight to observations close to the cutoff, but giventhe low number of observations in the sample, the uniform kernelremains the preferred choice. In general, the coefficients using a tri-angular kernel are in line with, or even larger in magnitude than, thepreferred specification. For enrollment, the effect is slightly weaker at13–15 percentage points. In addition, a few other outcomes becomesignificant.

3.6.4 Calendar effects

The treatment group arrives in Denmark later than the control groupby definition. One potential concern is therefore that any observedeffects depend on this difference rather than on the reform itself. Forexample, calender effects could potentially affect the results if thestate of the labor market differs between the points in time when thecontrol and the treatment group receive their asylum approvals.41 In2002, asylum seekers were not allowed to work until their applicationswere approved, which means that the difference in approval rates be-tween treatment and control needs to be considered. The distributionof approval dates for the treatment and control groups are quite sim-ilar, suggesting that there are no substantial differences in when thecontrol and treatment groups are allowed to enter the labor market.42

Because the approval dates of the two groups look rather similar(see Figure 3.23), it is possible that the processing times instead differ.The control group had a somewhat longer processing time, implyingthat these individuals spent more time in the asylum center awaitingtheir decision. If the time in processing matters, for example becauseof discouragement from a lack of meaningful activities or because alonger time spent in Denmark gave the control group an advantage

41Another concern could be if it was suspected that asylum seekers arriving be-tween November and February are inherently different compared to asylum seekersarriving in the spring. Potential differences in observed characteristics can be con-trolled for, but differences in unobserved characteristics cannot be controlled for.

42Descriptive graphs are available upon request.

222 CHAPTER 3

before entering the labor market, this could be of relevance. The dif-ferences are, however, not so large to be likely to impact the results.43Furthermore, Figure 3.6 shows that there is no discontinuity in theprocessing time at the cutoff.

The appointment of a new government, on November 27, 2001,was, however, clearly associated with stricter immigration policiesto come. Discussions of these policies started formally in January–February 2002 and there was media coverage on the intentions toimplement measures aimed at reducing immigration. This means thatimmigrants could have been aware of the intention to reform Danishasylum policies. Still, they would not have been able to foresee theexact timing of the reform.

43Descriptive graphs are available upon request.

3.7. OTHER OUTCOMES 223

3.7 Other Outcomes

So far, focus has been on outcomes related to the labor marketand human capital investments. These variables are most directlyrelated to the reform under study. However, there are several otherways in which the lower ex-ante probability of permanent residency(through a longer time period with temporary residency) may haveaffected individual behavior. In addition to educational investmentsand labor-market outcomes, it is explored whether the reformaffected crime rates, health, family composition, and the durationin Denmark. These outcomes are of interest to understand the fullimpact of the reform.

The criminal registers accessed in this study includes informationon whether individuals have been convicted of any crime, as well asproperty crimes separately, during their time in Denmark. The mea-sure of criminal activity captures the share of individuals that haveever been convicted during the twelve years that are observed.44 Thisvariable is included to consider potential outside options to the regularlabor market, as well as potential deterring effects of the prolongedtemporary status. The impact of the reform on health outcomes isalso assessed. In particular, the number of hospital/doctor visits overthe twelve-year period observed is considered.45 This adds a dimen-sion to the analysis as an individual’s health status may affect bothher current and future labor-market prospects. Furthermore, there isanecdotal evidence that the reform imposed stress on refugees. Thismakes health itself a relevant outcome if the direct impact on refugees’welfare are of interest and, more generally, it speaks to the potentialcosts to society of the reform. Finally, the effects on fertility behaviorare considered by studying the number of births during the first twelve

44From the KRAF register, information on charges and convictions are accessed.45This variable is created using data from the LPRPOP register on health care

utilization and diagnoses.

224 CHAPTER 3

years in Denmark.46 The intuition here is that increased uncertaintyabout the future in Denmark could have discouraged individuals fromhaving children.

3.7.1 Duration in Denmark

First, however, it is studied whether the reform had an impact onwhether asylum holders actually stayed in Denmark during the twelveyears they can be observed in the data. Individuals in the treatmentgroup faced the risk of losing their residence status for a longer timebefore they were eligible to apply for permanent residency. This couldlead to more individuals leaving Denmark, because their asylum claimwas no longer valid and they did not qualify for residency based onlabor-market attachment. Further, asylum holders may have left Den-mark by choice, due to the change in regulations. This highlights theimportance of looking at how long these individuals stay in Den-mark, since any effects on other outcome variables could potentiallybe driven by selection effects of individuals having to, or choosing to,leave Denmark. Estimation results in Table 3.4 show that there is nosignificant difference between the control and treatment group in theshare that is still registered in Denmark in 2015. This is true for thefull sample as well as the different subgroups, confirmed graphicallyin panel (e) of Figure 3.10, and it facilitates the interpretation of theother results. It is unlikely that any effects are driven by selection ineither the control or treatment group.47

46An indicator equal to one if the number of children in a family increases fromone year to the other is used as a measure of a birth.

47In Figure 3.22 regression discontinuity estimates of the share still in Denmarkfor each individual year up until twelve years after application are presented.These results confirm that there is no significant difference in the share staying inDenmark. Table 3.5 confirms that the groups are still relatively well balanced in2015.

3.7. OTHER OUTCOMES 225

3.7.2 Crime

A change in the permit structure could have affected crime rates asthe opportunity cost of criminal activity may have changed. The re-form studied could also have had a more direct effect on crime rates,since a criminal record reflected negatively on applications for pro-longed residency, implying that a longer time period with temporaryresidency in combination with these stricter rules may have deterredindividuals from committing crimes. In addition, the reform may haveimpacted the immigrant’s view of the host country and affected herwillingness to comply with its norms and laws.

The impact on convictions in general as well as on property crimespecifically is studied. This variable measures the share of individualsthat are ever convicted of any crime (for the general case) or of a prop-erty crime. Table 3.4 presents regression estimates for the full sampleand the subgroups. For the full sample, a negative but non-significantcoefficient of between 6 and 12 percentage points is documented. Con-sidering property crime rates instead, a significant (at the 5 percentlevel including controls) decrease of around 12 percentage points isestimated. This effect is mainly driven by a reduction of around 25percentage points for males (significant at the 1 percent level).48

The negative jump is confirmed in panels (a) and (b) in Figure3.10 as well as in panels (a) and (b) in Figure 3.14, showing the crimerates in the control and treatment group for the different years overtime. Panels (a) and (b) in Figure 3.22 show that there is no cleartrend in the estimated coefficients over time.

3.7.3 Fertility behavior and health

Finally, the impact on fertility behavior and health is studied. Thelower ex-ante probability of getting permanent residency in Denmark

48There is also a statistically significant difference in means between the twogroups, with lower rates for the treatment group.

226 CHAPTER 3

may have induced asylum holders to delay having children. A reasonfor this could have been to achieve a more stable situation before start-ing a family.49 The variable of interest is the total number of timesthat individuals have children during the years they are observed inthe data.

In Table 3.4, the coefficients on the number of children born in thefull sample are negative, but the effect is non-significant. For females,a significant (at the 5 percent level) decrease in the number of newbirths is estimated, with a coefficient of around -0.7. Panel (c) inFigure 3.10 confirms this picture for the discontinuity estimation, andpanel (c) in Figure 3.14 shows no clear graphical differences over time.Figure 3.22 shows that there is not a very clear pattern when lookingat the estimated effect at different years after arrival. If anything,there is a small upward trend in the coefficient in the later years.

Finally, anecdotal evidence suggests that the reform created amore stressful situation for asylum holders due to increased uncer-tainty about their future in Denmark. This is explored by studyingthe effect on the total number of visits to health care centers andhospitals (during the twelve years they can be observed). The resultssuggest no significant impact of the reform. This is true also lookingat the long run and at the estimated impact at different time horizons.

49This is in line with the model outlined by Ranjan (1999). He suggests thatthe irreversible aspect associated with childbearing, together with the ability topostpone, lead people to postpone childbearing when there is uncertainty aboutfuture income. The analogue to the context analyzed in this study is straightfor-ward, uncertainty about the future in Denmark may have lead asylum holdersto postpone childbearing. Gustafsson (2005) emphasizes that changes in fertilitybehavior do not need to be driven by a change in expected family sizes, but bedue to a changed timing of family formation. Postponing having children pays off,because women who have children later earn more, all else equal. This could play arole in this context, where the prolonged temporary period of the treatment groupmay have increased the value of labor-market attachment as an alternative way ofreceiving residency.



This study analyzes the effects of lowering the ex-ante probability ofreceiving permanent residency status on refugees’ outcomes. A Dan-ish reform in 2002, that prolonged the time period that a refugee wasrequired to have been a legal resident before being eligible to apply fora permanent residence permit, is exploited. In light of recent asylumpolicy changes in Europe and elsewhere, this is a question that hasassumed center stage in the policy debate. However, there is very littleevidence on how such reforms affect the integration of refugees andtheir labor-market prospects. While proponents of temporary protec-tion regimes often argue that stronger incentives to qualify for res-idency based on labor-market attachment will speed up the processof entering the labor market, this study finds no evidence of positiveeffects on labor-market outcomes. There is no difference between theemployment rates of individuals in the control and treatment groups.Similarly, there is no evidence of increased earnings. The estimatedcoefficient is negative but not significantly different from zero. Op-ponents to temporary protection argue that worsening the prospectsof staying in the new host country may deter country-specific humancapital investments since the expected payoff of doing so is discountedat a higher rate. This study finds evidence of the opposite, docu-menting large and significant effects on education enrollment rates,driven by females and low-skilled workers. These findings on invest-ments in education are in line with the predictions of the theoreticalmodel when the ex-ante probability of obtaining permanent residencyis more negatively affected for low-skilled individuals, which in theempirical analysis corresponds to individuals further from the labormarket.

Some limitations of the study should be emphasized. While thereform analyzed in this study is in many ways ideal for studying theeffect of prolonged temporary status, the setting also offers some chal-lenges. The sample size is limited for two reasons: (1) outcomes can

228 CHAPTER 3

only be studied for individuals that are actually given asylum and(2) the time interval needs to be restricted to four months before andafter the reform to avoid confounding policy changes. This meansthat the results have to interpreted carefully. Furthermore, the exter-nal validity of the results depends on the institutional setting. First,the composition of refugees is clearly time dependent and depends onmany things outside the control of the policy maker. Second, tem-porary protection schemes may be designed in many different ways,making them more or less comparable to the reform analyzed in thisstudy. Therefore, it is important to compare the results to the findingsof future studies of temporary permits analyzed in other settings. Fi-nally, this study focus on a variety of outcomes that are relevant butabstract from many others. For example, the role of intra-householdrelationships may be important in order to understand heterogeneousresponses of females and males. Exploring other outcomes and assess-ing potential mechanisms at work remain interesting tasks for futureresearch.

230 CHAPTER 3

3.9 Figures and Tables

Figure 3.6: General (predetermined) characteristics

(a) Male

.3

.4

.5

.6

-100 -50 0 50 100

Sample average within bin Polynomial fit of order 1

(b) No. of children at application

1

1.5

2

2.5

3

-100 -50 0 50 100


(c) Partner at application

.4

.45

.5

.55

.6

.65

-100 -50 0 50 100


(d) Age at application

29

30

31

32

33

-100 -50 0 50 100


(e) Processing time

500

600

700

800

900

-100 -50 0 50 100


(f) Predicted wage

0

.05

.1

.15

.2

-100 -50 0 50 100


Notes: The graphs are generated using evenly spaced bins, a linear polynomial (order 1), and a uniformkernel.

3.9. FIGURES AND TABLES 231

Figure 3.7: Danish language courses and education level (predetermined)

(a) Danish 1

0

.1

.2

.3

.4

-100 -50 0 50 100


(b) Danish 2

.2

.3

.4

.5

.6

-100 -50 0 50 100


(c) Danish 3

.2

.25

.3

.35

-100 -50 0 50 100


(d) Primary and secondary

.4

.45

.5

.55

-100 -50 0 50 100


(e) Higher

.15

.2

.25

.3

-100 -50 0 50 100



232 CHAPTER 3

Figure 3.8: Country of origin

(a) Afghanistan

.1

.2

.3

.4

.5

-100 -50 0 50 100


(b) Iraq

0

.05

.1

.15

.2

.25

-100 -50 0 50 100


(c) Former Yugoslavia

.05

.1

.15

.2

.25

-100 -50 0 50 100


(d) Somalia

0

.05

.1

.15

.2

.25

-100 -50 0 50 100


(e) Other nationalities

.1

.2

.3

.4

.5

.6

-100 -50 0 50 100




Figure 3.9: Education and labor-market outcomes

(a) Enrollment in education

.1

.15

.2

.25

.3

-100 -50 0 50 100


(b) Enrollment in university edu-cation

0

.05

.1

.15

-100 -50 0 50 100


(c) Employed

.3

.4

.5

.6

.7

.8

-100 -50 0 50 100


(d) Earnings 3Y

40000

50000

60000

70000

80000

-100 -50 0 50 100


(e) Earnings 7Y

60000

80000

100000

120000

-100 -50 0 50 100


Notes: The graphs are generated using evenly spaced bins, a linear polynomial (order 1), and a uniformkernel. Enrollment is a dummy variable equal to one if the individual at some point is enrolled in generaleducation. Enrollment in university education is the corresponding variable for university education.Employed is a dummy equal to one if the individual was ever employed in Denmark. Earnings is totallabor earnings from employment and/or self-employment after three and seven years, respectively.

234 CHAPTER 3

Figure 3.10: Crime, fertility behavior, and health outcomes

(a) Criminal conviction

.2

.3

.4

-100 -50 0 50 100


(b) Property crime

.05

.1

.15

.2

.25

-100 -50 0 50 100


(c) Giving birth

1

1.5

2

2.5

-100 -50 0 50 100


(d) Hospital visits

13

14

15

16

-100 -50 0 50 100


(e) Registered in Denmark, 2015

.6

.7

.8

.9

1

-100 -50 0 50 100


Notes: The graphs are generated using evenly spaced bins, a linear polynomial (order 1), and a uniformkernel. Registered in Denmark 2015 is a dummy equal to one if the individual is registered in Denmarkin the year 2015. Criminal conviction is a dummy equal to one if ever convicted of any crime. Propertycrime is equal to one if ever convicted of a property crime. Giving birth is the number of times a familyhas more children. Hospital visits is the number of doctor/hospital visits.


Figure 3.11: Education outcomes over time


0

.05

.1

.15

Shar

e

1 3 5 7 9 11Year relative to approval year

95% CI treat Mean treat95% CI control Mean control


0

.02

.04

.06

.08

Shar

e



Notes: The graphs show a quadratic polynomial and 95 percent confidence intervals.

