I. Introduction - MIT Economics

SOURCES OF GEOGRAPHIC VARIATION IN HEALTH CAREEVIDENCE FROM PATIENT MIGRATION

Amy Finkelstein

Matthew Gentzkow

Heidi Williams

We study the drivers of geographic variation in US health care utilizationusing an empirical strategy that exploits migration of Medicare patients toseparate the role of demand and supply factors Our approach allows us toaccount for demand differences driven by both observable and unobservablepatient characteristics Within our sample of over-65 Medicare beneficiarieswe find that 40ndash50 of geographic variation in utilization is attributable todemand-side factors including health and preferences with the remainderdue to place-specific supply factors JEL Codes H51 I1 I11

I Introduction

Health care utilization varies widely across the UnitedStates (Fisher et al 2003a 2003b) Adjusting for regional differ-ences in age sex and race health care spending for the averageMedicare enrollee in Miami FL was $14423 in 2010 but just$7819 for the average enrollee in Minneapolis MN The averageenrollee in McAllen TX spent $13648 compared with $8714 innearby and demographically similar El Paso TX1 Similar geo-graphic variation is observed in the frequency of specific treat-ments (Chandra Cutler and Song 2012) and in measures of total

We are grateful to Daron Acemoglu Raj Chetty David Cutler RebeccaDiamond Joe Doyle Liran Einav Ben Handel Nathan Hendren Pat KlineMatt Notowidigdo Adam Sacarny Jesse Shapiro Jonathan Skinner DougStaiger and seminar participants at the BUHarvardMIT health economics sem-inar Dartmouth the NBER Aging Meeting the NYC health economics seminarStanford the University of Chicago the University of Maryland MIT UtahWinter Business Economics Conference Wellesley and the 60th AnniversaryCongress of the Yrjo Jahnsson Foundation for comments to Lizi Chen GrantGraziani Yunan Ji Sara Kwasnick Tamar Oostrom Daniel Prinz DanielSalmon and Tony Zhang for excellent research assistance and to the NationalInstitute on Aging (Finkelstein R01-AG032449 P01-AG19783) and the NationalScience Foundation (Williams 1151497 Gentzkow 1260411) for financial supportGentzkow also thanks the Neubauer Family Foundation and the Initiative onGlobal Markets at the University of Chicago Booth School of Business

1 Authorsrsquo tabulations based on total Medicare Parts A and B reimburse-ments per enrollee from Dartmouth Atlas of Health Care httpwwwdartmouthatlasorgdownloadstablespa_reimb_hrr_2010xls

The Author(s) 2016 Published by Oxford University Press on behalf of Presidentand Fellows of Harvard College All rights reserved For Permissions please emailjournalspermissionsoupcomThe Quarterly Journal of Economics (2016) 1681ndash1726 doi101093qjeqjw023Advance Access publication on July 19 2016

1681Downloaded from httpsacademicoupcomqjearticle-abstract131416812468872by MIT Libraries useron 20 June 2018

health care utilization that adjust for regional variation in ad-ministratively set prices (Gottlieb et al 2010) Higher area-levelutilization is not generally correlated with better patientoutcomes2

Understanding what drives geographic variation in utiliza-tion has important implications for policy If high-utilizationareas like McAllen and Miami are different mainly becausetheir doctorsrsquo incentives or beliefs lead them to order excessivetreatments with low return policies that change those incentivesor beliefs could result in savings on the order of several percent-age points of GDP (Congressional Budget Office 2008 Gawande2009 Skinner 2011) On the other hand if patients in high-utilization areas are simply sicker or prefer more intensivecare such policies could be ineffective or counterproductive

In this article we exploit patient migration to separatevariation due to patient characteristics such as health or pref-erences from variation due to place-specific variables such as doc-torsrsquo incentives and beliefs endowments of physical or humancapital and hospital market structure As a shorthand we referto the former as lsquolsquodemandrsquorsquo factors and the latter as lsquolsquosupplyrsquorsquo fac-tors3 To see the intuition for our approach imagine a patient whomoves from high-utilization Miami to low-utilizationMinneapolis If all of the utilization difference between thesecities arises from supply-side differences like doctor incentivesor beliefs we would expect the migrantrsquos utilization to drop im-mediately following the move to a level similar to other patientsof the low-utilization doctors in Minneapolis If all of the utiliza-tion difference reflects the demand-side reality that residents ofMiami are sicker we would expect the migrantrsquos utilization toremain constant after the move at a level similar to the typicalperson in Miami Where the observed utilization change falls

2 See Skinner (2011) for an extensive discussion The Congressional BudgetOffice (2008) concludes that high-spending areas lsquolsquotend to score no better and insome cases score worse than other areas do on process-based measures of qualityand on some measures of health outcomesrsquorsquo and that more intensive treatment inhigh-spending areas lsquolsquoappear[s] to improve health outcomes for some types of pa-tients but worsen outcomes for othersrsquorsquo

3 This corresponds to the usual definitions of demand and supply in mostcases but the correspondence is not perfect For example peer effects or sociallearning will generally be captured in our framework as a place-specific(lsquolsquosupplyrsquorsquo) factor since the composition of peers can change when a patientmoves but it would be more natural to think of them as shifters of demandrather than supply

QUARTERLY JOURNAL OF ECONOMICS1682

Downloaded from httpsacademicoupcomqjearticle-abstract131416812468872by MIT Libraries useron 20 June 2018

between these two extremes identifies the relative importance ofdemand and supply factors

We implement this strategy using claims data for a 20 sam-ple of Medicare beneficiaries from 1998 to 2008 Our main out-come measure adjusts health care spending for geographic pricedifferences to create a quantity measure of utilization as inGottlieb et al (2010) We introduce a simple model of healthcare demand and supply which implies that the log of a patientrsquosannual health care utilization can be written as a combination ofa patient fixed effect a location fixed effect and a vector of time-varying controls including indicators for year relative to move formigrants This specification allows for the possibility that mi-grants have systematically different utilization levels fromnonmigrants and that these levels are correlated with the mi-grantrsquos origin and destination regions It also allows for arbitrarydifferences in utilization trends of migrants relative to nonmi-grants The key identifying assumption is that such differentialtrends do not vary systematically with the migrantrsquos origin anddestination

We begin with an event-study analysis of changes in log uti-lization around moves We observe a sharp change in the year of amove equal to about half of the difference in average log utiliza-tion between the origin and destination There is little systematictrend premove and no systematic adjustment postmove The on-impact effect is similar for moves from low-to-high and high-to-low utilization regions and is roughly linear in the absolutevalue of the originndashdestination difference in log utilization

Our estimated model exploits this variation to infer that 47of the difference in log utilization between above- and below-median areas is due to patient characteristics with the remain-der due to place-specific factors The shares are similar fordifferences between the top and bottom quartiles deciles or ven-tiles The share of the difference in log utilization due to patientsis also similar when we isolate differences between the very high-est utilization areas such as McAllen or Miami and the verylowest utilization areas such as El Paso or Minneapolis The re-sults appear inconsistent with patient effects arising primarilyfrom habit formation or persistence of treatments startedpremove and instead point toward heterogeneity in healthstatus or preferences that are fixed over the horizon of our data

Our decomposition can be interpreted in terms of a counter-factual by what share would the gap in utilization between areas

GEOGRAPHIC VARIATION IN HEALTH CARE 1683


fall if patients were randomly reallocated between themImportantly this is a partial equilibrium experiment in that itholds fixed supply-side characteristics such as stocks of physicaland human capital In the long run we would expect some ofthese characteristics to adjust endogenously to the change in pa-tient demand (Chandra and Staiger 2007) If this led to conver-gence on the supply side the long-run fall in geographic variationunder the counterfactual would be greater than our short-runestimates would suggest

We replicate our analysis for various components of total uti-lization All measures show sharp changes in the year of a movewith magnitudes implying patient shares ranging from 9 to 71Consistent with intuition we find large patient shares for outcomeswhere we might think patients have significant discretionmdashpreventive care and emergency room visits for examplemdashandsmaller patient shares for outcomes where we might think theyhave lessmdashdiagnostic tests imaging tests and inpatient care Wealso find some suggestive evidence that the patient share may belower at higher percentiles of the utilization distribution

In the final section we present evidence on the observablearea-level correlates of our estimated place and patient effectsThe potential correlates we consider include the number qualityand organizational form of hospitals survey-based measures ofdoctor beliefs about appropriate practice style and patient pref-erences over alternative practice styles average patient demo-graphics and average patient health status An importantchallenge arises with regard to the latter because standard mea-sures of underlying health status are derived from claims data agiven condition is more likely to be recorded in a high-intensityarea and standard measures of patient health therefore include alarge component of systematic place-specific measurement error(Song et al 2010 Welch et al 2011) To address this we extendour mover-based strategy to estimate the place-specific measure-ment error component and derive corrected health measurespurged of this measurement error

The correlations are broadly consistent with intuition fromour model and evidence from the existing literature On thesupply side we find that the place component of utilization isparticularly high in areas with a larger share of for-profit hospi-tals and a larger share of doctors who report a preference foraggressive care the latter is consistent with recent literatureemphasizing the importance of physician practice styles and



beliefs in driving geographic variation (eg Cutler et al 2015)We also find that the place component is higher in areas wherepatients are sicker which is consistent with past work arguingthat physical and human capital are likely to adjust endoge-nously to patient demand (Chandra and Staiger 2007) On thedemand side we find that the patient component of utilizationis higher where patients are sicker and of higher socioeconomicstatus consistent with patient health status and patient prefer-ences playing an important role Were we to take the correlationsbetween log utilization and corrected health measures as causalthey would imply that about a quarter of the geographic variationin log utilization (or equivalently about half of our estimate of thepatient share of this variation) may be explained by our correctedpatient health measures the remainder may reflect preferencesor unmeasured health

Studying Medicare patients is appealing due to the availabil-ity of high-quality rich data on large numbers of beneficiariesand the relatively uniform insurance environment Medicare ac-counts for a significant share of total US health spending 205as of 2011 (Moses et al 2013)4 Nevertheless extrapolating ourconclusions to other populations requires caution Although re-gional variation in utilization appears to be the norm5 the rela-tive importance of place and patient factors could differ in othersettings In private insurance markets moreover substantialcross-area differences in prices mean the correlates of area spend-ing may differ substantially from the correlates of area utilization(Chernew et al 2010 Dunn Shapiro and Liebman 2013 Cooperet al 2015)

Our work contributes to a large existing literature seeking toseparate the role of demand-side and supply-side factors in driv-ing geographic variation in health care utilization All of thesestudies infer the role of demand-side factors from the explanatorypower of patient observables Our main contribution is to develop

4 This number includes under-65 beneficiaries who we exclude from ourstudy According to Neuman et al (2015) beneficiaries under 65 accounted for22 of spending in traditional Medicare in 2011

5 Large regional differences have been documented in the US VeteransAffairs system (Ashton et al 1999 Subramanian et al 2002 CongressionalBudget Office 2008) in private insurance markets (Baker Fisher and Wennberg2008 Rettenmaier and Saving 2009 Chernew et al 2010 Philipson et al 2010Dunn Shapiro and Liebman 2013 Cooper et al 2015) and in other countries in-cluding the United Kingdom and Canada (McPherson et al 1981)



a strategy that exploits migration to capture observed and unob-served patient characteristics

Taken together the evidence from the prior literature sug-gests three conclusions (i) supply-side factors are a key driver ofgeographic variation (ii) patient preferences and characteristicsother than health status explain little variation (iii) differencesin health status may be important but the evidence is inconclu-sive because of endogenous measurement error6 Our findingsconfirm that supply-side factors are important while revealingthat patient preferences and health status together account fora large share of variation Once we address the endogenous mea-surement issue with patient health we find that roughly a quar-ter of the geographic variation in log health care utilization canpotentially be attributed to observable patient health Whetherthe remaining patient component reflects preferences or unmea-sured health remains an open question

Like past decompositions ours is not sufficient to drawstrong conclusions about the efficiency of observed geographicvariation Although supply-driven heterogeneity may reflectwaste stemming from disagreement among physicians regardingoptimal practice styles it could also reflect an optimal allocationof physical and human capital Conversely although our modelshows a formal sense in which demand-driven variation is likelyto be consistent with efficiency this need not be the case in aricher model We view our findings as a first step toward amore welfare-relevant understanding and a clarification of aninfluential body of existing evidence

