Guns and Butter? Fighting Violence with the
Promise of Development
Gaurav Khanna and Laura Zimmermann ∗
University of Michigan
January 2013
Abstract
There is a growing awareness in developing countries that government-fundedeconomic development programs may play an important role in the fight againstinternal security threats. This paper studies the impact of a large public-worksprogram, the National Rural Employment Guarantee Scheme (NREGS), on in-surgency-related violence in India. We set up a model of information procurementin which the government obtains information from civilians in exchange for thepromise of development. This allows the police force to effectively crack down oninsurgent activities. Empirically, we analyze the impact of NREGS by exploitingthe phasing-in of the program, and use Regression Discontinuity and Difference-in-Difference approaches. The results are robust across specifications and showthat in the short-run, fatalities and insurgent-related incidents increase when theprogram is introduced, which is consistent with the predictions of the model.
JEL: H12, H53, H56, I38
Keywords: public works program, National Rural Employment GuaranteeScheme, NREGA, NREGS, India, regression discontinuity design, terrorism, Nax-alites, Maoists, conflict, insurgency, civil war
∗email: [email protected] and [email protected]. We thank Raj Arunachalam and partici-pants of the University of Michigan Informal Development Seminar for valuable comments, feedbackand suggestions. We also thank Melissa Trebil for excellent research assistantship.
1
1 Introduction
Over 20% of countries experienced internal military conflicts over the course of the 1990s(Blattman and Miguel 2010). Violence remains a problem in a number of countriestoday and is concentrated in developing economies, where governments are looking forways to effectively deal with insurgencies. Traditionally, the main government strategyhas often been to rely on military and police strength, but there is a growing awarenessthat force alone may not be enough to solve internal conflicts in the long run: Insurgentsoften rely on the loyalty of the local population in their guerrilla tactics, and recruitmembers from economically marginalized groups of society. Cross-country data alsoshows a high correlation between poverty and conflict (Collier and Hoeffler 2007).
The direction of causality is unclear, however, as conflict can be an important reasonfor the persistence of poverty. And its impact on institutions, lives, physical and humancapital may contribute to the growing gap between richer and poorer nations (Collier,Elliot, Hegre et al. 2003). In this context, government programs aimed at povertyreduction and local development may play a crucial role. Yet, empirical work thatanalyzes the causal impact of government programs on the incidence of violence is stillquite rare.
In this paper, we analyze the effect of a large public-works program in India, theNational Rural Employment Guarantee Scheme (NREGS), on incidents of left-wingextremism in the country. NREGS is one of the largest anti-poverty programs in theworld, with annual expenditures of around one percent of the Indian GDP and a legalframework that guarantees 100 days of employment to all rural households (about70 percent of the population) willing to work at the minimum wage. In addition togenerating labor market opportunities and improving local infrastructure, one of thehopes of the government was to reduce incidents of violence by Naxalites, a group ofMaoist insurgents who are a major threat to internal security in India.
We first develop a simple model, closely related to Berman, Shapiro and Felter(2011) that incorporates key facts of the Naxalite conflict in India: The success of In-dian police forces depends critically on information about Naxalites provided by thelocal population since the insurgents practice guerrilla tactics and rely on the loyaltyof the local community. A government program like NREGS may therefore improvethe relationship between police and the people since it demonstrates the government’swillingness to help economically marginalized parts of the country, which in turn mayincrease the police’s efficiency in tracking down rebels. The model predicts that in-cidents of violence should increase in the short run because of larger military actionby the police as well as potential retaliation by the insurgents. It also claims thatthere should be more Naxalite fatalities and captures since the police have now betterinformation about the rebels’ whereabouts.
We then test the predictions of this model empirically. Since NREGS was rolledout non-randomly, we exploit institutional knowledge of the algorithm used to assigndistricts to different implementation phases, and use both regression discontinuity anddifference-in-difference approaches to analyze the empirical impact of the program.
2
We find that the number of incidents and fatalities increases substantially, by about75 percent from the baseline. This higher violence is concentrated among districtsthat received the program early. Naxalites and civilians are the most affected groups,whereas there is little impact on police casualties. These results are robust across anumber of different specifications.
Overall, these results are consistent with the predictions of the model and supportthe idea that governments may need to use a combination of force and programs aimedat economic development to fight internal security challenges. The paper contributes toour understanding of how actual government programs may work within the context ofnational security strategies. To the best of our knowledge, this is one of the first papersto analyze this question in a context in which the government is the main actor fightingagainst an internal security threat. Berman, Shapiro and Felter (2011) look at a similarquestion in Iraq, but the situation in Iraq with its heavy involvement of foreign powersand its emphasis on reconstruction programs is a relatively special situation that is notapplicable to many other developing countries facing internal security threats.
Our paper also contributes to the broader empirical literature on the relationshipbetween violence and development in developing countries. It complements existingresearch on within-country causes of conflict that make use of micro-data: Do and Iyer(2007), for example, find that a rugged and forested terrain is a strong predictor ofNepalese conflict, whereas Murshed and Gates (2005) find a strong correlation betweenconflict and low living standards. Barron, Kaiser and Pradhan (2004) present evidenceof local unemployment, inequality and natural disasters being correlated with localizedviolence in Indonesia. Dube and Vargas (2007) use exogenous price shocks to study civilviolence in Colombia, and Humphreys and Weinstein (2008) find that in Sierra Leonehigh poverty rates and low education are correlated with recruitment into militantgroups. Together these studies suggest that higher individual opportunity costs maylower the probability of participation in rebel groups.
Our empirical results are also consistent with the cross-country literature on develop-ment and violence. Collier and Hoeffler (1998) argue that while political dissatisfactionis a global phenomenon, only some countries witness civil wars. This is because eco-nomic incentives in these countries are such that they could lead to conflict. In Collierand Hoeffler (1998; 2007), slow contemporaneous economic growth and the presence ofnatural resources are strongly correlated with conflict. More education, on the otherhand leads to a lower probability of conflict. Consistent with their results, we knowthat Naxalites in India are found in the less developed portions of the country, and havea strong presence in regions with natural resources (Boroorah 2008). The presence ofnatural resources as a determinant of conflict is also seen in Bellows and Miguel’s (2008)study of diamond wealth in the Sierra Leone conflict. Fearon and Laitin (2003) findthat conditions favoring conflict, like rugged terrain, increase the likelihood of civil war.Similarly, Naxalites are strongly concentrated in forested areas where it is easier to hidefrom police forces. While Fearon and Laitin show that proxies for state institutionalcapacity are strong predictors of conflict, Collier and Hoeffler argue that political fea-tures, including ethnic diversity, inequality and democracy, are not. Many papers seek
3
to isolate exogenous variation in income to find causal impacts. Miguel, Satyanath andSergenti’s (2004) study of wars in sub-Saharan Africa uses annual rainfall growth asan instrument for income growth. They find that a drop in income growth leads to arobust increase in the likelihood of conflict the following year. Our paper exploits thenon-random phasing-in of the program to isolate causal impacts.
The remainder of this paper is structured as follows: Section 2 provides some back-ground on the Naxaite movement, NREGS and potential hypotheses of the impact ofNREGS on violence. Section 3 sets up a simple theoretical model, whereas section 4provides details on the empirical strategy and the data. Section 5 presents the mainresults as well as some extensions and robustness checks. Section 6 concludes.
2 Background
2.1 The Naxalite Movement
The Naxalite movement is one of India’s most severe threats to national security. PrimeMinister Manmohan Singh famously referred to it as “the single biggest internal securitychallenge ever faced by our country” in 20061. Members of the movement are typicallycalled Naxalites or Maoists. Official government documents often refer to affecteddistricts as Left-Wing Extremism (LWE) districts.
Naxalites have been operating since 1967 when an attack on a tribal villager bylandlords in the small village Naxalbari in West Bengal triggered an uprising. Bythe early 1970s, the movement had spread to Andhra Pradesh, Bihar and Orissa, butsplintered into more than 40 groups. In 2004, the two biggest previously competingNaxalite groups joined hands to form the Communist Party of India (Maoist). Thisis believed to have substantially exacerbated India’s problem with left-wing extremismand to have driven the recent growth in violence in important parts of the country(Kujur 2009, Lalwani 2011). The Indian Home Ministry believed the movement tohave around 15000 members in 2006, controlling about one fifth of India’s forests, andactive in 160 districts (Home Ministry 2006). Naxalites are most active in the easternparts of India, as Figure 1 shows. These areas are often referred to as the Red Corridor.
The Naxalites’ main goal is to overthrow the Indian state and to create a liberatedzone in central India since they believe that the Indian government neglects the lowerclasses of society and exclusively caters to the elites. Traditionally, the primary targetswere landlords and upper caste landowners, but attacks have recently turned into largerand better-planned strikes on government institutions and personnel (Singh undated).One common target of attacks are infrastructure projects such as railways, public busesand telecommunication towers (Ramana 2011). The strongholds of the Naxalites con-tinue to be in tribal areas that are typically characterized as being resource-rich areasbut at the same time places of chronic underdevelopment. Substantial mining activi-ties that lead to the displacement of the tribal population and the absence of economic
1Hindustan Times, April 13, 2006: Naxalism biggest threat: PM
4
development opportunities are believed to be an explanation for the continuing popu-larity of the movement among the local population (see e.g. Borooah 2008, Kujur 2009).Naxalites finance themselves mainly through levies on industries and forest contractors,both in return for protection, as well as in payment for engaging in illegal tree-fellingor mining (Sundar 2011).
The Indian government has been fighting the Maoists since the 1960s, but decadesof using force have been largely unsuccessful in suppressing the movement. While Indiaofficially subscribes to a population-centric approach to counterinsurgency, which relieson a mixture of force and winning over the local population through taking care oftheir grievances, a number of researchers note that India traditionally relies almostexclusively on military strength to fight the Naxalite movement (see e.g. Banerjee andSaha 2010, Lalwani 2011). The main responsibility in this fight rests with the civiland paramilitary forces of the state police in the affected areas, although they areoften supported by central paramilitary battalions. While some states have been moresuccessful at suppressing violence than others, the central government generally blamesthe overall failure to contain the Naxalites on inadequate training and equipment, aswell as on poor coordination between police forces of different states. Many observersalso refer to the often widespread disregard for local perceptions, however, as well asthe sometimes excessively brutal nature of police force behavior that also affects manycivilians. These destroy not only the trust of the local population in the Indian statebut also the opportunity for police forces to take advantage of information on insurgentsprovided by the people (Bakshi 2009, Lalwani 2011, Sundar 2011).
Yet, the view that military force alone may not be effective in solving the Naxaliteproblem in the long run seems to have grown in recent years: In 2007, for example,Prime Minister Manmohan Singh said ‘Development and internal security are two sidesof the same coin. Each is critically dependent on the other.’2 He also noted that manyMaoist recruits come from economically deprived and marginalized groups of society.The central government has therefore shown a growing interest in increasing economicdevelopment in underdeveloped areas of the country through anti-poverty programs, inthe hope that an improvement in the local population’s situation would lead to a reduc-tion in Naxalite violence (Ramana 2011). The National Rural Employment GuaranteeScheme is by far the most ambitious and largest anti-poverty program introduced bythe Indian government is, and some case studies suggest that the program may haveindeed helped to reduce violence in certain areas (see oneworld.net 2011 for a case studyof Balaghat in Madhya Pradesh).
Additionally, the Indian Home Secretary Gopal K. Pillai said in 2010 that the in-telligence gathering system of the police has improved over the last couple of years,making police forces more successful at catching Maoists.3 In his empirical analysis,Vanden Eynde (2011) shows that Naxalite violence against civilians increases after neg-ative rainfall shocks, which is consistent with his theoretical model in which Maoists try
2The Indian Express, December 20, 2007: Divide, uneven growth pose threat to our security: PM3Summary of a lecture given by Gopal K. Pillai on March 10, 2010, which is available at:
http://www.idsa.in/event/EPLS/Left-WingExtremisminIndia
5
to prevent the local population from being recruited as government informants duringbad economic times.
2.2 NREGS
The National Rural Employment Guarantee Scheme (NREGS)4 is often referred toas the largest government anti-poverty program in the world. The scheme providesan employment guarantee of 100 days of manual public-sector work per year at theminimum wage to all rural households. The legal right to this employment is laid downin the National Rural Employment Guarantee Act (NREGA) that was passed in theIndian Parliament in August 2005. Under the scheme, all households can apply for workat any time of the year as long as they live in rural areas and their members are preparedto do manual work at the minimum wage. To incentivize female NREGS employment,men and women are paid equal wages. At least one third of the NREGS workforce issupposed to be female, and child care arrangements need to be made if enough womenbring their children to the work site. Wages are to be paid within 15 days since thework was performed, otherwise the worker is eligible for an unemployment allowance.5
While the minimum wage is state-specific, NREGA specifies a floor minimum wagewhich was Rs. 60 per day at the introduction of the program. It has been raised overtime, and was Rs. 120 per day in 2009.
One of the main goals of NREGS is to combine flexible employment opportunities forrural households with the creation of local assets. Public works projects are in generalsuggested by village-level institutions and then approved by government officials atthe district level. To conform to official guidelines, proposed projects need to fall intocertain categories that include areas like anti-drought measures and infrastructure. Theintention is that local government institutions have a good idea of the most pressingproblems in their villages and can tailor NREGS projects to these challenges, althoughfield reports observe that in the end many local governments seem to lack the technicalexpertise to propose useful local projects (see e.g. NCAER-PIF 2009).
NREGS was rolled out non-randomly6 across the country in three phases: The200 poorest districts, according to a government algorithm, received the scheme inFebruary 2006 (Phase 1), and an additional 130 districts started implementation inApril 2007 (Phase 2). Since April 2008, the scheme has operated in all rural districtsin India (Ministry of Rural Development 2010).7 Districts that should have receivedthe program in the first or second phase according to the reconstructed government
4The program was renamed to Mahatma Gandhi National Rural Employment Guarantee Scheme(MGNREGS) in 2009, but the abbreviations NREGS and NREGA (for the act on which it is based)have stuck in the academic literature on this topic.
5For more details on the scheme see e.g. Dey et al. (2006), Government of India (2009), andMinistry of Rural Development (2010).
6Details on the procedures used to determine district assignment to phases will be provided below.7The scheme only excludes districts with a 100 percent urban population, and is active in 99 percent
of Indian districts.
6
algorithm (which will be explained in more detail below) are dark or light grey inFigures 2 and 3, respectively.
Many of the poorest Indian districts are also those heavily affected by left-wing ex-tremism, as can be seen when comparing the red corridor districts from Figure 1 to theleast developed districts predicted to receive NREGS in the earliest phase in Figure 2.One potential concern with development programs in these areas is that the presence oflocal governments is relatively weak and Naxalites may hinder or prevent the workingof these schemes in their fight against the government. Naxalites are indeed knownto have blocked a number of government development programs (Banerjee and Saha2010). Interference from extremists does not seem to be a major problem for the work-ing of NREGS, however: In contrast to a number of other government schemes, chiefsecretaries from the seven states most heavily affected by left-wing extremism believeNREGS to work relatively well in LWE districts8. And case studies by researchers inJharkhand, Chhattisgarh and Orissa also come to the conclusion that the Naxalitesare not blocking NREGS projects, with the exception of road construction projects(Banerjee and Saha 2010).
As one of the most ambitious government programs in the world, NREGS has gen-erated a lot of interest in the popular press, among policymakers and in academicresearch. A rapidly increasing number of papers in Economics analyze the working ofthe employment guarantee and its impact on various outcomes of interest. By now,a number of papers suggest that implementation issues may substantially limit theeffectiveness of the program: Dutta et al. (2012), for example, document widespreadrationing of NREGS employment since there is often excess demand, and show that thisis especially common in poorer states. Niehaus and Sukhtankar (2012a and 2012b) an-alyze the existence and characteristics of corruption in the implementation of NREGSin the Indian state Orissa, and find that an increase in the minimum wage was notpassed through to workers. NREGS seems to work much better in some other stateslike Andhra Pradesh, however: Johnson (2009a) finds that NREGS seems to providea safety net for rural households since take-up of the program increases after negativerainfall shocks. Johnson (2009b) finds that the working of NREGS does not seem tobe substantially affected by the specific party in power at the local panchayat level,suggesting that political pressures on NREGS in the state are not pervasive.
A number of papers have also focused on analyzing the impact of the employ-ment guarantee scheme on rural labor markets in India. Imbert and Papp (2012) use adifference-in-difference approach to look at the program’s impact on wages and employ-ment, comparing Phase 1 and Phase 2 districts to the Phase 3 districts that had not yetreceived the program in 2007/08 and therefore function as control districts. They findthat NREGA increases employment by 0.3 days per prime-aged adult and private-sectorwages by 4.5 percent, with the impacts concentrated during the agricultural off-season.Azam (2012) also uses a difference-in-difference approach, and finds that public sectoremployment increases by 2.5 percent while wages for males and females increase by
8The Times of India, April 14, 2010: Naxals backing NREGA?
7
1 and 8 percent, respectively. In another variation of the difference-in-difference de-sign, Berg et al. (2012) analyze the impact of NREGS on agricultural wages by usingmonthly information on wages from 2000 to 2011. The results in the paper suggestthat agricultural wages have risen by about 5 percent, but that it takes between 6 and11 months for these wage increases to be realized. All of these papers need to relyon some version of the parallel trend assumption in their analysis, which may well beviolated in practice since the NREGS roll-out was non-random.9 Zimmermann (2013)reconstructs the government algorithm used during the roll-out of the program anduses a regression-discontinuity framework instead, finding little evidence for any wageincreases, although the time frame of the analysis would be consistent with Berg et al.’sfinding that wage effects may take some time to come into effect. The paper also showsthat the difference-in-difference results are not very robust across different samples andempirical specifications.