Figure 3.12: Labor-market outcomes over time

(a) Employed

0

.2

.4

.6

Shar

e



(b) Earnings

20000

60000

100000

140000

180000

DKK




236 CHAPTER 3

Figure 3.13: (Heterogeneous) Earnings conditional on employment overtime

(a) Male

100000

200000

300000

400000

Aver

age

labo

r ear

ning

s (D

KK)



(b) Female

100000

200000

300000

400000

Aver

age

labo

r ear

ning

s (D

KK)



(c) High skilled

100000

200000

300000

400000

Aver

age

labo

r ear

ning

s (D

KK)



(d) Low skilled

100000

200000

300000

400000

Aver

age

labo

r ear

ning

s (D

KK)





Figure 3.14: Crime, fertility behavior, and health over time


0

.05

.1

.15

Shar

e



(b) Property crime

0

.02

.04

.06

Shar

e1 3 5 7 9 11

Year relative to approval year


(c) Giving birth

0

.1

.2

.3

.4

Num

ber o

f birt

hs



(d) Hospital visits

0

1

2

3

4

5

Num

ber o

f hos

pita

l vis

its



(e) Fraction still living in Den-mark

.8

.85

.9

.95

1

Shar

e




238 CHAPTER 3

Figure 3.15: General (predetermined) characteristics, RD coefficients bybandwidth

(a) Male

-.4

-.2

0

.2

.4

.6

20 30 40 50 60 70 80 90 100 110 120Bandwidth

95% CI upper/lower Parameter estimate

(b) No. of children at application

-2

-1

0

1

2

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(c) Partner at application

-.4

-.2

0

.2

.4

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(d) Age at application

-5

0

5

10

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(e) Processing time

-500

0

500

1000

1500

20 30 40 50 60 70 80 90 100 110 120Bandwidth



Figure 3.16: Danish language courses and education level (predetermined),RD coefficients by bandwidth

(a) Danish 1

-.4

-.2

0

.2

.4

.6

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(b) Danish 2

-.5

0

.5

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(c) Danish 3

-.6

-.4

-.2

0

.2

.4

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(d) Primary and secondary

-.6

-.4

-.2

0

.2

.4

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(e) Higher

-.4

-.2

0

.2

.4

20 30 40 50 60 70 80 90 100 110 120Bandwidth


240 CHAPTER 3

Figure 3.17: Country of origin, RD coefficients by bandwidth

(a) Afghanistan

-.6

-.4

-.2

0

.2

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(b) Iraq

-.4

-.2

0

.2

.4

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(c) Former Yugoslavia

-.2

0

.2

.4

.6

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(d) Somalia

-.6

-.4

-.2

0

.2

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(e) Other nationalities

-.2

0

.2

.4

.6

20 30 40 50 60 70 80 90 100 110 120Bandwidth



Figure 3.18: Education outcomes, RD coefficients by bandwidth


-.2

0

.2

.4

.6

20 30 40 50 60 70 80 90 100 110 120Bandwidth



-.3

-.2

-.1

0

.1

.2

20 30 40 50 60 70 80 90 100 110 120Bandwidth


242 CHAPTER 3

Figure 3.19: Labor-market outcomes, RD coefficients by bandwidth

(a) Employed

-1

-.5

0

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(b) Earnings 3Y

-150000

-100000

-50000

0

50000

100000

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(c) Earnings 7Y

-300000

-200000

-100000

0

20 30 40 50 60 70 80 90 100 110 120Bandwidth



Figure 3.20: Crime, fertility behavior, and health, RD coefficients by band-width


-.8

-.6

-.4

-.2

0

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(b) Property crime

-.6

-.4

-.2

0

.2

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(c) Giving birth

-1.5

-1

-.5

0

.5

1

20 30 40 50 60 70 80 90 100 110 120Bandwidth


(d) Hospital visits

-10

-5

0

5

10

15

20 30 40 50 60 70 80 90 100 110 120Bandwidth



-.4

-.2

0

.2

20 30 40 50 60 70 80 90 100 110 120Bandwidth


244 CHAPTER 3

Figure 3.21: Education and labor-market outcomes, RD coefficients overtime


-.1

0

.1

.2

1 2 3 4 5 6 7 8 9 10 11 12Years after arrival


(b) Employed

-.3

-.2

-.1

0

.1



(c) Earnings

-150000

-100000

-50000

0

50000




Figure 3.22: Crime, fertility behavior, and health, RD coefficients overtime


-.2

-.15

-.1

-.05

0

.05



(b) Property crime

-.1

-.05

0

.05

.1



(c) Giving birth

-.2

0

.2

.4



(d) Hospital visits

-.4

-.2

0

.2

.4




-.15

-.1

-.05

0

.05

.1



246 CHAPTER 3

Figure 3.23: Approval rates for lodged applications, November 2001 – June2002

0

.1

.2

.3

.4

.5

Nov Jan March May

Data source: Statistics Denmark.


Tab

le3.

2:Ed

ucationou

tcom

es

Outcome

Fullsample

Male

Fem

ale

Highskill

Low

skill

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Enrollm

ent

0.176∗

∗0.172∗

∗∗

0.094

0.114

0.278∗

∗∗

0.218∗

∗0.002

-0.046

0.180∗

0.157∗

(0.070)

(0.066)

(0.096)

(0.092)

(0.103)

(0.093)

(0.138)

(0.133)

(0.104)

(0.095)

N635

635

315

315

320

320

150

150

292

292

Enrollm

entun

iversity

0.066

0.059

0.066

0.070

0.071

0.038

-0.099

-0.138

0.119∗

0.107∗

(0.043)

(0.043)

(0.063)

(0.063)

(0.060)

(0.056)

(0.107)

(0.107)

(0.063)

(0.059)

N635

635

315

315

320

320

150

150

292

292

Covariates

NO

YES

NO

YES

NO

YES

NO

YES

NO

YES

Not

es:Regressions

areestimated

forthedifferentgrou

psusingapolyn

omialof

order1an

daun

iform

kernel.Highskilledis

defined

asha

ving

aun

iversity

education,

while

low

skilledha

veprim

aryor

second

aryeducationup

onarrivalin

Denmark.

Covariatesinclud

eageat

application,

gend

er,pa

rtner,

number

ofchildren,educationlevelan

ddu

mmiesforthemostcommon

nation

alities(A

fgha

nistan

,Former

Yug

oslavia,

Iraq

,an

dSo

malia).

Enrollm

entis

adu

mmyvariab

leequa

lto

oneif

theindividu

alat

somepoint

isenrolled

ingeneraleducation.

Enrollm

entun

iversity

isthecorrespon

ding

variab

leforun

iversity

education.

*,**

and***deno

tesign

ificancelevels

atthe10

percent,5percent

and1percent

levels,respectively.

248 CHAPTER 3

Tab

le3.

3:La

bor-marketou

tcom

es

Outcome

Fullsample

Male

Fem

ale

Highskill

Low

skill

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Employed

-0.041

-0.096

-0.002

0.008

-0.088

-0.137

-0.037

-0.134

-0.053

-0.135

(0.080)

(0.073)

(0.100)

(0.095)

(0.121)

(0.111)

(0.170)

(0.147)

(0.111)

(0.101)

N635

635

315

315

320

320

150

150

292

292

Earning

s3Y

-7,051

-22,536

-9,772

-12,544

-21,887

-27,372∗

∗7,385

-4,854

11,043

-10,544

(19,971)

(18,054)

(34,664)

(33,728)

(14,046)

(13,682)

(39,737)

(34,465)

(26,473)

(22,529)

N601

601

294

294

307

307

143

143

285

285

Earning

s7Y

-22,722

-41,671

-13,564

-16,631

-46,607∗

-53,064∗

∗-90,460

-125,518

∗∗

2,481

-17,994

(27,907)

(25,969)

(50,659)

(49,954)

(23,948)

(23,539)

(58,893)

(53,692)

(30,982)

(29,911)

N563

563

266

266

297

297

133

133

278

278

Covariates

NO

YES

NO

YES

NO

YES

NO

YES

NO

YES

Not

es:Regressions

areestimated

forthedifferentgrou

psusingapolyn

omialof

order1an

daun

iform


definedas

having

aun

iversity

education,

while

low

skilledha

veprim

aryor

second

aryeducationup

onarrivalin

Denmark.

Covariatesinclud

eageat

application,

gend

er,pa

rtner,

number


ddu


nation

alities(A

fgha

nistan

,Former

Yug

oslavia,

Iraq

,an

dSo

malia).

Employed

isadu

mmyequa

lto

oneiftheindividu

alwas

ever

employed

inDenmark.

Earning

sis

total

annu

allabor

earnings

inDKK

from

employ

mentan

d/or

self-employmentafterthreean

dsevenyears.

*,**

and***deno

tesign

ificance

levels

atthe10

percent,5percent

and1percent


3.9. FIGURES AND TABLES 249T

able

3.4:

Crime,

fertility

behavior,a

ndhealth

Outcome

Fullsample

Male

Fem

ale

Highskill

Low

skill

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

InDenmark2015

-0.009

0.003

0.047

0.061

-0.052

-0.041

-0.020

-0.005

-0.044

-0.024

(0.049)

(0.048)

(0.077)

(0.075)

(0.055)

(0.055)

(0.100)

(0.100)

(0.069)

(0.065)

N635

635

315

315

320

320

150

150

292

292

Cri

me

Criminal

conv

iction

-0.065

-0.119

-0.313

∗∗

∗-0.269

∗∗

0.103

0.054

0.001

-0.065

-0.044

-0.049

(0.076)

(0.072)

(0.110)

(0.107)

(0.010)

(0.099)

(0.166)

(0.143)

(0.111)

(0.103)

N635

635

315

315

320

320

150

150

292

292

Property

crim

e-0.105

∗-0.119

∗∗

-0.282

∗∗

∗-0.251

∗∗

∗0.057

0.031

-0.197

-0.222

∗-0.075

-0.071

(0.056)

(0.056)

(0.073)

(0.072)

(0.087)

(0.087)

(0.127)

(0.119)

(0.087)

(0.085)

N635

635

315

315

320

320

150

150

292

292

Fert

ility

beha

vior

Givingbirth

-0.311

-0.188

0.393

0.313

-0.729

∗∗

-0.661

∗∗

-0.498

0.030

-0.362

-0.292

(0.257)

(0.236)

(0.368)

(0.341)

(0.364)

(0.305)

(0.590)

(0.563)

(0.395)

(0.343)

N635

635

315

315

320

320

150

150

292

292

Hea

lth

stat

usHospitalvisits

-1.312

-0.000

0.755

-0.040

-1.130

0.281

1.091

3.627

-1.077

0.118

(2.098)

(1.983)

(2.235)

(2.217)

(3.412)

(3.431)

(6.819)

(6.328)

(2.681)

(2.324)

N635

635

315

315

320

320

150

150

292

292

Covariates

NO

YES

NO

YES

NO

YES

NO

YES

NO

YES

Not

es:Regressions

areestimated

forthedifferentgrou

psusingapolyn

omialof

order1an

daun

iform


definedas

having

aun

iversity

education,

while

low

skilledha

veprim

aryor

second

aryeducationup

onarrivalin

Denmark.

Covariatesinclud

eageat

application,

gend

er,pa

rtner,

number


ddu


nation

alities(A

fgha

nistan

,Former

Yug

oslavia,

Iraq

,an

dSo

malia).

InDenmark2015

isadu

mmyequa

lto

oneiftheindividu

alis

registered

inDenmarkin

theyear

2015.

Criminal

conv

iction

isadu

mmyequa

lto

oneifever

conv

ictedof

anycrim

e.Property

crim

eis

equa

lto

oneifever

conv

ictedof

aprop

erty

crim

e.Givingbirthis

thenu

mber

oftimes

afamilyha

smorechildren.Hospitalvisits

isthenu

mber

ofdo

ctor/h

ospitalvisits.*,

**an

d***deno

tesign

ificancelevels

atthe10

percent,5percent

and1percent


250 CHAPTER 3

Table 3.5: Comparison of means for individuals residing inDenmark 2015 (bw 119 days)

(1) (2) (2)–(1)Control Treatment Normalized difference

Demographic characteristicsMale 0.51 0.43 -0.16

(0.50) (0.50)No. of children 1.78 1.66 -0.06

(1.91) (2.05)Partner 0.53 0.57 0.08

(0.50) (0.50)Age 30.92 31.65 0.08

(9.15) (9.73)

EducationDanish 1 0.23 0.25 0.05

(0.42) (0.43)Danish 2 0.40 0.36 -0.08

(0.49) (0.48)Danish 3 0.26 0.28 0.04

(0.44) (0.45)Primary or secondary 0.48 0.47 -0.02

(0.50) (0.50)Higher 0.23 0.23 0.00

(0.42) (0.42)

Country of originAfghanistan 0.37 0.33 -0.08

(0.48) (0.47)Iraq 0.19 0.10 -0.26

(0.39) (0.30)Former Yugoslavia 0.10 0.16 0.18

(0.30) (0.37)Somalia 0.13 0.10 -0.09

(0.34) (0.30)Other 0.21 0.31 0.23

(0.41) (0.47)N 329 239

Notes: Values in parenthesis are (s.d.). This table shows the means and normal-ized difference for individuals in the sample that were still residing in Denmarkin 2015. Demographic characteristics are measured at application. Danish 1 -Danish 3 indicate the level of Danish courses assigned at approval, whereasprimary or secondary and higher education indicates the level of education ac-quired prior to applying for asylum in Denmark. The normalized difference isdefined as xt−xc√(

sd2t+sd2

c

)/2

.


Table 3.6: Regression discontinuity density test

bw 30 days bw 60 days bw 90 days bw 119 days(1) (2) (3) (4) (5) (6) (7) (8)

p-value 0.017 0.623 0.363 0.009 0.200 0.077 0.121 0.935

Degree of polynomial 1 2 1 2 1 2 1 2Notes: The test is implemented using the rddensity command in Stata, using the robust bias-correctedestimates. Reported values are p-values from this test.

Table 3.7: Placebo test: Education outcomes

Outcome Left of the cutoff Right of the cutoff(1) (2) (3) (4)

Enrollment 0.298 0.279 0.042 -0.093(0.187) (0.181) (0.208) (0.202)

Enrollment university 0.048 0.090 0.048 -0.066(0.114) (0.112) (0.048) (0.064)

Covariates NO YES NO YESNotes: Regressions are estimated for the full sample using a polynomialof order 1 and a uniform kernel. Covariates include age, gender, partner,number of children, education level (all measured at application) and dum-mies for the most common nationalities (Afghanistan, Former Yugoslavia,Iraq, and Somalia). The sample is split into two halves at the cutoff andthe regression is run on each sample using the median as the cutoff.

252 CHAPTER 3

Table 3.8: Placebo test: Labor-market outcomes


Employed -0.019 -0.090 0.211 0.874∗(0.221) (0.201) (0.366) (0.451)

Earnings 3Y -9,993 -27,485 53,063 43,639(78,314) (59,605) (41,749) (60,357)

Earnings 7Y 35,296 4,172 72,119 54,080(85,831) (73,671) (65,266) (76,644)

Covariates NO YES NO YESNotes: Regressions are estimated for the full sample using a poly-nomial of order 1 and a uniform kernel. Covariates include age,gender, partner, number of children, education level (all measuredat application) and dummies for the most common nationalities(Afghanistan, Former Yugoslavia, Iraq, and Somalia). The sampleis split into two halves at the cutoff and the regression is run oneach sample using the median as the cutoff.