Our empirical strategy relates to past work using changes inresidence or employment to separate effects of individual charac-teristics from geographic or institutional factors Most closely

6 See Skinner (2011) and Chandra Cutler and Song (2012) for reviews andCutler et al (2015) and Baker Bundorf and Kessler (2014) for more recent contri-butions Chandra Cutler and Song (2012) write lsquolsquoIn general the literature pointsto the importance of supply-side incentives over demand-side factors in drivingtreatment choicersquorsquo (p 425) and lsquolsquomost of the literature agrees that patient charac-teristics and preferences do not explain much of the differences across areasrsquorsquo(p 402) With regard to health status they write lsquolsquoSome researchers argue thatvariation is accounted for by population disease burden but other authors arguethat prevalence of diagnosis is itself endogenous across areasrsquorsquo (p 402) Skinner(2011) writes lsquolsquoWhile demand factors are importantmdashhealth in particularmdashthereremains strong evidence for supply-driven differences in utilizationrsquorsquo (p 46) Anexception to this consensus is Sheiner (2014) who argues that patients may explainmost or all of the variation



related are Song et al (2010) who look at how health measureschange around patient moves and Molitor (2014) who looks atcardiologist behavior changes around their moves Outside of thehealth care sector a number of papers beginning with Abowdet al (1999) use matched worker-firm data to separately identifyworker and firm fixed effects In this vein we draw especially onCard Heining and Klinersquos (2013) study of German workers andfirms Other work uses geographic or employment changes tostudy neighborhood effects on children (Aaronson 1998) culturalassimilation of immigrants (Fernandez and Fogli 2006) brandpreferences (Bronnenberg Dube and Gentzkow 2012) tax re-porting (Chetty et al 2013) teacher value added (ChettyFriedman and Rockoff 2014) and retirement savings decisions(Chetty et al 2014)

Section II introduces our model and estimation strategySection III describes our data and presents summary statisticsSection IV presents our main analysis of the role of demand andsupply factors in explaining geographic variation in health careutilization Section V explores the correlates of our estimated pa-tient and place effects Section VI concludes All appendix mate-rials are available in the Online Appendix

II Model and Empirical Strategy

IIA Model

We build a simple model of demand and supply for healthcare similar in spirit to the model in Cutler et al (2015) Ourgoals are to illustrate the demand and supply factors that driveequilibrium utilization and clarify the underlying assumptions ofour empirical specification

A population of patients i in year t utilizes health careyit 2 R

thorn Patients differ along three dimensions health statushit preferences i and geographic area j Higher values of hit

represent worse health the time-constant scalar preference pa-rameter i is defined so that higher values represent tastes formore aggressive care Some patients are lsquolsquononmoversrsquorsquo who live inone area j throughout the sample whereas others are lsquolsquomoversrsquorsquowhose area changes exactly once7 Patientsrsquo expected

7 We exclude patients who move more than once from the main analysis forsimplicity but we show that the results are robust to including them



continuation utility u yjhit ieth THORN frac14 12 y hiteth THORN

2thorn iy is maximized

at yit frac14 hit thorn i the level of care the patient would choose if theywere fully informed and faced a zero out-of-pocket price for careWe assume that the expectation of yit given the data observed bythe econometrician depends only on a patient fixed effect and avector of observables xit E yitj i j t xit

frac14 i thorn xit

Each patient living in area j in year t is matched to a repre-sentative physician who determines the patientrsquos care Weassume that a physician chooses yit to maximize the perceivedutility of her patients ~ujethyTHORN minus her (net private) costs of careprovision PCjtethyTHORN

8 Thus we can write

yit frac14 arg maxy

~uj yjhit ieth THORN PCjt yeth THORNeth1THORN

The difference between perceived patient utility ~ujethTHORN and truepatient utility uethTHORN captures potentially heterogeneous beliefs thatwould lead physicians to disagree about the appropriate level ofcare We assume ~uj yjhit ieth THORN frac14 u yjhit ieth THORN thorn ljy so that highervalues of lj represent relatively aggressive practice styles A va-riety of factors can affect PCjt including monetary incentivesorganizational rewards and physical and human capital Weassume PCjtethTHORN is linear in y and additively separable in j and t

Maximization of equation (1) yields our main estimatingequation for patients i living in area j throughout year t9

yijt frac14 i thorn j thorn t thorn xitthorn eijteth2THORN

where j thorn t frac14 lj PCjt0ethTHORN and our assumptions imply

Eetheijtjfi j t xitgTHORN frac14 0 In our main specification the quantity ywill be the log of total utilization which we define more pre-cisely in Section III and xit will consist of dummies for five-yearage bins and fixed effects r iteth THORN for movers where for a moverwho moves during year ti the relative year is r i teth THORN frac14 t ti Including these relative year effects allows for the possibilitythat the decision to move is correlated with shocks to health

8 For simplicity we assume that net provider costs have been scaled to thesame units as patient utility times its weight in the physicianrsquos objective functionLess compactly we can assume the physician maximizes ~ujethTHORN

~PCjtethTHORN where ~PCjtethTHORN

is measured in dollars and is the weight in the physicianrsquos objective assigned topatient utility then define PCjtethTHORN frac14

~PCjtethTHORN

9 We do not model outcomes for movers in year ti when they spend part of theyear in their origin area and part of the year in their destination When we estimateequation (2) we omit these observations



statusmdashfor example because sick patients sometimes move toseek care or at the other extreme because they are unable tobear the physical costs of moving We normalize r iteth THORN to zero fornonmovers

Our main goal is to decompose variation in average log uti-lization across regions into a demand-side component attribut-able to patients and a supply-side component attributable toplace To define this decomposition formally let yjt denote the

expectation of yit across patients living in area j in year t andlet yj denote the average of yjt across t Let yjt and yj denote the

analogous expectations of the patient-optimal level of careyit frac14 hit thorn i which by the foregoing assumptions is equal toi thorn xit Then the difference in average log utilization betweenany two areas j and j0 is the sum of the differences of the place andpatient components yj yj0 frac14 ethj j0 THORN thorn ethy

j yj0 THORN When we talk

about larger groups R that consist of multiple areas j we abusenotation by letting yR yR and R denote the simple averages ofyj yj and j across areas in R

We define the share of the difference between areas j and j0

attributable to place to be

Splace j j0eth THORN frac14j j0

yj yj0eth3THORN

and we define the share attributable to patients to be

Spat j j0eth THORN frac14yj yj0

yj yj0

Note that although Spat j j0eth THORN and Splace j j0eth THORN sum to 1 neitherneed be between 0 and 1 since it is possible that j j0 andyj yj0 have opposite signs We define Spat RR0eth THORN and Splace RR0eth THORN

to be the analogous shares for groups R and R0 We let yj denotethe sample analogue of yj Given consistent estimates j of j weform consistent estimates yj frac14 yj j of yj

IIB Discussion

Our stylized model clarifies the underlying economic factorsthat drive the patient and place components we estimate and theirrelationship to factors previously discussed in the literature

The patient component ethyj yj0 THORN is the difference in the util-ity-maximizing level of care yit for an average patient living in j



and an average patient living in j0 This in turn is the sum of thedifferences in average health hit and average preferences i Theformer will be driven by demographics such as age behavioralfactors such as diet exercise or smoking (Xu et al 2013) andgenetic predispositions to disease The latter captures the waypatients trade off the disutility of the pain suffering or inconve-nience of treatment against the value of improved health as wellas ethical or religious beliefs about the value of prolonging life(Barnato et al 2007)

The place component j j0

is the sum of the differencesbetween j and j0 in physiciansrsquo perceptions of marginal benefits lj

minus private marginal costs PC0ethTHORN Each of these nests a varietyof factors that have been fleshed out in more detail in the litera-ture For example differences in ~u0ethTHORN capture heterogeneous be-liefs about appropriate or effective treatment such as thelsquolsquocowboyrsquorsquo or lsquolsquocomforterrsquorsquo approaches to care documented in thesurvey evidence of Cutler et al (2015) Differences in privatemarginal costs PC0ethTHORN capture a number of factors including phy-siciansrsquo disutility or difficulty in delivering a given level of carewhich in turn reflects factors such as skill training or experience(as in Chandra and Staiger 2007) liability concerns (as exploredin Currie and MacLeod 2008) or the opportunity cost of physi-ciansrsquo time Private marginal costs may also be affected by orga-nizational features such as available physical capital theprevalence of nonprofit hospitals nonmonetary career incentivesinsurer constraints peer effects among doctors and organiza-tional culture (Lee and Mongan 2009)

One way to interpret the patient share Spat j j0eth THORN is in terms ofa counterfactual by what share would the gap in utilization be-tween the two areas fall if patients were randomly reallocatedbetween them Crucially this is a partial equilibrium experimentin that it holds fixed the existing perceptions incentives andphysical and human capital of physicians and organizations em-bedded in j In the medium or long run we would expect all ofthese factors to potentially adjust an area facing sicker patientsmight invest more in physical or human capital might specializein more intensive forms of care or might see shifts in its physi-ciansrsquo perceptions about what care is appropriate Chandra andStaiger (2007) provide evidence suggesting that such adjust-ments are quantitatively important and we stress that our esti-mates should be interpreted as partial equilibrium effects thatshut down adjustment along these margins



Under the assumptions of our model we can also relate ourpatient-place decomposition to conclusions about welfareSuppose that the social cost of care in location j and year t isSCjtethyTHORN so that a social planner would choose yit to maximizeuethyjhit iTHORN SCjtethyTHORN Then maintaining the other assumptions ofour model and assuming that we can write SCjt

0ethyTHORN frac14 ethj thorn

t THORN

equation (2) under the social planner solution would be identicalexcept that t thorn j would be replaced with t thorn

j For any two

locations j and j0 the patient component ethyj yj0 THORN is lsquolsquoefficientrsquorsquo inthe sense that it would remain unchanged under the social plan-ner In contrast the place component ethj j0 THORNmay or may not beefficient in this sense it will differ from the social planner solu-tion to the extent that physicians have inaccurate beliefs (lj 6frac14 0)or their net private marginal costs PC0ethTHORN differ from the socialmarginal cost SC0ethTHORN Of course these welfare implications requirestrong assumptions and our patient-place decomposition neednot have a tight relationship to welfare in a more generalmodel For example if variation in patient preferences (i)partly reflects misinformation or distorted beliefs (as inBaicker Mullainathan and Schwartzstein 2015) some compo-nent of the patient component could be inefficient as well10

IIC Identification

The model in equation (2) is only identified if the data includemovers If all patients were nonmovers there would be no way toseparate differences in the area fixed effects j from differences inthe average patient characteristics yj The key to separate iden-tification of these two components is the observed changes in uti-lization when patients move11

To build intuition consider a simplified version of our modelin which the t and xit are set to zero so utilization depends onlyon patient and place fixed effects plus the error term Normalize

10 The patient component could also fail to be efficient if the difference betweenthe private and social marginal cost of care SC0ethTHORN PC0ethTHORN varies with patient healthor preferences An example would be if out-of-pocket costs for different kinds oftreatments vary across patients andor areas Setting this aside seems a reasonableapproximation in the Medicare context since patients have relatively homoge-neous insurance although it may not be strictly true (for example because ratesof supplemental coverage vary across areas)

11 A sufficient condition for identification is that the number of movers be-tween any pair of areas j and j0 grows large as the total sample size approachesinfinity Abowd et al (2002) discuss weaker conditions for identification



the national mean of i to zero (in general we will not be able toidentify the national means of patient place and time effectsseparately but this will not affect our decompositions of geo-graphic variation) Suppose we observe a large number of pa-tients who move from area j0 to area j Then the difference j

j0

between their average yit in the years after the move and theyears before the move is a consistent estimator of j j0 If weobserve similar samples of patients moving between the otherareas in the sample along with the overall mean of log utilizationy we can form consistent estimates j of each j The yj wouldthen be consistently estimated by yj j