Overall, the empirical literature on NREGS therefore suggests that the program maynot be very effective, at least in the short run and in the early stages of implementation.Since we focus on a similar time interval in our empirical analysis, any impacts of thescheme on violence in this paper are therefore probably more driven by the promise ofgovernment support than any substantial labor market benefits or improvements in localinfrastructure. Since a stronger commitment to development projects and the welfareof the local population in the Naxalite-affected districts presents an important shiftfrom the traditional reliance on force in the conflict, NREGS could still have importantimpacts on the relationship between the local population, and the government and onNaxalite behavior.
2.3 The Relationship between Conflict and Development
The existing literature on the relationship between conflict and development providesseveral possible theories for the expected impact of a large public-works program likeNREGS on violence, although most of these apply only partially to our case. One hy-pothesis is that violence should fall after the introduction of the program: If Naxalitesare really fighting for “development”, then introducing NREGS should give them fewerreasons to continue their struggle. In this case, NREGS would solve a commitmentproblem and ‘complete the contract’ (Powell 2006) between rebels and the government,since the moral foundation of the Naxalites is heavily undermined once the govern-ment can credibly commit to developing the region. The employment guarantee couldprovide such a credible commitment, since the costs of dismantling NREGS may behigh because of constitutional obligations and the political popularity the scheme maygarner.10 On the other hand, there may be asymmetric information about these costs,and the rebels may expect the government to renege on the promise (Dal Bo and Pow-ell 2007). The Indian government has a long history of development programs that
9More details are provided in the next section.10Zimmermann (2013) finds that the duration of NREGS receipt may have had an important influ-
ence on national election outcomes, for example.
8
were often temporary and largely ineffective. So this experience with fleeting govern-ment programs may affect present expectations on the longevity of this scheme andcould weaken any conflict-reducing impacts. Additionally, as discussed above, thereis mounting evidence of severe implementation problems with NREGS and of, at best,small labor market impacts of the scheme. This further undermines the idea that actualdevelopment is taking place in these areas. There is thus little reason to believe that‘development’ itself has a direct impact on rebel behavior.
Alternatively, we might expect violence to increase after the introduction of NREGSif we have a model of competition for resources in mind (Hirshleifer 1989, Grossman1991, Skaperdas 1992): If the government program increases the wealth of a region, thenthe larger resource pie is worth fighting over and the incidence of violence will increase.Such contest models usually predict that when resources rise in a region in equilibrium,more effort will be put into fighting rather than production. Again, however, the factthat NREGS often does not seem to generate substantial economic benefits makes thismodel less applicable in our case. This is also true for the model variant developed byGrossman (1991) who introduces a participation problem which predicts that povertyincreases participation in soldiering.11 This means that the rebels may find it harderto recruit soldiers when the government offers civilians well-paid jobs. In the case ofNREGS, this model would only apply if the mere promise of job opportunities wasenough to raise opportunity costs, at least in the short run.
The strand of theoretical literature that is probably most applicable to our caseis one that recognizes the role played by civilians in internal conflicts, and the com-plex nature of insurgency and counterinsurgency efforts. Berman, Shapiro and Felter(2011) and Akerlof and Yellen (1994) introduce models where the two fighting partiescompete for information available to civilians. Akerlof and Yellen’s (1994) work onstreet-gangs discusses how the government can ‘buy’ information about gang-activityfrom the civilian population. The Berman et al. (2011) paper, on the other hand,looks at counterinsurgency efforts in Iraq. They argue that the government and the USmilitary used reconstruction spending to win the trust of the civilian population whothen helped the government by providing them with information about rebel activity.
11The participation problem is also closely related to literature on economic inequality and groupformation in the conflict literature. Grossman (1999) argues that incentives such as wages, opportuni-ties to loot and protection from danger are often used to motivate participation. In this view, economicinequality may lead to conflict because there is more to gain from victory (Fearon 2007). The Naxalitemovement, however, originally started out in 1967 as an agrarian revolution, and so the role playedby economic inequality is more subtle. Davies (1962) and Gurr (1971) argue that inequality generatesfrustration and discontent, which then may lead to agrarian revolutions of the kind seen in the 1960s.This however, raises the question of why non-violent means were not used to pursue socio-economicchange. The government has argued that in a democracy like India there is a political process that canbe used to address such grievances. The Naxalites, however, claim that this process has been hijackedby the rich, the elite and the landowners, necessitating violent upheaval. And Walter (2004) claimsthat the absence of effective non-violent alternatives may in fact incite rebellion. Furthermore, theNaxalite movement is an ‘ideological’ struggle as well, about the role of State and its obligation to thepeople. It thus attracts groups who adhere to such ideologies and provides a common identity andstronger commitment from its soldiers (Weinstein 2007).
9
This increases the efficiency of counterinsurgency measures due to the government’sgreater intelligence, and thus lowers the cost that the rebels can impose on the govern-ment through violent activities. Since the rebels cause less damage on the margin, theirincentives for attacking the government fall. The paper therefore predicts that govern-ment funding will lead to a reduction in rebel-related violence, which is consistent withthe empirical results in their paper.
In the next section, we adapt the Berman et al. model to the Indian context. As intheir paper, information provided by the civilians about rebel activity plays a crucialrole in our model, since the government is assumed to use NREGS to win over thecivilian population and thus obtain information about Naxalites. In contrast to theirmodel, however, the struggle for territory is important for the Naxalites, which fits inwith the competition for resources literature described above.
3 Model
As mentioned above, force is an important part of the government’s strategy in deal-ing with the Naxalite conflict. In analyzing the impact NREGS has had on Naxaliteviolence, we therefore need to take into account the police strategy and their success insuppressing Maoist violence. Furthermore, police forces are dependent on informationfrom the local population about the Naxalites, which in turn requires winning the trustof the people.
In this section, we present a simple model that allows for all of these complications.The model closely follows the work done by Berman, Shapiro and Felter (2011) oncounterinsurgency in Iraq, and the Akerlof and Yellen (1994) paper on street-gangs.The fundamental difference is that in our context the rebels are fighting for territory,whereas in Berman, Shapiro and Felter (2011) their goal is to impose costs on thegovernment. The Naxalites have strongholds in mineral-rich or forested areas, and thegovernment has been trying to wrest these areas back from them. Their continued sur-vival depends on help from civilians, who hide them and provide them with resources.They in turn claim to provide civilians with services like protection from exploitationfrom large mining conglomerates (Borooah 2007). The police forces, thus need to relyon information about Naxalites provided by civilians who have a better knowledge oftheir activities. We set up a game where the fight for territory depends on informationprovision by the civilians, and the government will try to obtain this information byimplementing development programs like NREGS.
3.1 Strategies, Sequence and Set-up
There are three players in the game. The government (G) and the Naxalites (N) arefighting for territory. The civilian community (C) can choose how much informationabout Naxalites to provide to the government. Since violence by the Naxalites and
10
military action by the police require preparation and time, it is assumed that thecivilians are the last movers. Play proceeds in the following order:
• Nature chooses norms n that only civilians can observe. These norms determinehow much the civilians want Naxalites to be in control of the territory. Let n bedrawn from a uniform distribution U [nl, nu]
• Governments choose whether to provide the civilians with NREGS, g, and howmuch military action to take against the Naxalites, m. The government can onlyeffectively implement NREGS if it has control of the territory. Simultaneously,Naxalites determine how much violence, v, to attempt against the government,and whether to provide the civilians with services, s, or to retaliate against themfor sharing information, r. The services could be in the form of fighting for theinterests of the civilians. Assume that nl ≤ v + g − (r + s) ≤ nu, so that neitherG nor N can single-handedly decide the outcome of information provision by C.
• Civilians, C, decide how much information, i, to share with G.
• Payoffs occur, and territorial control is determined.
If the government controls the territory, we represent it as a = 1, else a = 0. Theprobability of controlling it is determined by the following factors: (a) the extent ofmilitary actionm by the government, (b) the amount of militant action by the Naxalites,v, and (c) the amount of information the civilians share with the government, i. 12
P (a = 1) = p(m, v, i) (1)
Military action by the government increases their hold on the territory, whereasmilitant action by the Naxalites will lower it. Thus, pm > 0 and pv < 0.13 Informationsharing helps the government, pi > 0. Furthermore, the more information the gov-ernment has, the less the impact of Naxalite violence, pvi < 0, since the governmentcan better defend themselves from the Maoists’ attacks. Similarly, the government’smilitary action is more effective the more information they have pmi > 0.
3.2 Payoffs
The community’s payoffs are as follows:
UC = u(c+ g − n− r)a+ u(c− v + s)(1− a)
12This technology of control of territory is different from the Berman, Shapiro and Felter (2011)model. They assume P (a = 1) = h(m)i and does not depend on militancy by the rebels. In ourcontext, however, it is more realistic that rebel violence can lower the probability of governmentcontrol in the territory.
13We assume pmm < 0, pii < 0 and pvv > 0. For an interior solution we need to bound the amountof militant violence v < vmax. And p(m, vmax, i) ≥ 0
11
If the government is in control (a = 1), then the community consumes c, andNREGS g. But the Naxalites may choose to retaliate against the community with r. Ifthe Naxalites are in control, (a = 0), then the community benefits from their servicess, but suffers from their violence v (either because of collateral damage or empathizingwith the police who are the targets). NREGS is only implemented in the regions wherethe government has control (or alternatively the government can use NREGS as abargaining chip and only provide it to regions that help them win control). Because ofuncertainty in who has control, the community’s expected utility function is:
EUC = u(c+ g − n− r)p+ u(c− v + s)(1− p) (2)
The Naxalites wish to control the territory. They maximize their expected payoff:
EUN = E[(1− a)]− B(v)− S(s) = (1− p(m, v, i))− B(v)− S(s) (3)
The costs of militancy B(v) and of providing services S(s) are increasing and convex.The government fights for the same territory and bears (increasing and convex)
costs of military action D(m) and NREGA provision H(g).
EUG = E[a]−D(m)−H(g) = p(m, v, i)−D(m)−H(g) (4)
The government is not trying to maximize social welfare. Even if civilians wanted theNaxalites to be in power, the government would still fight for the territory rather thanhand over the land to the rebels. The concavity of the utility functions will ensure anexistence of a Nash equilibrium of the game.
The payoffs for the rebels and the government are also different from Berman,Shapiro and Felter (2011). In their context (Iraq), the rebels maximize the cost ofharm done by violence, whereas the government tries to minimize this cost. In thecontext of Naxalite violence, however, it is a question of controlling territory.
3.3 Equilibrium
The pure-strategies sub-game perfect Nash Equilibrium can be solved by backwardinduction. Civilians choose how much information to provide, i, in order to maximizetheir expected payoffs EUC = u(c+ g− n− r)p+ u(c− v + s)(1− p). Their first ordercondition is
∂EUC
∂i= u(c+ g − n− r)pi(m, v, i)− u(c− v + s)pi(m, v, i) ≤ 0
For non-trivial solutions, their best response function is:
i∗ =
{
1 if u(c+ g − n− r) > u(c− v + s) ↔ n < g + v − (r + s)0 if u(c+ g − n− r) ≤ u(c− v + s) ↔ n ≥ g + v − (r + s)
(5)
NREGS provision and militancy violence will incentivize civilians to aid the govern-
12
ment, whereas the threat of retaliation and Naxal services will reduce information flowto the authorities. Therefore, the expected information flow is
E(i∗) = P (n < g + v − (r + s)) =[g + v − (r + s)]− nl
nu − nl
The last equality comes from the fact that community norms are assumed to follow auniform distribution.
Solving backwards, the government and Naxalites simultaneously choose their beststrategies. The government maximizes their expected payoff (p(m, v, i)−D(m)−H(g))by choosing m and g. Their first-order conditions are 14:
∂EUG
∂m= pm(m, v, i∗)−D′(m) ≤ 0
and∂EUG
∂g= pi(m, v, i∗)
∂i
∂g−H ′(g) ≤ 0
Since E(i∗) = [g+v−(r+s)]−nl
nu−nl
, we know ∂i∂g
= 1nu−nl
> 0.
Information from civilians makes government military action more effective, pmi(m, v, i) >0. Therefore,
∂2EUG
∂m∂g
∣
∣
∣
∣
v,r,s
= pmi(m, v, i)∂i
∂g> 0
By the implicit function theorem ∂m∂g
∣
∣
∣
v,r,s> 0 indicating that the government may in-
crease militant action along with service provision (i.e. they are strategic complements).
Naxalites choose s ≥ 0, v ≥ 0 and ru ≥ r ≥ 0, where ru is the maximum amountof retaliation possible. Since retaliation depends on C’s choice, we can assume r and i
are chosen simultaneously. In such a case, r = 0 when a = 0 and r = ru when a = 1.So there is a corner solution for r. When NREGS is implemented, it increases i andtherefore P (a = 1), thus causing r to rise (holding constant m, v, s). The Naxalitesmaximize their expected payoff (1 − p(m, v, i)) − B(v) − S(s), and their first ordercondition with respect to v is
∂EUN
∂v= −pv(m, v, i)− pi(m, v, i)
∂i
∂v− B′(v) ≤ 0
If we assume that information makes rebel violence less effective, then pvi ≤ 0. This isbecause the government can use that information to protect themselves from militant
14The second order conditions are satisfied if p(m, v, i) is concave in m and g. In general, becausepmm ≤ 0 and pii ≤ 0 we get the second derivatives of EUG with respect to m and g to be negative.In order to make stronger statements about concavity, we need assumptions on pmi. For example, ifp = mαv−βiγ , with 0 ≤ m, i and 0 < v ≤ vmax and α+ γ ≤ 1 is one of many possible functional forms
13
violence and thus preserve their hold on the territory. Therefore,
∂2EUN
∂v∂g
∣
∣
∣
∣
m,r,s
= −pvi(m, v, i)∂i
∂g− pii
(
∂i
∂g
)2
> 0
By the implicit function theorem ∂v∂g
∣
∣
∣
m,r,s> 0, indicating that militant violence, in the
short run, rises when the government tries to win the territory by providing publicgoods g. 15
This model generates certain testable implications on the ‘incidence of conflict’ inthe region. Empirically, our measure of the ‘incidence of conflict’ are fatalities, injuriesand the number of incidents. Let the incidence of conflict increase with (a) governmentmilitary action m, (b) rebel militancy v, and (c) retaliation r. As we saw above, allthese three (m, v and r) increase with g, indicating that the incidence of conflict, inthe short run, will increase when NREGS is implemented. Furthermore, g will increasethe flow of information to the government making their fight against the rebels moreeffective. Therefore we should expect to see more Naxalites dying or being capturedwhen NREGS is implemented. The long run impacts may be very different, however.If the government is more effective at catching the Naxalites, it may speed up the endof the conflict, and in the long run violence may fall because one side will emerge asthe victor. The testable implications that we study are:
• When NREGS is implemented, there is an increase in government military action,Naxalite activity and retaliation. This results in more Naxalite related incidents,deaths, injuries and abductions.
• Using the better information they have, the police are more effective in capturingor killing the Naxalites.
4 Identification Strategy, Data and Empirical Spec-
ification
4.1 NREGS Roll-out and the Assignment Algorithm
The challenge in analyzing the impact of NREGS empirically is that the program wasrolled out non-randomly to less economically developed districts first. The Indian gov-ernment used an algorithm, however, to determine which districts would start imple-menting the program and also which of three phases the roll-out would take place in.
15This result is the exact opposite of what Berman, Shapiro and Felter (2011) show. This is due toa couple of reasons: (a) in their context of the Iraqi conflict, the rebels seek to inflict violence ratherthan capture territory. And the government seeks to lower the cost of violence rather than gain controlof the territory. In our context, it is more of a territorial dispute. (b) Furthermore, in our model, theprobability of winning control also depends on militant violence.
14
Zimmermann (2013) reconstructs the likely algorithm from available information onthe NREGS rollout and institutional knowledge about the implementation of develop-ment programs in India. The algorithm has two stages: First, the number of treatmentdistricts that are allocated to a given state in a given phase is determined. It is pro-portional to the prevalence of poverty across states, which ensures inter-state fairnessin program assignment.16 Second, the specific treatment districts within a state arechosen based on a development ranking, with poor districts being chosen first.
We use this procedure in our empirical analysis. The ‘prevalence of poverty’ measureused in the first step of the reconstructed algorithm is the state headcount ratio timesthe rural state population, which provides an estimate of the number of below-the-poverty-line people living in a given state and shows how poverty levels compare acrossstates. In the first step of the algorithm, a state is therefore assigned the percentage oftreatment districts that is equal to the percentage of India’s poor in that state. For thecalculations, we use the headcount ratios calculated from 1993-1994 National SampleSurvey data17.
The development index used to rank districts within states comes from a PlanningCommission report from 2003 that created an index of ‘backwardness’, which is a termoften used in India to refer to economic underdevelopment. The index was createdfrom three outcomes for the 17 major states for which data was available: agriculturalwages, agricultural productivity, and the proportion of low-caste individuals (ScheduledCastes and Scheduled Tribes) living in the district (Planning Commission 2003).18 Dataon these outcomes was unavailable for the remaining Indian states, and it is unclearwhether a comparable algorithm using different outcome variables was used for them.We therefore restrict our empirical analysis to these 17 states. Districts were rankedon their index values.