Table 3.9: Placebo test: Crime, fertility behavior, andhealth


In Denmark 2015 0.104 0.116 0.226 0.222∗(0.101) (0.089) (0.146) (0.124)

CrimeCriminal conviction -0.214 -0.209 0.271 -0.002

(0.200) (0.194) (0.203) (0.217)Property crime -0.144 -0.064 0.163 0.147

(0.168) (0.147) (0.194) (0.168)Fertility behaviorGiving birth -0.575 -0.295 -1.803∗∗ 1.881∗∗∗

(0.687) (0.669) (0.891) (0.664)Health statusHospital visits -2.654 -5.605 -1.059 -16.570∗∗

(7.311) (5.629) (5.799) (7.757)Covariates NO YES NO YES

Notes: Regressions are estimated for the full sample using a polynomialof order 1 and a uniform kernel. Covariates include age, gender, partner,number of children, education level (all measured at application) and dum-mies for the most common nationalities (Afghanistan, Former Yugoslavia,Iraq, and Somalia). The sample is split into two halves at the cutoff andthe regression is run on each sample using the median as the cutoff.

REFERENCES 253

References

Adda, J., Dustmann, C., and Görlach, J.-S. (2016). The Dynam-ics of Return Migration, Human Capital Accumulation and, WageAssimilation. Mimeo.

Andersen, L. H., Dustmann, C., and Landersø, R. (2019). LoweringWelfare Benefits: Intended and Unintended Consequences for Mi-grants and their Families. Discussion Paper Series 05/19, CReAM.

Baker, S. R. (2015). Effects of Immigrant Legalization on Crime.American Economic Review: Papers & Proceedings, 105(5):210–213.

Bennett, P., la Cour, L., Larsen, B., and Waisman, G. (2015). Nega-tive Attitudes, Networks and Education. Working paper 01-2015,Copenhagen Business School, Department of Economics.

Blomqvist, N., Skogman Thoursie, P., and Tyrefors Hinnerich, B.(2018). Restricting Residence Permits: Short-Run Evidence from aSwedish Reform. Mimeo.

Calonico, S., Cattaneo, M. D., and Titiunik, R. (2014). Robust Data-Driven Inference in the Regression Discontinuity Design. The StataJournal, 14(4):909 – 946.

Cascio, E. U. and Lewis, E. (2017). How Much Does AmnestyStrengthen the Safety Net? Evidence from the Immigration Reformand Control Act of 1986. Mimeo.

Chen, J., Kosec, K., and Mueller, V. (2016). Temporary and Per-manent Migrant Selection: Theory and Evidence of Ability-SearchCost Dynamics. Discussion Paper Series 9639, IZA.

Chiswick, B. R. (1978). The Effect of Americanization on the Earningsof Foreign-Born Men. Journal of Political Economy, 86(5):897–921.

254 CHAPTER 3

Clausen, J., Heinesen, E., Hummelgaard, H., Husted, L., andRosholm, M. (2009). The Effect of Integration Policies on the TimeUntil Regular Employment of Newly Arrived Immigrants: Evidencefrom Denmark. Labour Economics, 16:409–417.

Cortes, K. E. (2004). Are Refugees Different from Economic Immi-grants? Some Empirical Evidence on the Heterogeneity of Immi-grant Groups in the United States. The Review of Economics andStatistics, 86(2):465–480.

Devillanova, C., Fasani, F., and Frattini, T. (2017). Employment ofUndocumented Immigrants and the Prospect of Legal Status: Ev-idence from an Amnesty Program. Industrial and Labor RelationsReview (forthcoming), pages 1–29.

Diamond, P. A. (1984). A Search Equilibrium Approach to the Mi-crofoundations of Macroeconomics. The MIT Press, Cambridge,Massachusetts.

Duleep, H. O. and Regets, M. C. (1999). Immigrants and Human-Capital Investment. The American Economic Review, 89(2):186–191.

Dustmann, C., Fasani, F., Frattini, T., Minale, L., and Schönberg,U. (2017a). On the Economics and Politics of Refugee Migration.Economic Policy, 32(91):497–550.

Dustmann, C., Fasani, F., and Speciale, B. (2017b). Illegal Migrationand Consumption Behavior of Immigrant Households. Journal ofthe European Economic Association, 15(3):654–691.

Ersbøll, E. and Gravesen, L. K. (2010). The INTEC Project: Integra-tion and Naturalisation Tests: the New Way to European Citizen-ship. Country report denmark, Centre for Migration Law, RadboudUniversity Nijmegen.

REFERENCES 255

Fasani, F. (2015). Understanding the Role of Immigrants’ Legal Sta-tus: Evidence from Policy Experiments. CESifo Economic Studies,61(3–4):722–763.

Fasani, F. (2018). Immigrant Crime and Legal Status: Evidence fromRepeated Amnesty Programs. Journal of Economic Geography,4(1):887–914.

Fasani, F., Frattini, T., and Minale, L. (2018). (the Struggle for)Refugee Integration into the Labour Market: Evidence from Eu-rope. Discussion Paper Series 11333, IZA.

Gustafsson, S. (2005). Having Kids Later. Economic Analyses forIndustrialized Countries. Review of Economics of the Household,3:5–16.

Huynh, D. T., Schultz-Nielsen, M.-L., and Tranæs, T. (2007). Em-ployment Effects of Reducing Welfare to Refugees. Study Paper 15,Rockwool Foundation Research Unit.

Imbens, G. W. and Lemieux, T. (2008). Regression DiscontinuityDesign: A Guide to Practice. Journal of Econometrics, 142:615–635.

Imbens, G. W. and Woolridge, J. M. (2009). Recent Developmentsin the Econometrics of Program Evaluation. Journal of EconomicLiterature, 47:5–86.

Kolesár, M. and Rothe, C. (2016). Inference in Regression Discontinu-ity Designs with a Discrete Running Variable. American EconomicReview (forthcoming).

Lozano, F. and Sørensen, T. A. (2011). The Labor Market Value toLegal Status. Discussion Paper 5492, IZA.

256 CHAPTER 3

Mansouri, F., Leach, M., and Nethery, A. (2010). Temporary Pro-tection and the Refugee Convention in Australia, Denmark andGermany. Refuge: Canada’s periodical on refugees, 26(1):135–147.

Mastrobuoni, G. and Pinotti, P. (2015). Legal Status and the Crim-inal Activity of Immigrants. American Economic Journal: AppliedEconomics, 7(2):175–206.

Mortensen, D. T. and Pissarides, C. A. (1994). Job Creation andJob Destruction in the Theory of Unemployment. The Review ofEconomic Studies, 61(3):397–415.

Nekoei, A. and Weber, A. (2017). Does Extending Unemploy-ment Benefits Improve Job Quality? American Economic Review,107(2):527–561.

Orrenius, P. M. and Zavodny, M. (2015). The Impact of TemporaryProtected Status on Immigrants’ Labor Market Outcomes. Amer-ican Economic Review: Papers & Proceedings, 105(5):576–580.

Pinotti, P. (2017). Clicking on Heaven’s Door: The Effect of Im-migrant Legalization on Crime. American Economic Review,107(1):138–168.

Ranjan, P. (1999). Fertility Behaviour Under Income Uncertainty.European Journal of Population, 15(1):25–43.

Rosholm, M. and Toomet, O. (2005). A Search Model of Discourage-ment. Discussion Paper Series 1633, IZA.

Rosholm, M. and Vejlin, R. M. (2010). Reducing Income Transfersto Refugee Immigrants: Does Starthelp Help You Start? LabourEconomics, 17(1):258–275.

Sarvimäki, M. and Hämäläinen, K. (2016). Integrating Immigrants:The Impact of Restructuring Active Labor Market Programs. Jour-nal of Labor Economics, 34(2):479–508.

REFERENCES 257

The Danish Immigration Service (2003). Statistical Overview 2002.

258 CHAPTER 3

Appendices

3.A The Danish Asylum Process

The process of applying for asylum in Denmark is governed by theAliens Act from 1983, to which several changes have been made overthe years. This section will briefly describe the Danish asylum sys-tem in effect around 2002 and builds upon information in The DanishImmigration Service (2003). An asylum seeker arriving in Denmarkunder these conditions should report to the police once at the bor-der. The application is filed either at a local police station or at thecenter in Sandholm. First, the Danish Immigration Service (DIS) willconfirm that Denmark is responsible for processing the asylum ap-plication. Asylum seekers that are not rejected at the border will besent to a registration center and, once identity and travel routes toDenmark have been established, to one of several accommodation cen-ters in the country. During the time when the application for asylumis processed, accommodation and financial support for the asylumseeker are provided by the DIS.50 Most asylum seekers will be accom-modated at a residence center until the final decision has been made,but after six months from the application date, the asylum seeker isallowed to find own housing until the claim has been processed (butthey are not allowed to buy property). During the time when the ap-plication is processed, the asylum seeker is not allowed to accept anypaid work. Voluntary activities are provided, and there are also somecompulsory activities.

The asylum application is handled by the DIS, which is the firstinstance of decision and they will determine if the asylum applicationfalls under the provisions of the Geneva Convention or the DanishAliens Act. The assessment is made using information provided by

50In cooperation with the Danish Red Cross, the Danish Emergency Manage-ment unit and the municipalities of Hansthom and Brovst.

3.A. THE DANISH ASYLUM PROCESS 259

the asylum seeker as well as information collected by the DIS on theasylum seeker’s country of origin. Convention refugee status is regu-lated by the UN 1951 refugee convention. Asylum seekers who do notdirectly qualify as refugees according to the definition of the RefugeeConvention but who risk the death penalty or being subjected to tor-ture or inhuman or degrading treatment or punishment in case ofreturn to his or her country of origin get protection status. This cat-egory extends the refugee status beyond the UN refugee convention,to individuals with "asylum reasons similar to those in the conven-tion". Prior to the reform in 2002 this would fall under the de factorefugee status in the Danish Aliens Act. As part of the reform, a newStatus B was introduced with much stricter criteria to get protectionstatus. There are two different procedures, the normal procedure andthe manifestly unfounded procedure.51 If an application is rejectedby the DIS under the normal procedure, it is automatically appealedto the Refugee Appeals Board52 (whose decision is final), in order tospeed up the process. If the application is rejected, the individual canstill obtain a residence permit for humanitarian reasons or for otherexceptional reasons (in which case decisions are made by the Ministryof Integration, later the Ministry for Foreigners, Integration and Hous-ing), although very few individuals are considered for these types ofresidence permits. These decisions are final, and cannot be appealed.If granted asylum for humanitarian reasons, one can only stay in thecountry for as long as those reasons still exist. What constitutes hu-manitarian reasons has varied over time and, for example, used toinclude families with small children from countries at war and indi-

51The manifestly unfounded procedure is applied when it is clear that the ap-plication cannot be approved. In this case, there is no possibility to appeal, andthe applicant has to leave Denmark immediately. However, it is required that theDanish Refugee Council (an NGO) agrees with the DIS’s assessment. If the DanishRefugee Council uses its veto, the case will instead be processed under the normalprocedure.

52An independent body with representatives from the government and the Dan-ish Bar and Law Society.

260 CHAPTER 3

viduals suffering the effects of torture. Nowadays, it is only grantedto individuals with life-threatening illness who cannot get treatmentin their own country. Once recognized as a refugee, social benefitsare given on the same conditions as for Danish citizens. According tothe 1999 integration plan, the DIS required that the refugee residedin a specific municipality during a three-year integration program.Refugees would be assigned a municipality based on a quota systemdesigned to achieve even distribution, with considerations to circum-stances related to the municipality and the refugee. The integrationprogram consisted of Danish culture courses, language classes, andvocational training. After three years (prior to the reform), perma-nent residence permits were conditioned on the performance in theintegration program.

3.B Other Reform Components

The other components of the reform package is briefly described be-low:

1. Access to the Danish welfare state was limited. Following thereform, individuals were required to have been a resident inDenmark for seven out of the eight most recent years to get thestandard level of benefits. For others, benefits were lowered by35 percent. This part of the reform also applied to native Daneswho had lived abroad. The change applied to all asylum seekerswho got their applications granted after July 1, 2002.

2. Family reunification of refugees was discouraged in several ways.First, by disallowing reunification for spouses under the age of24 (both spouses had to be 24 years of age or older). In addition,if a Danish citizen wanted to sponsor a spouse, the couple hadto prove that their “ties” were stronger to Denmark than to thecountry of the non-Danish spouse. Further, a Danish citizen

3.B. OTHER REFORM COMPONENTS 261

could no longer sponsor a parent aged 60 years or older. Thischange applied to applications for family reunion lodged afterJuly 1, 2002.

3. The de facto refugee status was abolished. This status previ-ously implied that an individual could get asylum even if thecriteria of the UN Geneva Convention from 1951 were not sat-isfied. This was no longer possible. Instead, a new status B wasintroduced with a more narrow scope. This change applied toall refugees who lodged their applications after July 1, 2002.

4. Prior to the reform, it was possible to apply for asylum in Den-mark at a Danish embassy or consulate abroad. This possibilitywas abolished by the reform.53 The possibility to lodge an ap-plication abroad was removed as of July 1, 2002.

The control and treatment groups are defined to make sure that theseother changes do not interfere with the component of interest for thisstudy.

53During the first six months of 2002, 354 individuals lodged their asylumapplications from abroad. In 2001 that number was 1,933 with a vast majorityof the applications (1,669) coming from the embassy in Afghanistan (The DanishImmigration Service, 2003).

262 CHAPTER 3

3.C Model Details

3.C.1 Derivatives

To study the impact on θLT of changing ρLE and ρLU , the expression inequation (3.13) is evaluated around ρLE = ρLU :

dθLTdρLE

=

(r+a+σ+d

r+a+σ+ρLE+d

(yL−kθLPr+a+σ −

yL−kθLTr+a+σ+d

)+

r+a+dr+a kθLP−kθ

LT

r+a+d+ρLU

)((r + a+ σ + ρLE + d)(1− α)2(θLT )−α + 1

)k

> 0

dθLTdρLU

= −

(1

r+a+ρLE+d

)(1r+a(r + a+ d)kθLP − kθLT

)((r + a+ σ + ρLE + d)(1− α)2k(θLT )−α + k

) < 0.

The impact on labor-market tightness of a change in ρLE is positive,whereas the impact of a change in ρLU is negative. Next, the impactof changes to ρLE and ρLU on wages in the transitory state, wLT , isconsidered. The following is obtained:

dwLTdρLE

=− 0.5[( r+a+d

r+a )θLPk − θLTkr + a+ d+ ρLU

−r+a+σ+dr+a+σρLE+d

(yL−θLP kr+σ+a −

yL−θLT kr+σ+a+d

)+ θLP k( r+a+d

r+a )−kθLTr+a+d+ρLU

(r + a+ σ + ρLE + d)2(1− α)(θLT )−α) + 1

]≷ 0,

dwLTdρLU

= 0.5k[ θLP (r + a+ d)(r + a)(r + a+ d+ ρLE)

− θLTr + a+ d+ ρLE

][1− 1

(r + a+ σ + ρLE + d)2(1− α)(θLT )−α + 1

]> 0.

This means that the impact on wages of a change in ρLE is indeter-minate, whereas the effect of a change in ρLU is positive. Finally, the

3.C. MODEL DETAILS 263

impact of an increase in dE = dU = d (for skill group S) is considered:

dθSTdd

= −2k(θST )1−α

(r + a+ σ + d+ ρSE)(1− α)2k(θST )−α + k< 0.

It can be seen that the labor-market tightness decreases as the de-portation risk increases, implying that labor-market conditions dete-riorate.