Identification in the full model is similar Identifying the t

and age coefficients is standard and does not rely on moversAdding the relative year effects r iteth THORN to xit has a more substantial

effect It allows for arbitrary changes in log utilization for moverspre- and postmove with the restriction that these changes are thesame regardless of the origin and destination In the full modeltherefore observing only movers from j0 to j is not enough to iden-

tify j j0 because jj0 would also depend on the difference be-

tween the postmove and premove r iteth THORN Identification in this case

comes from the differences in the changes across movers withdifferent origins and destinations If we have movers from j0 to jand also movers from j to j0 for example we can estimate j j0

consistently as

j

j0

j0

j

2 Importantly our model permits movers to differ arbitrarily

from nonmovers in both levels of log utilization (via the i) andtrends in log utilization around their moves (via the r iteth THORN) Thelatter would allow for moves to be associated with either positiveor negative health shocks for example In principle we canallow substantially more flexibility including area- or individ-ual-specific trends different fixed effects by subperiods and in-teractions between j and patient observables We can also addflexibility by using data for movers only in the years just before orafter their move in the spirit of a regression discontinuity Weexplore robustness to specifications along these lines later

Our model is nevertheless restrictive in several importantways First we cannot allow for shocks to utilization that coincideexactly with the timing of the move and that are correlated withutilization in the origin and destination In the example abovesuppose that for movers from j0 to j the conditional expectation of



health hit in years just after the move is strictly greater than for

movers from j to j0 This would inflate jj0 relative to j0

j and lead

j

j0

j0

j

2 to be an overestimate of j j0 As a concrete example this

could occur if patients who receive adverse health shocks respondby moving to relatively high-utilization areas The result in thiscase would be that we would attribute some of the health shock tothe effect of moving and thus overstate the role of places relativeto patients

Although we cannot rule out such bias entirely the pattern ofresults provides some comfort that it is unlikely to be largeDeterioration in health status that occurs gradually and is corre-lated with utilization in the destination and origin would tend toshow up as pretrends in our event study analysis (Section IVA)In fact we do find a positive pretrend but its magnitude is smalland we show that the results are robust to restricting the data to asmall window around the move Sudden health events thatprompt a move to systematically higher utilization places couldpotentially cause bias without a pretrend However such shockswould tend to produce a postmove spike in our event studies thatdissipates over time (assuming treatment for acute conditions ismost intense immediately after they occur) and this is not thepattern we observe

Second our specification assumes that i and j are addi-tively separable in the equation for log utilization We see thisas an attractive assumption economically It has the intuitiveimplication that patient and place characteristics affect thelevel of utilization multiplicatively and thus that the (level) uti-lization of patients who are sick or prefer intensive care (ie havehigh i) will vary more across places than that of patients who arehealthy or rarely seek care (ie have low i)

12 We also see the logmodel as appealing on econometric grounds given utilizationrsquosskewed cross-sectional distribution and large secular trend

That said the log specification nevertheless imposes someimportant restrictions It rules out for example variationacross places that causes an equal level shift for all patients

12 To take a concrete example suppose that patients have either one or twochronic conditions and that places spend either $5000 or $10000 per chronic con-dition This would imply a model additive in logs with exp ethiTHORN 2 f12gexp ethjTHORN 2 f510g and the log of utilization yij equal to i thorn j



regardless of their i This could occur for example if some placesmandate flu shots or other preventive treatments with similarcost for all patients More subtly our decompositions of geo-graphic variation in log utilization give relatively more weightto differences in the bottom part of the utilization distributionthan a decomposition in levels would In Section IV we presenta variety of specification and robustness checks that bear on theseissues

The assumption that patient and place effects are addi-tively separable also rules out the possibility that differenttypes of patients seek out different types of health care withina place This can be relaxed by allowing interactions between j

and patient observables as we explore later It can also bepartly addressed by examining whether the results vary whenarea j is defined at higher and lower levels of geography whichwe also explore later But our specification does not capture richermodels of behavior in which within appropriately defined areasobservationally similar patients seek out different types ofproviders

Third our approach relies fundamentally on the assumptionthat the j that are relevant for movers are the same as those thatare relevant for nonmovers If movers differ in ways that changethe relevant place effects our decompositions would apply only tovariation in utilization among movers rather than to the popula-tion as a whole

Finally our model does not allow for the possibility that i ina given period is a function of past values of yit If for examplepatients in high-utilization areas become accustomed to visitingthe doctor frequently and receiving a large number of tests whenthey do they might continue to demand these services postmoveIn this case variation across areas in current i could partly becaused by the influence of j in the past We discuss the possibilityof such habit formation and evidence that suggests it may besmall in Section IVA However the fact that we focus on olderpatients means that we cannot rule out habit formation over longhorizons If i depends on consumption of health care early in lifefor example some of what we attribute to patients may reflectsupply-side differences in the past The long-term effect of supply-side changes could then be larger than our estimates wouldsuggest



IID Event-Study Representation

To visualize the way utilization changes when patientsmove we define an alternative lsquolsquoevent-studyrsquorsquo representation ofequation (2)

To build intuition it again helps to start with the simple casewhere t and xit are set to zero and where our panel of movers isbalanced in the sense that each mover is observed for the samenumber of years pre- and postmove If all movers had the sameorigin j0 and destination j we could construct an event study bysimply plotting the average of y for movers by relative year r iteth THORNWhen origins and destinations vary however this plot would notbe very informative If the flow from any j0 to j were equal to theflow from j to j0 for example we would expect the graph to showno change around the move even if the absolute values of theunderlying changes on move were large

To produce a more informative plot we would like to scale yso that the direction and magnitude of the jump on move areinformative regardless of the origin and destination For amover i whose origin and destination areas are o ieth THORN and d ieth THORN re-spectively we denote by i the difference in average log utilizationbetween the moverrsquos destination and origin

i frac14 yd ieth THORN yo ieth THORNeth4THORN

and we let Siplace frac14 Splace d ieth THORN o ieth THORNeth THORN and Si

pat frac14 Spat d ieth THORN o ieth THORNeth THORN Fol-lowing Bronnenberg Dube and Gentzkow (2012) we define formover i

yscaledit frac14

yit yo ieth THORN

i

Note that yscaledit will be 0 if the moverrsquos utilization is equal

to the average in his origin 1 if it is equal to the average inhis destination and between 0 and 1 if the moverrsquos utilizationfalls between the two If the model is correct the expectation ofyscaled

it should be flat before and after move and the jump onmove will be equal to the average value of Si

place acrossmovers Plotting the averages of yscaled

it by relative year wouldthus produce an event-study figure with a direct interpretationin terms of the model quantities of interest The larger thejump in yscaled

it on move the greater the share of geographicvariation we would attribute to place and the smaller theshare we would attribute to patients



To implement this in the full model we must deal with threeadditional complications First we need to allow for the controlst and xit Second our panel is not balanced and so changes in thecomposition of movers could introduce pre- or post-trends into theevent-study figure To avoid this we need to control for the indi-vidual fixed effects i explicitly Third the difference i can bevery small in some cases which would make the simple averageof yscaled

it poorly behaved This leads us to prefer a regression im-plementation that avoids dividing by i

Observe that we can rewrite equation (2) for movers as

yit frac14 i thorn o ieth THORN thorn Ir iteth THORNgt0Siplacei thorn t thorn xitthorn eiteth5THORN

where Ir iteth THORNgt0 is an indicator variable for relative yeargreater than zero Combining i thorn o ieth THORN into a single patientfixed effect ~i replacing i with its sample analogue i (calcu-lated based on both movers and nonmovers in the destinationand the origin) and parameterizing the interaction with i as aflexible function of relative year yields

yit frac14 ~i thorn r iteth THORNi thorn t thorn xitthorn eiteth6THORN

This is the event-study equation we take to the data The rel-ative-year specific coefficients r iteth THORN are the parameters of inter-est they measure changes in yit in years around the movescaled relative to i If the sampling error in i is ignorable13

and heterogeneity in Siplace is orthogonal to the other variables

in the model the plot of the r iteth THORN will have a precise interpre-tation similar to that of the average yscaled

it in the simple casethe plot should be flat before and after move and jump on moveby a weighted average of Si

place

III Data and Summary Statistics

IIIA Data and Variable Definitions

Our primary data source is a 20 random sample ofMedicare beneficiaries (lsquolsquopatientsrsquorsquo) from 1998 through 200814

These data contain approximately 13 million patients For each

13 Because the number of nonmovers we observe in each HRR is large sam-pling error in i is small We show in Online Appendix Section 45 that accountingexplicitly for noise in i has no impact on our event-study results

14 The sample is a panel defined by taking all Medicare beneficiaries in eachyear whose Social Security number ends in either 0 or 5 The sample thus varies



patient we observe information on all Medicare claims for inpa-tient care outpatient care and physician services For eachclaim the data include information on the diagnosis type andquantity of care provided and the dollar value reimbursed byMedicare We also observe demographic information for each pa-tient including age sex race and zip code of residence definedas the address on file for Social Security payments as of March 31of each year To match the timing with which we observe patientsrsquoresidence we define all outcome variables for year t to be aggre-gates of claims from April 1 of year t through March 31 of yeart + 115

Our primary outcome variable is based on an index of overallhealth care utilization by individual by year which we refer tosimply as lsquolsquoutilizationrsquorsquo To construct the measure we followGottlieb et al (2010) in adjusting total annual expenditure forregional variation in prices Online Appendix Section 2 describesthe construction of the measure in detail We prefer to focus onutilization quantities and set aside variation in administrativelyset prices because the drivers of the latter are different and betterunderstood16

In our main specifications we define the outcome yit to be thelog of utilization plus 1 which we refer to simply as lsquolsquolog utiliza-tionrsquorsquo As discussed in Section IIC we prefer a log specificationboth economically and econometrically We explore other func-tional forms in the robustness section later We also examine anumber of other outcome measures including subcategories ofutilization and indicators for particular treatments which aredefined in more detail shortly

Our geographic unit of analysis is a Hospital Referral Region(HRR) as defined by the 1998 Dartmouth Atlas of Health CareThe 306 HRRs are collections of zip codes designed to approxi-mate markets for tertiary hospital care17 Consistent with the

from year to year but a given patient remains in the sample as long as they areenrolled in Medicare

15 We include data from the first few months of 2009 to compute outcomes forour final sample year (t = 2008) which runs from April 2008 to March 2009

16 We show in the robustness analysis that our main conclusions areunchanged if we use total annual expenditure in place of utilization

17 See wwwdartmouthatlasorgdownloadsgeographyziphsahrr98xls andhttpwwwdartmouthatlasorgdownloadsmethodsgeogappdxpdf Each HRRconsists of a collection of zip codes that contain at least one hospital that performsmajor cardiovascular procedures and neurosurgeries Zip codes are grouped into an



existing literature we define average log utilization and otheroutcomes for an HRR j to include all claims by residents of jregardless of the location of the claims themselves On averageabout 16 of claims occur outside a patientrsquos HRR of residence

We define patients to be lsquolsquononmoversrsquorsquo if their HRR of resi-dence is the same throughout our sample period We define pa-tients to be lsquolsquomoversrsquorsquo if their HRR of residence changes exactlyonce Our baseline analysis excludes patients whose HRR of res-idence changes more than once We show in Online AppendixSection 43 that including multiple movers does not substantivelychange our estimates

In some of our analyses we compare movers to a matchedsubsample of nonmover patient years chosen to match as closelyas possible the characteristics of our mover sample For eachmover in our data in each calendar year we randomly draw anonmover in the same year in the moverrsquos origin HRR who sharesthe moverrsquos gender race and five-year age bin The union of theselected nonmover patient-years forms the lsquolsquomatched sample ofnonmoversrsquorsquo we refer to below

IIIB Sample Restrictions and Summary Statistics

From our original sample of 13 million patients we retain a25 random sample of nonmovers along with all movers We thenrestrict the sample to the 88 of patient-years where patients arebetween 65 and 99 years old exclude 20 of the remainingpatient-years for patients enrolled in Medicare Advantage (forwhom we do not observe claims) and exclude the remaining 7of patient-years for patients who do not have Medicare Part A orB coverage in all months (including for example patients whoenroll mid-year in the year they turn 65) Finally among patientswhose HRR of residence changes at least once we exclude the18 whose HRR of residence changes more than once as wellas the 35 of the remaining movers whose share of claims intheir destination HRR among claims in either their origin or

HRR based on where the highest proportion of cardiovascular procedures are re-ferred Each HRR must have a population of at least 120000 We drop roughly 2 ofpatient-years whose zip codes do not match the 1998 HRR definitions



destination HRR is not higher by at least 075 in the postmoveyears relative to the premove years18

When we compute HRR averages such as the sample ana-logue yj of yj we omit movers in their move year and we weightnonmovers by four to account for our sampling procedure AllHRR averages are computed by first averaging across individualsin the HRR in each year and then taking a simple average acrossyears