Because of the two-step procedure of the algorithm, the resulting cutoffs for treat-ment assignment in a given phase are state-specific. Since implementation proceededin three phases, two cutoffs can be empirically identified: the cutoff between Phase 1and Phase 2, and the cutoff between Phase 2 and Phase 3. These cutoffs correspond toPhase 1 and Phase 2 NREGS rollout, respectively. We exploit both cutoffs.
Empirically, we will be exploiting this assignment algorithm to predict NREGStreatment in a given phase, and use it in a regression discontinuity (RD) framework
16In practice this provision also ensures that all states (union territories are usually excluded fromsuch programs) receive at least one treatment district.
17We use the rural state headcount ratios from Planning Commission (2009), since the original head-count ratio calculations do not have estimates for new states that had been created in the meantime.Since these are official Planning Commission estimates, they seem like the best guess of the informationthe Indian government would have had access to at the time of NREGS implementation. NSS data isnationally representative household survey data. The newest available information on headcount ratiosat the time would have been the 1999-2000 NSS data, but these were subject to data controversies andtherefore not used.
18The purpose of the index was to identify especially underdeveloped districts for wage and self-employment programs and, as mentioned above, it was used in pre-NREGS district initiatives, althoughthose programs were much less extensive than NREGS and usually envisioned as temporary programs.
15
as well as to form treatment and control groups for a difference-in-difference (DID)approach. Since treatment cutoffs differ by state in the RD analysis, ranks are madephase- and state-specific for the empirical analysis and are normalized so that a districtwith a normalized state-specific rank of zero is the last district in a state to be eligiblefor receiving the program in a given phase. This allows the easy pooling of data acrossstates, since the treatment effect can then be measured at a common discontinuity ineach phase. Negative numbers are assigned to districts with lower ranks than the cutoffrank, whereas positive numbers are assigned to the districts that are too developed tobe eligible according to the district ranking and will function as control districts in ourempirical analysis.
Table 1 reports how well the algorithm predicts NREGS receipt in Phase 1 andPhase 2 for the 17 major states.19 The first column reports the number of non-missingrank districts per state. Columns 2 and 3 show the actual number of NREGS districtsof each state in Phase 1 and Phase 2, respectively, whereas columns 4 and 5 provide theprediction success rate of the algorithm for Phases 1 and 2. Figures 2 and 3 also showthe prediction success of the algorithm graphically for Phase 1 and Phase 2, respectively,with light grey districts being correctly predicted NREGS districts in a given phase,whereas dark grey districts are those that should have received NREGS but did notactually get the program in that phase.
As Table 1 shows, the overall prediction success rate of the assignment algorithm isabout 84 percent in Phase 1 and about 82 percent in Phase 2.20. This means that thereis some slippage in treatment assignment in both phases, and there is considerableheterogeneity in the performance of the algorithm across states. Nevertheless, thealgorithm performs quite well in almost all of the 17 states and especially in the Phase2 assignment. Overall, Table 1 therefore suggests that the proposed algorithm worksquite well for predicting Phase 1 and Phase 2 district allocations.
To achieve internal validity, the RD framework crucially relies on the idea thatbeneficiaries were unable to perfectly manipulate their treatment status, so that obser-vations close to the treatment cutoff value are plausibly similar on unobservables anddiffer only with respect to their treatment status (Lee 2008). But potential politicalmanipulation of the algorithm would also be a threat to the parallel trend assumptionthat the DID approach makes since internal validity in the DID framework relies onthe assumption that treatment and control districts would have had the same trend inoutcomes in the absence of the program.
Therefore, it is important to rule out that districts were able to manipulate the
19Rank data in the 17 major Indian states is complete for all districts classified as rural by thePlanning Commission in their report, so there is no endogeneity in the availability of data in thesestates. Urban districts in the Planning Commission report are districts that either include the statecapital or that have an urban agglomeration of more than one million people. Rank data is availablefor 447 of 618 districts in India. Data for the index creation was unavailable in some states, in mostcases because of internal stability and security issues during the early 1990s when most of the datawas collected. We exclude these states from the analysis.
20Prediction success rates for Phase 2 are calculated after dropping Phase 1 districts from theanalysis.
16
algorithm to their advantage. In the case of the two-step RD, this means that districtsshould not have been able to manipulate their predicted status under the algorithm ineither step. This seems plausible: As mentioned above, the headcount poverty ratioused to calculate the number of treatment districts for a state in the first step of thealgorithm used data from the mid-1990s, which had long been available by the time theNREGS assignment was made.21
Like the information used for the first step, it also seems unlikely that it was possibleto tamper with the data used in the second step of the algorithm: The ‘backwardness’index was constructed from outcome variables collected in the early to mid-1990s, elimi-nating the opportunity for districts to strategically misreport information. Additionally,the suggestion of the original Planning Commission report had been to target the 150least developed districts, but NREGS cutoffs were higher than this even in Phase 1(200 districts received it in Phase 1). Therefore, districts would have had an incentiveto be among the 150 poorest districts but not to be among the 200 or 330 least devel-oped districts for Phase 1 and Phase 2, respectively. Lastly, the Planning Commissionreport lists the raw data as well as the exact method by which the development indexwas created, again eliminating room for districts to manipulate their rank. Overall, ittherefore seems like manipulation of the rank variable is not a major concern.22
Figures 4 and 5 look more closely at the distribution of index values over state-specific ranks. Ideally, the assignment variable should be continuous at the cutoff, sincediscontinuities at the cutoffs are typically taken as signs of potential manipulation(McCrary 2008). The figures therefore plot the relationship between the PlanningCommission’s index and the normalized state-specific ranks for the Phase 1 and Phase2 cutoffs, respectively. For most states, the poverty index values seem pretty smoothat the cutoff of 0, again suggesting that manipulation of the underlying poverty indexvariable is not a big concern.
Finally, we need to verify that there really is a discontinuity in the probability of re-ceiving NREGS at the state-specific cutoff values. Figures 6 and 7 show the probabilityof receiving NREGS in a given phase for each bin, as well as fitted quadratic regressioncurves and corresponding 95 percent confidence intervals on either side of the cutoff.The graphs demonstrate that the average probability of receiving NREGS jumps downat the discontinuity, although this discontinuity is much stronger in Phase 2 than inPhase 1. This suggests that there is indeed a discontinuity in the probability of be-
21The algorithm also uses state rural population numbers from the 2001 Census to transform head-count ratios into absolute numbers, but those figures were also long publicly available at the time.The RD may be potentially fuzzier than it really is because of some potential for measurement errorintroduced into the algorithm at this step since the exact numbers the government used in this stepare not known, but this should not introduce systematic bias into the empirical analysis.
22This does not mean that actual treatment assignment was not subject to political pressures sinceTable 1 shows that compliance with the algorithm is often below 100 percent. Zimmermann (2013)shows that deviations from the algorithm are correlated with party affiliation. This is also consistentwith Gupta (2006) who analyzes the relationship between rule deviations and party affiliation for anearlier program. This paper ignores the fact that the program most likely also used the two-stepalgorithm, however, could substantially affect the results.
17
ing treated with the employment guarantee scheme at the cutoff. Since the treatmentdiscontinuity for Phase 1 is relatively small, we later use a doughnut-hole approachas a robustness check, which is discussed in more detail in the empirical specificationsection.
4.2 Data and Variable Creation
The primary source of data used in this paper comes from the South Asian TerrorismPortal (SATP). This is a website managed by a non-profit organization called theInstitute of Conflict Management in New Delhi. The SATP aggregates news reportson Naxalite-related incidents and summarizes them. The summary usually contains(a) the location of the incident (district), (b) the date of the incident, (c) number ofcasualties (Naxalites, civilians, or police), and (d) the number of injuries, abductionsor surrenders. For example the following summary: “Andhra Pradesh, 2007: February
7 Police kill two CPI-Maoist cadres near Karampudi village in the Guntur District.’ ’is coded as two Naxalite deaths in Guntur district in February. The source also codesthe incident as ‘minor’ or ‘major.’
Using this information we can construct variables at the district-month level. ‘Noincidents’ are coded as 0. If some information is unclear, we search for the source newsreports online and verify the information. For now, we use data between the years 2005and 2007. This is a period of special interest since phase 1 of NREGS was implementedin early 2006 and phase 2 in early 2007. Thus, we have enough data both before andafter implementation of the program to be able to analyze changes with the introductionof NREGS. This dataset is then merged with information on the poverty index rankfrom the 2003 Planning Commission Report.
There are certain limitations to using this dataset, however. The number of Nax-alites killed or injured is difficult to verify, and security forces may have an incentiveto overstate their accomplishments by inflating the numbers. On the other hand, someminor incidents in remote areas may not reach the newspapers. Sundar (2011) pointsout two further issues: First, the data source misses the deaths of the members ofcertain vigilante groups like the Salwa Judum. Second, Naxalites are often blamed forcivilians killed by security forces to protect the integrity of the police.
These limitations introduce measurement error into the analysis but should notsystematically bias our regression discontinuity (RD) or difference-in-differences (DID)results. For a systematic bias, there would have to have been a systematic change inthe reporting of Naxalite incidents when NREGS was (or rather should have been)introduced in a certain district. We have no prior reason to believe that any suchsystematic change in reporting occurred for a specific district.
Table 2 shows some summary statistics for our primary variables of interest. Ourdataset records 1296 incidents, covering a total of 1978 fatalities. 262 of these incidentswere coded by the SATP source as ‘major’. Furthermore, in this 36-month period,2220 people were either injured or abducted, or they surrendered to the police. Onaverage, in any given red-corridor district, there are about 0.5 deaths a month related
18
to Naxalite activities, and about 0.33 incidents a month.We also collect data regarding the police force from the Home Ministry of India.
This data is at the state level and contains information on number of police officers,police posts and stations, as well as some other measures of police strength.
4.3 Empirical Specification
To test the robustness of our results across different specifications, we use two differentempirical identification strategies. Since NREGS was rolled out based on an algorithmthat assigned state-specific ranks to districts, this algorithm can be used as a runningvariable in an RD framework. As the number of observations near the cutoff is limited,we use parametric regressions to estimate the impact of NREGS on Naxalite violenceempirically.23 To test the robustness of the estimates, all main result tables for the RDspecification show the estimated coefficients for linear and quadratic regression curvesin the running variable with and without constraining the slope of the curves to be thesame on either side of the cutoff.24 Since the algorithm only generates a fuzzy RD,we use a two-stage least squares specification where actual NREGS receipt in a givenphase is instrumented with predicted NREGS treatment according to our algorithm.To increase the precision of our estimates, we control for the baseline outcome variable.To ensure that the RD results are not affected by observations far away from the cutoff,we run results separately by cutoff, and drop the observations that should not affectthe treatment effect at a given cutoff: For Phase 1 district assignment, we only includePhase 1 and Phase 2 districts in the analysis, and drop Phase 3 districts. Similarly,for Phase 2 cutoff regressions, we drop Phase 1 observations and only consider theremaining districts.
The equation below shows the regression equation for one of the specifications werun, which is linear in the running variable but does not constrain the coefficients tobe the same on either side of the cutoff:
yij= β0 + β1ranki + + β2nregsi + β3nregs ∗ ranki + β4baselineyij+ ηj + ǫij
where the subscripts refer to district i in month j since one observation is a district-month. y is an outcome variable of interest. rank is a district’s rank based on thestate-specific normalized index, and η are time fixed effects. The coefficient of interestis β2, and nrega is actual NREGS receipt in a given phase, which is instrumented withpredicted NREGS receipt. Standard errors are clustered at the district level.
23This is in line with the advice in Lee and Lemieux (2009) in such a situation.24F-tests reject the null hypothesis that higher-order polynomials add important flexibility to the
model. More flexible models also tend to be unstable in the second stage of the two-stage leastsquares estimation procedure, although the estimated coefficients are often qualitatively similar to thequadratic results. The quadratic flexible specification is also usually outperformed statistically by thelinear flexible specification, so we do not include the flexible quadratic specification in our main resultstables at this point.
19
In addition to the RD results, we also present all results using a difference-in-difference strategy. The DID coefficients estimate the average treatment effect acrossall observations rather than the treatment effect at the cutoff as in the RD framework,and may therefore be interesting in their own right. Additionally, the RD specificationmay be subject to some power issues because of the limited number of observations,which is much less of an issue in DID specifications. The DID estimations thereforealso tell us something about how robust our results are across specifications.
The main downside of the DID approach is that the parallel trend assumption thatunderlies the empirical strategy is often a strong assumption. The parallel trend as-sumption requires that treated and untreated districts in a given phase would have hadthe same trend in outcome variables in the absence of NREGS. While we cannot testthis assumption explicitly, knowledge of the algorithm allows us to deal with one poten-tial source of violations of this assumption. As Table 1 shows, there was some slippagefrom the algorithm in almost all states, with some districts receiving the program thatshould not have been eligible for it. These districts are likely to either have powerfulpolitical connections, or to have high expected benefits from the program, and couldquite likely have experienced differential trends in violence from other districts even inthe absence of NREGS. Rather than using actual treatment receipt, we therefore runan intent-to-treat specification of the DID approach, where the treatment group arethose districts predicted to receive NREGS in a given phase based on the algorithm.We refer to these results as ITT-DID estimates. We then go on to show, using a graph-ical analysis, that there is little to reason to expect differential trends in the ITT-DIDapproach.
As in the RD specifications, we run regressions for Phase 1 and Phase 2 assignmentsseparately and include time fixed effects. In addition, the ITT-DID framework alsoallows us to run stacked regressions that exploit the time dimension of our data morefully and include both Phase 1 and Phase 2 district assignments. For these, we estimatethe following equation:
yij = β0 treatedij + δi + θj + ǫij
where treatedij is an indicator of whether district i should have access to NREGSat time j (stretching both Phase 1 and Phase 2), whereas δ and θ are district and timefixed effects, respectively. Standard errors are again clustered at the district level. Thecoefficient of interest here is β0, which is the average change in outcome of interest y
between receiving and not receiving NREGS.As the theoretical model points out, the introduction of NREGS may influence
the efficiency of the police in capturing Naxalites because of a better information flowbetween the local population and the police, since public goods provision and militaryaction are strategic complements. While we cannot explicitly test this channel with ouravailable data at this point, one indirect test of higher efficiency of police actions is tocompare the regression coefficients that control for police force strength to those thatdo not. In the absence of such controls, increases in violence and a higher likelihood of
20
deaths or captured Naxalites could be driven both by increases in the police force andby improvements in the efficacy of the existing policemen. Controlling for the strengthof the police force therefore keeps this channel constant, and more directly measureswhether the same number of policemen now lead to different outcomes than before.Since we do not have data on the actual police force in a district, we estimate it usingstate-level data from the Indian Home Ministry. Any change in the police force for agiven state is assumed to be attributable to NREGS districts only. In reality, thesestate-level estimates most likely overemphasize the change in the police force and maytherefore provide us with conservative estimates of the impact of NREGS.
In our main results, we therefore add police force controls and later compare theseestimates to specifications without such controls. Additionally, we include two typesof robustness checks: First, we use a doughnut-hole approach to test the robustness ofour RD results to small changes in the composition of observations close to the cutoffs.Because the first stage of the algorithm relies on the best guess of the procedure used bythe government, rather than on the actual one, the algorithm introduces measurementerror into predicted NREGS treatment. This particularly affects the observations closeto the cutoff. We therefore re-estimate our results by dropping observations with state-specific ranks in the immediate vicinity of the cutoffs.
Second, we exploit the panel version of our data to analyze whether pre-treatmentdifferential trends in the outcome variables are likely to bias our DID results.
5 Results
5.1 Main Results
Our model predicts that on the introduction of NREGS, we should expect a rise inthe incidence of Naxalite-related violence. Tables 3 and 4 show the main results of theimpact of NREGS on Naxal incidents using both the RD and the ITT-DID approaches.Table 3 presents the RD results for the five main outcome variables (a) deaths, (b)injuries, abductions or surrenders, (c) minor incidents, (d) major incidents, and (e)total incidents related to Naxal violence. Panel A shows the impact on Phase 1 districtsonly, whereas Panel B looks at the effect on Phase 2 districts. Each panel has threedifferent specifications (i) linear, (ii) linear with a flexible slope, and (iii) quadratic.The results are robust across specifications. As mentioned above, specifications in thistable control for (estimated) police force changes.25
Looking at Panel A in Table 3, we see that there was an increase in fatalities inPhase 1 when NREGS is introduced. Depending on the specification, there is a rise ofabout 0.33 to 0.43 deaths per month in a given district. At a mean of about 0.5 deathsper month in a Red Corridor district, this amounts to about a 75% increase from thebaseline level. Similarly, there are about 0.2 more injuries/abductions/surrenders in a
25Not controlling for police force changes does not change the results substantially. These resultsare presented in Tables 8 and 9.
21
given district-month unit. The number of total incidents rises by about 0.18 per month,which is about a 55% increase from the baseline mean. These results are robust acrossall specifications.