3.C.2 Impact of policy

Case 1: In the first case a reduction in the likelihood of obtain-ing permanent residency is considered only for unemployed low skillworkers, dρLU < 0 (around ρHE = ρLE = ρHU = ρLU ). It is found that:

de

dρLU={

1r + a+ d+ ρLU

[θLT −

(r + a+ d)θLPr + a

]

−dθLT

dρLU

}k(r + a)r + a+ ρHU

1c′(e) > 0.

Noting that c′(e) < 0, there is a positive relationship between e andρLU . Since the policy change implied a decrease in ρLU , it is concludedthat the impact of the policy is a decrease in e and thus an increasein the share of individuals that acquire education (remember that eis the share of uneducated individuals).

Case 2: In the second case a reduction in the likelihood of obtainingpermanent residency is considered only for employed low skill workers,

264 CHAPTER 3

dρLE < 0 (around ρHU = ρLU ). It is found that:

de

dρLE= − (r + a)k

r + a+ ρHU

dθLTdρLE

1c′(e) > 0.

There is a positive relationship between the share of uneducatedindividuals and the probability of permanent residency. This meansthat a decrease in ρLE will increase the investments in education.

Case 3: Finally, in the third case, a decrease in the likelihood ofobtaining permanent residency for educated individuals is considered,dρHE = dρHU < 0. It is found that:

de

dρH= 1c′(e)

{(r + a)

[dθHTdρH

(r + a+ ρH) + (θHP − θHT )]

+ d

r + a+ d+ ρLU

[(r + a)θLT + ρLUθ

LP

]} k

(r + a+ ρH)2 < 0.

This implies that when the probability of obtaining permanent resi-dency falls for educated individuals, fewer individuals invest in edu-cation.

3.D. DETAILS ON THE DATA 265

3.D Details on the Data

3.D.1 Mapping from DISCO codes to skill levels

The Danish version of the International Standard Classifications ofOccupations (DISCO) is used to map working functions, observed inthe data, into different skill levels. A skill level is defined by pickingone occupation per year and person. If a person has more than onejob, the job with the highest skill level is considered. If there is morethan one job with the same skill level, the most common job withinthat skill level is considered. Finally, if a person has two jobs fromdifferent occupations that have the same skill level in a given year, theoccupation with the highest ranking according to the DISCO code isconsidered. If a person has an equal number of jobs in two occupationswith the same skill level and one of the occupations is in the armedforces, the civilian occupation is considered as the main occupation.This variable is used to see if there is a change in average skill levelover time.

Code Description Skill level

1 Managers 3 + 4a

2 Professionals 43 Technicians and Associate Professionals 34 Clerical Support Workers 25 Services and Sales Workers 26 Skilled Agricultural, Forestry and Fishery Workers 27 Craft and Related Trades Workers 28 Plant and Machine Operators and Assemblers 29 Elementary Occupations 10 Armed Forces Occupations 1, 2 + 4b

aLevel 3 for managers in Hospitality, Retail and Service. Other managers haveskill level 4.

bMilitary officers are level 4, other occupations are at level 1. Non-commissioned Officers count as skill level 2.

266 CHAPTER 3

Chapter 4

The Effects ofPerformance Based Bonuses inthe Swedish Language-TrainingProgram for Immigrants∗

∗I thank Peter Fredriksson, Eskil Wadensjö, Hans Grönqvist, Jonas Ceder-löf, Niklas Blomqvist, and Karin Edmark for helpful comments. This researchbenefited from financial support from Handelsbanken’s Research Foundations. Allerrors are my own.

268 CHAPTER 4

4.1 Introduction

Migration flows have increased over the last decades and today morethan five million permanent migrants settle in the OECD countries ev-ery year. A non-negligible share of these are people seeking protectionfollowing the humanitarian crisis in Syria and recent developments inother parts of the world.1 It is well-documented that immigrants per-form worse on the labor market compared to natives and that they,to a higher extent, rely on social welfare. For example, on average, ittakes about 20 years for refugees in Europe to reach the same employ-ment level as natives (OECD, 2016). This puts pressure on decisionmakers to implement policies that promote integration of newcomers.Speeding up the integration process holds the promise of not only im-proving immigrants’ welfare but also decreasing public expenditures.For this reason, many countries offer (or require) newcomers to par-ticipate in integration programs. Language training is usually a keycomponent of these programs.

While proficiency in the host country language is crucial for suc-cessful integration, many immigrants never fully learn the new lan-guage. In the European Union, self-reported language knowledge ofimmigrants arriving during the last ten years suggest that only 24percent of refugees and 54 percent of other non-EU born immigrantshave at least advanced host-country language knowledge. Looking atcorresponding numbers for those who have spent more than 10 yearsin their host country, the shares increase to 49 and 69 percent respec-tively (OECD, 2016). At the same time, several studies show thathaving language-skills is associated with higher wages and increasedemployment probability.2 In addition, it is generally argued that lan-

1In 2017, the estimated number of permanent migrants was just above fivemillion and this represents the first decrease in migration flows since 2011. Thenumber of asylum claims summed up to 1.23 million in 2017 (OECD, 2018).

2See for example Chiswick (2008) who finds that language proficient immi-grants have about 15 percent higher earnings than non-proficient immigrants and


guage skills facilitate social as well as political inclusion. Improvinglanguage skills therefore seems a relevant task and a motivated compo-nent of integration programs. Still, despite the potentially large gainsfrom improving immigrants’ language skills, little is known about theoptimal design of language training programs.

This study analyzes a specific component of a language-trainingprogram, designed to improve students’ performances. In 2010,performance-based bonuses were introduced in the Swedishlanguage-training program for immigrants, Svenska för invandrare,henceforth referred to as Sfi education. The introduction of thebonus system was motivated by what was perceived as deficienciesin the language-training program, with a high drop-out rate andlong period of studies before receiving a passing grade.3 A financialreward of up to 12,000 SEK (approximately 1,666 USD) for passinga bonus-entitling was introduced with the purpose of strengtheningimmigrants’ incentives to complete their language training. Thereward was conditional on receiving a passing grade within twelvemonths from starting Sfi education and 15 months from arrival toSweden.

The effects are estimated using a sharp regression discontinuitydesign, exploiting the fact that within the targeted population, eligi-bility to the bonus system was completely determined by the date ofimmigration to Sweden. There is no evidence of manipulation aroundthe cutoff date, and it is highly unlikely that individuals could orwould want to postpone their date of registration in order to becomeeligible for a bonus given the other benefits associated with beingregistered in a municipality. The first outcome of interest is the en-

Kennerberg and Åslund (2010) who document a five percentage point higher em-ployment rate after ten years in Sweden among immigrants who complete theirlanguage training. Although one should be careful with interpreting these effectsas causal because of the inherent selection problem, they do confirm that there isan association between language proficiency and labor market outcomes.

3For the full motivation behind the introduction of the bonus system, see theproposal by the Swedish government (Prop. 2009/10:188).

270 CHAPTER 4

rollment rate measured four, eight, and twelve months after arrival toSweden. The second outcome of interest is the pass rate, overall andfor bonus-entitling courses specifically. The pass rate will be measuredas the share with a passing grade, in the courses of interest, withinthe aforementioned time requirements.

The results do not indicate any effect on the enrollment rate mea-sured twelve months after arrival, which may be explained by a highenrollment rate of almost 80 percent already before the reform wasimplemented. Even if the bonus system did not impact the overall en-rollment rate, it could have incentivized immigrants to enroll earlierin order to complete courses in time to qualify for the bonus. How-ever, the absence of an effect on the enrollment rate measured closerto arrival (after four and eight months) suggests that the timing ofenrollment was not affected either.

Turning to the results for the pass rate, there is no evidence ofan impact on the average pass rate for all courses. As the bonus sys-tem implied that completion was only rewarded for certain courses,and not all, the pass rate on bonus-entitling courses is also analyzedseparately. The results are inconclusive. A positive effect of aboutone to four percentage points cannot be excluded, but the results areimprecisely estimated and only significant at the ten percent levelat the broadest possible bandwidth. A definitive conclusion regard-ing the impact on bonus-entitling courses is therefore not possible. Ifwe were to trust the estimates, the magnitude of a few percentagepoints is considerable given the baseline of 17 percent of the controlgroup completing a bonus-entitling course within the stipulated timerequirements.

Before the nationwide implementation of the bonus system in 2010it was introduced small scale in a quasi-experimental setting in 2009.The Institute for Evaluation of Labour Market and Education Pol-icy (IFAU) designed the experiment which was evaluated by Åslundand Engdahl (2018) using a differences-in-differences approach. In line


with this study, they find no impact on average student achievement.When studying bonus-entitling courses specifically, they estimate asignificant positive effect. While this study contributes with addi-tional evidence on the impact of the bonus system, it cannot be con-firmed whether these effects persisted also when the bonus system wasmade permanent in 2010. Advantages with this study is that bonus-eligibility can be identified at the individual level, unlike Åslund andEngdahl (2018) who cannot determine which immigrants were tar-geted by the policy. Further, the permanent introduction rules outany risk of Hawthorne effects and enables an analysis of the effects inmunicipalities that did not volunteer to implement the bonus system.However, the imprecise results prohibits conclusive inference, suggest-ing that further analysis using a different empirical approach mightbe a fruitful continuing of this study.

To think about the potential impact of introducing monetary re-wards in language training, the theoretical framework presented byChiswick (2008) is useful. He models language proficiency as humancapital, and assumes that individuals compare the costs and bene-fits from learning the new language. The costs are associated withthe investment that has to be made to acquire the language skillsand the benefits include better expected outcomes on the local la-bor market, enhanced political empowerment, and an extended socialnetwork. Chiswick suggests that there are three main determinantsof language proficiency: a) exposure to the new language, b) the easewith which it can be acquired, and c) the economic incentives to learnthe language. The design of the bonus system is an attempt to affectc). By rewarding certain behavior, such as good performance in Sfieducation, theory predicts that individuals should be more likely toperform well.

Implicit in the introduction of the bonus system in Sfi educationis the notion that immigrants do not put enough effort into theirlanguage studies. By introducing a financial reward, the incentives

272 CHAPTER 4

to spend time on language studies are strengthened. If immigrantsunderestimate the returns to language skills or neglect potential pos-itive externalities of language proficiency, the reward could improvetheir outcomes. On the other hand, if immigrants can use their timemore productively in other activities, the reward could lead to worseoutcomes. In addition, for the bonus system to work, it must be thecase that an increased effort level by the Sfi students leads to betterachievements. Fryer (2011) suggests two other possible effects of finan-cial rewards in an educational context: a) if the education productionfunction has important complements that the students cannot control,for example good teachers, an increased effort level by the studentsalone might not lead to better achievements, or b) the financial re-ward might undermine the intrinsic motivation of the students andthereby have a negative impact on their academic achievements.

Although teachers in the Swedish language-training programwarned that monetary bonuses may have an effect opposite tothe intended, there is no evidence to support the notion that thebonus had a negative impact on student achievement. Rather, theresults are unable to discriminate between null and positive effects.Åslund and Engdahl (2018) find positive effects, driven by a highercompletion rate specifically on the bonus-entitling courses at thehighest level in the program. At this level, participants are in generalmore highly educated to begin with. This might indicate that onlythose individuals responded to economic incentives or that only theywere able to transform increased effort to a passing grade becausethe time requirements were too restrictive for the less educated.Because of the data-demanding empirical strategy used in thisstudy, estimations on sub samples is not feasible. It is left for futurestudies to analyze the question using other techniques, in order toget a more decisive answer to the question of whether the bonussystem had a positive impact on students’ performances or not.

In addition to the literature on language-skills, this study is also


related to the broader empirical literature on the efficacy of financialrewards in an educational context. There are several studies using anexperimental design to estimate effects of various kinds of financialrewards. Leuven et al. (2010) present evidence from a randomizedexperiment with a design similar to the bonus system introduced inSfi education. University students in the Netherlands could earn afinancial reward if passing all first-year requirements within the firstyear. They find that introducing economic incentives have a positiveeffect on high-ability students but a negative effect on low-abilitystudents. The positive effect dominates, resulting in improved averagestudent performance. Several other studies have also found positiveeffects, see for example Angrist et al. (2009), Angrist and Lavy (2009),Pallais (2009), Kremer et al. (2009), and Dearden et al. (2009), butoften with important heterogeneity between groups.

There are also studies suggesting that the effects are more limited,or finding no effects at all. Fryer (2011) uses a series of experimentsconducted in schools in the U.S., implementing three different incen-tive schemes. The results suggest that there are no significant effectson student achievement. Other studies that find limited effects areAngrist et al. (2014) and Bettinger (2012). Furthermore, evidencefrom a randomized trial in Nepal, by Sharma (2010), suggest that fi-nancial rewards do not undermine the intrinsic motivation of studentsbut they reallocate their effort between subjects to maximize their re-turn. Although this suggests that rewards affect the behavior of thestudents, it might not be a desirable outcome if students "study-for-the-test" rather than accumulate human capital. This relates to theresults of this study. The bonus system rewards participants with amonetary bonus only for some courses, but not others. The estimatedeffects indicate that if anything, the positive effect is limited to thesecourses. However, due to the design of Sfi education, this may beless problematic than in other settings as the course structure is se-quential. In other words, enrolling in one course requires passing the

274 CHAPTER 4

previous course, so reallocation between courses is not possible.The remainder of this chapter is organized as follows. Section 4.2

discuss the institutional setting and the reform, followed by Section4.3 describing the data. The empirical strategy is outlined in Section4.4 while Section 4.5 presents the results. Section 4.6 concludes.


This section describes the Swedish language-training program for im-migrants and the bonus system implemented in 2010.

4.2.1 The Swedish Language-Training Program

Sfi education is part of the public adult-education system in Swe-den and was introduced in the 1960’s when immigration to Swedenincreased sharply (Kennerberg and Sibbmark, 2005). The main goalis to give adult immigrants basic knowledge in the Swedish languageand provide literacy training for those lacking such skills.4 Since 1986,municipalities have been responsible for the provision of Sfi educationbut they can outsource the implementation to other providers. Par-ticipation is free of charge and the program is financed by local taxesor state grants.5

In 2010, the program enrolled 96,136 students. At this point intime, a majority of the students were in the age of 25-39 and almost60 percent were women. There was a lot of heterogeneity in theireducational backgrounds; while a majority had more than ten yearsof schooling, about 15 percent had less than three years of schooling(Swedish National Agency for Education, 2011). 25 percent partici-

4Sfi education is regulated by Skollag 2010:800 and SKOLFS 2009:2. Thissection builds upon the regulations of Sfi education if no other references aregiven.

5State grants are given to compensate the municipalities for expenses relatedto refugees and this generally includes expenses related to language training.


pated in literacy training. The total cost of the program in 2012 wasabout 2 billion SEK (approximately 2.8 million USD) but has sincethen increased sharply by 52 percent (Statistics Sweden, 2017).6

The requirements for immigrants who want to enroll in Sfi educa-tion is that they lack skills in the Swedish language, have turned 16years, and are registered in a Swedish municipality in the populationregister at the Swedish Tax Agency. An exception is immigrants profi-cient in Norwegian or Danish for whom language training in Swedish isnot deemed necessary. Further, Finnish citizens working in a Swedishmunicipality can be entitled to Sfi education even if settled in Finland(close to the Swedish boarder).