Our final sample includes 25 million patients of whom ap-proximately 05 million are movers Table I reports summary sta-tistics separately for movers and nonmovers The characteristicsof the two groups are broadly similar although there are somedifferences Relative to nonmovers movers are slightly morelikely to be female white and older and more likely to live ini-tially in the South or West rather than the Midwest or NortheastAverage annual utilization in both groups is roughly $7500 peryear with a standard deviation of about $10000 and 6 of ob-servations equal to 0 Health care utilization is notoriously right-skewed the median across both groups is about $4300 and the90th percentile is almost $18000

There are a variety of reasons that individuals may enter orexit the sample including death entering or exiting MedicareAdvantage and entering or exiting our 65ndash99 age window Theaverage nonmover in our sample is observed for 63 years (out of apossible 11) and the average mover for 75 years The differenceis partly mechanical due to the fact that we must observe a pa-tient for at least two years to classify them as a mover About athird of patients die during our sample period and about 20enter or exit at some point due to enrollment in Medicare

18 The change in claim share is not defined for movers who do not have at leastone claim both pre- and postmove We exclude these cases if (i) they have no post-move claims and a premove destination claim share greater than 005 (ii) they haveno premove claims and a postmove destination claim share less than 095 Theclaims data suggest several explanations for why some movers do not satisfy ourchange in claim share criterion In a large share of cases the geographic distribu-tion of claims remains roughly the same before and after the recorded move sug-gesting that the patient changed the address on file with Social Security withoutchanging their residence This could occur if they decided to have their SocialSecurity checks sent to a child who was handling their finances for example Inother cases patients appear to have multiple residences both before and after themove with the share of claims in the destination increasing postmove by an amountless than our 075 threshold We show in Online Appendix Section 43 that ourresults are robust to alternative ways of defining movers



Advantage In our robustness analysis we discuss possible biasesdue to selective attrition and show that our results are robust tosome alternative ways of handling it Mortality rates are broadlysimilar for movers and nonmovers alleviating concern that ouranalysis of movers might miss end-of-life expenditures

Figure I shows the distribution of average annual utilizationacross HRRs The mean HRR has average utilization of $6629per person per year with a standard deviation of $779 The rank-ing of HRRs by utilization is reasonably stable over time thecorrelation between an HRRrsquos rank in the first half of our

TABLE I

SUMMARY STATISTICS

(1) (2)Nonmovers Movers

Female 057 060White 086 088Age first observed

65ndash74 067 05975ndash84 024 03185 009 009

First observed residenceNortheast 020 017South 039 041Midwest 026 019West 016 023

Annual utilizationMean $7796 $7399Std dev $12690 $9567Share of patient-years with zero 006 006

Number of chronic conditionsMean 298 330Std dev 215 206Share of patient-years with zero 018 015

Average of years observed 626 745Share who die during sample 035 032Share of patient-years excluded because

patient is in Medicare Advantage that year 018 020 of patients 2033096 497097 of patient-years 12730766 3702189

Notes Rows for female white age first observed and first observed residence report the shares ofpatients with the given characteristics among movers and nonmovers Patient-years in MedicareAdvantage are excluded from the baseline sample The denominator for the row lsquolsquoShare of patient-yearsexcluded because patient is in Medicare Advantage that yearrsquorsquo is the sample of all movers and 25 ofnonmovers before any other sample restrictions In all other rows the sample is the baseline sample of allmovers and 25 of nonmovers (N = 16432955 patient-years)



sample (1998ndash2003) and the second half of our sample (2004ndash2008) is 09 We show in Online Appendix Figure 9 that if wedivide HRRs into quintiles by utilization the evolution of utiliza-tion for the different quintiles is roughly parallel These facts areconsistent with prior literature showing patterns of geographicvariation in health care utilization have been relatively stablesince the early 1990s (Weinstein et al 2004 Rettenmaier andSaving 2009 Chandra Sabik and Skinner 2011)

4666 minus 59205920 minus 64156415 minus 68086808 minus 73017301 minus 9985

00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000

Average Utilization in HRR

Mean = 6629SD = 779

FIGURE I

Distribution of Utilization across HRRs

Map shows the distribution of level utilization in quintiles Lower andupper limit of each quintile are displayed in the legend The sample is allmovers and nonmovers (N = 16432955 patient-years) Histogram displays thedistribution of average utilization by HRR We first average utilization acrossindividuals within each HRR-year upweighting nonmovers by four and thentake a simple average within HRR across years



Online Appendix Section 31 presents additional summarystatistics for movers The average distance moved is 588 mileswith a median of 357 miles and a standard deviation of 616 milesRoughly 68 of moves cross state boundaries and 50 crosscensus division boundaries Moves to Florida account for 12 ofall moves and moves to Arizona or California account for an ad-ditional 12 we show in Online Appendix Table 9 that our re-sults are robust to excluding moves to Florida Arizona andCalifornia We also show the distribution of movers across differ-ent destination HRRs The median HRR receives 1133 moversthe range of movers into an HRR is from 135 to 12797

Finally we examine the time-varying correlates of movingOnline Appendix Figure 4 shows that moving is correlated withan increase in utilization including a spike in utilization in theyear of move We also report evidence from the Health andRetirement Study (HRS) on the reasons older Americans moveThe study is a nationally representative (approximately bian-nual) longitudinal survey of Americans over the age of 50 Welimit the HRS sample to individuals aged 65 and over anddefine movers as individuals who move across HRRs The mostcommon self-reported reasons for moving are to be lsquolsquoNearwithchildrenrsquorsquo (31) lsquolsquoHealth problems or servicesrsquorsquo (13) and to belsquolsquoNearwith relatives or friendsrsquorsquo (10) Analysis of the HRS paneldata shows that significant predictors of moving include beingwidowed and retiring Declines in self-reported health status donot predict moving in the panel

IV Main Results Patient versus Place

IVA Event Study

We begin with two figures that illustrate the variation drivingour event study Figure II shows a moverrsquos claims in her destina-tion HRR as a share of those in either her origin or her destina-tion by relative year The figure shows a sharp change in the yearof the move with only a small share of claims in the destinationpremove or in the origin postmove19 The claim share in the year ofthe move (relative year 0) is close to 05 consistent with moves

19 In Online Appendix Section 43 we show that our results are robust to ad-justing for the small amount of apparent measurement error in the timing of movesand to a range of alternative definitions of movers



being roughly uniform throughout the year Figure III shows thedistribution of i the average log utilization in a moverrsquos destina-tion minus the average log utilization in her origin The meanvalue of i is close to zero and the distribution is roughly symmet-ric implying that moves from low- to high-utilization HRRs are ascommon as moves from high to low The standard deviation is025 and there are a significant number of moves for which theabsolute value of the difference is greater than 05

As a first look at the way utilization changes around movesFigure IV plots the change in log utilization (the average two tofive years postmove minus the average two to five years premove)against the destinationndashorigin difference in log utilization i If allgeographic variation were due to place effects we would expectthis plot to have a slope of 1 If all variation were due to patienteffects we would expect this plot to have a slope of 0 One minusthe actual slope is an estimate of a weighted average of the pa-tient share Si

patFigure IV shows that the slope is in fact 063 suggesting an

average patient share of roughly 037 The relationship is sym-metric above and below zero and strikingly linear This providesstrong support for our additively separable model which implies

02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR

minus10 minus9 minus8 minus7 minus6 minus5 minus4 minus3 minus2 minus1 0 1 2 3 4 5 6 7 8 9

Year Relative to Move

FIGURE II

Share of Claims in Destination by Relative Year

Figure shows the share of a moverrsquos claims located in their destinationHRR among those in either their origin or their destination HRR Thesample is all movers (N = 3702189 patient-years)



that the absolute change in log utilization when patients movefrom j to j0 should be the same as when patients move from j0 to jThese patterns are also consistent with the relative importance ofpatients being similar across origin-destination pairs

We also plot with an the average change in log utilizationover the same period for our matched sample of nonmovers towhom we assign i frac14 020 That this point and all points for movershave y values greater than zero reflects the positive time and agetrends in utilization That the point for nonmovers lies below theones for movers with iamp0 shows that moving is associated withan increase in utilization on average This main effect of movingwill be absorbed by our relative year indicators r iteth THORN We presentadditional descriptive evidence on the main effects of moving inOnline Appendix Section 31

Figure V shows how premove utilization of movers comparesto utilization of nonmovers in their origin HRRrsquos The plot is iden-tical to Figure IV except that the variable on the y-axis is now theaverage difference between log utilization of movers two to fiveyears premove and that of their matched nonmovers in the same

00

20

40

60

8

Sha

re o

f Mov

ers

minus15 minus1 minus5 0 5 1 15

DestinationminusOrigin Difference in Average Log Utilization

Mean = minus0010SD = 0251

FIGURE III

Distribution of Destination-Origin Difference in Log Utilization

Figure shows the distribution across movers of the difference i in averagelog utilization between their origin and destination HRRs The sample is allmovers (N = 3702189 patient-years)

20 See notes to Figure IV for details on this matching



years The plot has a small upward slope suggesting that pa-tients who will move to a high-utilization HRR have relativelyhigher premove utilization than those who will move to a low-utilization HRR This slope is an order of magnitude smallerthan the slope in Figure IV Any systematic differences of thiskind in the average utilization of movers will be absorbed by ourpatient fixed effects i

02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5


Movers Matched NonminusMovers

Slope=63

FIGURE IV

Change in Log Utilization by Size of Move

Figure shows the change in log utilization before and after move For eachmover we calculate the difference i in average log utilization between theirorigin and destination HRRs then group i into ventiles The x-axis displaysthe mean of i for movers in each ventile The y-axis shows for each ventileaverage log utilization two to five years postmove minus average log utilizationtwo to five years premove The line of best fit is obtained from simple OLSregression using the 20 data points corresponding to movers and its slope isreported on the graph The sample is all mover years between two and fiveyears premove and between two and five years postmove (N = 1919137 pa-tient-years) For comparison we also compute the average change in log utili-zation for a sample of matched nonmovers which we show with the markeron the graph Specifically for each mover in our data in each calendar year werandomly draw a nonmover in the same year in the moverrsquos origin HRR whoshares the moverrsquos sex race and five-year age bin the union of the selectednonmover patient-years forms the matched sample



Our main event-study results are shown in Figure VI whichplots estimated coefficients r iteth THORN from equation (6)21 Since thesecoefficients are only identified up to a constant term we normal-ize the value for r i teth THORN frac14 1 to 0 The figure shows a sharp dis-continuous jump at the time of the move from 0 to approximately05 As discussed 1 minus the size of this jump can also be inter-preted as an estimate of a weighted average of Si

pat This figurethus implies a patient share of roughly a half

minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove

minus5 0 5DestinationminusOrigin Difference in Average Log Utilization

Slope=096

FIGURE V

Premove Differences in Log Utilization

Figure shows the level of premove log utilization for movers relative tononmovers by the size of their subsequent move i For each mover we calcu-late the difference i in average log utilization between their origin and desti-nation HRRs then group i into ventiles The x-axis displays the mean of i formovers in each ventile The y-axis shows for each ventile the average of differ-ence in log utilization between mover and matched nonmover patient-years twoto five years premove In Figure IV we describe the construction of the matchedsample of nonmovers The line of best fit is obtained from simple OLS regres-sion using the 20 data points and its slope is reported on the graph Thesample is all mover years between two and five years premove (N =1048843 patient-years)

21 For computational ease all of the event studies we report are estimated onthe sample of movers only We show in Online Appendix Figure 16 that includingnonmovers does not affect the analysis



Under the assumptions of our model the plot should be flat inthe years before and after the move In practice the plot shows nopost-trend and a small but statistically significant pretrend Thistrend could reflect systematic changes in log utilization of moversrelative to nonmovers Because our model restricts both HRR andpatient effects to be time constant it could also pick up HRR-specifictrends that are the same for movers and nonmovers but happen tobe correlated with moversrsquo i In our robustness analysis we exploreextensions that allow our fixed effects to change over time We alsoallow arbitrary pre- and postmove trends for movers by using dataonly from the years just before or after the move in the spirit of a

minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study

Figure shows the coefficients ~ r iteth THORN estimated from equation (6) The coeffi-cient for relative year 1 is normalized to 0 The dependent variable yit is logutilization xit consists of indicator variables for five-year age bins The dashedlines are upper and lower bounds of the 95 confidence interval We constructthis confidence interval using a two-step procedure In the first step for eachHRR j we construct the asymptotic distribution of yj which is a normal dis-tribution with mean j and standard deviation j calculated from the data Inthe second step we bootstrap equation (6) with 50 repetitions drawn at thepatient level making a random draw from the distribution of yj for eachmoverrsquos origin and destination to construct their i for each repetition Thesample is all movers (N = 3702189 patient-years)



regression discontinuity Finally we show that the event studieslook similar when estimated on various balanced panels