The Phase 2 results in Panel B tend to be smaller and much less precisely estimated.With the exception of the number of major incidents, all coefficients are positive asin Panel A, but no coefficient is statistically significant at conventional levels. Onepotential reason for this is that the most affected Naxalite districts received NREGSin Phase 1. As Figures 1 and 2 reveal there is a large overlap between NREGS Phase1 district assignment and being in the Red Corridor as many Naxalite districts areeconomically underdeveloped. Additionally, the time interval used in our analysis mayplay a role if impacts take some time to materialize: Our data spans the years 2005 to2007, which implies that we have only 9 months of post-NREGS treatment for Phase 2districts, but 12 months for Phase 1 districts.
Table 4 presents the corresponding ITT-DID results. While the RD results werecomparing districts on either side of the cutoffs, the ITT-DID results compare the im-pact on the average district that should have got NREGS, with the average district thatshould not have received it. Panel A shows the impact of NREGS on Naxalite incidentson Phase 1 and 2 simultaneously, using a ‘stacked regression’, whereas Panels B and Creport the impacts on just Phase 1 and just Phase 2, respectively. The results in Table4 are consistent with the model and the RD results in Table 3: There are about 0.06more incidents a month, which is an increase of about 18% on average from the baseline.There is also a significant increase in the number of injuries/abductions/surrenders.Once again, this impact is concentrated in Phase 1 districts, with little or no effect onPhase 2 districts.
Overall, both our Regression Discontinuity (RD) and Difference-in-Differences (ITT-DID) results therefore show that after the introduction of NREGS there is an increasein the incidence of violence. This impact is concentrated in the Phase 1 districts, withlittle or no impact on Phase 2 districts.
5.2 Who is Affected?
Since the data allows us to distinguish between civilians, Naxals and the police force, wecan study the impact of NREGS on each of these groups in terms of fatalities, injuriesand abductions. According to the theoretical model, the police should now have betterinformation to catch the Naxalites and the Naxalites, in turn, may want to retaliateagainst civilians for helping the police. Thus, we should see the bulk of the impactconcentrated on Naxals and civilians. Tables 5 to 7 report the empirical results of thisanalysis.
Table 5 presents the RD results for fatalities classified by each of these groups.Civilian casualties rise by about 0.2 deaths a month, whereas Naxal casualties increaseby around 0.15 deaths a month after the introduction of the NREGS. The police forcedoes not see a statistically significant increase in fatalities, and the magnitudes are alsomuch smaller. The results are once again robust across specifications and concentrated
22
in Phase 1.While Table 5 looked only at casualties, in Table 6 we create a more comprehensive
variable called ‘affected’ and use it as an outcome measure in the RD analysis. Thisvariable captures deaths, injuries, captures, surrenders and abductions for each of thethree populations. As Table 6 shows, once again Naxals and civilians are affected to alarge degree, whereas there is no significant impact on the police force.
Table 7 presents the corresponding ITT-DID results. Again, Naxalites are mostlikely to be affected when NREGS is introduced in the district.
Overall, the results therefore all point to Naxalites being the most affected by theincrease in incidents. This is consistent with the mechanism in the model that thepolice force may be using information from civilians to effectively target the rebels.Alternatively, it could mean that the rebels are committing more desperate acts ofviolence to hold on to their territory.
5.3 Robustness Checks
The results presented so far are consistent with our model. Both the RD and theITT-DID results show an increase in the incidence of violence - captured by a rise infatalities, injuries/abductions/surrenders, and the number of incidents overall. Whilethe RD results show more Naxal as well as civilian casualties, the ITT-DID results saythat it is mostly the Naxalites who were affected. While the impacts are not large inabsolute terms, they are large relative to the baseline means. Furthermore, we are onlycapturing the short-term impact of this program. In the long run, a more effectivepolice strategy may bring about a quicker end to the war, and we would then expectto see a fall in the incidence of violence. To analyze the robustness of our results, weperform a number of robustness checks.
All of the results presented so far in Tables 3 through 7 have controlled for thestrength of the police force, so the increases in incidents of violence are not driven byan increase in police officers engaged in anti-Naxalite operations. The controls for thepolice force include variables such as the change in sanctioned police force per capita,change in state police force per capita, change in number of police posts, and the changein number of police stations. Tables 8 and 9 reproduce the main results from Tables 3and 4 but do not control for the police variables, showing how the main results change ifwe allow the strength of the police force to increase as well. The results are qualitativelysimilar and continue to be statistically significant.
A concern with RD specification is that there may be measurement error in therank variable. We provide a robustness check by using a doughnut-hole approach anddropping the districts with state-level ranks lying between -1 and 1 (the cutoff is at astate-specific rank of 0). These results are presented in Table 10. Again, they are similarin magnitude and statistical significance to the other RD results presented, implyingthat the results are not driven by measurement error of the observations close to thecutoff.
One worry that arises with Difference-in-Differences analyses is the existence of
23
differential trends or Ashenfelter’s Dips. If NREGS districts had a rising incidenceof violence over time, for example, we would get biased estimates of the impact ofintroducing NREGS. In our case, we have relatively little reason to expect a ‘politicalmanipulability’ bias since our ITT-DID specification is using information collected manyyears prior to implementation to construct treatment and control groups. However, bydesign this makes treatment districts poorer than control districts, and poorer districtsmay have had differential trends in Naxal violence than richer districts even in theabsence of NREGS, which would bias the DID results. One way of analyzing whetherdifferential trends are likely to be a problem is to exploit the time dimension of ourdata. In the following equation, we normalize the treatment period to be 0, so that θjis a dummy for period j relative to the treatment period.
Yij = θ−35 + θ−34 + ...+ θ−1 + θ1 + ...θ22 + δi + α ∗ Trend++γ ∗ Police+ ǫij (6)
We estimate all θ fixed effects and plot them. In the absence of differential trends,pre-treatment periods should have fixed effect coefficients close to zero, whereas wewould expect the coefficients to be greater than 0 post treatment. We present graphsfor total number of incidents and total number of Naxals affected. Other graphs showsimilar pre-treatment (absence of) trends. These graphs are presented in Figures 8and 9. The vertical line in the graph is the period of treatment. The pre-treatmentestimates are indeed small in magnitude and statistically insignificant from 0, so it doesnot look like pre-existing differential trends or Ashenfelter’s Dips are a major concernin our case. The impact of introducing NREGS is also apparent by looking at the post-treatment period - both the number of incidents and of Naxals affected see a sharp risein the post-treatment period.
6 Conclusion
This paper has analyzed the impact of introducing a large public-works program inIndia, the National Rural Employment Guarantee Scheme (NREGS), on incidents ofleft-wing violence by the Naxalite movement. We exploited the fact that the programwas phased in over time according to an algorithm that prioritized economically un-derdeveloped districts, to use both regression discontinuity (RD) and intent-to-treatdifference-in-difference (ITT-DID) approaches. The results are robust across a numberof different specifications and show a substantial rise in fatalities and incidents in thedistricts that received NREGS. Fatalities occur primarily amongst Naxalites and civil-ians, with little impact on the police force, and are concentrated among the districtsthat received NREGS in the first phase of the rollout.
These results contribute to a large literature on the relationship between conflictand development, although the exact nature of this connection is still a contested ques-tion. When thinking about the role played by development, research usually assumesthat the impact will be felt directly by the rebels: Either the rebels fight for the larger
24
resource pie, or they end their struggle and join in productive activities due to increas-ing opportunity costs. This neglects the role played by a third group - the civilians.Civilians play a key role in the conflict between the government and the Naxalites sincethey are usually well-informed about rebel activity, but do not voluntarily surrenderthis information. Building on the Berman et al. (2011) model, we argued in this paperthat a plausible model of the Indian situation is one where the direct impact of NREGSis primarily on the civilians, since they are the biggest potential beneficiaries of the in-creased levels of employment and wages and improved infrastructure that the NREGSpromises. In return for this stability and prosperity, civilians may choose to providethe government with the information necessary to tackle rebel activity. In addition tobetter-targeted police attacks on the Naxalites, the Naxalites may retaliate and try toprotect their territory. This is consistent with the overall rise in violence that we seein the data as well as the fact that the Naxalites are the most affected group. Manyother theoretical models have difficulties with fitting both the specifics of the Indiansituation and these empirical patterns. A large part of the problem in testing compet-ing theoretical models more effectively is the challenge that secrecy pertaining to thepolice strategy is inherently built in. The secrecy protects the civilians who provideinformation, and make enhance the effectiveness of police strikes. Overall, however, thistheoretical framework may help answer Max Weber’s (1965) question on whether theState should use force or development to tackle internal conflict. It is possible that forthe Indian government, which has been trying to fight the Naxalites for over 30 years,the best strategy may be to combine both force and development.
In the specific case of NREGS, two qualifications to the results are important,however: First, a growing body of literature questions the effectiveness of NREGS asa tool for actual development, at least in the short run. This implies that what maywin over the rural community at least initially is the promise of development ratherthan substantial actual changes. Second, while our results indicate that the incidenceof violence rises when NREGS is introduced, this impact may just be a short-run effect.In the long run, a more efficient police force should lead to a quicker end to the war, andtherefore a fall in violence. On the other hand, the mere promise of development maynot be a credible enough tool to ensure the aid received from the civilian populationover a longer period of time. Once civilians realize that the program is not deliveringon its promises, this may not only stop civilian aid in exchange for the benefits of theprogram, but may lead to distrust in government programs in general. It is imperative,therefore, that the Indian government take steps to ensure that NREGS is implementedeffectively and the promise of development is indeed fulfilled.
25
References
[1] Akerlof, George and Janet Yellen. 1994. ‘Gang Behavior, Law Enforcement, andCommunity Values.’ Issue 53 of CIAR Program in Economic Growth and Policyreprint series
[2] Azam, Mehtabul. 2012. ‘The Impact of Indian Job Guarantee Scheme on LaborMarket Outcomes: Evidence from a Natural Experiment.’ IZA Discussion Paper6548.
[3] Bakshi, G D. 2009. ‘Left Wing Extremism in India: Context, Implications andResponse Options.’ Manekshaw Paper No. 9.
[4] Barron, Patrick, Kai Kaiser and Menno Pradhan. 2004. ‘Local Conflict in Indone-sia: Measuring Incidence and Identifying Patterns.’ World Bank Policy ResearchWorking Paper, 3384.
[5] Banerjee, Kaustav and Partha Saha. 2010. ‘The NREGA, the Maoists and theDevelopmental Woes of the Indian State.’ Economic and Political Weekly, XLV(28):42-48.
[6] Bellows, John and Edward Miguel. 2006. ‘War and Institutions: New Evidence fromSierra Leone.’ American Economic Association, Papers and Proceedings, 96:2, pp.394-99.
[7] Berg, Erlend, Bhattacharyya, Sambit, Durgam, Rajasekhar, and Manjula Ra-machandra. Can Rural Public Works Affect Agricultural Wages? Evidence fromIndia. CSAE Working Paper WPS/2012-05, 2012.
[8] Berman, Eli, Shapiro, Jacob N., and Joseph H. Felter. 2011. ‘Can Hearts and MindsBe Bought? The Economics of Counterinsurgency in Iraq.’ Journal of Political
Economy, 119(4): 766-819
[9] Blattman, Chris and Edward Miguel. 2010. ‘Civil War.’ Journal of Economic Liter-ature 48:1, 3-57
[10] Borooah, Vani K. 2008. ‘Deprivation, Violence and Conflict: An Analysis of ‘Nax-alite’ Activity in the Districts of India. International Journal of Conflict and Vio-
lence, 2: 317-333.
[11] Collier, Paul, V. L. Elliot, Hvard Hegre, Anke Hoeffler, Marta Reynal-Querol, andNicholas Sambanis. 2003.‘ Breaking the Conflict Trap: Civil War and DevelopmentPolicy.’ Washington,DC: World Bank and Oxford University Press.
[12] Collier, Paul and Anke Hoeffler. 2007. ‘Civil War,”’ in Handbook of Defense Eco-nomics. K. Hartley and T. Sandler eds: Elsevier North Holland.
26
[13] Collier, Paul and Anke Hoeffler. 1998. ‘On economic causes of civil war.’ OxfordEconomic Papers, 50:4, pp. 563-73.
[14] Dal B, Ernesto and Robert Powell. 2007. ‘Conflict and Compromise in Hard andTurbulent Times.’ UC Berkeley Department of Political Science Working Paper.
[15] Davies, J. C. 1962. ‘Toward a Theory of Revolution.’ American Sociological Review,27:1, pp.5-19.
[16] Dey, Nikhil, Jean Dreze, and Reetika Khera. 2006. Employment Guarantee Act: A
Primer. (Delhi: National Book Trust, India)
[17] Do, Quy-Toan and Lakshmi Iyer. 2007. ‘Poverty, Social Divisions and Conflict inNepal.’ Unpublished working paper, Harvard Business School.
[18] Dutta, Puja, Murgai, Rinku, Ravallion, Martin, and Dominique van deWalle. 2012.‘Does India’s Employment Guarantee Scheme Guarantee Enployment? World BankPolicy Research Working Paper 6003.
[19] Fearon, James D. 2007. ‘Economic development, insurgency, and civil war,’ inInstitutions and Economic Performance. Elhanen Helpman ed. Cambridge: HarvardUniversity Press.
[20] Fearon, James D. and David D. Laitin. 2003. ‘Ethnicity, Insurgency and Civil War.’American Political Science Review, 97:1, pp. 75-90.
[21] Government of India. 2009. ‘The National Rural Employment Guarantee Act.’
[22] Grossman, Herschell I. 1991. ‘A General Equilibrium Model of Insurrections.’ TheAmerican Economic Review, 81:4, pp. 912-21.
[23] Grossman, Herschell I. 1999. ‘Kleptocracy and revolutions.’ Oxford Economic Pa-pers, 51, pp.267-83.
[24] Gupta, Santanu. 2006. ‘Were District Choices for NFFWP Appropriate?’ Journalof Indian School of Political Economy, 18(4): 641-648.
[25] Gurr, Ted R. 1971. ‘Why Men Rebel.’ Princeton: Princeton University Press.
[26] Hirshleifer, Jack. 1989. ‘Conflict and rent-seeking functions: Ratio versus differencemodels of relative success.’ Public Choice, 63, pp. 101-12.
[27] Imbert, Clement, and John Papp. 2012. ‘Equilibrium Distributional Impacts ofGovernment Employment Programs: Evidence of India’s Employment Guarantee.’http://www.parisschoolofeconomics.eu/docs/imbert-clement/2012-03-19-pse-working-paper-equilibrium-distributional-impacts-of-government-programs–imbert-papp.pdf.
27
[28] Johnson, Doug. 2009a. ‘Can Workfare Serve as a Substitute for Weather Insurance?The Case of NREGA in Andhra Pradesh.’ Institute for Financial Management andResearch, Centre for Micro Finance, Working Paper 32.
[29] Johnson, Doug. 2009b. ‘How Do Caste, Gender and Party Affiliation of LocallyElected Leaders Affect Implementation of NREGA?’ Institute for Financial Man-agement and Research, Centre for Micro Finance Working Paper 33.
[30] Kujur, Rajat Kumar. 2009. ‘Naxal Conflict in 2008: An Assessment’. IPCS Issue
Brief, New Delhi.
[31] Lalwani, Sameer. 2011. ‘India’s Approach to Counterinsurgency and the NaxaliteProblem.’ CTC Sentinel, 4(10): 5-9.
[32] Lee, David S. 2008. ‘Randomized Experiments from Non-Random Selection in U.S.House Elections.’ Journal of Econometrics, 142(2): 675-697.
[33] McCrary, Justin. 2008. ‘Manipulation of the Running Variable in the RegressionDiscontinuity Design: A Density Test.’ Journal of Econometrics, 142(2): 698-714.
[34] Miguel, Edward, Shanker Satyanath and Ernest Sergenti. 2004. ‘Economic Shocksand Civil Conflict: An Instrumental Variables Approach.’ Journal of Political Econ-omy, 112:4, pp. 725-53.
[35] Ministry of Home Affairs, Government of India. 2006. ‘Annual Report 2006-2007.’
[36] Ministry of Rural Development, Department of Rural Development, Governmentof India. 2010. ‘Mahatma Gandhi National Rural Employment Guarantee Act 2005- Report to the People 2nd Feb 2006 - 2nd Feb 2010.’
[37] Murshed, S. Mansoob and Scott Gates. 2005. ‘Spatial-Horizontal Inequality andthe Maoist Insurgency in Nepal.’ Review of Development Economics, 9:1, pp. 121-34.
[38] NCAER-PIF. 2009. ‘Evaluating the performance of the National Rural Employ-ment Guarantee Act.’
[39] Niehaus, Paul, and Sandip Sukhtankar. 2012a. ‘The Marginal Rate of Corrup-tion in Public Programs.’ http://www.dartmouth.edu/ sandip/niehaus-sukhtankar-marginal-corruption.pdf.
[40] Niehaus, Paul, and Sandip Sukhtankar. 2012b. ‘Corruption Dynamics: TheGolden Goose Effect.’ http://www.dartmouth.edu/ sandip/niehaus-sukhtankar-golden-goose.pdf.
[41] oneworld.net. 2011. ‘Community MGNREGS Programme for Naxalite AffectedAreas.’ Report.
28
[42] Planning Commission. 2003. ‘Report of the Task Force: Identification of Districtsfor Wage and Self Employment Programmes.’
[43] Planning Commission. 2009. ‘Report of the Expert Group to Review the Method-ology for Estimation of Poverty.’