Teaching is offered at three different levels, each of which consistsof two courses. The structure is described in Table 4.1. At whichlevel an individual starts is mainly determined by her educationalbackground. The last course of one level has the same content as thefirst course of the next level, but the way of teaching is adjusted to suitthe ability of the students at each level. If one completes, for example,1B and continues to the next level, one will start at 2C (as 2B has thesame content as 1B). Regardless of which level an individual startsat, she can continue to the next level until she reaches the last courseon the last level (3D). The municipalities should offer at least 15lecture hours per week and there is a benchmark of 525 lecture hoursper course.7 Three different grades were used: pass with distinction(VG), pass (G) and fail (IG).8 On the final course of each level, thegrading was based on standardized tests.

6The approximation of the costs in US dollar is made using the average ex-change rate between the US dollar and the Swedish krona during 2010.

7At the time of the implementation of the bonus, there was confusion regardingwhether the recommendation of 525 hours referred to the course, level or totalnumber of hours in Sfi education. Today, there is no such recommendation. In thenext section, the average time spent in each course will be examined in the data.

8A new grading system has been implemented using a scale from A to E,according to SKOLFS 2012:13, but during the time period of interest for thisstudy the old grading system was used in accordance with SKOLFS 2009:2

276 CHAPTER 4

Table 4.1: Course Structure in Sfi Education

Level Courses BonusSfi 1 1A, 1B 6 000 SEK for completing 1BSfi 2 2B, 2C 8 000 SEK for completing 2CSfi 3 3C, 3D 12 000 SEK for completing 3D

Notes: This table presents the courses and bonusescorresponding to each level in Sfi education.

An immigrant who wants to enroll should be offered a place in theprogram within three months. About one third of the participants inSfi education are refugees and they generally enroll in Sfi education aspart of an introductory program for which they receive an allowance(the Swedish Schools Inspectorate, 2010). For these individuals, par-ticipation and achievements are already subject to financial incentivesas the allowance is conditional on their participation in the introduc-tory program. Other immigrants enroll voluntarily in the program.

A number of reports have stressed the shortcomings of the Swedishlanguage-training program, for example the Swedish National AuditOffice (2008), the Swedish Schools Inspectorate (2010, 2011), and theSwedish Agency for Public Management (2009). The latter reportpresents that one third of all Sfi students do not pass any coursewithin three years from enrollment. Among the other two thirds, onlyhalf of the Sfi students completes the highest level within three years.Other conclusions from these evaluations is that the municipalitieshave trouble finding good teachers, that they find it hard to adjustthe teaching to the different backgrounds and abilities of the studentsand that many municipalities lack routines for evaluating the resultsin Sfi education. The high drop-out rate is discussed as a potentialproblem by both the Swedish Agency for Public Management and theSwedish National Audit Office. The latter also highlights the lack ofrigorous studies of the results in Sfi education and the methodologicalproblems that has prevented the undertaking of such studies.


4.2.2 The Bonus System

The bonus system was first introduced in a quasi-experimental wayin 2009. Out of 35 municipalities that expressed an interest in partici-pating, 13 pairs of municipalities were selected based on their compa-rability with each other.9 Within pairs, municipalities were randomlyassigned to either the control or the treatment group. The quasi-experimental introduction of the bonus system lasted from July 1,2009 to June 30, 2010. The evaluation of the experiment made byIFAU was published in December 2012, but at this point the bonussystem was already made permanent and implemented nationwide.This decision was taken in May 2010 and the new law came into ef-fect on August 1, 2010, making all immigrants that arrived from July1, 2010 and onward eligible for the reward, subject to meeting therequirements for the bonus.10 As the law came into effect in August,no bonuses were paid out before this month. Figure 4.1 displays thetime-line of the implementation.

The bonus system is based on the students’ performances in spe-cific courses. To qualify for a bonus, the individual has to:

– be registered in a Swedish municipality no earlier than July 1,2010,

– have a residence permit in accordance with the Aliens Act,Chapter 5, §§1, 2, 3, 3a, 4, 5 or 6, Chapter 12, §18, Chapter21, or Chapter 22,

– be of age 18 to 64 at the time when applying for the bonus, and9In total, the following 28 municipalities participated: Stockholm, Göteborg,

Uppsala, Södertälje, Huddinge, Haninge, Borås, Jönköping, Växjö, Kalmar, Sand-viken, Gävle, Nacka, Täby, Sollentuna, Solna, Halmstad, Helsingborg, Karlstad,Västerås, Trelleborg, Landskrona, Härnösand, Örnsköldsvik, Uddevalla, Trollhät-tan, Katrineholm and Nyköping. 13 pairs adds up to 28 municipalities as one halfof a pair consists of three municipalities (Gothenburg, Södertälje and Uppsala) tomatch the largest municipality (Stockholm).

10When time of immigration or time of arrival are used in this study it refers tothe date of registration in a Swedish municipality as this is the date that determinesbonus eligibility.

278 CHAPTER 4

Figure 4.1: Time-line of the Implementation of the Bonus system

Mar

2010

Apr

2010

May

2010

Jun

2010

Jul2010

Aug

2010

Sep

2010

Oct

2010

Nov

2010

Dec

isio

nta

ken

toim

ple

men

tth

e

bon

us

syst

emna-

tion

wid

e

Law

com

esin

to

effec

t

Subje

ctto

the

bon

us

syst

em

– receive a passing grade in one of the bonus-qualifying courses(1B, 2C, 3D) within 15 months from immigration and withintwelve months from starting Sfi education.

To receive the bonus, an individual has to fulfill all the requirementsand then send an application to the municipality within three monthsfrom receiving the passing grade. The amount received in bonus de-pends on which course were completed. 1B gives 6,000 SEK, 2C gives8,000 SEK, and 3D gives 12,000 SEK (see Table 4.1).11 An individ-ual can receive more than one bonus if she completes more than onebonus-entitling course, but the maximum amount that can be paidout to one individual is 12,000 SEK. The municipalities get reim-bursed from the state for their expenses related to the bonus system.More details can be found in the law that regulates the bonus system(SFS 2010:538). The requirements on the type of residence permitmeans that it is mainly refugees and family reunification immigrantsthat are eligible for the bonus system.

Up until December 31, 2011, there were 1,005 bonuses granted(Swedish National Agency for Education, 2012).12 62 percent of thesewere granted for passing grades on the highest level (3D) and only 11

11This corresponds to approximately $833, $1249 and $1666 using the averageexchange rate between the US dollar and the Swedish krona during 2010.

12These numbers only include bonuses paid out to individuals in the municipal-ities that participated in the experiment and who arrived to these municipalitiesduring the time period for the experiment. This means that individuals who re-

4.3. DATA 279

percent were due to passing grades on the first level (1B). Accordingto Åslund and Engdahl (2018), the average number of years in gen-eral education for students on the first level in Sfi education during2009/2010 was 4.5 years. For students on the third level, the corre-sponding number was 14 years. This suggests that it was mainly well-educated immigrants who received bonuses for their performances inSfi education. This is not the same as saying that the bonus systemonly affected well-educated individuals. It could be the case that thesewell-educated individuals would have performed as well in absence ofthe bonus system. Moreover, the achievements of low-educated indi-viduals could be affected even if they do not reach the requirementsfor receiving the reward.

4.3 Data

All data used in this study was delivered by Statistics Sweden. Threedifferent registers are used: the Longitudinal Database for IntegrationStudies (STATIV), the Historical Population Register, and the Sfiregister. STATIV is an annual data set consisting of all individuals inSweden.13 It contains demographic information such as year of birth,sex, marital status, and municipality of residence as well as the highestlevel of education achieved. For immigrants there is also informationon the country of origin and the ground for residence permit as well ashours in Sfi education and any bonuses paid out from the Sfi program.

The Sfi register adds more information about Sfi participationincluding the level and course taken as well as the start and end dateof each course. For those not completing courses, there is an indicatorof the reason for dropping out, while the grade is observed for thosecompleting the course. The Sfi register also records the hours spent in

ceived and qualified for a bonus in accordance with the regulation of the permanentbonus system introduced in 2010 are not included.

13Formally, it consists of all individuals who are registered in a Swedish munic-ipality (folkbokförd).

280 CHAPTER 4

education per course for each individual. Lastly, from the HistoricalPopulation Register, the date of immigration to Sweden is observedfor all immigrants living in Sweden during any of the years 2008-2011.

The population of interest is defined as individuals, aged 18-64,that immigrated to Sweden during 2010 with a bonus-entitling res-idence permit.14 Immigrants from Norway, Denmark, Finland, andIceland are excluded.15 Further, immigrants who settled in a munici-pality that participated in the experimental introduction of the bonussystem in 2009 are excluded.16 Given the design of the bonus system,student performance during the first 15 months after arrival to Swe-den is of interest. As the data ends in December 2011, this impliesthat the estimation sample has to be restricted to immigrants who ar-rived before October 2010. A symmetrical window around the reformdate therefore implies that the estimation sample will be restricted toimmigrants arriving to Sweden during the time period of March 31,2010 to September 30, 2010 (a bandwidth of 92 days). 93 individualswho fulfill these requirements are dropped due to inconsistencies inthe data.17 This results in an estimation sample of 6,560 individuals,

14It should be mentioned that information about the type of residence permitwas only available on a grouped level. To determine bonus eligibility at the indi-vidual level, a list of the bonus-entitling residence permits was given to StatisticsSweden. Based on this list, Statistics Sweden, with access to more detailed in-formation, generated a dummy variable for immigrants holding a bonus-entitlingresidence permit which is used in this paper.

15As mentioned previously, immigrants proficient in Norwegian or Danish aregenerally not entitled to Sfi education and Finnish citizens are sometimes entitledto Sfi education even if not settled in Sweden. As, in the data, immigrants fromNorway, Denmark, Finland, and Iceland are grouped together in one region, theseobservations are excluded.

16In these municipalities, all individuals who fulfill the requirements and immi-grated during 2010 are eligible to the bonus system either under the experimentalimplementation which lasted until June 30, 2010, or the permanent implementa-tion which begun at July 1, 2010.

17According to the data, these individuals enroll in Sfi education before beingregistered in a Swedish municipality. This should not be possible, suggesting thateither of these two dates are mistaken. Since both date variables are important forthe analysis, these individuals are dropped.

4.3. DATA 281

divided into a control and treatment group based on whether theyarrived prior to or post the reform date of July 1, 2010.

Table 4.2 presents summary statistics of predetermined variablesfor individuals in the estimation sample. All information on predeter-mined characteristics are from STATIV. The average immigrant is 32years old and women are in majority. Many have a partner in Swe-den, which is not surprising given that about 60 percent of the sampleare granted residency for family reunification. The vast majority ofthe remaining immigrants are people seeking protection in Sweden.About half of the sample are settled in a metropolitan municipalityand the largest groups immigrated from Asia and Africa. The educa-tional level is heterogeneous (and unobserved for a large share), with24 percent having at most a primary education and about equallymany having a higher education (defined as university-level educa-tion). This lines up well with the characteristics of the full body ofstudents in Sfi education, as reported by the Swedish National Agencyfor Education (2011).

282 CHAPTER 4

Table 4.2: Summary Statistics of Predetermined Char-acteristics for the Estimation Sample

Mean SD Min Max

Demographic characteristicsAge 31.62 9.67 18.00 64.00Share female 0.58 0.49 0.00 1.00Share with partner 0.63 0.48 0.00 1.00Metropolitan municipality 0.47 0.50 0.00 1.00

Type of residence permitFamily reunification 0.61 0.49 0.00 1.00Refugees 0.11 0.31 0.00 1.00Humanitarian reasons 0.02 0.13 0.00 1.00In need of protection 0.27 0.44 0.00 1.00Other 0.00 0.05 0.00 1.00

Region of birthEU27, N. America, Oceania 0.07 0.26 0.00 1.00EUR excl. EU27 0.13 0.33 0.00 1.00Africa 0.36 0.48 0.00 1.00S. America 0.03 0.17 0.00 1.00Asia 0.41 0.49 0.00 1.00

Level of educationPrimary 0.24 0.43 0.00 1.00Secondary 0.15 0.36 0.00 1.00Higher 0.25 0.44 0.00 1.00Unknown 0.35 0.48 0.00 1.00

Observations 6560Notes: This table presents the mean, standard deviation, minimum, andmaximum of each variable. The population is restricted to immigrantsof age 18-64 with a bonus-entitling residence permit who settled in amunicipality that did not participate in the experimental introduction ofthe bonus system and were registered in Sweden between March 31, 2010and September 30, 2010. Immigrants from Norway, Denmark, Finland,and Iceland are excluded.

The outcome variables are constructed based on information fromthe Sfi register and the Historical Population Register. The first out-come of interest is the enrollment rate. Figure 4.2 shows the shareof immigrants enrolling in Sfi education, separately for the treatmentand control group. One can see that a clear majority of both the treat-ment and the control group enrolls in Sfi education at some point in

4.3. DATA 283

time. The enrollment rate is slightly lower for individuals in the treat-ment group, but enrollment is only observed up until December 2011.As individuals in the treatment group by definition arrived to Swe-den later than those in the control group, they had less time to enrollbefore this point in time.

Figure 4.2: Enrollment in Sfi Education, by Treatment Status

0

.2

.4

.6

.8

Enro

llmen

t in

Sfi e

duca

tion

Control group Treatment group

Notes: This figure shows the share of immigrants enrolling in Sfi education, by treatment status. Thepopulation is restricted to immigrants of age 18-64 with a bonus-entitling residence permit whosettled in a municipality that did not participate in the experimental introduction of the bonussystem and were registered in Sweden between March 31, 2010 and September 30, 2010. Immigrantsfrom Norway, Denmark, Finland, and Iceland are excluded. Enrollment in Sfi education is a dummyvariable equal to one if the individual enrolled in Sfi education at any point in time up untilDecember 2011.

Rather than measuring the enrollment rate at a specific date, theoutcome variables will be observed m months after arrival to Sweden.More specifically, they are defined as:

Yi,m ={

1 if enrolled within m months from immigration0 otherwise,

where m ∈ 4, 8, 12. Enrollment is measured at three different pointsin time as the reform might affect the enrollment rate as well asthe timing of enrollment. The earliest measure is four months after

284 CHAPTER 4

enrollment in order to avoid that the results are influenced by theseasonal pattern in Sfi course starts. Figure 4.7 in Appendix 4.A showsthe seasonality pattern, and a spike in enrollment is clearly visible inAugust. Still, individuals are guaranteed a seat within three monthsfrom stating their interest in enrolling in Sfi education. After fourmonths, those interested in enrolling early should therefore have beenable to start their language-training.

The other set of outcomes studied is related to the students’ per-formances. This is also measured by a dummy variable defined as:

Yi,c ={

1 for passing grade in course c in time0 otherwise

where course c is defined as any course (1A-3D) or a bonus-entitlingcourse (1B, 2C, 3D). A passing grade in time refers to the time re-quirements which specify that the passing grade should be receivedwithin 15 months from immigration and twelve months from enroll-ment in order to qualify for the bonus. As a baseline, for comparison,Table 4.5 in Appendix 4.A presents summary statistics for the out-come variables for individuals in the control group (i.e., those in theestimation sample arriving to Sweden prior to July 1, 2010).