One source of the patient heterogeneity we measure could behabit formation in the sense of Becker and Murphy (1988) patientpreferences today could be a function of patientsrsquo utilization in thepast For example patients who build a habit of getting regularcheckups or flu shots may continue to do so wherever they go Moremechanically patients who receive diagnoses or begin treatmentin high-utilization areas may continue their treatment even afterthey move to low-utilization areas This would affect the interpre-tation of our results since patient characteristics today wouldpartly reflect the impact of place characteristics in the past

Several features of Figures IV and VI suggest that the role ofhabit formation may be limited First stories such as continuingaggressive treatments started premove would tend to predict a lotof persistence for those moving from high- to low-utilization areasand less persistence for those moving from low to high In factFigure IV shows that for any given magnitude jij of the differencebetween origin and destination log utilization changes in log uti-lization look symmetric for moves up and down22 Second a signa-ture of most models of habit formation is that utilization shouldcontinue to adjust toward average behavior in the destination inthe years following a move This is the key pattern that identifieshabit formation in Bronnenberg Dube and Gentzkow (2012) forexample However Figure VI shows remarkably little evidence ofpostmerge convergence23 Log utilization jumps discretely on movebut remains almost perfectly flat for up to nine years thereafterand this remains true whether we look at moves from low to high-utilization areas or moves from high to low (see Online AppendixFigure 11) Finally many habit-formation models would predictless adjustment for older patients since they have accumulatedlarger stocks of past experience we show in Online AppendixFigure 12 that there is no evidence of such systematic differencesby age As already mentioned however none of this evidence rulesout habit formation occurring at younger ages or over longer timespans than we observe in our data

22 As further evidence Online Appendix Figure 11 shows similar changes inutilization on move in event-study plots separately for moves up eth i gt 0THORN and movesdown eth i lt 0THORN

23 Online Appendix Figure 8 shows that this remains true when we estimateour event study using a balanced panel



IVB Model Estimates

We exploit the variation captured in Figure VI to estimateequation (2) We use the estimates to quantify the roles of pa-tients and of places in explaining geographic variation in log uti-lization We present three main types of decompositions

Table II which we consider the central set of results inthe article presents an additive decomposition of the differencebetween high- and low-utilization areas For different sets ofhigh- and low-utilization HRRs R and R0 we report the sample

analogue of the patient share Spat RR0eth THORN frac14yRyR0yRyR0

as well as the

components yR yR0 R R0 and yR yR0 Column (1) decomposes the difference between above-median

and below-median HRRs We find that 47 of the difference inaverage log utilization is due to patients This estimate is fairlyprecise we can reject a role for patients of more than about 52 orless than about 41

Other partitions of HRRs result in a similar patient sharePatients account for 41 of the difference between the top and

TABLE II

ADDITIVE DECOMPOSITION OF LOG UTILIZATION

(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis

Difference in average log utilizationOverall 0283 0456 0664 0817 0587 0667Due to place 0151 0271 0406 0461 0374 0466Due to patients 0132 0185 0258 0356 0213 0200

Share of difference due toPatients 0465 0405 0388 0435 0363 0300

(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700

Notes Table is based on estimation of equation (2) where the dependent variable yijt is log utilizationand the controls xit are indicators for age in five-year bins The adjusted R-squared from estimatingequation (2) is 0503 Each column defines a set of areas R and R0 In columns (1)ndash(4) these are basedon percentiles of average utilization yj The first row reports the difference in average utilization overallbetween the two areas yR yR0

the second row reports the difference due to place R R0

the third

row reports the difference due to patients yR yR0

The fourth row reports the share of the difference inaverage utilization between the two areas due to patient

Spat RR0eth THORN

which is the ratio of the third row to

the first row The last row reports the share of the difference in average utilization between the two areasdue to place

Splace RR0eth THORN

which is the ratio of the second row to the first row Standard errors (in

parentheses) are calculated using a bootstrap with 50 repetitions at the patient level The sample ismovers and nonmovers excluding relative year 0 (N = 16031875 patient-years)



bottom quartiles (column (2)) 39 of the difference between thetop and bottom deciles (column (3)) and 44 of the differencebetween the top and bottom 5 (column (4))24 The final two col-umns look at two cases discussed in the introduction McAllenrelative to El Paso and Miami relative to Minneapolis Herewe find that patients account for 36 and 30 of the differencesrespectively though precision naturally falls with these smallersamples

The magnitudes are consistent with the event-study analy-sis which suggested a patient share of 50 based on the jump inlog utilization from relative year 1 to 1 as well as with the slopeof Figure IV That the estimates are not identical reflects the factthat the additive decomposition is a slightly different experi-mentmdashanalyzing differences between two groups of HRRsrather than averaging Si

pat across all movers imdashthat the modeluses all pre- and postmove years rather than the on-impact effectof the move and that the model is estimated on both movers andnonmovers The stability of the patient share across differentpartitions is consistent with the linear relationship shown inFigure IV which implies that Spatethj j0THORN is not strongly correlatedwith yj yj0

We present a second alternative decomposition in Table IIIHere we ask what share of the cross-HRR variance in log utiliza-tion would be eliminated in a counterfactual where average pa-tient characteristics yj were equalized across HRRs This is

Svarpat frac14 1

Var j

Var yj

eth7THORN

Similarly the change if area fixed effects were equalized is

Svarplace frac14 1

Var yj

Var yj

Note that unlike Spat and Splace this is not an additive decompo-sition the sum of Svar

pat and Svarplace will not be 1 so long as Covethyj jTHORN

is nonzero In estimating the relevant variances and covarianceswe correct for sampling error using a split-sample approach25

24 Online Appendix Figure 19 shows the corresponding event studies for thevarious partitions shown in columns (1) through (4)

25 We randomly assign movers within each origin-destination pair and non-movers within each HRR to two approximately equal-sized subsamples and



We find that 56 of variance would be eliminated if patienteffects were equalized We find that 72 of variance would beeliminated if place effects were equalized We also find thatthere is a positive correlation between yj and j with patientswith a high demand for health care tending to sort to higher-uti-lization areas Because this correlation is positive Svar

pat and Svarplace

sum to more than 1The results thus far decompose geographic variation in log

utilization As discussed in Section IIC we believe modeling uti-lization in logs is appealing both economically and econometri-cally However even if our model is correctly specifiedquantifying the drivers of geographic variation in logs is a differ-ent exercise than quantifying the drivers of variation in levels asthe implicit weights on low- and high-utilization observations willbe different If the relative importance of demand and supplyfactors varied substantially across the distribution of

TABLE III

VARIANCE DECOMPOSITION OF LOG UTILIZATION

(1)

Cross-HRR variance of averageLog utilization 0035HRR effects 0015Patient effects 0010

Correlation of average HRR and patient effects 0353(0052)

Share variance would be reduced ifHRR effects were made equal 0717

(0014)Patient effects were made equal 0558

(0013)

Notes Results based on estimates of equation (2) The first row reports variance of y j which isestimated using the same specification as in Table II The second third and fourth rows report thevariance of j variance of yj and the correlation between j and yj respectively using a split-sampleapproach to correct for the (correlated) measurement error in j and yj The last two rows report theshare of the variance in cross-HRR utilization that would be reduced if HRR effects were made equalacross areas (S

var

place) and if patient effects were made equal across areas (Svar

pat ) Standard errors (in pa-rentheses) are calculated using a bootstrap with 50 repetitions at the patient level The sample size is thesame as in Table II

estimate equation (2) separately on each subsample We compute the variance of j

(or yj ) as the covariance between the j rsquos (or yj rsquos) estimated from the two subsam-ples The correlation between j and yj is computed from the variances of j and yj and the covariance between j and yj which we estimate as the average of thecovariances between j from one subsample and yj from the other subsample



utilizationmdashfor example because patient preferences were eithermore or less important in big-ticket end-of-life expenditures com-pared with low-cost routine caremdashlog and level decompositionscould give very different answers

To assess the importance of this issue we present a thirddecomposition Here we ask what our estimated (log) modelimplies about the drivers of geographic variation measured inlevelsmdashhow differences in level utilization would change ifeither yj or j were equalized across places The details of thisexercise and a complete set of results are presented in OnlineAppendix Section 42 A limitation to this exercise is that ourlsquolsquoadditive decompositionrsquorsquo into Spat RR0eth THORN and Splace RR0eth THORN is nolonger additive the difference between the high- and low-utilizationareas in levels is the product rather than the sum of the patientand place components and so the percentage changes when weequalize one or the other need no longer sum to 1 The resultssuggest that equalizing patient characteristics across areaswould reduce geographic differences by 27 whereas equalizingplace characteristics would reduce them by 72 As a separateexercise we show in Online Appendix Table 8 that simply esti-mating the model in levels also yields a somewhat lower patientshare (23) We also show in the same table that if we define theoutcome to be whether the patient is in the top X of the nationaldistribution of utilization the patient share ranges from 17 to51 with some trend toward lower patient shares at the top ofthe distribution Comparing these results directly to our mainestimates is difficult but they suggest that the relative impor-tance of patients may be somewhat lower at higher percentiles ofthe utilization distribution

Robustness We explore the robustness of our results along anumber of dimensions in Online Appendix Section 4 Here webriefly summarize the main conclusions

First we show that our results are robust to using observa-tions for movers only in small windows around the move year inthe spirit of a regression discontinuity This suggests that our re-sults are not driven by differential utilization trends for moversthat are related systematically to the origin and destination

Second consistent with our assumption that patient andplace effects are time-constant we show that our conclusionsare similar if we estimate the model separately on subperiods of



the data or allow explicitly for patient and place-specific trendsin utilization

Third we relax our additive separability assumption by allow-ing different place effects j for each quartile of patient age We seethis as a step toward a model where j and i interact since patientage is the one of the strongest observable predictors of patientdemand This model yields similar results As further support foradditive separability we follow Card Heining and Kline (2013)and show that the increase in R2 when we fully saturate the modelwith patient-place fixed effects is relatively small

Fourth we show that the qualitative conclusions are robustto excluding all observations for patients who exit or enter thesample due to death or HMO status suggesting that any biasfrom selective attrition is likely small

Fifth to assess robustness to our market definition we showthat results are similar if we define markets at higher or lowerlevels of geography and if we include only movers who cross statelines or census region boundaries

Finally we explore robustness to other implementation de-cisions This includes using alternative definitions of movers(such as including individuals who move multiple times withinour sample period or varying the criteria used to define validmoves) using alternative dependent variables (such as expendi-ture rather than utilization or other functional forms for utiliza-tion) excluding nonmovers from the estimation altogetherdropping age and relative year as covariates and excludingmoves to Florida Arizona and California

IVC Other Outcomes

In Table IV we replicate our main decomposition results forthe following alternative components of annual utilization dum-mies for whether a patient has (i) seen a primary care physician(ii) seen a specialist (iii) been hospitalized or (iv) visited the emer-gency room the log of 1 plus (i) the number of diagnostic tests thepatient received (ii) the number of imaging tests the patient re-ceived (iii) the number of preventive care measures the patientreceived (iv) the number of different doctors the patient saw (v)inpatient utilization (vi) outpatient utilization (vii) emergencyroom utilization and (viii) other utilization Detailed definitionsof these measures are provided in Online Appendix Section 22For each row we reestimate equation (2) with yijt defined to be the



measure in question We then report the sample mean of the out-come measure the difference in the mean of the outcome measurebetween above- and below-median HRRs and the share of thisdifference due to patients where the partitions into above- andbelow-median are defined based on the outcome measure in ques-tion and so vary across rows