[44] Powell, Robert. 2006. ‘War as a Commitment Problem.’ International Organiza-tion, 60, pp.169-203
[45] Ramana, P.V. 2011. ‘India’s Maoist Insurgency: Evolution, Current Trends, andResponses.’ in India’s Contemporary Security Challenges, edited by Michael Kugel-man, Woodrow Wilson International Center for Scholars Asia Program.
[46] Singh, Prakash. undated. ‘Naxal Threat and State Response.’ Mimeo.
[47] Skaperdas, Stergios. 1992. ‘Cooperation, Conflict, and Power in the Absence ofProperty Rights.’ American Economic Review, 82:4, pp. 720-39.
[48] Sundar, Nandini. 2011. ‘At War with Oneself: Constructing Naxalism as India’sBiggest Security Threat.’ in India’s Contemporary Security Challenges, edited byMichael Kugelman, Woodrow Wilson International Center for Scholars Asia Pro-gram.
[49] Vanden Eynde, Oliver. 2011. ‘Targets of Violence: Evidence from India’s NaxaliteConflict.’ Mimeo.
[50] Walter, Barbara F. 2004. ‘Does Conflict Beget Conflict? Explaining RecurringCivil War.’ Journal of Peace Research, 41:3, pp. 371-88.
[51] Weber, Max. 1965. ‘Politics as a Vocation.’ Minneapolis:Fortress Press.
[52] Weinstein, Jeremy M. 2007. ‘Inside Rebellion.’ Cambridge: Cambridge UniversityPress
[53] Zimmermann, Laura. 2013. ‘Labor Market Impacts of a Large-Scale Public WorksProgram: Evidence from the Indian Employment Guarantee Scheme.’ Mimeo. Ear-lier version of the paper: IZA Discussion Paper No. 6858
[54] Zimmermann, Laura. 2013. ‘Jai Ho? The Impact of a Large Public Works Programon the Government’s Election Performance in India.’ Mimeo.
29
Table 1: Prediction Success of Algorithm for Major Indian Statesactual NREGS prediction success rate
district N Phase 1 Phase 2 Phase 1 Phase 2Andhra Pradesh 21 13 6 0.9048 0.7500Assam 23 7 6 0.9130 0.7500Bihar 36 22 14 0.8056 1.0000Chhattisgarh 15 11 3 0.7333 1.0000Gujarat 20 6 3 0.8000 0.9286Haryana 18 2 1 0.7222 0.9375Jharkhand 20 18 2 0.8500 1.0000Karnataka 26 5 6 0.8846 0.5238Kerala 10 2 2 0.7692 1.0000Madhya Pradesh 42 18 10 0.7619 0.8750Maharashtra 30 12 6 0.9333 0.5556Orissa 30 19 5 0.7333 0.9091Punjab 15 1 2 1.0000 0.9286Rajasthan 31 6 6 0.9032 0.7200Tamil Nadu 26 6 4 0.8846 0.9500Uttar Pradesh 64 22 17 0.8750 0.7857West Bengal 17 10 7 0.7647 1.0000Total 447 180 100 0.8412 0.8202
Note: Table includes all districts with non-missing development index rank for17 major Indian states (the only missing districts in these states are urban districtsaccording to the Planning Commission report definition from 2003 and therefore includeeither the state capital or an urban agglomeration of at least one million people).Column 1 provides the number of non-missing rank districts in each state. Columns 2and 3 give the actual number of treatment districts per state in a given phase of NREGSrollout. Columns 4 and 5 give the success rate of the algorithm in predicting a district’streatment status (NREGS or no NREGS) in a given phase. The proposed algorithmstates that the number of treatment districts a state is assigned in a given phase isproportional to the percent of India’s poor living in that state, and that this quotashould be filled with the least developed districts according to the Indian PlanningCommission’s ranking of districts (Planning Commission 2003). In the first phase,districts on an existing official list of districts majorly affected by left-wing terrorismwere prioritized regardless of their rank.
30
Table 2: Summary Statistics
Mean Mean TotalVariable Red Corridor All districts All districts
Deaths 0.5078 0.1229 1978Injuries/Abductions 0.57 0.1379 2220
Minor Incidents 0.2655 0.0642 1034Major Incidents 0.0673 0.0163 262Total Incidents 0.3327 0.0805 1296
Note: A unit of observation is a district in a given month and year. There are 447districts and 36 months. Red corridor refers to the districts marked as Naxalite districtsin Figure 1. They are defined as any district that had at least one Naxalite incident inthe 36 months of the data.
31
Table 3: RD Impact of NREGS on Naxalite Incidents
Panel A: Phase 1
Specification Fatalities Injuries Minor# Major# Total#
Linear 0.393*** 0.224* 0.0793* 0.0813*** 0.189**(0.149) (0.124) (0.0475) (0.0294) (0.0797)
R-squared 0.455 0.354 0.362 0.390 0.491
Linear Flexible Slope 0.327** 0.197** 0.0641 0.0699*** 0.161**(0.140) (0.0947) (0.0456) (0.0267) (0.0768)
R-squared 0.456 0.354 0.364 0.392 0.494
Quadriatic 0.433** 0.244* 0.0595 0.0967*** 0.185**(0.178) (0.141) (0.0517) (0.0364) (0.0937)
R-squared 0.454 0.354 0.364 0.388 0.491
Panel B: Phase 2
Specification Fatalities Injuries Minor# Major# Total#
Linear 0.0791 0.0833 0.122 -0.00108 0.118(0.0772) (0.0962) (0.104) (0.00399) (0.104)
R-squared 0.030 0.075 0.064 0.014 0.072
Linear Flexible Slope 0.111 0.164 0.158 -0.00147 0.155(0.0969) (0.137) (0.133) (0.00329) (0.130)
R-squared 0.022 0.069 0.053 0.015 0.062
Quadriatic 0.117 0.166 0.161 -0.000576 0.159(0.0969) (0.138) (0.133) (0.00257) (0.131)
R-squared 0.022 0.069 0.053 0.015 0.062
Controls include baseline averages of each dependent variable, time fixed effectsand estimated police-force changes. Phase 1 contains 2772 observations across231 clusters, and Phase 2 contains 2128 observations across 237 clusters. Thelast three columns enumerate the number of incidents in a month. Unit ofobservation is district-month. Standard errors clustered at district level.
32
Table 4: ITT-DID Impact of NREGS on Naxalite Incidents
Panel A: Stacked
Fatalities Injuries Minor# Major# Total#
NREGS 0.0396 0.103*** 0.0450*** 0.00710 0.0521***(0.0388) (0.0384) (0.0160) (0.00624) (0.0200)
Observations 16,085 16,085 16,085 16,085 16,085R-squared 0.342 0.211 0.339 0.268 0.425
Panel B: Phase 1
Fatalities Injuries Minor# Major# Total#
NREGS 0.0971 0.190*** 0.0460*** 0.0194 0.0654**(0.0721) (0.0713) (0.0161) (0.0126) (0.0259)
Observations 12,067 12,067 12,067 12,067 12,067R-squared 0.340 0.208 0.342 0.291 0.429
Panel C: Phase 2
Fatalities Injuries Minor# Major# Total#
NREGS -0.0113 0.0197 0.0500 -0.00651* 0.0435(0.0348) (0.0422) (0.0414) (0.00354) (0.0409)
Observations 10,001 10,001 10,001 10,001 10,001R-squared 0.121 0.088 0.195 0.104 0.230
Controls include time fixed effects, district fixed effects and estimatedpolice-force changes. The last three columns enumerate the number ofincidents in a month. Unit of observation is district-month. Standarderrors clustered at district level.
33
Table 5: RD Impact on Fatalities: Who is Affected?
Panel A: Phase 1
Specification Civilians Killed Police Killed Naxals Killed
Linear 0.214** 0.0184 0.192**(0.0901) (0.0290) (0.0970)
R-squared 0.351 0.227 0.226
Linear Flexible Slope 0.194** 0.0120 0.151*(0.0982) (0.0221) (0.0808)
R-squared 0.351 0.227 0.228
Quadriatic 0.245** 0.0227 0.158(0.115) (0.0344) (0.0976)
R-squared 0.350 0.227 0.227
Panel B: Phase 2
Specification Civilians Killed Police Killed Naxals Killed
Linear 0.0659 0.0272 -0.00480(0.0455) (0.0172) (0.0239)
R-squared 0.000 0.011 0.055
Linear Flexible Slope 0.0870 0.0335 0.00536(0.0586) (0.0229) (0.0247)
R-squared 0.000 0.000 0.059
Quadriatic 0.0867 0.0348 0.0108(0.0585) (0.0227) (0.0228)
R-squared 0.000 0.006 0.057
Controls include baseline averages of each dependent variable, time fixedeffects and estimated police-force changes. Phase 1 contains 2772 observa-tions across 231 clusters, and Phase 2 contains 2128 observations across 237clusters. The last three columns enumerate the number of incidents in amonth. Unit of observation is district-month. Standard errors clustered atdistrict level.
34
Table 6: RD Impact on Being ‘Affected’: Who is Affected?
Panel A: Phase 1
Specification Naxals affected Civilians affected Police affected
Linear 0.116* 0.278** 0.130(0.0671) (0.136) (0.0979)
R-squared 0.322 0.366 0.109
Linear Flexible Slope 0.0887* 0.268* 0.100(0.0484) (0.152) (0.0712)
R-squared 0.322 0.366 0.109
Quadriatic 0.0550 0.332* 0.139(0.0482) (0.179) (0.116)
R-squared 0.322 0.366 0.109
Panel B: Phase 2
Specification Naxals affected Civilians affected Police affected
Linear 0.0180 0.148 0.0316(0.0799) (0.0916) (0.0379)
R-squared 0.074 0.000 0.015
Linear Flexible Slope 0.0554 0.203 0.0422(0.122) (0.125) (0.0453)
R-squared 0.070 0.000 0.006
Quadriatic 0.0636 0.200 0.0455(0.122) (0.123) (0.0454)
R-squared 0.073 0.000 0.010
Controls include baseline averages of each dependent variable, time fixed effectsand estimated police-force changes. Phase 1 contains 2772 observations across231 clusters, and Phase 2 contains 2128 observations across 237 clusters. The lastthree columns enumerate the number of incidents in a month. Unit of observationis district-month. Standard errors clustered at district level.Affected means “killed, injured, abducted, surrendered or captured”
35
Table 7: ITT-DID: Who is Affected?
Panel A: Stacked
Naxals affected Civilians affected Police affected
NREGS 0.0545** 0.0768 0.0116(0.0255) (0.0502) (0.0182)
Observations 16,085 16,085 16,085R-squared 0.178 0.200 0.177
Panel B: Phase 1
Naxals affected Civilians affected Police affected
NREGS 0.111*** 0.158 0.0181(0.0417) (0.106) (0.0328)
Observations 12,067 12,067 12,067R-squared 0.174 0.205 0.128
Panel C: Phase 2
Naxals affected Civilians affected Police affected
NREGS -0.00699 0.0151 0.000247(0.0381) (0.0359) (0.0156)
Observations 10,001 10,001 10,001R-squared 0.081 0.099 0.048
Controls include time fixed effects, district fixed effects and estimatedpolice-force changes. The last three columns enumerate the number ofincidents in a month. Unit of observation is district-month. Standarderrors clustered at district level.Affected means “killed, injured, abducted, surrendered or captured”
36
Table 8: RD Impact of NREGS on Naxalite Incidents (No Police Controls)
Panel A: Phase 1
Specification Fatalities Injuries Minor# Major# Total#
Linear 0.364** 0.207* 0.0675 0.0760*** 0.169**(0.147) (0.120) (0.0462) (0.0288) (0.0795)
R-squared 0.444 0.354 0.349 0.385 0.472
Linear Flexible Slope 0.317** 0.188** 0.0545 0.0669** 0.146*(0.143) (0.0906) (0.0453) (0.0267) (0.0788)
R-squared 0.444 0.354 0.350 0.386 0.475
Quadriatic 0.489** 0.209 0.0593 0.0992*** 0.188*(0.191) (0.135) (0.0562) (0.0376) (0.104)
R-squared 0.442 0.354 0.349 0.382 0.471
Panel B: Phase 2
Specification Fatalities Injuries Minor# Major# Total#
Linear 0.0505 0.104 0.0850 -0.00200 0.0805(0.0538) (0.0711) (0.0723) (0.00284) (0.0721)
R-squared 0.029 0.061 0.066 0.008 0.074
Linear Flexible Slope 0.0823 0.165 0.122 -0.00250 0.119(0.0771) (0.106) (0.105) (0.00305) (0.103)
R-squared 0.004 0.033 0.033 0.006 0.041
Quadriatic 0.0926 0.174 0.131 -0.00174 0.129(0.0811) (0.115) (0.111) (0.00240) (0.110)
R-squared 0.019 0.055 0.052 0.008 0.060
Controls include baseline averages of each dependent variable, time fixed effects.Phase 1 contains 2772 observations across 231 clusters, and Phase 2 contains2128 observations across 237 clusters. This table does not control for estimatedpolice force changes. The last three columns enumerate the number of incidentsin a month. Unit of observation is district-month. Standard errors clustered atdistrict level.
37
Table 9: ITT-DID Impact of NREGS on Naxalite Incidents (No Police Con-trols)
Panel A: Stacked
Fatalities Injuries Minor# Major# Total#
NREGS 0.0303 0.108*** 0.0349*** 0.00686 0.0417**(0.0382) (0.0372) (0.0134) (0.00636) (0.0173)
Observations 16,085 16,085 16,085 16,085 16,085R-squared 0.341 0.211 0.338 0.267 0.424
Panel B: Phase 1
Fatalities Injuries Minor# Major# Total#
NREGS 0.0869 0.187*** 0.0393** 0.0189 0.0582**(0.0715) (0.0680) (0.0152) (0.0122) (0.0250)
Observations 12,067 12,067 12,067 12,067 12,067R-squared 0.339 0.208 0.340 0.290 0.427
Panel C: Phase 2
Fatalities Injuries Minor# Major# Total#
NREGS -0.0574 0.0458 0.0347 -0.00976* 0.0249(0.0399) (0.0467) (0.0350) (0.00523) (0.0345)
Observations 10,001 10,001 10,001 10,001 10,001R-squared 0.114 0.086 0.193 0.102 0.227
Controls include time fixed effects, district fixed effects. This table doesnot control for estimated police force changes. The last three columnsenumerate the number of incidents in a month. Unit of observation isdistrict-month. Standard errors clustered at district level.
38
Table 10: RD: Doughnut Hole Approach
Panel A: Phase 1
Specification Fatalities Injuries Minor# Major# Total#
Linear 0.317** 0.0865 0.0792* 0.0633*** 0.166**(0.143) (0.0767) (0.0427) (0.0238) (0.0683)
R-squared 0.465 0.374 0.383 0.419 0.521
Linear Flexible Slope 0.300** 0.0939 0.0709 0.0569** 0.154**(0.145) (0.0699) (0.0431) (0.0239) (0.0708)
R-squared 0.465 0.374 0.384 0.419 0.521
Quadriatic 0.334** 0.0775 0.0618 0.0688** 0.155**(0.164) (0.0747) (0.0454) (0.0270) (0.0776)
R-squared 0.465 0.374 0.384 0.418 0.521
Panel B: Phase 2
Specification Fatalities Injuries Minor# Major# Total#
Linear 0.103 0.0836 0.139 -0.000372 0.140(0.0973) (0.115) (0.131) (0.00489) (0.130)
R-squared 0.030 0.064 0.054 0.019 0.058
Linear Flexible Slope 0.147 0.152 0.176 -0.000151 0.180(0.122) (0.138) (0.163) (0.00404) (0.161)
R-squared 0.000 0.032 0.028 0.019 0.031
Quadriatic 0.157 0.158 0.182 0.000894 0.187(0.124) (0.139) (0.165) (0.00335) (0.164)
R-squared 0.015 0.053 0.040 0.020 0.043
Controls include baseline averages of each dependent variable, time fixed effectsand estimated police-force changes. Phase 1 contains 2232 observations across186 clusters, and Phase 2 contains 2029 observations across 226 clusters. Thelast three columns enumerate the number of incidents in a month. Unit ofobservation is district-month. Standard errors clustered at district level.This table drops districts with rank -1, 0 and +1 as a robustness check againstmeasurement error and slippage.
39
Figure 1: Red Corridor Districts
Note: Red corridor districts are all districts that had at least one Naxalite incidentin the 36 months of the data used in this paper.
40
Figure 2: Predicted NREGS Phase 1 Assignment
Note: All colored districts are predicted to receive NREGS in the first phase basedon the algorithm. Light grey districts actually receive NREGS in Phase 1, whereasdark grey districts do not.
41
Figure 3: Predicted NREGS Phase 2 Assignment
Note: All colored districts are predicted to receive NREGS in the second phase basedon the algorithm. Light grey districts actually receive NREGS in Phase 2, whereas darkgrey districts do not.
42
Figure 4: Distribution of Index over State-Specific Ranks (Phase 1 Treatment Assign-ment)
0.5
11.
52
Pla
nnin
g C
omm
issi
on In
dex
−40 −20 0 20 40Normalized State Rank Phase 1
Figure 5: Distribution of Index over State-Specific Ranks (Phase 2 Treatment Assign-ment)
.51
1.5
22.
5P
lann
ing
Com
mis
sion
Inde
x
−20 −10 0 10 20Normalized State Rank Phase 2
43
Figure 6: Discontinuity of treatment status for Phase 10
.2.4
.6.8
1
−20 0 20 40normalized state−specific rank for Phase 1
95% CIprobability of receiving NREGS in Phase 1
Note: The used bin size is 1, so each individual rank.