To further understand what the bonus system implied, Figure 4.3displays histograms of the number of days spent in language trainingbefore receiving a passing grade, by course, for individuals who werenot subject to the bonus system and eventually received a passinggrade. On average, students spend about 100-150 hours per course,implying that completing one level would take about 200-300 days.A back-of-the-envelope calculation suggest that the average student,without adjusting her effort level, should be able to pass the bonus-requirement for the final course at the entry level given no delaybetween completing the first course and starting the bonus-course.However, this assumes that the individual enrolls immediately afterarrival to Sweden, but in fact, the average number of days until en-

4.3. DATA 285

rollment in Sfi education is 145 days. This would still mean that thebonus course could be completed in about 15 months from arrival,and the twelve month requirement is measured from enrollment inSfi education. However, this makes a strong assumption about nowaiting times, is based on the performances of students that even-tually receives a passing grade and ignores the heterogeneity acrosscourses that is clearly visible in Figure 4.3. Still, it gives some intu-ition regarding the difficulty of meeting the requirements for receivinga bonus.

Figure 4.3: Course Duration for Individuals in the Control Group

(a) Course 1A

Mean = 139.46

0

.002

.004

.006

Den

sity

0 20 40 60 80 100

120

140

160

180

200

220

240

260

280

300

320

340

360

Duration in days, course 1A

(b) Course 2B

Mean = 152.9

0

.002

.004

.006

.008

Den

sity

0 20 40 60 80 100

120

140

160

180

200

220

240

260

280

300

320

340

360

Duration in days, course 2B

(c) Course 3C

Mean = 141.99

0

.002

.004

.006

Den

sity

0 20 40 60 80 100

120

140

160

180

200

220

240

260

280

300

320

340

360

Duration in days, course 3C

(d) Course 1B

Mean = 131.77

0

.002

.004

.006

.008

Den

sity

0 20 40 60 80 100

120

140

160

180

200

220

240

260

280

300

320

340

Duration in days, course 1B

(e) Course 2C

Mean = 103.05

0

.002

.004

.006

.008

.01

Den

sity

0 20 40 60 80 100

120

140

160

180

200

220

240

260

280

300

320

340

Duration in days, course 2C

(f) Course 3D

Mean = 95.17

0

.002

.004

.006

.008

Den

sity

0 20 40 60 80 100

120

140

160

180

200

220

240

260

280

300

320

340

Duration in days, course 3D

Notes: This figure presents histograms of the number of days before receiving a passing grade in eachcourse. The population is restricted to immigrants of age 18-64 with a bonus-entitling residencepermit who settled in a municipality that did not participate in the experimental introduction of thebonus system and were registered in Sweden between March 31, 2010 and June 30, 2010. Immigrantsfrom Norway, Denmark, Finland, and Iceland are excluded.

286 CHAPTER 4


The main empirical challenge in policy evaluation is how to establishthe causal effect of a certain policy. Ideally, one would like to ran-domize treatment among individuals and compare the outcomes ofthe treated and non-treated, which would give an unbiased estimateof the causal effect. When treatment is not randomized, the worry isthat there might be systematic differences between the treated andnon-treated. In this context, the concern is that individuals who ar-rived to Sweden after July 1, 2010, might be different compared tothose who arrived before that date. This would invalidate a simplecomparison-of-means strategy.

In this study, a sharp regression discontinuity design is used tohandle this problem. The bonus system implemented in the Swedishlanguage-training program was not designed to randomize treatment,but the mechanism that determines who is eligible to receive a bonusis known. It is completely determined by the date of registration in aSwedish municipality, with everyone in the targeted population reg-istered from July 1, 2010 and onward being entitled to a bonus uponreceiving a passing grade within the required time. The regressiondiscontinuity framework exploits the knowledge of this assignmentmechanism and estimates the difference in outcomes between thosearriving just before this date and those arriving just after.

Let xi denote the date of registration (the assignment variable)and x0 the date of implementation of the bonus system, July 1, 2010,(the cutoff point). Individuals registering on July 1, 2010, or later areeligible whereas those who registered prior to that date are not. Thisimplies that the treatment status, defined as a dummy variable (Ti),changes discontinuously from 0 to 1 when xi reaches the cutoff pointx0. Hence, Ti is a deterministic and discontinuous function of xi:

Ti ={

1 if xi ≥ x00 if xi < x0.


The regression discontinuity approach is distinguished from otherevaluation methods as there is no value of xi for which both treatedand untreated observations exist. This implies that the regression dis-continuity approach relies on the willingness to extrapolate across co-variate values, at least in the neighborhood of x0. The basic regressionspecification is:

Yi = α+ βTi + f(xi) + εi, (4.1)

where Yi is the outcome of interest and β is the coefficient of interest,capturing the effect of treatment. The assignment variable is includedto distinguish the discontinuous function Ti = 1[xi ≥ x0] from theunderlying smooth relationship between xi and Yi. f(xi) specifies thefunctional form of this relationship.

If the relationship is linear, f(xi) can be replaced by γxi to obtainunbiased estimates. Using that specification when the relationshipis non-linear could result in the unaccounted-for non-linearity beingmistaken for a discontinuity. The estimates would then be biased.There are two ways to approach a non-linear relationship. The firstis to use local linear regression which only uses observations in anarrow bandwidth around the cutoff point. If the bandwidth is setsmall enough, the relationship should be approximately linear. Theother approach is to use a broader bandwidth and correctly modelthe relationship between xi and Yi.

The choice between a narrow and broad bandwidth is discussed inImbens and Lemieux (2008) as well as Lee and Lemieux (2010). Thenumber of observations are seldom enough to set the bandwidth smallenough to obtain entirely unbiased estimates (unless the relationshipis precisely linear). Extending the bandwidth will also impose somebias unless the specification of f(xi) is a very good approximationof the true underlying relationship. Without knowledge about thisrelationship it is not possible to determine which approach minimizes

288 CHAPTER 4

the bias. Generally, using a broad bandwidth increases the efficiencyof the estimations while a more narrow bandwidth is less likely tobe biased. The data available impose a restriction on the bandwidth,implying that it cannot be broader than 92 days, giving an estimationsample of about 6,000 observations. The main results will be presentedfor bandwidths of 46, 69 and 92 days as well as using a data-drivenapproach for optimal bandwidth selection following Calonico et al.(2014). To test the robustness of the results, Appendix 4.D will showhow the coefficients respond to varying the bandwidth two days atthe time, from 10 to 92 days.

The exact specification of the regression equation behind the re-sults presented in Section 4.5 is:

Yi = α+ βTi + f(xi) + δ(Ti · xi) + σ′Zi + εi. (4.2)

The assignment variable is normalized, so that xi = xi−x0, to ensurethat β can be directly interpreted as the local average treatment effect.The interaction between treatment and the assignment variable (Ti ·xi) is included to allow for different trends on each side of the cutoffpoint. The preferred specification uses a polynomial of order one,but all tables include a specification where f(xi) is specified as apolynomial of order two. The results are also presented both with andwithout the inclusion of Zi, a vector of covariates, as the inclusion ofZi could increase the efficiency of the estimator.18

The effects of the reform are identified by estimating the magni-tude of the discontinuity in the outcome variable at the cutoff point.The identifying assumption behind this design is that of continuity.If individuals are as good as randomly assigned to treatment near thecutoff, the running variable as well as predetermined characteristics

18The results are presented without clustered standard errors. While clusteringhas been the standard approach, this study follows Kolesár and Rothe (2016) andreports only heteroskedasticity robust errors. All estimations have been repeatedwith clustering on the running variable and this will be commented on in Section4.5.


should be continuous at this point. The relevant question to ask iswhether individuals were able to decide whether to be registered be-fore or after July 1, 2010, in order to impact their eligibility to thebonus system. If so, the identifying assumption would be violated. Inabsence of manipulation, the variation of treatment around the cutoffpoint is random.19 In other words, the probability of being treated isthe same for individuals near the cutoff point on each side.

It is highly unlikely that individuals were able to perfectly manip-ulate the date of registration in a Swedish municipality. Furthermore,it is questionable whether the bonus system provides strong enoughincentives for immigrants to postpone their date of registration, asbeing registered has several advantages (including benefit eligibility)that are likely to be more important than the prospect of receivinga bonus from the Sfi program. Figure 4.4 shows the number of im-migrants by date of registration. In the presence of manipulation, anunusually high registration load would be expected just after the cut-off point. There is nothing in Figure 4.4 to suggest that manipulationoccurred.

A more formal way of testing for manipulation of the assignmentvariable, suggested by McCrary (2008), is to look at the change inthe density of the distribution of the assignment variable at the cut-off point. A discontinuity in the density at the cutoff point wouldindicate the presence of manipulation. In this study, such a test is im-plemented following Calonico et al. (2018), testing the null hypothesisof continuity in the running variable at the cutoff point.20 This testyields a p-value of 0.152, suggesting that the null hypothesis cannotbe rejected at any conventional significance level. A graphical repre-

19According to Lee and Lemieux (2010), this is one of the features that dis-tinguishes the regression discontinuity design from other evaluation methods inwhich the randomization is an assumption that has to be made. When the regres-sion discontinuity design is implemented correctly, randomization of treatment inthe neighborhood of x0 is a consequence.

20Available in Stata as rddensity.

290 CHAPTER 4

Figure 4.4: Number of Immigrants by Date of Registration, 2010

0

20

40

60

80

100

Freq

uenc

y

-180 -150 -120 -90 -60 -30 0 30 60 90 120 150 180Normalized date of registration

Notes: This figure shows the number of immigrants registering in a Swedish municipality by date ofregistration. The population is restricted to immigrants of age 18-64 with a bonus-entitling residencepermit who settled in a municipality that did not participate in the experimental introduction of thebonus system and were registered in Sweden during 2010. Immigrants from Norway, Denmark,Finland, and Iceland are excluded. The date variable is normalized relative to the cutoff point, whichis marked by the vertical red line.

sentation of the McCrary (2008) test is provided in Appendix 4.B,Figure 4.8.

Returning to the inclusion of Zi, the continuity assumption shouldhold also for predetermined characteristics (Lee and Lemieux, 2010).Discontinuities in covariates of predetermined characteristics are anindication that treatment cannot be considered random in the neigh-borhood of x0 and that the validity of the regression discontinuityapproach is threatened. For example, if immigrants registering in amunicipality in July 2010 are more well-educated than those register-ing in June 2010, one would suspect that well-educated individualswere able to delay their registration in a municipality to achieve el-igibility to the bonus system. If including Zi in the regression equa-tion changes the point estimates, there is reason to worry about thecontinuity-assumption being violated. To further inspect the conti-nuity of predetermined characteristics, Figure 4.9-4.11 in Appendix

4.5. RESULTS 291

4.C show the average outcome of the covariates for individuals ineach 5-day bin using a bandwidth of 92 days, with the cutoff pointmarked by the vertical line. The sample appears well balanced aroundthe cutoff point. Taken together, there is no evidence suggesting thatmanipulation occurred.

4.5 Results

This section presents and discuss the results from the regression dis-continuity analysis. The figures presented here display data for thefull bandwidth of 92 days with a regression using a linear fit and nocontrol function. In the regression tables, coefficients are presentedfor bandwidths of 46, 69 and 92 days as well as using optimal band-width selection. Furthermore, results are presented with and withoutthe inclusion of the control function and with a linear and quadraticpolynomial.

4.5.1 Enrollment

The financial incentives provided by the bonus system could affect theshare of immigrants enrolling in Sfi education. Table 4.3 shows theeffect on the enrollment rate measured four, eight and twelve monthsafter immigration. The coefficients are in general small and through-out all specifications insignificant. The absence of an effect also seemsplausible, looking at the graphical evidence presented in Figure 4.5.The results are robust to varying the bandwidth (see Figure 4.12 inAppendix 4.D). There is thus no evidence of the reform increasingthe enrollment rate during the first year after arrival to Sweden. Itshould be noted that the baseline is relatively high, with almost 80percent of individuals in the control group enrolling in Sfi educationwithin twelve months from arrival to Sweden. Still, results in Table4.3 also suggest that the reform did not induce individuals to enroll

292 CHAPTER 4

in Sfi education earlier, as there is no impact on enrollment measuredearlier either (after four and eight months).

Figure 4.5: Effects on the Enrollment Rate

(a) Within 4 months

0

.1

.2

.3

.4

.5

.6

.7

.8

.9

1

Enro

llmen

t rat

e


(b) Within 8 months

0

.1

.2

.3

.4

.5

.6

.7

.8

.9

1

Enro

llmen

t rat

e


(c) Within 12 months

0

.1

.2

.3

.4

.5

.6

.7

.8

.9

1

Enro

llmen

t rat

e


Notes: This figure present the average enrollment rate in each 5-day bin, represented by the blue dots.The blue lines are the fitted values from the regression with no control function and a polynomial oforder one. The population is restricted to immigrants of age 18-64 with a bonus-entitling residencepermit who settled in a municipality that did not participate in the experimental introduction of thebonus system. The bandwidth is 92 days, implying that all individuals immigrated to Sweden duringthe time period March 31, 2010 to September 30, 2010. The outcome variables is defined as a dummyequal to one if the individual had enrolled in Sfi education (any course) within four months (panel a),eight months (panel b), or twelve months (panel c) respectively.

4.5.2 Pass Rate

The main goal of the reform was to increase the share of studentscompleting their coursework in Sfi education within a reasonable time.

4.5. RESULTS 293

The prospect of receiving a monetary reward increases incentives topass the courses within the time requirements of twelve months ofenrolling and 15 months from immigration. Table 4.4 presents theestimated effects for all courses and the bonus-entitling courses sepa-rately. Panel A suggest that the reform did not impact the overall passrate, but when looking at the bonus-entitling courses specifically, theresults are less conclusive. The coefficients are positive throughout allspecifications, and significant at the ten percent level for the broadestbandwidth using a linear specification.21 The magnitude varies be-tween zero and four percentage points, depending on the bandwidth,with higher coefficients for broader bandwidths. With a baseline of17 percent, this effect is potentially large, but the imprecise estimatesprohibit a clear interpretation. To further examine whether the re-sults seem plausible, Figure 4.6 shows the graphical evidence. Thereis no visible discontinuity in panel A showing the impact on all coursestogether, but if anything a positive discontinuity in terms of bonus-entitling courses.

21When standard errors are clustered at the running variable, the linear speci-fication including control variables is no longer significant.

294 CHAPTER 4

Figure 4.6: Effects on the Pass Rate

(a) Any course

0

.1

.2

.3

.4

.5

Pass

rate


(b) Bonus-entitling course

0

.1

.2

.3

.4

.5

Pass

rate


Notes: This figure present the average pass rate in each 5-day bin, represented by the blue dots. Theblue lines are the fitted values from the regression with no control function and a polynomial of orderone. The population is restricted to immigrants of age 18-64 with a bonus-entitling residence permitwho settled in a municipality that did not participate in the experimental introduction of the bonussystem. The bandwidth is 92 days, implying that all individuals immigrated to Sweden during thetime period March 31, 2010 to September 30, 2010. The outcome variables is defined as a dummyequal to one if the individuals had passed any course (panel a) or a bonus-entitling course (panel b)within 15 months from immigration and twelve months from course start.