The results suggest that the patient share varies from a lowof 009 for diagnostic tests to a high of 071 for emergency roomvisits26 A natural hypothesis is that this variation reflects thedegree of patient involvement in decision making the outcomesfor which we find a large patient sharemdashpreventive care andemergency room visits for examplemdashtend to be ones where wemight think patients have a significant amount of discretionwhereas the outcomes for which we find a smaller patientsharemdashdiagnostic tests imaging tests and inpatient care for ex-amplemdashtend to be ones where we might think more discretion lieswith physicians

We can make this hypothesis more precise through a slightrevision of our model In equation (1) we assume that physiciansunilaterally choose the level of care yit based on their perception~ujethTHORN of the patientrsquos utility Suppose instead that for a particularoutcome m (diagnostic tests emergency room visits etc) thequantity chosen maximizes a combination of the physicianrsquosand the patientrsquos incentives

ymit frac14 arg max

ymu yjhitieth THORN thorn 1 meth THORN ~uj yjhitieth THORN PCjt yeth THORN

eth8THORN

Here the parameter m reflects the weight of the patient relativeto the physician in decision making or bargaining The patientshare Spat j j0eth THORN is increasing in m holding other parameters ofthe model constant as long as Spat j j0eth THORN 2 01frac12 For outcomessuch as emergency room visits where the patient may decidewhether to go unilaterally we might expect m to be high con-sistent with the high observed patient share for outcomes suchas diagnostic or imaging tests we would expect m to be lower

26 Online Appendix Figure 10 shows event-study graphs parallel to Figure VIfor each of these outcomes As with our main utilization measure we observe ineach case large discontinuous changes on move and relatively small trends pre- orpostmove The size and direction of the pre- and post-trends vary somewhat acrossoutcomes and so as a robustness check Online Appendix Table 7 shows the resultsare similar in magnitude when we limit the estimation sample for movers to oneyear pre- or postmove



consistent with the low observed patient share None of thesemodifications substantively change our conclusions

V Correlates of Patient and Place Effects

Both the patient and place components of utilization couldreflect a range of underlying economic primitives In SectionIVA we argued that the evidence looks inconsistent with patienteffects being primarily driven by either habit formation or persis-tence of specific treatments started premove Here we provideadditional evidence on mechanisms by exploring the observablecorrelates of our estimated place effects eth jTHORN and average patienteffects ethyj THORN

We focus on observables that proxy for the main demand andsupply factors suggested by the model in Section IIA We presentdetailed definitions data sources and summary statistics forthese measures in Online Appendix Section 32

For places we are interested in proxies for net private costs(PCjt) and physician beliefs (lj) The literature discussed inSection IIB suggests a number of economic drivers of theformer including physical capital human capital and organiza-tional form As proxies we use hospital beds per capita primarycare physicians per capita specialists per capita hospital qualityof care scores and the share of hospitals that are nonprofit Tocapture physician beliefs about appropriate practice style wedraw on survey-based measures from Cutler et al (2015)27

They present a sample of physicians with patient vignettes andask them to rate the likelihood they would recommend differentcourses of action We use the shares of primary care physiciansand cardiologists respectively in each HRR who recommendlevels of follow-up care greater (lsquolsquohigh follow-uprsquorsquo) or less (lsquolsquolowfollow-uprsquorsquo) than clinical guidelines suggest as well as the respec-tive shares who recommend aggressive (lsquolsquocowboyrsquorsquo) or less aggres-sive (lsquolsquocomforterrsquorsquo) end-of-life care

For patients we are interested in proxies for average treat-ment preferences ethiTHORN and health status ethhitTHORN in each HRR Weinclude survey-based measures of preferences collected byCutler et al (2015) the shares of patients in each HRR whowould request additional tests or specialist referrals even if not

27 We are grateful to the authors for sharing these data as well as the patientsurvey data discussed below



recommended by their primary care physician (lsquolsquohave unneededtestsrsquorsquo and lsquolsquosee unneeded cardiologistrsquorsquo) and the shares that wouldchoose relatively more or less aggressive end-of-life care (lsquolsquoaggres-siversquorsquo and lsquolsquocomforterrsquorsquo patient shares) We also use average pa-tient age race and sex from the Medicare claims data as wellas median household income and high school completion ratesfrom census data These demographics could proxy for both pref-erences and health Finally we include standard health mea-sures derived from the diagnoses recorded in Medicare claimsThese are the log of the Hierarchical Condition Categories(HCC) score and the log of 1 plus the count of a patientrsquos chronicconditions the Charlson Comorbidity Index and the count of apatientrsquos Iezzoni Chronic Conditions

An important challenge arises with interpreting these stan-dard health measures Although they are intended to capture theunderlying health status of a patient Song et al (2010) documentevidence that they in fact include a large measurement errorcomponent that varies systematically by place In places thattreat more aggressively a given underlying condition is morelikely to be diagnosed and more likely to be recorded in claimsconditional on being diagnosed This makes interpreting the cor-relations of the raw health measures with patient and place com-ponents difficult and it has led the literature on geographicvariation to be cautious about inferring the role of patient health

The empirical strategy developed in this article gives us away to extend the approach developed by Song et al (2010) andpurge these measures of the place-specific measurement errorcomponent We model a given observed health measure hmeas

ijt asthe sum of true health hit and a place-specific measurement error13ijt that in turn depends on place and year fixed effects and anorthogonal mean-zero error term This yields a model with thesame functional form as equation (2) where the dependent var-iable is now hmeas

ijt Estimates of this model allow us to differenceout the place-specific measurement error and recover a correctedestimate hit of patient health Details of this exercise and theassociated event-study figure showing changes in measuredhealth around moves are in Online Appendix Section 1 we esti-mate that about 50 of the geographic variation in measuredhealth reflects place-specific measurement error We use the cor-rected health measures throughout the analysis that follows

Before turning to the results we note that because the cor-relations we are looking at are cross-sectional they are likely to



Comforter PatientAggressive Patient

See Unneeded CardiologistsHave Unneeded Tests

Adjusted Log Iezzoni

Adjusted Log CharlsonAdjusted Log Chronic Conditions

Adjusted Log HCC Score

Average EducationMedian Family Income

Percent Female

Percent BlackAverage Age

Patient Characteristics

Comforter CardiologistsCowboy Cardiologists

Low Followminusup CardiologistsHigh Followminusup Cardiologists

Comforter PCPCowboy PCP

Low Followminusup PCPHigh Followminusup PCP

NonminusProfit Hospitals

Hospital Beds Per CapitaPCP Per Capita

Specialists Per CapitaHospital Compare Score

Place Characteristics

minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII

Correlates of Average Place Effects

Figure shows bivariate OLS regression results (left panel) and post-Lasso mul-tivariate regression results (right panel) of HRR-level place effects on a set of HRR-level characteristics All covariates have been standardized to have mean 0 andstandard deviation 1 To obtain the post-Lasso estimates we first run a Lasso re-gression on the full set of covariates with the penalty level chosen by a 10-fold cross-validation to minimize mean squared error We then run an OLS regression on theset of covariates chosen by the Lasso regression The sample in both panels is the 96HRRs for which all covariates are available Horizontal bars show 95 confidenceintervals Hospital Compare Score approximates hospital quality using timely andeffective care measures publicly reported by CMS Specialists per capita PCP percapita and hospital beds per capita count specialists primary care physicians andhospital beds per thousand residents respectively Non-profit hospitals is the per-cent of hospitals that are nonprofit Physician preference measures are drawn fromsurvey responses of PCPs and cardiologists from Cutler et al (2015) physiciansclassified as high follow-up or low follow-up recommend follow-up visits more (orless) frequently than clinical guidelines suggest physicians classified as cowboyrecommend care more intensive than guidelines suggest and those classified ascomforter recommend palliative care for severely ill patients Average age percentblack and percent female are computed among all patients in our baseline sample ofMedicare beneficiaries Median family income is the median income of householdsacross zip codes in each HRR taken from census data Average education is thepercent of the 25 and older population with a high school degree as computed fromcensus data The health variables are all the estimated patient components of aseries of health measures as described in Online Appendix Section 1 The patientpreferences variables are drawn from Cutler et al (2015) and detail Medicare ben-eficiariesrsquo survey responses to desired care in hypothetical cases have unneededtests and see unneeded cardiologists are the fraction of patients who would desiresuch treatment regimens aggressive patient provides the fraction of patients whowould like aggressive end-of-life care and comforter patient provides the fraction ofpatients who would like palliative end-of-life care even if it shortens their life



pick up the kind of long-term endogenous response of physicaland human capital to patient demand documented in Chandraand Staiger (2007) For example places that start out with sickerpatients may build more hospitals and hire doctors trained inaggressive treatments This would imply that the place compo-nent of utilization could be correlated with patient proxies such ashealth (as seen in Figure VII) and that the patient componentcould be correlated with place proxies such as number of hospitalsor share of cowboy physicians (as we will now see)

Figure VII summarizes the correlates of the estimated placeeffects ( j) Each row represents a different proxy variable Thepoints in the left panel are coefficients from separate bivariateOLS regressions The points in the right panel are coefficientsfrom post-Lasso multivariate OLS where a subset of the vari-ables shown in the left panel were selected in a first-stageLasso regression (Belloni and Chernozhukov 2013)28 All vari-ables are standardized so that the coefficients report the associ-ation between a 1 standard deviation change in the covariate andthe respective outcome The sample is limited to the 96 HRRs(representing about 60 of our baseline sample) for which themeasures of patient preferences and physician beliefs are avail-able from Cutler et al (2015) We present results using the full setof HRRs and omitting the Cutler et al (2015) measures in OnlineAppendix Table 13

The results show that HRRs with more hospital beds percapita a higher share of cowboy cardiologists and a lowershare of nonprofit hospitals are all associated with statisticallysignificantly higher place effects HRRs with more female pa-tients less educated patients and sicker patients by any measureare also associated with higher place effects The post-Lasso mul-tivariate regressions show that the shares of cowboy cardiologistsand the number of nonprofit hospitals remain significant as doespatient health The evidence is therefore consistent with pastliterature highlighting physician beliefs (Cutler et al 2015) andendogenous responses to patient demand (Chandra and Staiger2007) as key drivers

28 In the first stage we select variables using Lasso regression with a penaltychosen by 10-fold cross-validation to minimize the mean squared error in thesecond stage we report the coefficients and standard errors from multivariableOLS on the selected covariates Online Appendix Figures 17 and 18 show the setof covariates that would have been chosen for alternative values of the penalty



Figure VIII shows analogous results for the estimated averagepatient effects (yj ) Consistent with intuition we see that HRRswith older patients sicker patients (by any of the health mea-sures) female patients and higher socioeconomic status patients(as measured by income or education) are associated with statisti-cally significant higher average patient demand The first two ofthese seem clearly related to health status the latter two couldreflect differences in preferences We also see that places wherehospitals with higher quality of care (lsquolsquoComparersquorsquo) scores are alsoassociated with statistically significant higher average patient ef-fects an association that might reflect reverse causality The post-Lasso multivariate regression confirms the intuitive pattern forthe patient characteristics with sex education and health all re-maining significant several place characteristics also have signif-icant coefficients perhaps again partly reflecting reverse causality

One of the clearest takeaways from this analysis is that pa-tient health status is a strong predictor of both components ofutilization Were we to take the leap and view these relationships







Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII

Correlates of Average Patient Effects

Figure shows bivariate OLS regression results (left panel) and post-Lassomultivariate regression results (right panel) of HRR-level patient effects on aset of HRR-level characteristics Procedure and explanatory variables are thesame as in Figure VII



as causal the bivariate correlation in Figure VIII would suggestthat equalizing the chronic conditions measure acrossareas would reduce the gap in the patient component of log utili-zation between above- and below-median areas from 013 to00429 Recalling that the overall gap in utilization is 028(Table II) this implies that 013004

028 or 32 of the overall gapcould be attributed to the demand-side differences in observablepatient health

Following recent literature (eg Zuckerman et al 2010) wecan refine this estimate by using individual-level rather thancross-area variation to determine the coefficient linking healthstatus to utilization That is rather than using the bivariateOLS coefficient from Figure VIII we can use the coefficientfrom an individual-year-level panel regression of log utilizationof nonmovers on the adjusted health measure in question alongwith HRR and year fixed effects30 Here the main difference rel-ative to past work based on similar regressions is the ability tocorrect for measurement error in observed health