Figure 7: Discontinuity of treatment status for Phase 2
0.2
.4.6
.81
−20 −10 0 10 20normalized state−specific rank for Phase 2
95% CIprobability of receiving NREGS in Phase 2
Note: Figure 7 excludes Phase 1 districts. The used bin size is 1, so each individualrank.
44
Figure 8: Total Number of Incidents
Figure 9: Number of Naxalites Affected: Killed/injured/captured/surrendered
45
1
Hate Crimes in India: An Economic Analysis of Violence and Atrocities against Scheduled
Castes and Scheduled Tribes
Smriti Sharma1
Department of Economics
Delhi School of Economics
This version: May 2012
Abstract
Crimes against the historically marginalized Scheduled Castes and Scheduled Tribes (SC/ST) by the
upper castes in India represent an extreme form of prejudice and discrimination. In this paper, we
investigate the effect of changes in relative material standards of living between the SC/ST and
upper castes, as measured by consumption expenditures, on changes in the incidence of crimes
against SC/ST. Using official district level crime data for the period 2001-‐10, we find a positive
association between crimes and expenditure of SC/ST vis-‐à-‐vis the upper castes suggesting that a
widening of the gap between groups is associated with a decrease in caste-‐based crimes. Moreover,
this effect seems to be driven by the upper castes’ responding to changes in status quo. The results
are robust to changes in specifications and modeling assumptions.
Keywords: Hate Crimes, Castes, India
JEL: J15, K42
1 Email:[email protected]. Address for correspondence: Department of Economics, Delhi School of Economics, University of Delhi, Delhi-‐110007. I am grateful to my advisors Ashwini Deshpande and Parikshit Ghosh for their excellent guidance. Thanks are also due to Jeffrey Nugent, J.V. Meenakshi, Abhiroop Mukhopadhyay, Jared Rubin, Saurabh Singhal and participants at the IRES 2012 Workshop, ISNIE 2012, WEAI 2012, Delhi School of Economics, Indian Statistical Institute and Jawaharlal Nehru University for their comments, and to officials at the National Crimes Records Bureau for providing the data and clarifying my queries. All remaining errors are mine.
2
1. Introduction
In India, ex-‐untouchable castes and several tribal groups are victims of discrimination, economic and
social exclusion and a stigmatized identity. Additionally, similar to hate crimes in other parts of the
world, these groups have been victims of bias-‐motivated crimes and atrocities at the hands of the
upper castes. Atrocities against lower castes routinely take the form of rape of women, abuse by
police personnel, harassment of lower caste village council heads, illegal land encroachments, forced
evictions and so on (Human Rights Watch, 1999). These instances are in blatant violation of the
Indian constitution that abolished untouchability and upholds the ideal of equality among all
citizens. Subsequently, there have been other provisions such as the Scheduled Castes and
Scheduled Tribes (Prevention of Atrocities) Act, 1989, which specifically target such hate crimes. In
2006, acknowledging the gravity of the problem, Indian Prime Minister Manmohan Singh equated
the practice of untouchability to that of apartheid2.
In this paper, we analyze crimes against the historically disadvantaged Scheduled Castes and
Scheduled Tribes (ex-‐untouchables and marginalized tribes, SC and ST respectively) by the upper
castes to understand the mechanisms that cause crimes based on group identity to occur
repeatedly. Since the caste system has meant a hierarchy such that the upper castes have
traditionally been economically better-‐off than the lower castes with resulting social dominance, the
objective of this study is to analyze whether regional variations in the incidents of violence are
systematically linked to variations in relative group economic outcomes between the upper and
lower castes and tribes. The motivation is to examine whether crimes committed by the upper
castes against the lower castes are a means of asserting their superiority or expressing their
frustration at the shift in status quo.
This paper is among the first to analyze data on crimes committed by upper castes against SCs and
STs. This is largely facilitated by the fact that starting 2001, official data on such crimes became
available at the level of the district. While there is some literature on general violent crimes in India
(Dreze and Khera, 2000; Prasad, 2009; Chamarbagwala and Sharma, 2010) and crimes against
women (Iyer et al, forthcoming; Sekhri and Storeygard, 2011), the only existing piece of research
studying crimes against SC/ST groups is Bros and Couttenier (2010). Using cross-‐sectional district-‐
level crime data for 2001, they find crimes against SC/ST groups to be higher in districts that have
greater commonality of water sources. Common water sources imply water sharing between castes
which is considered ritually polluting for the upper castes—more so in rural areas—and is often 2 Rahman, Maseeh. “Indian Leader Likens Caste System to Apartheid Regime”. The Guardian, Dec. 28, 2006. Available at http://www.guardian.co.uk/world/2006/dec/28/india.mainsection (accessed January 5, 2012)
3
countered with an act of violence against the lower castes. Our study investigates a different
hypothesis and exploits the panel structure of the data through which fixed unobservable factors
can be controlled.
Our paper can be considered closest in terms of motivation to Mitra and Ray (2010) inasmuch as
they too consider the hypothesis that the economic empowerment of a minority or oppressed group
leads to a violent backlash against them. They develop a theoretical model to determine how
differences in economic progress between religious groups lead to inter-‐group conflict and test it
using Hindu-‐Muslim riots data for India. They find that an improvement in Muslims’ well-‐being leads
to an increase in Hindu-‐Muslim riots while Hindus’ well-‐being has no significant effect. However,
there are two crucial differences between the two studies. Firstly, they analyze communal riots,
which represent violence involving a large group of people, while we study individually targeted
caste-‐based violence. Secondly, and more importantly, their data do not allow separation of
perpetrators and victims by religion, except by inference, whereas in our data, the identification
between victims (SC/ST) and offenders (non-‐SC/ST) is clear from the start. Thus, this study is a new
contribution to the discussion of group-‐based violence in the Indian context. However, there is an
extensive social science literature from the United States that has studied racial violence and this
paper builds on that literature.
Using district level data on such crimes and expenditures as a proxy for material standard of living,
we find that the incidence of caste violence is positively correlated with the expenditure of the lower
castes and tribes relative to the expenditure of the upper castes. Dividing the crimes into pre-‐
dominantly violent crimes and non-‐violent crimes, we find that changes in relative material
standards of living between groups lead to changes in violent crimes aimed at extracting some form
of economic surplus or property from the victims.
Although discrimination has largely been discussed in the context of labor markets and access to
public goods, this is among the first studies to quantitatively analyze the phenomenon of crimes
targeted at the SC/ST groups. Since crimes committed by non-‐SC/ST individuals against individuals
belonging to SC/ST groups fall under the broad category of hate crimes, this paper will be nested in
that literature while also drawing from the general crime literature.
The remainder of this paper is organized as follows. Section 2 provides a background on the caste
system, existing inequalities and a brief overview of the hate crimes literature. Section 3 describes
4
the dataset and the summary statistics. Section 4 presents the results and section 5 discusses and
concludes.
2. Related Literature
2.1 The Indian caste system
The ‘caste system’ is an arrangement of the Hindu population into several thousand groups called
jatis (castes). These groups have emerged from the ancient varna system (also translated as caste)
according to which society was divided into initially four, later five, hereditary, endogamous,
mutually exclusive and occupation-‐specific groups. At the top of the varna system were the
Brahmins (priests and teachers) and the ‘Kshatriya’ (warriors and royalty), followed by ‘Vaishya’
(traders, merchants and moneylenders) and finally the ‘Shudra’ (engaged in menial labor and low-‐
end jobs). Over time, the Shudras split into two tiers, with those engaged in the most menial and
dirty jobs being called the ‘Ati-‐Shudras’. The Ati-‐Shudras were considered untouchable, such that
any contact with them was seen as polluting. They were forced to live in segregated housing, denied
access to schools and places of worship attended by upper castes, and required to maintain physical
distance from upper castes in order to not pollute them. Additionally, there are the indigenous tribes
(or the Adivasis) who on account of geographical isolation, primitive agricultural practices and
distinct lifestyle and customs have been socially distanced and face large-‐scale exclusion from
mainstream Indian society.
In 1950, the Constitution notified the untouchable jatis and the Adivasis as ‘Scheduled Castes’ and
‘Scheduled Tribes’ respectively, in order to remedy their extreme social, educational and economic
backwardness3. Affirmative action was extended to them in the form of reservations or quotas in
national and state legislatures, local village councils and institutions of higher education and
government jobs. In addition to the SCs and STs, there is a third category to which reservations have
been extended since the early 1990s. This group, known as the ‘other backward classes’ (OBC) while
not burdened with the stigma of untouchability, were socially and educationally backward and
suffered from a persistent lack of opportunity and poor socio-‐economic outcomes4.
While the reservation policy has made a discernible positive impact in some dimensions, gaps
remain between the SC/ST and non-‐SC/ST groups. Empirical studies continue to find evidence of
3 While ‘Scheduled Castes’ is the official nomenclature, ex-‐untouchables prefer to self-‐identify themselves as Dalit (meaning the oppressed) as a term of pride. We will use both terms depending on the context. 4 Starting late 1990s, large-‐scale datasets use four social group categories: SC, ST, OBC and ‘others’. ‘Others’ is a reasonable approximation of the upper caste category.
5
caste-‐based discrimination in labor markets and ‘pre-‐market’ discrimination in terms of access to
public goods5. Shah et al (2006) in a survey of 565 villages across 11 large states in 2001-‐02
document the widespread practice of untouchability in a significant proportion of villages in the
form of denial of entry to Dalits into non-‐Dalit homes and places of worship, blocking access to use
common water sources, separate seating for Dalit students in schools etc. Moreover, there is a
burden of a ‘stigmatized ethnic identity’ (Berreman, 1971) that even richer Dalits continue to live
with. The following account by a Dalit surgeon effectively summarizes the sentiment behind
‘stigmatized ethnic identity’: “I am a micro-‐surgeon specializing in hand and spinal reconstruction,
and am [a Member of Legislative Assembly] from Bihar, but I still remain very much a dalit—a dhobi,
to be precise— open to routine humiliation from the upper castes.”6 This stresses the fact that
despite significant policy initiatives, notions about caste rigidities are deeply ingrained and upward
economic mobility has not necessarily ensured social integration and tolerance. Violence against
lower castes is only the most severe manifestation of that intolerance.
2.2 Hate crimes
The term hate crime refers to “unlawful, violent, destructive, or threatening conduct in which the
perpetrator is motivated by prejudice toward the victim’s putative social group” (Green et al., 2001,
pg.480). The most crucial difference between a hate crime and a similar non-‐hate crime is the
underlying motivation. While a conventional crime might be motivated by a desire to expropriate
resources from the victim for the personal gain of the offender, in the case of hate crimes, there is a
deliberate intention to victimize an individual because of his membership in a certain social group.
A review of the literature, most of which comes from the United States and Europe, indicates that
among other things, relative economic position of the dominant group vis-‐à-‐vis the subaltern group
is an important determinant of hate crimes. Beck and Tolnay (1990) find that during the period
1882-‐1930, mob violence against blacks in the southern states of USA increased during the years
when economic competition intensified. Economic competition was greater during periods when
cotton prices fell and labor demand declined thereby leading to greater competition for the reduced
number of jobs between blacks and whites. Price et al (2008) find that during 1882-‐1920, the
probability of being lynched was positively associated with the victim’s former slave status thereby 5 For excellent detailed discussions on economic discrimination in India, see Deshpande (2011) and Thorat and Newman (2010). 6 Kanaujia, Ramsunder Ram. “Surgeon Second, Dhobi First”, Tehelka, Feb. 3, 2007, available at http://www.tehelka.com/story_main26.asp?filename=Cr020307Shadow_lines.asp (Accessed Sept 14, 2011)
6
indicating the importance of racial stigma in white-‐black relations. Further, they also find the
lynching probability to be inversely related to the cotton prices in the southern US counties. In a
more contemporary setting, Jacobs and Wood (1999) investigate the relationship between economic
and political competition and interracial murders for US cities. As economic competition for jobs
increases between blacks and whites, murders of blacks by whites increase. Additionally, cities with
a black mayor experience more murders of blacks by whites. Eitle et al (2002) find that economic
competition—measured by the ratio of white and black unemployment rate—has a positive effect
on crimes committed by whites against blacks. On the other hand, political threat—ratio of black to
white voters—has no significant effect on the interracial crimes. Gale et al (2002) use American
state-‐level data for 1992-‐1995 and find unemployment rates and black-‐white income gaps to have a
positive effect on hate crimes committed by whites against blacks. Krueger and Pischke (1997) find
no relationship between incidence of anti-‐foreigner crimes and unemployment rate or wages in
post-‐unification Germany during 1991-‐93. The percentage of foreigners has no effect on ethnic
crimes in the western Germany but in eastern Germany, foreigner victimization rate falls as their
relative number increases. However, Falk et al (2011) find unemployment rates to positively affect
violent and non-‐violent right wing extremist crimes in Germany with the effect on non-‐violent crimes
being greater.
3. Data
3.1 Crime data
The crime data used in this paper are from the annual publication ‘Crime in India’ by National Crime
Records Bureau (NCRB), Government of India. This data is based on complaints or ‘first information
reports’ filed with the police and not the cases convicted7. A First Information Report (FIR) is a
written document prepared by the police when they receive information about the commission of a
‘cognizable’ offence from either the victim or by someone on his on her behalf. It is only after the FIR
is registered in the police station that the police take up investigation of the case. While earlier years
reported crimes at the state level, since 2001, data on crimes against SC/ST under various categories
have become available at the district level. The distinctive feature of this dataset is its classification
7 As defined by the Code of Criminal Procedure of India, a ‘cognizable’ offence is one in which the police is empowered to register an FIR, investigate, and arrest an accused without a court issued warrant. A ‘non-‐cognizable’ offence is an offence in which police cannot register an FIR, investigate or arrest without the prior permission from the court. NCRB data records only cognizable offences.
7
system. The data are defined in such a way that the victim belongs to the SC/ST group and the
offender to a non-‐SC/ST group. For this study, we use the crime data from 2001 to 2010 for 415
districts that make up the following 18 large states: Haryana, Himachal Pradesh, Punjab, Uttar
Pradesh, Uttarakhand, Karnataka, Andhra Pradesh, Kerala, Tamil Nadu, Bihar, Jharkhand, Orissa,
West Bengal, Madhya Pradesh, Chhattisgarh, Gujarat, Rajasthan and Maharashtra.
There are two main types of crimes: those reported under the Indian Penal Code (IPC) and those
that are registered under the Special and Local Laws (SLL). IPC crimes include: i) murder, ii) rape, iii)
physical assault or hurt, iv) kidnapping, v) robbery, vi) arson, vii) dacoity and viii) other classified IPC
crimes. Other classified IPC crimes is a residual category that includes crimes such as assaulting
public servants, killing cattle, criminal trespass and intimidation etc. Crimes under SLL are: i)
Protection of Civil Rights Acts, 1955, and ii) The Scheduled Castes and Scheduled Tribes (Prevention
of Atrocities) Act, 1989. The SLL categories constitute special social enactments to safeguard the
interests of SC/ST groups. The Prevention of Atrocities Act specifies provisions for rehabilitation and
compensation of victims and setting up of special courts to expedite the trial of cases. Examples of
crimes included under SLL are: denying admission to Dalits into places of recreation and worship,
educational institutions and hospitals; denying Dalits access to water sources; wrongfully occupying
land owned by SC/ST; stripping them naked; practice of untouchability; compelling them to do
bonded labor or scavenging jobs and so on8. Broadly, the IPC crimes include acts of overt force and
aggression and are predominantly violent crimes. On the other hand, SLL crimes are untouchability
related offences with the intention of humiliating members of the lower castes, with some amount
of violence. Hence, they are largely non-‐violent crimes.
The issue of under-‐reporting of crime is a standard limitation of most official data on crime, even for
developed countries. For a hate crime, under-‐reporting is expected since there is courage required
on the part of the victims to report the crime due to a fear of reprisal. Moreover, the victim is likely
to feel ashamed in reporting a crime where he has been humiliated on account of his social identity.
Ideally, one would like to use victimization surveys—that ask random samples of individuals whether
they have been victims of certain types of crimes over a fixed recall period as compared to police
data that is a measure of crimes that get reported by victims to the police—to study crimes of this
nature. However, in the absence of such data, this paper makes use of best available nationally
representative data and we believe that is a good starting point, especially since quantitative
evidence on crimes against SCs and STs is limited. Moreover, with the fixed effects in our regression,
we are able to control for the district-‐specific time-‐invariant component of under-‐reporting.
8 The complete list of SLL crimes against SC/ST is in appendix 1.
8
3.2 Explanatory variables
Since our unit of analysis is the district, district-‐level information on the explanatory variables is
calculated from the large-‐scale household surveys conducted once in five years by the National
Sample Survey Organization (NSSO). Since our crime data spans the period 2001-‐2010, we use NSS
data from the ‘Consumer Expenditure Survey’ and ‘Employment-‐Unemployment Survey’ modules of
1999-‐2000 (55th round) and 2004-‐05 (61st round). Districts in both rounds of NSS data have been
made comparable using the weights provided by Kumar and Somanathan (2009).
In order to utilize the entire time series of district-‐level crime data and also to match it with the two
rounds of district-‐level data from NSS, for the first period, we aggregate crimes for years 2001 to
2005 and for the second period, we aggregate crimes for years 2006 to 2010. Therefore, we have a
two period panel with 415 districts in each period.