Figure 4.13 in Appendix 4.D shows how the coefficient changes asthe bandwidth is increased by two days at the time, from 10 days tothe full bandwidth of 92 days. The figure show that the estimated co-efficients are positive for the full range of bandwidths, close to zero atnarrow bandwidths but around 4-5 percentage points when the band-widths exceed about 30 days. With broader bandwidths, the numberof observations increases and the coefficients are borderline signifi-cant at the broader bandwidths. A positive impact on bonus-entitlingcourses would be in line with the results presented in Åslund and En-gdahl (2018). They estimate a smaller impact, about two percentagepoints, however, the results are not directly comparable. Their esti-mation sample include all immigrants, not just those eligible for thebonus, implying that not all individuals in their sample had incentivesto improve their performance. If the positive effect persisted when thebonus system was introduced nationwide, a larger impact would beexpected in this study as the effects are estimated only on individu-

4.5. RESULTS 295

als targeted by the bonus system (assuming that potential spill-overeffects were smaller than the direct effect).

Still, the evidence in favor of a positive effect is weak, to say theleast. While increasing the bandwidth implies more power in the esti-mations, it also increases the risk of bias. The main worry in this studywould be that the broader bandwidth implies that the control grouparrives prior to or during the beginning of the summer, and thereforewould have to wait longer before they could enroll in Sfi education.This could very well impact whether they were able to complete theircourses within the time requirements for receiving the bonus. Oneargument against this is the absence of effects on the enrollment rateand the initial courses that has to be completed in order to continueto the bonus-entitling courses. Nevertheless, the results are inconclu-sive and further analysis is needed in order to understand whether thepositive effect shown in Åslund and Engdahl (2018) persisted duringthe nationwide introduction of the bonus system.

296 CHAPTER 4

Tab

le4.

3:Eff

ects

ontheEn

rollm

entRa

te

Pan

elA

.Optim

albw

(MSE

)Ban

dwidth

=46

Ban

dwidth

=69

Ban

dwidth

=92

(1a)

(1b)

(1c)

(2a)

(2b)

(2c)

(3a)

(3b)

(3c)

(4a)

(4b)

(4c)

Enrollin

4m-0.030

-0.025

-0.054

0.000

0.006

-0.028

0.007

0.015

-0.018

-0.041

-0.033

0.025

(0.056)

(0.055)

(0.066)

(0.034)

(0.034)

(0.051)

(0.027)

(0.027)

(0.042)

(0.024)

(0.024)

(0.036)

Linear

√√

√√

√√

√√

√√

√√

Qua

dratic

√√

√√

Con

trols

√√

√√

Ban

dwidth

1717

2746

4646

6969

6992

9292

N1350

1350

2166

3535

3535

3535

5148

5148

5148

6560

6560

6560

Pan

elB

.

Enrollin

8m-0.022

-0.013

-0.070

-0.018

-0.010

-0.010

-0.011

0.001

-0.022

-0.030

-0.016

-0.007

(0.048)

(0.046)

(0.055)

(0.030)

(0.029)

(0.044)

(0.024)

(0.024)

(0.036)

(0.021)

(0.021)

(0.032)

Linear

√√

√√

√√

√√

√√

√√

Qua

dratic

√√

√√

Con

trols

√√

√√

Ban

dwidth

1617

2746

4646

6969

6992

9292

N1270

1350

2166

3535

3535

3535

5148

5148

5148

6560

6560

6560

Pan

elC

.

Enrollin

12m

0.005

0.013

-0.011

-0.018

-0.010

0.011

-0.017

-0.004

-0.009

-0.038

-0.025

-0.002

(0.044)

(0.043)

(0.051)

(0.028)

(0.027)

(0.041)

(0.023)

(0.023)

(0.034)

(0.020)

(0.020)

(0.030)

Linear

√√

√√

√√

√√

√√

√√

Qua

dratic

√√

√√

Con

trols

√√

√√

Ban

dwidth

1717

2846

4646

6969

6992

9292

N1350

1350

2269

3535

3535

3535

5148

5148

5148

6560

6560

6560

Not


theestimated

regression

discon

tinu

ity

coeffi

cients.Heteroskeda

sticity-robu

ststan

dard

errors

inpa

rentheses.

All

regression

sinclud

eapolyn

omialof

firstor

second

degree.The

controlfunction

includ

esindicators

forage,

gend

er,pa

rtner,

settled

ina

metropolitan

mun

icipality,

region

ofbirth,

typeof

residencepermit,an

dlevelof

education.

The

pop

ulationis

restricted

toim

migrantsof

age

18-64withabon

us-entitling

residencepermit

who

settledin

amun

icipalitythat

didno

tpa

rticipatein

theexperim

entalintrod

uction

ofthe

bon

ussystem

.Im

migrantsfrom

Norway,Denmark,

Finland

,an

dIcelan

dareexclud

ed.The

outcom

evariab

leis

anindicatorforwhether

the

immigrant

hadenrolled

inSfi

education(any

course)withfour

mon

ths(pan

elA),

eigh

tmon

ths(pan

elB)or

twelve

mon

ths(pan

elC).

*p<

0.1,

**p<

0.05,***p<

0.01.

4.5. RESULTS 297T

able

4.4:

Effects

onthePa

ssRa

te

Pan

elA

.Optim

albw

(MSE

)Ban

dwidth

=46

Ban

dwidth

=69

Ban

dwidth

=92

(1a)

(1b)

(1c)

(2a)

(2b)

(2c)

(3a)

(3b)

(3c)

(4a)

(4b)

(4c)

Any

course

-0.044

-0.052

-0.027

-0.021

-0.015

-0.010

0.007

0.015

-0.015

0.00

40.010

-0.003

(0.052)

(0.054)

(0.060)

(0.034)

(0.034)

(0.051)

(0.027)

(0.027)

(0.042)

(0.024)

(0.024)

(0.036)

Linear

√√

√√

√√

√√

√√

√√

Qua

dratic

√√

√√

Con

trols

√√

√√

Ban

dwidth

1918

3246

4646

6969

6992

9292

N1477

1412

2528

3535

3535

3535

5148

5148

5148

6560

6560

6560

Pan

elB

.

Bon

us-cou

rse

0.007

0.003

0.013

0.030

0.028

0.010

0.041

0.037

0.031

0.045∗

0.041∗

0.029

(0.042)

(0.042)

(0.044)

(0.027)

(0.026)

(0.040)

(0.021)

(0.021)

(0.032)

(0.019)

(0.019)

(0.028)

Linear

√√

√√

√√

√√

√√

√√

Qua

dratic

√√

√√

Con

trols

√√

√√

Ban

dwidth

1818

3646

4646

6969

6992

9292

N1412

1412

2876

3535

3535

3535

5148

5148

5148

6560

6560

6560

Not


theestimated

regression

discon

tinu

itycoeffi

cients.Heteroskeda

sticity-robu

ststan

dard

errors

inpa

rentheses.

All

regression

sinclud

eapolyn

omialof

firstor

second

degree.The

controlfunction

includ

esindicators

forage,

gend

er,pa

rtner,

settled

ina

metropolitan

mun

icipality,

region

ofbirth,

typeof

residencepermit,an

dlevelof

education.

The

pop

ulationis

restricted

toim

migrantsof

age18-64withabon

us-entitling

residencepermit

who

settledin

amun

icipalitythat

didno

tpa

rticipatein

theexperim

entalintrod

uction

ofthebon

ussystem

.Im

migrantsfrom

Norway,Denmark,

Finland

,an

dIcelan

dareexclud

ed.The

outcom

evariab

leis

anindicatorforwhether

theim

migrant

hadpa

ssed

any(pan

elA)or

abon

us-entitling

(pan

elB)course

withintw

elve

mon

thsfrom

course

startan

d15

mon

thsfrom

immigration

.*p<

0.1,

**p<

0.05,***p<

0.01.

298 CHAPTER 4


This study evaluates the introduction of performance-based bonusesin the Swedish language-training program. The bonus system was in-troduced nationwide in Sweden in 2010, targeting adult immigrantswho were mainly approved residency in Sweden as refugees or dueto family reunification. With high levels of immigration and lan-guage proficiency being a key determinant of successful integration,understanding how to improve newcomers’ performance in language-training should be of high interest to policy makers.

To qualify for a bonus of up to 12,000 SEK, one had to pass abonus-entitling course within a year from enrollment in Sfi educationand within 15 months from arrival to Sweden. Using a regression dis-continuity approach, the impact on enrollment and course completionis studied. The estimated effects suggest no impact on the enrollmentrate, measured four, eight, and twelve months after arrival to Sweden.

Similarly, no impact on the overall pass rate is found. The resultson the pass rate on bonus-entitling courses is, however, inconclusive.The estimated coefficients are positive throughout all specifications,preventing the conclusion that the bonus system had no impact. Still,the magnitude of the coefficients is sensitive to the regression speci-fication and only significant at the ten percent level when using thebroadest possible bandwidth. The evidence is thus too weak to confi-dently conclude that there was a positive impact. Increasing the band-width increases the risk of introducing bias, where the main worry isthat the seasonality in Sfi course starts would impact the results.However, the null effects on the enrollment rate and overall pass ratespeaks against this as an explanation for why the coefficients are pos-itive when looking at the pass rate of bonus-entitling courses. All inall, this study is unable to make a confident statement regarding theimpact on bonus-entitling courses.

If the results were to be trusted, a positive effect of a few per-centage points is substantial, given the baseline pass rate on bonus-


entitling courses of about 17 percent in the control group. Such effectswould also be consistent with basic economic theory, predicting thatperformance would be improved on the courses where a passing gradeis rewarded, and with the results presented in the evaluation by Ås-lund and Engdahl (2018) of the quasi-experimental introduction ofthe bonus-system. While the gains from language proficiency seemlarge, an unanswered question is whether immigrants substitute timeaway from more productive activities in order to focus on learningthe Swedish language. To understand the full impact of this reform,more research is needed, focusing not only on the outcomes withinSfi education but also on activities that might be given lower priorityduring the time in language-training and activities where the gainsfrom increased language proficiency might be realized.

300 CHAPTER 4

References

Angrist, J., Lang, D., and Oreopoulos, P. (2009). Incentives andServices for College Achievement : Evidence from a RandomizedTrial. American Economic Journal: Applied Economics, 1(1):136–163.

Angrist, J. and Lavy, V. (2009). The Effect of High Stakes HighSchool Achievement Awards: Evidence from a Randomized Trial.American Economic Review, 99(4):301–331.

Angrist, J., Oreopoulos, P., and Williams, T. (2014). When Opportu-nity Knocks, Who Answers? New Evidence on College AchievementAwards. Journal of Human Resources, Summer 2014, 49(3):572–610.

Åslund, O. and Engdahl, M. (2018). The value of Earning for Learn-ing : Performance Bonuses in Immigrant Language Training. Eco-nomics of Education Review, 62(2018):192–204.

Bettinger, E. (2012). Paying to Learn: The Effect of Financial Incen-tives on Elementary School Test Scores. The Review of Economicsand Statistics, 94(3):686–698.

Calonico, S., Cattenao, M. D., and Titiunik, R. (2014). Robust Data-driven Inference in the Regression Discontinuity Design. The StataJournal, 14(4):909–946.

Calonico, S., Jansson, M., and Ma, X. (2018). Manipulation TestingBased on Density Discontinuity. The Stata Journal, 18(1):234–261.

Chiswick, B. R. (2008). The Economics of Language: An Introductionand Overview. Discussion Paper 3568, IZA.

Dearden, L., Emmerson, C., Frayne, C., and Meghir, C. (2009). Con-ditional Cash Transfers and School Dropout Rates. The Journal ofHuman Resources, 44(4):827–857.

REFERENCES 301

Fryer, R. G. (2011). Financial incentives and student achievement:Evidence from randomized trials. Quarterly Journal of Economics,126(4):1755–1798.

Imbens, G. W. and Lemieux, T. (2008). Regression discontinuitydesigns: A guide to practice. Journal of Econometrics, 142(2):615–635.

Kennerberg, L. and Åslund, O. (2010). Sfi och Arbetsmarknaden.Rapport 2010:10, Institute for Evaluation of Labor Market andEducation Policy.

Kennerberg, L. and Sibbmark, K. (2005). Vem Deltar i Svenska förInvandrare? Rapport 2005:13, Institute for Evaluation of LaborMarket and Education Policy.

Kolesár, M. and Rothe, C. (2016). Inference in Regression Discontinu-ity Designs with a Discrete Running Variable. American EconomicReview (forthcoming).

Kremer, M., Miguel, E., and Thornton, R. (2009). Incentives to Learn.Review of Economics and Statistics, 91(1):437–456.

Lee, D. S. and Lemieux, T. (2010). Regression Discontinuity Designsin Economics. Journal of Economic Literature, 48(2):281–355.

Leuven, E., Oosterbeek, H., and van der Klaauw, B. (2010). TheEffect of Financial Rewards on Students’ Achievement: Evidencefrom a Randomized Experiment. Journal of the European EconomicAssociation, 8(6):1243–1265.

McCrary, J. (2008). Manipulation of the running variable in the re-gression discontinuity design: A density test. Journal of Economet-rics, 142(2):698–714.

OECD (2016). How Are Refugees Faring on the Labour Market inEurope? Working Paper 1/2016.

302 CHAPTER 4

OECD (2018). International Migration Outlook 2018. OECD Pub-lishing, Paris.

Pallais, A. (2009). Taking a Chance on College: Is the Tennessee Edu-cation Lottery Scholarship Program a Winner? Journal of HumanResources, 44(1):199–222.

Sharma, D. (2010). The Impact of Financial Incentives on AcademicAchievement and Household Behavior: Evidence from a Random-ized Trial in Nepal. SSRN Electronic Journal.

Statistics Sweden (2017). Kostnader för utbildningsväsendet 2012-2016.

Swedish Agency for Public Management (2009). Sfi – Resultat,Genomförande och Lärarkompetens. En Utvärdering av Svenskaför Invandrare. 2009:2.

Swedish National Agency for Education (2011). Elever och Studiere-sultat i Sfi 2010. Dnr 71-2011:14.

Swedish National Agency for Education (2012). Redovisning avregeringsuppdrag. Dnr 2011:13.

Swedish National Audit Office (2008). Svenskundervisning för Invan-drare (Sfi). En Verksamhet med Okända Effekter. RiR 2008:13.

Swedish Schools Inspectorate (2010). Svenskundervisning för Invan-drare (Sfi).

Swedish Schools Inspectorate (2011). Ändamålsenlighet och Resultati Svenskundervisningen för Invandrare.

4.A. ADDITIONAL DESCRIPTIVE EVIDENCE 303

Appendices

4.A Additional Descriptive Evidence

Figure 4.7: Seasonality in Sfi Enrollment, 2008-2011

0

2000

4000

6000

Freq

uenc

y

2007

m10

2008

m1

2008

m4

2008

m7

2008

m10

2009

m1

2009

m4

2009

m7

2009

m10

2010

m1

2010

m4

2010

m7

2010

m10

2011

m1

2011

m4

2011

m7

2011

m10

2012

m1

2012

m4

Month of enrollment

Notes: This figure shows the number of immigrants enrolling in Sfi education by month of enrollment.The population is restricted to immigrants of age 18-64 with a bonus-entitling residence permit whosettled in a municipality that did not participate in the experimental introduction of the bonussystem and were registered in Sweden during 2008-2010. Immigrants from Norway, Denmark, Finland,and Iceland are excluded.