The results of this decomposition are shown in Table V Toemphasize the importance of our measurement error correctionthe first four rows show results using the raw unadjusted healthmeasures They suggest that failing to account for measurementerror would lead us to attribute between 40 and 80 of the logutilization difference between above- and below-median areas topatient health (The bottom end of this range is close to the estimatein Zuckerman et al 2010) The bottom four rows show results usingour adjusted measures We find that between 22 and 37 of thelog utilization difference can be explained by patient health

29 The standardized bivariate coefficient on log chronic conditions in FigureVIII is 0074 Converting this to unstandardized units by dividing by the standarddeviation of the health measure (0042) yields a coefficient of 176 The average ofthe log chronic conditions measures is 125 in above-median HRRs and120 in belowmedian HRRs so we predict that this gap falls by 176 (125 120) or 0088

30 That is for the sample of nonmovers we estimate

yitfrac14obsjthornobs

t thornhitthorneobsijt

where obsj obs

t and obsijt are respectively area fixed effects year fixed effects and

an error term distinct from those in equation (2) and hit is one of our correctedhealth measures constructed as described in Online Appendix Section 1 We thenestimate the share of the utilization gap between two areas R and R0 attributable to

patient health asu hRhR0

yRyR0

where hR is the average of hit in R with the average

computed analogously to yR



We have already stressed that this is a predictive rather thancausal relationship A further important caveat is that our mea-surement error correction only adjusts for place-level sources oferror Our corrected measures could still include other forms ofmeasurement error In particular patients with a preference forvisiting the doctor more often may be more likely to have theirconditions recorded in claims leading our adjusted measure topartly reflect differences in preferences

Our findings on health may be of some interest beyond thestudy of geographic variation We provide a way to quantify andcorrect for the measurement error identified by Song et al (2010)

TABLE IV

COMPONENTS OF UTILIZATION

(1) (2) (3)

Utilization measure

Mean ofutilizationmeasure

Abovebelowmedian

difference inutilizationmeasure

Share dueto patients

(1) Baseline log(utilization) 7193 0283 0465 (0027)(2) Seen a primary care physician 0884 0042 0452 (0027)(3) Seen a specialist 0815 0051 0322 (0024)(4) Any hospitalization 0226 0037 0410 (0034)(5) Any emergency room visit 0346 0045 0714 (0031)(6) Log( of diagnostic tests) 1449 0550 0092 (0008)(7) Log( of imaging tests) 0842 0220 0142 (0014)(8) Log( of preventive care measures)a 1376 0098 0611 (0018)(9) Log( of different doctors seen) 1525 0113 0392 (0016)(10) Log(inpatient utilization)b 2004 0340 0242 (0035)(11) Log(outpatient utilization)b 6890 0193 0358 (0031)(12) Log(emergency room utilization)b 2296 0352 0639 (0031)(13) Log(other utilization)b 3430 0957 0124 (0010)

Notes Table reports the share of the difference in utilization between above and below median HRRsdue to patients analogous to column (1) of Table II with the dependent variable yijt defined to be variouscomponents of utilization The partition of HRRs into above and below median groups is based on theutilization of individuals in the baseline sample and differs in each row according to the definition ofutilization used Column (1) reports the mean of the utilization measure for the given sample Column (2)reports the difference in the average utilization measure between above and below median HRRs(yR yR0 ) Column (3) reports the share of the difference in column (2) that is due to patients(Spat RR0eth THORN) All log outcome measures are the log of the outcome plus one Online Appendix Table 11shows the percent with zero for each of these outcomes Standard errors (in parentheses) are calculatedusing a bootstrap with 50 repetitions at the patient level The sample size is the same as in Table II

a of preventive care measures is a count of the number of the following preventive treatments thepatient received in the past year ambulatory care eye screening hemoglobin test lipid screen cardioscreen diabetes management pelvic screen bone mass test colorectal cancer screening and flu shot or inthe past two years mammogram Pap test and prostate cancer screening

bThese four measures are mutually exclusive and exhaustive



and show that the endogenous error component accounts for alarge share of the geographic variation in measured health Ourcorrected measures may have other applications in the large lit-erature that uses health status measures as inputs into riskadjustments

VI Conclusion

Looking at over-65 Medicare beneficiaries we find robustevidence that 40ndash50 of geographic variation in the log ofhealth care utilization is due to fixed characteristics of patientsthat they carry with them when they move The remaining 50ndash60 of variation is due to place-specific factors Patients mattermore for outcomes such as emergency room visits where theyhave substantial discretion and they matter less for outcomessuch as diagnostic and imaging tests where the physician isthe main decision maker

Our analysis of mechanisms suggests that a substantialshare of the patient component can in turn be attributed to dif-ferences across areas in average health status The remainder

TABLE V

VARIATION IN LOG UTILIZATION EXPLAINED BY PATIENT HEALTH

(1)Share of abovebelow

median utilization differencedue to patient health

Raw health measure(1) Log(HCC score) 0435(2) Log(Charlson Comorbidity Index) 0483(3) Log( of Iezzoni chronic conditions) 0483(4) Log( of chronic conditions) 0794

Corrected health measure(5) Log(HCC score) 0220(6) Log(Charlson Comorbidity Index) 0242(7) Log( of Iezzoni chronic conditions) 0256(8) Log( of chronic conditions) 0371

Notes Table reports shares of the difference in average log utilization between above-median andbelow-median utilization HRRs explained by observable patient health (S

obs

pat RR0eth THORN) Rows (5)ndash(8) use thepatient component (hit) of health measures estimated from Online Appendix equation (1) All log outcomemeasures are the log of the outcome plus 1 except the HCC score which is simply the log of the outcome(there are no zeros) Online Appendix Table 11 shows the percent with zero for each of these outcomesThe sample size is the same as in Table II in rows (1)ndash(7) In row (8) the sample also excludes the year1998 as chronic conditions are not observed in that year (N = 14598443 patient-years)



likely reflects a combination of both unmeasured health and pref-erences We find intriguing correlations with our estimated placeeffects which suggest among other things that areas withhigher place components of log utilization also have physicianswho believe in a more aggressive practice style fewer nonprofithospitals and a sicker population

These findings both reinforce and refine the conclusions ofthe existing literature on geographic variation On the one handthey confirm that supply-side variation plays an important roleand argue against claims that the role of the supply side is neg-ligible On the other hand they clarify that a large component ofvariation is in fact due to differences in patient health and sug-gest that both patient preferences and unmeasured health differ-ences may play a larger role than conventional wisdom wouldsuggest

Our results do not permit us to draw strong conclusionsabout welfare Variation on the supply side may be but neednot be inefficient The correlation of our place component withthe health of the local population suggests that at least some ofthe former may be due to endogenous responses of human andphysical capital In the presence of such adjustments we wouldexpect nontrivial supply-side variation even in a first-best worldConversely demand-side variation need not be fully efficient pa-tient demand may reflect misinformation or behavioral biasesthat could potentially be corrected by policy Continuing to drilldown on the efficiency implications of geographic variation re-mains an important goal for future work

Even without direct welfare implications some of the pat-terns of changes we observe have potential implications forpolicy For example the event-study evidence that place effectshave immediate rather than gradual effects on movers impliesthat changing supply-side factors such as doctor practice stylescould have large effects in the short run At the same time thelack of postmove convergence suggests that policies aimed atchanging demand-side factors such as patient preferences mayhave at best very gradual effects at least among the 65 and olderpopulation (Moses et al 2013)

Massachusetts Institute of Technology and National

Bureau of Economic Research

Stanford University and National Bureau of Economic

Research





Supplementary Material

An Online Appendix for this article can be found at QJEonline (qjeoxfordjournalsorg)

References

Aaronson Daniel lsquolsquoUsing Sibling Data to Estimate the Impact of Neighborhoodson Childrenrsquos Educational Outcomesrsquorsquo Journal of Human Resources 33(1998) 915ndash946

Abowd John M Robert H Creecy and Francis Kramarz lsquolsquoComputing Person andFirm Effects Using Linked Longitudinal Employer-Employee Datarsquorsquo USCensus Bureau Center for Economic Studies Longitudinal Employer-Household Dynamics Technical Papers 2002-06 2002

Abowd John M Francis Kramarz and David N Margolis lsquolsquoHigh Wage Workersand High Wage Firmsrsquorsquo Econometrica 67 (1999) 251ndash333

Ashton Carol M Nancy J Peterson Julianne Souchek Terri J Menke Hong-Jen Yu Kenneth Pietz Marsha L Eigenbrodt Galen BarbourKenneth Kizer and Nelda P Wray lsquolsquoGeographic Variations in UtilizationRates in Veterans Affairs Hospitals and Clinicsrsquorsquo New England Journal ofMedicine 340 (1999) 32ndash39

Baicker Katherine Sendhil Mullainathan and Joshua SchwartzsteinlsquolsquoBehavioral Hazard in Health Insurancersquorsquo Quarterly Journal of Economics130 (2015) 1623ndash1667

Baker Laurence C M Kate Bundorf and Daniel P Kessler lsquolsquoPatientsrsquoPreferences Explain a Small but Significant Share of Regional Variation inMedicare Spendingrsquorsquo Health Affairs 33 (2014) 957ndash963

Baker Laurence C Elliott S Fisher and John E Wennberg lsquolsquoVariations inHospital Resource Use for Medicare and Privately Insured Populations inCaliforniarsquorsquo Health Affairs 27 (2008) w123ndashw134

Barnato Amber R M Brooke Herndon Denise L Anthony PatriciaM Gallagher Jonathan S Skinner Julie P W Bynum and ElliottS Fisher lsquolsquoAre Regional Variations in End-of-Life Care IntensityExplained by Patient Preferences A Study of the US MedicarePopulationrsquorsquo Medical Care 45 (2007) 386ndash393

Becker Gary S and Kevin M Murphy lsquolsquoA Theory of Rational Addictionrsquorsquo Journalof Political Economy 96 (1988) 675ndash700

Belloni Alexandre and Victor Chernozhukov lsquolsquoLeast Squares after ModelSelection in High-Dimensional Sparse Modelsrsquorsquo Bernoulli 19 (2013) 521ndash547

Bronnenberg Bart J Jean-Pierre H Dube and Matthew Gentzkow lsquolsquoTheEvolution of Brand Preferences Evidence from Consumer MigrationrsquorsquoAmerican Economic Review 102 (2012) 2472ndash2508

Card David Joerg Heining and Patrick M Kline lsquolsquoWorkplace Heterogeneity andthe Rise of West German Wage Inequalityrsquorsquo Quarterly Journal of Economics128 (2013) 967ndash1015

Chandra Amitabh David Cutler and Zirui Song lsquolsquoWho Ordered That TheEconomics of Treatment Choices in Medical Carersquorsquo in Handbook of HealthEconomics vol 2 Mark V Pauly Thomas G McGuire and Pedro P Barroseds (Amsterdam Elsevier 2012)

Chandra Amitabh Lindsay Sabik and Jonathan S Skinner lsquolsquoCost Growth inMedicare 1992 to 2006rsquorsquo in Explorations in the Economics of Aging DavidA Wise ed (Chicago University of Chicago Press 2011)

Chandra Amitabh and Douglas O Staiger lsquolsquoProductivity Spillovers in HealthCare Evidence from the Treatment of Heart Attacksrsquorsquo Journal of PoliticalEconomy 115 (2007) 103ndash140



Chernew Michael Lindsay Sabik Amitabh Chandra Theresa Gibson andJoseph Newhouse lsquolsquoGeographic Correlation between Large-FirmCommercial Spending and Medicare Spendingrsquorsquo American Journal ofManaged Care 16 (2010) 131ndash138

Chetty Raj John N Friedman Soren Leth-Petersen Torben Nielsen andTore Olsen lsquolsquoActive vs Passive Decision and Crowd-Out in RetirementSavings Accounts Evidence from Denmarkrsquorsquo Quarterly Journal ofEconomics 129 (2014) 1141ndash1219

Chetty Raj John N Friedman and Jonah E Rockoff lsquolsquoMeasuring the Impacts ofTeachers II Teacher Value-Added and Student Outcomes in AdulthoodrsquorsquoAmerican Economic Review 104 (2014) 2633ndash2679

Chetty Raj John N Friedman and Emmanuel Saez lsquolsquoUsing Differences inKnowledge across Neighborhoods to Uncover the Impacts of the EITC onEarningsrsquorsquo American Economic Review 103 (2013) 2683ndash2721

Congressional Budget Office lsquolsquoGeographic Variation in Health Care SpendingrsquorsquoWashington DC Congress of the United States CBO 2008