The primary variable of interest is the material standard of living of SC/ST relative to that of the
upper castes. In order to capture standard of living, we use data on consumption expenditure from
the NSS Consumer Expenditure Survey. Thus, our principal variable is defined as the logarithm of
ratio of expenditure of SC/ST and expenditure of upper castes. Ideally, one would like to use some
measure of income to study differences in standard of living but that is notoriously hard to collect in
a developing country like India. In developing countries, expenditure serves as a good proxy for
income for several reasons. Firstly, at low levels of income, savings are negligible resulting in a close
correspondence between income and consumption expenditure. Secondly, wage or earnings data,
even when reliable, do not account for days of employment and seasonality of work. Moreover,
wage or earnings data in the NSS is reported only for those who are employed in the regular salaried
sector and not for those who are self-‐employed or casual workers. Thirdly, payment is often in kind
and wage data typically accounts for the monetary component of earnings. Finally, like in the case of
agricultural households, a household is both a production and consumption unit and it is difficult to
distinguish between receipts and outflows (Deaton, 1997).
Among other regressors, we control for percentage of SC/ST in the district and its squared term.
District-‐level average per capita expenditure accounts for overall prosperity. Expenditure-‐based Gini
coefficient accounts for overall inequality. We control for percentage of population living in rural
areas since caste-‐based crimes are likely to be a predominantly rural phenomenon. Educational
attainment is controlled for by introducing dummies for different levels of education: illiterate,
primary, secondary, higher secondary and above. Unemployment is an important determinant of
crime since unemployed people with no legal income are more likely to engage in illegal activities as
9
a way of earning an income. However, in developing countries, the underemployment rate is a more
accurate measure of time utilization. Underemployment is commonly defined as the underutilization
of labor time or skills of the employed either due to seasonality of work or lack of sufficient work.
Percentage of males in the 15-‐24 age groups in the population represents the size of the group that
is most likely to engage in criminal activity.
We also control for political competition at the state-‐level by using effective number of parties
(Laakso and Taagepara, 1979) that is calculated using data from the state assembly election reports
from the Election Commission of India9. The idea is that in the event of smaller number of parties
competing, since a larger share of votes is required for the winning party, parties need to build
broad alliances spanning all social groups and cannot explicitly cater to the interests of a particular
group, while in case of a larger number of parties, since the required winning margin is smaller,
parties rely on particular social groups for support. Based on this reasoning, we would expect states
with greater electoral competition (larger effective number of parties) to be more sympathetic to
the cause of the SC/ST groups thereby leading to lesser violence against them10.
3.3 Descriptive Statistics
We define SC/ST total crime rate as the number of total crimes against SC/ST per 100000 SC/ST
population. SC/ST crime rates using IPC and SLL crimes against SC/ST are also measured per 100000
SC/ST population. Over the period 2001-‐10, SC/ST crime rates have registered a decline. IPC crimes
account for approximately 61 percent of total crimes against SC/ST whereas SLL crimes constitute
the remaining 39 percent. In terms of broad state-‐level statistics, Rajasthan has the highest SC/ST
total crime rate averaged over the period (29.82). Other states with high SC/ST crime rates are
Madhya Pradesh (25.83), Andhra Pradesh (23.25), Uttar Pradesh (16.57) and Bihar (16.26). The
lowest crime rates are recorded in West Bengal (0.12) and Punjab (1.81). In terms of IPC crimes
against SC/ST, Rajasthan (24.2), Madhya Pradesh (23.8) and Andhra Pradesh (14.21) have high crime
rates whereas in terms of SLL crimes against SC/ST, Bihar (10.63), Karnataka (9.66) and Andhra
Pradesh (9.04) are the states reporting high rates.
9 The formula for effective number of parties is n= 1/∑pi
2 where pi is the proportion of votes received by the ith
party in the state elections. Instead of using the total number of parties, this measure places greater weight on parties that have a higher share of votes as compared to those with a low vote share. 10 Wilkinson (2004) finds that Indian states with higher effective number of parties experience fewer Hindu-‐Muslim riots.
10
Table 1 contains the summary statistics of the district-‐level data for the period 2001-‐10. Of the
average 430 total crimes against SC/ST per district, approximately 288 are IPC crimes and 142 are SLL
crimes. The SC/ST total crime rate is 100 while the SC/ST IPC crime rate is 67 and SC/ST SLL crime
rate is 33. Among the IPC crimes, we make a distinction between crimes against body and non-‐body
crimes. Body crimes are the sum of murder, rape, kidnapping and physical assault/hurt. Non-‐body
crimes are the sum of dacoity, robbery, arson and other classified IPC crimes. The SC/ST body crime
rate is 23.07 and the non-‐body crime rate is 43.47. The general crime rate which measures general
crimes where the victims are non-‐SC/ST—defined as total IPC crimes in the district less IPC crimes
against SC/ST per 100000 non-‐SC/ST population—is 1544. On average, crime rates against SC/ST and
general crime rates registered a decline between the first and second period.
The district-‐level average expenditure is Rs. 535 and inequality as measured by expenditure-‐based
Gini is 0.25. The SC/ST average expenditure is Rs. 433, while it is Rs. 679 and Rs. 523 for the upper
castes and OBC respectively. Between the two periods, per capita expenditures of all social groups
increased although the rate of increase was slowest for the SC/ST groups, thereby leading to a
decline in SC/ST expenditure relative to upper castes’ expenditure. On average, SC/ST expenditure is
68 percent of the upper caste expenditure.
SC/ST account for 29 percent of the district-‐level population and 79 percent of the population is in
rural areas. The underemployment rate is around 16 percent. Males in the 15-‐24 age groups make
up 9 percent of the population. 46 percent of the population is illiterate, 20 percent have completed
primary education, 24 percent have completed secondary education and only 10 percent has
completed higher secondary and higher levels of education. The state-‐level effective number of
parties is around 4.6.
4. Analysis
4.1 Results
We use a linear fixed effects regression model. District fixed effects are included to control for
district-‐specific time-‐invariant unobservable factors that may influence the relationship between
crime and the explanatory variables. Among other factors, district fixed effects control for the time-‐
invariant district-‐specific under-‐reporting of crime. A time dummy is included for the second period.
Standard errors are clustered at the district level to account for possible correlated shocks to
district-‐level crimes over time.
11
The general form of the estimating equation is:
Ydt= α1+ β*edt+ ∑k µkXkdt +δd+ ϒt+ εdt
Where ydt is the logarithm of the SC/ST crime rate in district d in time period t11, edt is logarithm of
the relative expenditure between SC/ST and upper castes, Xkdt is the vector of k controls in district d
at time t, δd are district fixed effects, ϒt is a time dummy for the second period and εdt is the error
term. We expect the coefficient of the relative expenditure term, β, to be positive.
Table 2 presents the main results. In column 1, the dependent variable is the SC/ST total crime rate.
We use the following explanatory variables: percentage SC/ST, percentage SC/ST squared, percent
rural, Gini coefficient, underemployment rate, education dummy variables, percentage of young
males and effective number of parties. The coefficient of the relative expenditure term is positive
and significant. Since crime rates and relative expenditures show a downward trend between the
two periods over which we analyze the data, this implies that a 1 percent decrease in the relative
expenditures or widening of the gap between lower and upper castes is associated with a 0.3
percent decrease in violence. Percentage SC/ST and its quadratic term are negative and positive
respectively, suggesting that an increase in percentage of SC/ST is associated with a decrease in
victimization and the decrease is slower as the percentage SC/ST increases. Across all regressions,
we obtain similar results for the SC/ST percentage term. Becker (1957) suggests that the effect of
numbers of the minority groups can go in either direction: an increase in numbers could either
reduce prejudice and hostility on account of greater interaction and familiarity or could have an
adverse effect by fuelling fears that the minority group is trying to challenge the dominant group.
In column 2 of Table 2, the dependent variable is the SC/ST IPC crime rate. We use the same
explanatory variables as in column 1. Results are qualitatively similar to column 1. A 1 percent
decrease in the relative expenditure is associated with a 0.35 percent decrease in IPC crimes
committed by the upper castes against the SC/ST groups. Overall higher inequality in the district, as
measured by the Gini, is positively associated with IPC crime rates. This result is in accordance with
other literature that finds inequality to be significant determinant of violent crimes in society (Kelly,
2000; Fajnzylber et al, 2002). In column 3 of Table 2, the dependent variable is the SC/ST SLL crime
rate. In this regression, the relative expenditure term is insignificant thereby indicating that the
11 Results are qualitatively similar if crime rates are calculated per 100000 total population instead of SC/ST population.
12
relative economic position of SC/ST vis-‐à-‐vis the upper castes is not associated with the SLL crime
rate. The coefficient on the effective number of parties was insignificant in columns 1 and 2, it is now
negative and significant implying that as the number of parties competing in the state increases, SLL
crimes against SC/ST register a decline.
To understand the effects of group-‐wise economic progress on the incidence of caste violence, in
Table 3, instead of using relative expenditures, we enter the logarithm of expenditures group-‐wise:
expenditure of SC/ST, expenditure of other backward classes (OBC) and the expenditure of upper
castes (UC). Since we have group-‐wise expenditures, we do not control for overall expenditure. In
column 1, the dependent variable is the SC/ST total crime rate. While OBC expenditure and SC/ST
expenditure have no effect on crime rate, the upper castes’ expenditure coefficient is negative and
significant implying that a 1 percent increase in their expenditure is associated with a 0.34 percent
decrease in crime rates. The most plausible mechanism at play is that of opportunity cost. This idea
is crucial to models of crime originating with Becker’s influential work (1968). With an increase in the
material standard of living, each unit of time spent in committing crimes becomes more costly for
the perpetrators. As we show later, upper castes’ expenditure has no effect on the incidence of
general crimes implying that it is the lower castes that are differentially targeted with changes in
relative economic positions over time. Moreover, as our data indicate, the average expenditure
increased most rapidly for the upper castes and slowest for the SC/ST groups, thereby increasing the
gap between the two groups and diminishing the perceived threat associated with economic
position of subaltern group relative to the dominant group12.
In column 2, the dependent variable is the SC/ST IPC crime rate. Again, while SC/ST and OBC
expenditure is insignificant, a 1 percent increase in upper castes’ expenditure is associated with a
0.56 percent decrease in IPC crime rates. In column 3, the dependent variable is the SC/ST SLL crime
rate. All the group-‐wise expenditure terms are insignificant. The coefficient on the effective number
of parties is negative and significant. As in the relative expenditure specification in of Table 2, the
Gini coefficient is positively associated with IPC crime rate but uncorrelated with SLL crime rate.
12 Blumer (1958) posits that perceived threat among the dominant group manifests in the following ways: i) a feeling of superiority; ii) a feeling that the subordinate group is intrinsically different; iii) a feeling of exclusive claim over certain privileges; iv) a fear that the subordinate group desires a greater share of the dominant group’s prerogatives.
13
Results from Tables 2 and 3 jointly show that firstly, while relative expenditure is an important
determinant of caste-‐based crimes, it is the perpetrator (upper caste) characteristics and not the
victim (SC/ST) characteristics driving the results. Secondly, while IPC crimes are correlated with
relative expenditure and upper castes’ expenditure, SLL crimes are not. This indicates that changes
in relative economic status of groups are associated with changes in largely violent crimes where the
intention is to expropriate or wrest economic surplus from the victims rather than crimes that seek
to insult and humiliate victims on account of their lower social status. This seems to be consistent
with findings from field surveys that report increases in violent acts by upper castes whenever lower
castes try to assert their rights or demand their fair share by way of wages, forest rights or basic
human rights. SLL crimes that are largely non-‐violent occur on a more routine basis as a result of
long term social attitudes, for instance beliefs about hierarchy or the “right” order of the world,
place of the Dalits in the social hierarchy and therefore might not be as closely related to changes in
economic status.
In Table 4, we decompose the IPC crimes into two mutually exclusive categories: crimes against
body, and non-‐body crimes. Body crimes are the sum of murder, rape, kidnapping and physical
assault/hurt. Non-‐body crimes are the sum of dacoity, robbery, arson and other classified IPC crimes
and are largely property crimes. In columns 1 and 2, the dependent variable is the SC/ST body crime
rate and we report results of the relative expenditure specification and the group-‐wise expenditure
specification respectively. Neither relative expenditure nor the upper castes’ expenditure is
associated with the incidence of body crimes. In columns 3 and 4, the dependent variable is the
SC/ST non-‐body crime rate. Column 3 indicates a positive association between relative expenditure
of SC/ST and upper castes and non-‐body crimes. In column 4, upper castes’ expenditure is negatively
associated with non-‐body victimization. The coefficient on the Gini is positive and significant for the
non-‐body crimes but not for the body crimes. These results suggest that it is the non-‐body crimes
component of the IPC crimes against SC/ST that is responsive to changes in relative expenditure and
upper castes’ expenditure. This indicates that IPC crimes against SC/ST occur as crimes with the
objective of depriving victims of their material property rather than inflicting physical bodily harm.
These findings further strengthen our inference from Tables 2 and 3 that crimes by upper castes are
committed with the objective of grabbing the economic surplus and destruction of material property
of the SC/ST groups.
In Table 5, we add a control variable to capture how crime-‐prone the district is in general. While the
relationship between hate crimes and non-‐hate motivated crimes has not been clearly established in
14
the literature since the underlying motivation for hate crimes and similar non-‐hate crimes is
different, it is plausible that areas with a culture of violence or higher level of general crimes are
more susceptible to the occurrence of hate crimes on account of poorer law enforcement
machinery. We measure how crime-‐prone a district is by defining a variable called the general crime
rate that measures the general criminality in the district. It is measured as total IPC crimes in the
district less IPC crimes against SC/ST per 100000 non-‐SC/ST population. In columns 1 and 2, we use
the SC/ST total crime rate and in columns 3 and 4, we use the SC/ST IPC crime rate13. Results from
Tables 2 and 3 are robust to controlling for logarithm of general crimes rate. Moreover, in all
specifications, we find that the coefficient on general crime rate is positive and significant suggesting
that more crime-‐prone districts do in fact experience greater victimization of the SC/ST community.
In Tables 6a and 6b, we model the number of total crimes against SC/ST and IPC crimes against SC/ST
respectively as count data and employ a negative binomial regression model. We add the logarithm
of the SC/ST population on the right hand side. Since the relative expenditure and group-‐wise
expenditures are in log form, the coefficients can be interpreted as elasticities. In column 1, the main
explanatory variable of interest is the relative expenditure between SC/ST and upper castes, the
coefficient of which is positive and significant. In column 3, we use the group-‐wise expenditures
specification and find that the upper castes’ expenditure is negatively associated with violence. In
columns 2 and 4, we also control for logarithm of general crime rate in the district and the results
are qualitatively similar. Hence, our results are fairly robust to alterations in the modeling
assumptions.
4.2 Some Further Questions
This section discusses some questions and concerns that might follow from the results section and
addresses how we mitigate these concerns. One of the concerns is that the crimes against SC/ST
could be a part of the overall trend of general crimes in the district. The idea is that if the relative
economic status of caste groups is also correlated with general crimes in the district, then we cannot
conclude that it is only crimes against SC/ST that are uniquely linked to relative group economic
positions. In order to check for this, in Table 7a, we present results of regressions where the
dependent variable is logarithm of general crime rate. If the coefficients on our expenditure
variables turn out to be insignificant, we would have ruled out this concern. In column 1, the
coefficient of the relative expenditure term is insignificant, as are the group-‐wise expenditures in
column 2, which stress the fact that differences in material standard of living between caste groups 13 Results from regressions where the dependent variable is the SC/ST SLL crime rate (not shown) are similar to results in Tables 2 and 3.
15
uniquely affect crimes against SC/ST groups and are not associated with general crimes in society. As
a robustness check, I model the general crimes as a count variable in Table 7b and use the negative
binomial regression model. Results are qualitatively similar to those in Table 7a.
A second concern with the results could be that particular high crime states such as Rajasthan,
Madhya Pradesh or Uttar Pradesh might be driving the results, such that excluding observations
from that state might affect our results. In order to check for this, we iteratively run the entire set of
regressions dropping one state at a time and find that our results are robust to such exclusions14.
A third concern is that of reverse causality. Targeted crimes of a violent nature against a specific
community could be a debilitating force leading to reduced earnings and expenditures, which would
make them further worse-‐off compared to the upper castes. If this reverse causality exists, then our
effects are under-‐estimated and provide a lower bound for the true estimates. To check for this, we
regress log of SC/ST expenditures in 2004 on log of SC/ST crime rate in 200415. We find the
coefficient on the crime rate term to be insignificant implying that crimes against SC/ST do not
negatively affect their standard of living.
A fourth possible concern could be out-‐migration of SCs and STs from their districts to other districts
on account of such targeted violence. While, we cannot control for the possibility of migration in our
regression analysis since the NSS data does not allow us to identify migration, we cite findings from
other data sources to investigate this issue. Bhagat (2009) using 2001 Indian Census data documents
that 62 percent of the internal migration in India is in the form of intra-‐district migration. Inter-‐
district and inter-‐state migration account for 24 percent and 13 percent respectively of total internal
migration. For males and females, employment and marriage respectively are the primary reasons
for migration indicating that migration is on account of reasons other than violence. More crucially,
since our unit of analysis is the district and the largest stream of migration is intra-‐district, our
results will not be affected.