304 CHAPTER 4

Table 4.5: Summary Statistics of OutcomeVariables for Individuals in the Control Group

Mean SD Min Max

Enrolled in Sfi within4 months 0.44 0.50 0.00 1.008 months 0.74 0.44 0.00 1.0012 months 0.79 0.41 0.00 1.00

Pass rateany course 0.43 0.50 0.00 1.00bonus course 0.17 0.38 0.00 1.00

Observations 3605Notes: This table presents the mean, standard deviation,minimum, and maximum of each variable. The population isrestricted to immigrants of age 18-64 with a bonus-entitlingresidence permit who settled in a municipality that did notparticipate in the experimental introduction of the bonussystem and were registered in Sweden between March 31,2010 and June 30, 2010. Immigrants from Norway, Denmark,Finland, and Iceland are excluded.

4.B. TEST FOR MANIPULATION 305

4.B Test for Manipulation

Figure 4.8: McCrary density test

0

20

40

60

80

100

-90 -80 -70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70 80 90Normalized date variable

Notes: This figure illustrates the McCrary density test. The population is restricted to immigrants ofage 18-64 with a bonus-entitling residence permit who settled in a municipality that did notparticipate in the experimental introduction of the bonus system and were registered in Swedenbetween March 31, 2010 and September 30, 2010. Immigrants from Norway, Denmark, Finland, andIceland are excluded. The blue dots show the number of immigrant registration by (normalized) datefor a bandwidth of 92 days. The blue lines are the fitted values from the regression and the grey linesshow the confidence intervals.

306 CHAPTER 4

4.C Regression Discontinuity Graphs of Co-variates

Figure 4.9: Demographic Characteristics

(a) Age

20

25

30

35

40


(b) Female

.4

.5

.6

.7

.8

.9

1


(c) Partner

.4

.5

.6

.7

.8

.9

1


(d) Metropolitan municipality

.4

.5

.6

.7

.8

.9

1


Notes: This figure present the average value of predetermined characteristics in each 10-day bin. Theblue line represent fitted values from the regression. The population is restricted to immigrants of age18-64 with a bonus-entitling residence permit who settled in a municipality that did not participate inthe experimental introduction of the bonus system and were registered in Sweden between March 31,2010 and September 30, 2010. Immigrants from Norway, Denmark, Finland, and Iceland are excluded.

4.C. REGRESSION DISCONTINUITY GRAPHS OFCOVARIATES 307

Figure 4.10: Level of Education

(a) Primary education

0

.1

.2

.3

.4

.5


(b) Secondary education

0

.1

.2

.3

.4

.5


(c) Higher education

0

.1

.2

.3

.4

.5


(d) Education level unknown

0

.1

.2

.3

.4

.5


Notes: See Figure 4.9.

308 CHAPTER 4

Figure 4.11: Region of Birth

(a) EU27, N. America, Oceania

0

.1

.2

.3

.4

.5

.6


(b) Europe excl. EU27 and theNordic

0

.1

.2

.3

.4

.5

.6


(c) South America

0

.1

.2

.3

.4

.5

.6


(d) Africa

0

.1

.2

.3

.4

.5

.6


(e) Asia

0

.1

.2

.3

.4

.5

.6


Notes: See Figure 4.9.

4.D. VARYING THE CHOICE OF BANDWIDTH 309

4.D Varying the Choice of Bandwidth

Figure 4.12: Estimated Effects on the Enrollment Rate for Varying Band-widths

(a) Within 4 months

-.2

-.1

0

.1

10 20 30 40 50 60 70 80 90Bandwidth


(b) Within 8 months

-.3

-.2

-.1

0

.1

10 20 30 40 50 60 70 80 90Bandwidth


(c) Within 12 months

-.15

-.1

-.05

0

.05

.1

10 20 30 40 50 60 70 80 90Bandwidth


Notes: This figure presents the estimated coefficient and the 95 percent confidence interval for eachoutcome for a bandwidth starting at 10 days and increasing by 2 days until reaching the fullbandwidth of 92 days. The population is restricted to immigrants of age 18-64 with a bonus-entitlingresidence permit who settled in a municipality that did not participate in the experimentalintroduction of the bonus system. Immigrants from Norway, Denmark, Finland, and Iceland areexcluded.

310 CHAPTER 4

Figure 4.13: Estimated Effects on the Pass Rate for Varying Bandwidths

(a) Any course

-.2

-.1

0

.1

10 20 30 40 50 60 70 80 90Bandwidth


(b) Bonus course

-.1

-.05

0

.05

.1

10 20 30 40 50 60 70 80 90Bandwidth


Notes: This figure presents the estimated coefficient and the 95 percent confidence interval for eachoutcome for a bandwidth starting at 10 days and increasing by 2 days until reaching the fullbandwidth of 92 days. The population is restricted to immigrants of age 18-64 with a bonus-entitlingresidence permit who settled in a municipality that did not participate in the experimentalintroduction of the bonus system. Immigrants from Norway, Denmark, Finland, and Iceland areexcluded.

Sammanfattning

Den här avhandlingen består av fyra fristående uppsatser på tematutbildning och integration. De två första kapitlen studerar vad sompåverkar föräldrars val av grundskola och vilka konsekvenser det fårför fördelningen av elever mellan skolor. De två följande kapitlenstuderar integrationsprocessen av nyanlända invandrare genom attutvärdera deras gensvar på förändringar i integrationspolitiken.

Valfrihet har blivit ett allt vanligare inslag i dagens utbildningssys-tem. Medan välbeställda familjer alltid haft viss möjlighet att väljaskola åt sina barn, genom att helt enkelt flytta till ett attraktivtskolområde, har andra haft svårare att påverka vilken skola deras barnblivit tilldelad. De senaste decennierna har det dock blivit allt van-ligare med centraliserade skolvalsprogram och allt fler föräldrar stårinför beslutet att välja skola åt sina barn. En växande forskningslit-teratur studerar konsekvenserna av denna utveckling och försökerbesvara frågan om hur det optimala skolvalsprogrammet ser ut.

I det första kapitlet i denna avhandling, Are Parents Unin-formed? The Impact of School Performance Information onSchool Choices and School Assignments, som är författat till-sammans med Dany Kessel, studerar vi huruvida tillgång till in-formation om skolors prestation på de nationella proven påverkarbenägenheten att välja högpresterande skolor. Studien använder datafrån ett randomiserat experiment genomfört i Linköpings kommun år

311

312 SWEDISH SUMMARY

2016. Ett antal slumpmässigt utvalda hushåll, vars barn förväntadesbörja årskurs 7 till höstterminen, tillhandahölls information om allahögstadieskolors prestation på de nationella proven samt huruvida de-ras genomsnittliga poäng var över eller under förväntan givet skolanselevkomposition.

Vi finner inga effekter på hur många eller vilka hushåll som ak-tivt söker en annan skola än deras garantiskola, men efterfrågan påde högpresterande skolorna i kommunen ökar med cirka fem procen-tenheter. Detta förklaras av att högutbildade och inhemska hushåll istörre utsträckning börjar söka dessa skolor istället för de genomsnit-tligt presterande skolorna. Med hjälp av simuleringar undersöker visedan hur detta skulle påverka matchningen mellan elever och skolorunder antagandet att alla hushåll hade fått tillgång till samma in-formation. Som förväntat ökar andelen elever i högpresterande skolortill följd av det högre söktrycket. Magnituden motsvarar cirka tvåprocentenheter, vilket innebär att kapacitetsbegränsningar hindrarskolorna från att möta den ökade efterfrågan fullt ut. Det ökade sök-trycket från nya elevgrupper innebär också att elever med utländskbakgrund i viss mån knuffas ut från dessa skolor. Detta leder till lägresegregation mellan skolor med avseende på elevernas migrationsbak-grund då elever med utländsk bakgrund är överrepresenterade på debäst presterande skolorna.

I det andra kapitlet, School Choice, Admission Rules andSegregation in Primary Schools, också det samförfattat medDany Kessel, fortsätter vi att studera skolvalet men nu med fokuspå de antagningsregler som används. Med hjälp av data från ettskolvalsprogram i Botkyrka kommun analyserar vi hur skolsegrega-tionen påverkas av hur elevernas prioritet till skolor bestäms. Vi hartillgång till föräldrarnas skolval för fyra kohorter av barn som bör-jade förskoleklass mellan 2011 och 2014. För att kunna studera kon-trafaktiskta utfall börjar vi med att modellera föräldras preferenserför skolor i syfte att kunna bestämma deras rangordning av skolor i

313

olika scenarion. Därefter simulerar vi hur matchningen av elever tillskolor skulle se ut under tre olika typer av antagningsregler samt idet fall då elever tilldelas skola utifrån skolors upptagningsområdenistället för föräldrars önskemål.

Vi finner att när förtur till skolor baseras på närhet till skolan,så att elever som bor närmre skolan får högre prioritet, liknarsorteringen av elever den som uppstår när upptagningsområdenanvänds. När elevernas prioritet bestäms slumpmässigt, genomlotten, minskar skolsegregationen med avseende på elevernassocioekonomiska bakgrund. Det tredje antagningssystemet ger förturtill elever som bidrar till mångfalden på skolan. Detta minskarskolsegregationen med avseende på eleveras socioekonomiskabakgrund ytterligare. Kostnaderna av att frångå avståndsbaseradeprioritetsregler till förmån för ett lotteri eller antagning baseradpå kvoter, i termer av välfärd eller rangordning av den tilldeladeskolan, är förhållandevis små. Denna studie tyder därför på attförtursreglernas utformning påverkar segregationen mellan skolor.Resultaten tyder dock också på att en stor del av skolsegregationenberor på bostadssegregationen, vilket innebär att lösningen påproblemet med en alltmer segregerad skola troligen delvis liggerbortom skolans värld.

I de två följande kapitlen lämnar jag frågan om skolval bakommig och fokuserar istället på integrationsprocessen av nyanlända in-vandrare. Utvecklingen i världen de senaste åren har inneburit attmånga människor flytt sina hemländer för att söka skydd på andraplatser. Detta, tillsammans med en generellt hög invandring till Eu-ropa, har satt integrationsfrågan högt upp på dagordningen. Fleraländer i Europa har svarat på det ökade antalet asylsökande genomatt strama åt sin migrationspolitik. Ett exempel på detta är övergån-gen från permanenta till temporära uppehållstillstånd, vilket intro-ducerar en osäkerhet vad gäller tidshorisonten i det nya hemlandet.

Det tredje kapitlet, Should I Stay or Must I Go? Temporary

314 SWEDISH SUMMARY

Refugee Protection and Labor-Market Outcomes, samförfat-tat med Matilda Kilström och Birhe Larsen, analyserar denna fråga.Specifikt tittar vi på en reform som implementerades i Danmark år2002, som innebar en förlängning av tiden med temporära uppehåll-stillstånd (det vill säga tiden innan det är möjligt att söka ett perma-nent uppehållstillstånd). Flyktingar som anlände till Danmark frånoch med den 28 februari, 2002, bedömdes enligt de nya reglerna ochvar därmed tvugna att vara bosatta i Danmark i sju år innan de kundeansöka om permanent uppehållstillstånd. Innan reformen räckte detatt ha varit bosatt i Danmark i tre år. Vi fokuserar på hur dennareform påverkade individers investeringar i humankapital och arbets-marknadsutfall. Sättet som reformen implementerades på ger upphovtill en diskontinuitet som vi kan utnyttja för att estimera effekternamed hjälp av en regression discontinuity design. Eftersom uppehåll-stillstånd också kan fås genom uppvisande av en stabil anknytning tillarbetsmarknaden tror vi att effekterna kan variera med avseende påhur långt från arbetsmarknaden individerna befinner sig vid ankomsttill Danmark. För att förstå vilka mekanismer som ligger bakom hurreformen påverkade olika individer konstruerar vi en teoretisk searchand matching-modell, med heterogenitet i utbildningsnivå.

Vi estimerar en signifikant positiv effekt på andelen individer in-skrivna i utbildning. Effekten är starkast för kvinnor och individerutan universitetsutbildning. Detta ligger i linje med modellen, somförutsäger att individer som står längre ifrån arbetsmarknaden kom-mer att investera i utbildning för att skaffa sig en bättre ställning pådensamma. När det gäller arbetsmarknadsutfall, som inkomst ellerandelen i arbete, estimerar vi inte några signifikanta effekter för helaurvalet av individer. Vi studerar även några andra, relevanta, utfall.Reformen hade en negativ effekt på brottslighet, särskilt egendoms-brott, drivet av män. Trots anekdotisk evidens om att reformen or-sakade stress och hade en negativ inverkan på flyktingarnas hälsakan vi inte bekräfta detta genom några signifikanta skillnader mellan

315

grupperna i antalet läkarbesök. Vi finner visst stöd för att den ökadeosäkerheten om tidshorisonten i Danmark avskräckte flyktingar frånatt skaffa barn, eller ledde till att man valde att vänta längre med attbilda familj.

I det fjärde, och sista, kapitlet i denna avhandling, The Effectsof Performance Based Bonuses in the Swedish LanguageTraining Program for Immigrants, studerar jag effekterna avett prestationsbaserat bonus-program inom ramen för Sfi (Svenskaför invandrare). Att lära sig sitt nya hemlands språk är en nyckeltill framgångsrik integration. Trots detta är det många invandraresom aldrig fullt ur lär sig att bemästra sitt nya språk. 2010 försökteman i Sverige ändra på detta genom att inför ett bonussystem i Sfi-undervisningen. Ett godkänt betyg i vissa kurser gav rätt till en bonuspå upp till 12,000 SEK.

Implicit i införandet av bonusystemet ligger antagandet att in-vandrare ansträngde sig för lite när det kom till deras språkinlärn-ing. Detta skulle kunna bero på att de underskattar avkastningenav att lära sig det nya språket eller bortser från de positiva exter-naliteter som språkfärdighet ger upphov till. Utsikten om en monetärbonus skulle då kunna leda till bättre utfall eftersom incitamentenatt klara språkkurserna stärks. Om invandrare å andra sidan kan an-vända denna tid mer produktivt till andra aktiviteter skulle bonusenkunna leda till sämre utfall. Dessutom finns en risk att monetära in-citament tränger undan den inneboende motivationen för att lära sigspråket. I slutändan är därför effekterna av denna typ av policy enempirisk fråga.

I den här studien analyseras effekterna av bonussystemetsinförande med hjälp av en regression discontinuity design. Huruvidaen invandrare var behörig att få en bonus, givet slutförande av deuppställda kriterierna, bestämdes till fullo av datumet för ankomsttill Sverige samt på vilken grund uppehållstillståndet beviljats.Resultaten visar inte på någon effekt på inskrivningsgraden i

316 SWEDISH SUMMARY

Sfi-undervisning, vilket indikerar att bonussystemet inte fick flerindivider att påbörja språkträning. Detta skulle kunna förklaras avden redan höga inskrivningsgraden innan reformen genomfördes, påcirka 80 procent. Estimeringar av effekten på färdigställande avkurser generellt tyder inte heller på några positiva effekter, men näreffekten estimeras specifikt på de kurser som var bonus-kvalificerandeär resultaten inte entydiga. En positiv effekt på upp till fyraprocenenheter kan inte utslutas, men resultaten är känsliga för denempiriska specifikationen vilket försvårar slutsatserna.

'UUC[UQP'FWECVKQPCN%JQKEGU CPF+PVGITCVKQP individuals...

Documents

Transcript of 'UUC[UQP'FWECVKQPCN%JQKEGU CPF+PVGITCVKQP individuals...