Cooper Zack Stuart V Craig Martin Gaynor and John Van Reenen lsquolsquoThe PriceAinrsquot Right Hospital Prices and Health Spending on the Privately InsuredrsquorsquoNBER Working Paper 2015

Currie Janet and Bentley MacLeod lsquolsquoFirst Do No Harm Tort Reform and BirthOutcomesrsquorsquo Quarterly Journal of Economics 123 (2008) 795ndash830

Cutler David Jonathan Skinner Ariel Dora Stern and David WennberglsquolsquoPhysician Beliefs and Patient Preferences A New Look at RegionalVariation in Health Care Spendingrsquorsquo Harvard Business School WorkingPaper 15-090 2015

Dunn Abe Adam Hale Shapiro and Eli Liebman lsquolsquoGeographic Variation inCommercial Medical-Care Expenditures A Framework for DecomposingPrice and Utilizationrsquorsquo Journal of Health Economics 32 (2013) 1153ndash1165

Fernandez Raquel and Alessandra Fogli lsquolsquoFertility The Role of Culture andFamily Experiencersquorsquo Journal of the European Economic Association 4(2006) 552ndash561

Fisher Elliott S David E Wennberg Therese A Stukel Daniel J Gottlieb FL Lucas and Etoile L Pinder lsquolsquoThe Implications of Regional Variations inMedicare Spending The Content Quality and Accessibility of Care Part 1rsquorsquoAnnals of Internal Medicine 138 (2003a) 273ndash287

mdashmdashmdash lsquolsquoThe Implications of Regional Variations in Medicare Spending TheContent Quality and Accessibility of Care Part 2rsquorsquo Annals of InternalMedicine 138 (2003b) 288ndash298

Gawande Atul lsquolsquoThe Cost Conundrum What a Texas Town Can Teach Us aboutHealth Carersquorsquo The New Yorker 2009

Gottlieb Daniel J Weiping Zhou Yunjie Song Kathryn G Andrews JonathanS Skinner and Jason M Sutherland lsquolsquoPrices Donrsquot Drive Regional MedicareSpending Variationsrsquorsquo Health Affairs 29 (2010) 537ndash543

Lee Thomas H and James J Mongan Chaos and Organization in Health Care(Cambridge MA MIT Press 2009)

McPherson Klim P M Strong Arnold Epstein and Lesley Jones lsquolsquoRegionalVariations in the Use of Common Surgical Procedures Within andBetween England and Wales Canada and the United States of AmericarsquorsquoSocial Science amp Medicine A Medical Psychology amp Medical Sociology 15(1981) 273ndash288

Molitor David lsquolsquoThe Evolution of Physician Practice Styles Evidence fromCardiologist Migrationrsquorsquo University of Illinois at Urbana-ChampaignWorking Paper 2014 available at httpwwwbusinessillinoisedudmolitormoverspdf

Moses Hamilton David H M Matheson E Ray Dorsey Benjamin P GeorgeDavid Sadoff and Satoshi Yoshimura lsquolsquoThe Anatomy of Health Care in theUnited Statesrsquorsquo Journal of the American Medical Association 310 (2013)1947ndash1964

Neuman Tricia Juliette Cubanski Jennifer Huang and Anthony Damico lsquolsquoTheRising Cost of Living Longer Analysis of Medicare Spending by Age forBeneficiaries in Traditional Medicarersquorsquo Kaiser Family Foundation Report2015



Philipson Tomas J Seth A Seabury Lee M Lockwood Dana P Goldman andDarius N Lakdawalla lsquolsquoGeographic Variation in Health Care The Role ofPrivate Marketsrsquorsquo Brookings Papers on Economic Activity 41 (2010)325ndash361

Rettenmaier Andrew J and Thomas R Saving lsquolsquoPerspectives on the GeographicVariation in Health Care Spendingrsquorsquo National Center for Policy Analysis2009

Sheiner Louise lsquolsquoWhy the Geographic Variation in Health Care Spending CanrsquotTell Us Much about the Efficiency or Quality of Our Health Care SystemrsquorsquoBrookings Papers on Economic Activity (2014) 1ndash72

Skinner Jonathan lsquolsquoCauses and Consequences of Regional Variations in HealthCarersquorsquo in Handbook of Health Economics vol 2 Mark V Pauly ThomasG McGuire and Pedro P Barros eds (Amsterdam Elsevier 2011)

Song Yunjie Jonathan Skinner Julie Bynam Jason Sutherland andElliott Fisher lsquolsquoRegional Variations in Diagnostic Practicesrsquorsquo New EnglandJournal of Medicine 363 (2010) 45ndash53

Subramanian Usha Morris Weinberger George J Eckert Gilbert J LrsquoItalienPablo Lapuerta and William Tierney lsquolsquoGeographic Variation in Health CareUtilization and Outcomes in Veterans with Acute Myocardial InfarctionrsquorsquoJournal of General Internal Medicine 17 (2002) 604ndash611

Weinstein James N Kristen K Bronner Tamara Shawver Morgan and JohnE Wennberg lsquolsquoTrends and Geographic Variations in Major Surgery forDegenerative Diseases of the Hip Knee and Spinersquorsquo Health Affairs 23(2004) var81ndashvar89

Welch H Gilbert Sandra M Sharp Dan J Gottleib Jonathan S Skinner andJohn E Wennberg lsquolsquoGeographic Variation in Diagnosis Frequency and Riskof Death among Medicare Beneficiariesrsquorsquo Journal of the American MedicalAssociation 305 (2011) 1113ndash1118

Xu Fang Machell Town Lina Balluz William Bartoli Wilmon MurphyPranesh Chowdhury William Garvin Carol Pierannunzi Yuna ZhongSimone Salandy Candace Jones and Carol Crawford lsquolsquoSurveillance forCertain Health Behaviors Among States and Selected Local AreasmdashUnitedStates 2010rsquorsquo Morbidity and Mortality Weekly Report (MMWR)Surveillance Summaries Centers for Disease Control and Prevention 2013

Zuckerman Stephen Timothy Waldman Robert Berenson and Jack HadleylsquolsquoClarifying Sources of Geographic Differences in Medicare Spendingrsquorsquo NewEngland Journal of Medicine 363 (2010) 54ndash62



health care utilization that adjust for regional variation in ad-ministratively set prices (Gottlieb et al 2010) Higher area-levelutilization is not generally correlated with better patientoutcomes2

Understanding what drives geographic variation in utiliza-tion has important implications for policy If high-utilizationareas like McAllen and Miami are different mainly becausetheir doctorsrsquo incentives or beliefs lead them to order excessivetreatments with low return policies that change those incentivesor beliefs could result in savings on the order of several percent-age points of GDP (Congressional Budget Office 2008 Gawande2009 Skinner 2011) On the other hand if patients in high-utilization areas are simply sicker or prefer more intensivecare such policies could be ineffective or counterproductive

In this article we exploit patient migration to separatevariation due to patient characteristics such as health or pref-erences from variation due to place-specific variables such as doc-torsrsquo incentives and beliefs endowments of physical or humancapital and hospital market structure As a shorthand we referto the former as lsquolsquodemandrsquorsquo factors and the latter as lsquolsquosupplyrsquorsquo fac-tors3 To see the intuition for our approach imagine a patient whomoves from high-utilization Miami to low-utilizationMinneapolis If all of the utilization difference between thesecities arises from supply-side differences like doctor incentivesor beliefs we would expect the migrantrsquos utilization to drop im-mediately following the move to a level similar to other patientsof the low-utilization doctors in Minneapolis If all of the utiliza-tion difference reflects the demand-side reality that residents ofMiami are sicker we would expect the migrantrsquos utilization toremain constant after the move at a level similar to the typicalperson in Miami Where the observed utilization change falls

2 See Skinner (2011) for an extensive discussion The Congressional BudgetOffice (2008) concludes that high-spending areas lsquolsquotend to score no better and insome cases score worse than other areas do on process-based measures of qualityand on some measures of health outcomesrsquorsquo and that more intensive treatment inhigh-spending areas lsquolsquoappear[s] to improve health outcomes for some types of pa-tients but worsen outcomes for othersrsquorsquo

3 This corresponds to the usual definitions of demand and supply in mostcases but the correspondence is not perfect For example peer effects or sociallearning will generally be captured in our framework as a place-specific(lsquolsquosupplyrsquorsquo) factor since the composition of peers can change when a patientmoves but it would be more natural to think of them as shifters of demandrather than supply

































IIA Model











frac14 i thorn xit


8 Thus we can write

yit frac14 arg maxy










~PCjtethTHORN













yj yj0eth3THORN



yj yj0



IIB Discussion














t THORN


j For any two


IIC Identification








j0








consistently as

j

j0

j0

j








j and lead

j

j0

j0

j
























yscaledit frac14

yit yo ieth THORN

i



















place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References

















































































IIA Model











frac14 i thorn xit


8 Thus we can write

yit frac14 arg maxy










~PCjtethTHORN













yj yj0eth3THORN



yj yj0



IIB Discussion














t THORN


j For any two


IIC Identification








j0








consistently as

j

j0

j0

j








j and lead

j

j0

j0

j
























yscaledit frac14

yit yo ieth THORN

i



















place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References






















































frac14 i thorn xit


8 Thus we can write

yit frac14 arg maxy










~PCjtethTHORN













yj yj0eth3THORN



yj yj0



IIB Discussion














t THORN


j For any two


IIC Identification








j0








consistently as

j

j0

j0

j








j and lead

j

j0

j0

j
























yscaledit frac14

yit yo ieth THORN

i



















place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References




























































yj yj0eth3THORN



yj yj0



IIB Discussion














t THORN


j For any two


IIC Identification








j0








consistently as

j

j0

j0

j








j and lead

j

j0

j0

j
























yscaledit frac14

yit yo ieth THORN

i



















place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References




























































t THORN


j For any two


IIC Identification








j0








consistently as

j

j0

j0

j








j and lead

j

j0

j0

j
























yscaledit frac14

yit yo ieth THORN

i



















place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References




















































j0








consistently as

j

j0

j0

j








j and lead

j

j0

j0

j
























yscaledit frac14

yit yo ieth THORN

i



















place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References





















































j and lead

j

j0

j0

j
























yscaledit frac14

yit yo ieth THORN

i



















place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References

































































yscaledit frac14

yit yo ieth THORN

i



















place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References





























































place




































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References














































































TABLE I

SUMMARY STATISTICS



65ndash74 067 05975ndash84 024 03185 009 009











00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References





















































00

51

15

2

Sha

re o

f HR

Rs

4000 6000 8000 10000


Mean = 6629SD = 779

FIGURE I








IVA Event Study









02

55

75

1

Sha

re o

f Cla

ims

in D

estin

atio

n H

RR



FIGURE II








00

20

40

60

8

Sha

re o

f Mov

ers




FIGURE III







02

55

75

1

Cha

nge

in L

og U

tiliz

atio

n on

Mov

e

minus5 0 5



Slope=63

FIGURE IV







minus5

minus2

50

25

5

Mov

ersminus

Non

minusM

over

s D

iffer

ence

inA

vera

ge L

og U

tiliz

atio

n B

efor

e M

ove


Slope=096

FIGURE V







minus2

50

25

57

51

Log

Util

izat

ion

(Coe

ffici

ent)



FIGURE VI

Event Study











IVB Model Estimates




as well as the




TABLE II


(1) (2) (3) (4) (5) (6)Abovebelow

median

Top ampbottom

25

Top ampbottom

10

Top ampbottom

5McAllen amp

El PasoMiami amp

Minneapolis



(0027) (0029) (0026) (0025) (0161) (0088)Place 0535 0595 0612 0565 0638 0700



the third



Spat RR0eth THORN



Splace RR0eth THORN









Svarpat frac14 1

Var j

Var yj

eth7THORN


Svarplace frac14 1

Var yj

Var yj









pat and Svarplace



TABLE III


(1)





(0013)


var



















IVC Other Outcomes







ymit frac14 arg max


eth8THORN





























Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VII



















Percent Female











minus05 0 05 1

Bivariate OLS

minus05 0 05 1

PostminusLasso

FIGURE VIII













where obsj obs




yRyR0







TABLE IV


(1) (2) (3)

Utilization measure


Abovebelowmedian










VI Conclusion



TABLE V







obs











Research







References



















































I. Introduction - MIT Economics

Documents

Transcript of I. Introduction - MIT Economics