Fifth, one can claim that the effects we observe are really those of changes in reporting of crimes
rather than changes in actual incidence of crime. We argue that is not the case, primarily on two
grounds. Firstly, in regressions using the group-‐wise expenditure specification in table 3, crimes are
correlated with the upper castes’ expenditure and not with lower castes’ expenditure indicating that
14 Results are not shown here in the interest of space, but are available with the author. 15 Since district-‐level crime data for 2000 is not available, we can only check for reverse causality using the cross-‐section for 2004. Results are available with the author.
16
it is the perpetrator characteristics driving such crimes. If it were a case of reporting, then it is the
victim characteristics that should have been correlated with crimes. Secondly, the SLL crimes are the
purely caste-‐based crimes motivated solely by the lower caste status of the victims and this should
be the category that is most likely to be sensitive to reporting by victims16. However, as our
regressions indicate, SLL crimes are not associated with changes in relative economic positions.
Finally, and more in the nature of a caveat, is the fact that since the analysis is conducted at the level
of the district, nothing can be definitively said about the nature of individual motivations that leads
to the incidence of such crimes. This means that theories that make predictions about individual
incentives to engage in such behavior that do not vary across districts, cannot be tested. Having said
that, with this available data, these results provide suggestive evidence that variations in relative
group economic positions are linked to variations in violence levels.
5. Discussion and Conclusion
This paper provides one of the first analyses of crimes against Scheduled Castes and Tribes in India
with a view to understanding the effect of a change in the gap between upper and lower castes’
standard of living on the victimization of the SC/ST community. We find that changes in relative
economic position between the lower castes and upper castes are positively correlated with changes
in the incidence of hate crimes, such that a widening of the gap in expenditures between the lower
and upper castes is associated with an decrease in crimes committed by the upper castes against the
SC/ST. We interpret this as the upper castes’ responding to a changes in threat perception created
by changes in the relative standard of living of a group that has been historically subordinated. There
is ample evidence that suggests that upper castes use and justify various forms of violence as tools
to ensure adherence to caste-‐based norms and traditions by the lower castes. Further, between the
IPC and SLL crimes it is the violent IPC crimes that are responsive to economic gaps. As a re-‐
affirmation of this conjecture, we find that among the largely violent crimes, it is the property
related crimes—crimes that seek to deprive the victim of his property symbolic of his material
progress—that are affected by the relative standards of living. Even though the magnitudes of the
effects we obtain are small, just the incidence of such crimes instills a sense of apprehension among
lower castes. This reflects the fact that even though affirmative action has led to some visible
16 Iyer et al (forthcoming) find that there is better reporting of SLL crimes after lower castes obtain mandated representation in local councils.
17
changes in some dimensions of economic condition of SC/ST groups, they have not been truly
empowered since notions of caste hierarchies remain deeply entrenched in society.
Although the incidence of such crimes is usually treated as a law and order problem by the system, it
is more broadly a question of social justice. Attacks often take the form of collective punishment,
whereby entire communities are punished for the perceived transgressions of individuals who seek
to alter established norms or demand their rights. Dalit women, occupying the bottom of both the
caste and gender hierarchies, are uniquely susceptible to violence. Over and above the occurrence
of such crimes, the working of the criminal justice system perpetuates the problem. There is flagrant
violation of justice in the form of police resistance in filing complaints; low conviction rates leading
to easy acquittals for perpetrators; high pendency due to only a few special courts operating; and
poor implementation of economic relief to victims. Newspaper reports frequently find that
judgment on cases is delayed by several years due to the lax performance of the courts and the
apathetic attitude of the legal machinery. A report discussing the performance of the SC/ST
Prevention of Atrocities Act, 1989 finds that at the end of 2007, 79 percent of cases remained
pending for trial at criminal courts showing no significant improvement over a pendency rate of 82.5
percent in 200117. Moreover, the pendency rate is approximately the same for all crimes under the
Prevention of Atrocities Act, 1989, Protection of Civil Rights Act and IPC, indicating that the provision
for “speedy trials” under Prevention of Atrocities Act, 1989 is not being duly followed. The failure to
investigate, file, and pursue cases involving crimes against SC/ST groups has a deleterious effect not
only on the individuals harmed in each instance of violence, but more broadly on the communities in
general—such failures empower potential perpetrators by signaling that crimes against lower castes
will go unpunished and also further disempowers marginalized communities by eroding their trust in
the legal system.
While our analysis uses the lowest level of disaggregated data that are available, which is a good
starting point, a study at the village level would make the analysis much richer since such events are
highly localized and dependent on village level dynamics that we cannot observe in our data. Future
research should aim at studying the occurrence of such violence and atrocities at the household
level through victimization surveys in order to better understand individual motivations.
17 20 years Scheduled Castes and Scheduled Tribes (Prevention of Atrocities) Act: Report Card by National Coalition for Strengthening SCs and STs (Prevention of Atrocities) Act (accessed on January 5, 2012 at http://idsn.org/fileadmin/user_folder/pdf/New_files/India/SCST_PoA_Act_20_years_report_card_-‐_NCDHR.pdf)
18
References
Beck, E.M., Stewart Tolnay (1990). The Killing Fields of the Deep South: the Market for Cotton and the Lynching of Blacks, 1882–1930. American Sociological Review, 55, 526–539 Becker, Gary (1957). The Economics of Discrimination. Chicago: University of Chicago Press Becker, Gary (1968). Crime and Punishment: An Economic Approach, Journal of Political Economy, 76(1), 169-‐217 Berreman, Gerald (1971). Self, Situation and Escape from Stigmatized Ethnic Identity, available at http://www.eric.ed.gov/ERICWebPortal/search/detailmini.jsp?_nfpb=true&_&ERICExtSearch_SearchValue_0=ED058344&ERICExtSearch_SearchType_0=no&accno=ED058344 (accessed June 5, 2010) Bhagat, Ram. B (2009). Internal Migration in India: Are the Underclass More Mobile? Available at http://iussp2009.princeton.edu/download.aspx?submissionId=90927 (accessed November 20, 2011) Blumer, Herbert (1958). Race Prejudice as a Sense of Group Position. Pacific Sociological Review, 1(1), 3-‐7 Bros, Catherine and Mathieu Couttenier (2010). Is Blood Thicker than Water? Untouchability and Public Infrastructure. Working Paper Charmarbagwala, Rubaina and Gunjan Sharma (2010). Crime and Inequality in India. Working Paper, Indiana University. Deaton, Angus (1997) The Analysis of Household Surveys: A Micro-‐econometric Approach to Development Policy. Baltimore and London: Johns Hopkins University Press Deshpande, Ashwini (2011). The Grammar of Caste: Economic Discrimination in Contemporary India. New Delhi: Oxford University Press Dreze, Jean and Reetika Khera (2000). Crime, Gender and Society in India: Insights from Homicide Data. Population and Development Review, 26(2), 335-‐352 Eitle, David, Stewart J. D'Alessio and Lisa Stolzenberg (2002). Racial Threat and Social Control: A Test of the Political, Economic, and Threat of Black Crime Hypotheses. Social Forces, 81(2), 557-‐576 Election Commission of India. Statistical Reports of Assembly Elections, various years. Fajnzylber, Pablo, Daniel Lederman and Norman Loayza (2002). What causes Violent Crime? European Economic Review, 46 (7), 1323-‐1357 Falk, Armin, Andreas Kuhn and Josef Zweimuller (2011). Unemployment and Right Wing Extremist Crime. Scandinavian Journal of Economics, 113, 260-‐285 Gale, Lewis, Will Carrington Heath and Rand Ressler (2002). An Economic Analysis of Hate Crime. Eastern Economic Journal, 28 (2), 203-‐216 Green Donald, Laurence McFalls and Jennifer Smith (2001). Hate Crime: An Emergent Research Agenda. Annual Review of Sociology, 27, 479-‐504 Human Rights Watch (1999). Broken People: Caste Violence against India’s Untouchables. New York: Human Rights Watch Iyer, Lakshmi, Anandi Mani, Prachi Mishra, and Petia Topalova (forthcoming). The Power of Political Voice: Women’s Political Representation and Crime in India. American Economic Journal: Applied Economics Jacobs, D., Wood, K., (1999). Interracial conflict and interracial homicide: do political and economic rivalries explain white killings of blacks or black killings of whites? American Journal of Sociology, 105, 157–190
19
Kanaujia, Ramsunder Ram. Surgeon Second, Dhobi First. Tehelka, Feb. 3, 2007, available at http://www.tehelka.com/story_main26.asp?filename=Cr020307Shadow_lines.asp (Accessed Sept 14, 2011) Kelly, Morgan (2000). Inequality and Crime. The Review of Economics and Statistics, 82(4), 530-‐539 Krueger, Alan and Pischke, Jorg-‐Steffen (1997). A Statistical Analysis of Crime against Foreigners in Unified Germany. Journal of Human Resources, 32(1), 182-‐209 Kumar, Hemanshu and Rohini Somanathan (2009). Mapping Indian Districts across Census Years, 1971 to 2001. Economic and Political Weekly, XLIV (41), 69-‐88 Laakso, Markku, and Rein Taagepara (1979). Effective Number of Parties: A Measure with Application to West Europe. Comparative Political Studies, 12, 3-‐27 National Crime Records Bureau, Crime in India, Government of India, multiple years National Sample Survey Organization, Employment-‐Unemployment Survey and Consumer Expenditure Survey, multiple years Prasad, Kislaya (2009). Economic Liberalization and Violent Crime in India. Working Paper, University of Maryland. Price, Gregory. N, William A. Darity Jr. and Alvin A. Headen Jr. (2008). Does the stigma of slavery explain the maltreatment of blacks by whites? The case of lynchings. Journal of Socio-‐Economics, 37 (1), 167-‐193 Rahman, Maseeh. “Indian Leader Likens Caste System to Apartheid Regime”. The Guardian, Dec. 28, 2006, available at http://www.guardian.co.uk/world/2006/dec/28/india.mainsection (accessed January 5, 2012) Sekhri, Sheetal and Adam Storeygard (2011). The Impact of Climate Variability on Crimes Against Women: Dowry Deaths in India. Working Paper Shah, Ghanshyam, Harsh Mandar, Sukhadeo Thorat, Satish Deshpande and Amita Baviskar (2006). Untouchability in Rural India, Sage Publications: New Delhi Thorat, Sukhadeo and Katherine Newman (Eds.) (2010). Blocked by Caste: Economic Discrimination in Modern India. New Delhi: Oxford University Press. Wilkinson, Steven (2004). Votes and Violence: Electoral Competition and Ethnic Riots in India. Cambridge: Cambridge University Press
20
Table 1 Descriptive Statistics
Variable N Mean SD Crime variables:
Total crimes against SC/ST 830 429.949 419.273 General IPC crimes in the district 830 21321 20156.51
IPC crimes against SC/ST 830 288.131 351.523 SLL crimes against SC/ST 830 141.818 175.935 IPC Body crimes against SC/ST 830 94.984 112.963 IPC Non-‐body crimes against SC/ST 830 193.147 280.945 SC/ST total crime rate 829 99.843 135.11
SC/ST IPC crime rate 829 66.543 107.06 SC/ST SLL crime rate 829 33.299 52.68 SC/ST body crime rate 829 23.071 42.076
SC/ST non-‐body crime rate 829 43.472 75.431 General crime rate 830 1544.339 1492.724
Explanatory variables: SCST MPCE/UC MPCE 828 0.677 0.166
Population 830 2144479 1375044 Percent SC/ST 830 29.796 15.245 Percent Rural 830 79.39 17.285
Percent Underemployed 830 16.019 7.552 Percent Young male 830 9.414 1.854
Gini 830 0.252 0.056 Average expenditure 830 535.138 186.327 SCST expenditure 829 433.173 128.257 UC expenditure 829 679.067 274.219 OBC expenditure 820 523.106 179.112
Effective Number of parties 830 4.593 1.383 Percent Illiterate 830 45.872 15.417 Percent Primary 830 19.916 6.363
Percent Secondary 830 24.064 9.085 Percent Higher secondary and above 830 10.148 5.756
Note: Crime rate means crimes per 100000 SC-‐ST population. IPC crimes are the sum of murder, rape, kidnap, hurt, dacoity, robbery, arson and other IPC crimes. SLL crimes are the sum of crimes registered under the Prevention of Atrocities Act and the Protection of Civil Rights Act. Body Crimes are the sum of murder, rape, kidnapping and physical assault. Non-‐body crimes are the sum of dacoity, robbery, arson and other IPC crimes. General crime rate is total general IPC crimes less IPC crimes against SC/ST per 100000 non-‐SC/ST population. Young males refer to males in the 15-‐24 age groups.
21
Table 2 Effect of relative expenditure on total, IPC and SLL crime rates
(1) (2) (3) Crime rate
(Total Crimes) Crime rate (IPC Crimes)
Crime rate (SLL Crimes)
Ln (SCST exp/UC exp) 0.304** 0.355** 0.298 (0.118) (0.149) (0.294) Ln (exp) -‐0.008 -‐0.378 0.413 (0.283) (0.389) (0.591) SCST% -‐0.087*** -‐0.1*** -‐0.08*** (0.0096) (0.011) (0.022) SCST%-‐sq 0.0007*** 0.0009*** 0.0006** (0.0001) (0.0001) (0.0003) Rural% 0.005 0.0004 0.005 (0.00409) (0.00525) (0.00978) Gini 1.05 2.656** -‐0.378 (1.009) (1.239) (1.942) Underemployed 0.003 0.003 0.011 (0.00434) (0.00578) (0.00899) Young Males% -‐0.014 0.022 -‐0.031 (0.0163) (0.0241) (0.0343) Illiterate -‐0.01 -‐0.001 -‐0.019 (0.00964) (0.0131) (0.0217) Primary ed 0.008 0.007 0.005 (0.0129) (0.0165) (0.0240) Secondary ed -‐0.0021 -‐0.0083 0.001 (0.0132) (0.0177) (0.0272) Effective Parties 0.0344 0.134 -‐0.386** (0.0716) (0.0977) (0.162) Time Yes Yes Yes N F-‐statistic R-‐squared
828 16.45
0.34
828 9.96 0.22
828 4.53 0.12
Note: Standard errors in parentheses, clustered at district level. *, **, *** indicate significance at 10%, 5% and 1% respectively. All regressions include district fixed effects, time dummy and constant term. All dependent variables are in logs. Crime rate is crimes per 100000 SC/ST population. IPC crimes are the sum of murder, rape, kidnap, hurt, dacoity, robbery, arson and other IPC crimes. SLL crimes are the sum of crimes registered under the Prevention of Atrocities Act and the Protection of Civil Rights Act. For the education dummy variables, ‘higher secondary and above’ is the omitted category.
22
Table 3 Effect of group-‐wise expenditures on total, IPC and SLL crime rates
(1) (2) (3)
Crime rate (Total Crimes)
Crime rate (IPC Crimes)
Crime rate (SLL Crimes)
Ln (SCST exp) 0.138 -‐0.134 0.496 (0.191) (0.309) (0.474)
Ln (UC exp) -‐0.341** -‐0.563*** -‐0.158 (0.143) (0.185) (0.341)
Ln (OBC exp) -‐0.0393 0.112 -‐0.107 (0.196) (0.319) (0.454) SCST % -‐0.091*** -‐0.097*** -‐0.091*** (0.009) (0.0109) (0.0251) SCST %-‐sq 0.0008*** 0.0009*** 0.0008** (0.0001) (0.0001) (0.0003) Rural % 0.0052 -‐0.0004 0.006 (0.00411) (0.00527) (0.0102) Gini 0.978 2.652** -‐0.393 (0.936) (1.142) (1.898)
Underemployed 0.0009 0.002 0.0109 (0.00435) (0.00567) (0.00949)
Young Males% -‐0.0163 0.0269 -‐0.0448 (0.0158) (0.0253) (0.0349)
Illiterate -‐0.0077 0.0031 -‐0.0195 (0.00979) (0.0131) (0.0211) Primary ed 0.0092 0.0093 0.0045 (0.0133) (0.0164) (0.0239) Secondary ed 0.0045 0.0028 0.0043 (0.0128) (0.0166) (0.0276) Effective Parties 0.0428 0.151 -‐0.396** (0.0707) (0.0975) (0.164) Time Yes Yes Yes N F-‐stat R-‐squared
818 15.81 0.34
818 9.54 0.22
818 4.25 0.13
Note: Standard errors in parentheses, clustered at district level. *, **, *** indicate significance at 10%, 5% and 1% respectively. All regressions include district fixed effects, time dummy and constant term. All dependent variables are in logs. Crime rate is crimes per 100000 SC/ST population. IPC crimes are the sum of murder, rape, kidnap, hurt, dacoity, robbery, arson and other IPC crimes. SLL crimes are the sum of crimes registered under the Prevention of Atrocities Act and the Protection of Civil Rights Act. For the education dummy variables, ‘higher secondary and above’ is the omitted category.
23
Table 4 Decomposing IPC crimes into Body Crimes and Non-‐body crimes
(1) (2) (3) (4) Crime rate
(Body crimes) Crime rate
(Body crimes) Crime rate
(Non-‐Body crimes) Crime rate
(Non-‐Body crimes) Ln (SCST exp/UC exp) 0.0559 0.417** (0.154) (0.186) SCST% -‐0.0897*** -‐0.0929*** -‐0.1*** -‐0.0985*** (0.0113) (0.0105) (0.0127) (0.0146) SCST%-‐sq 0.0007*** 0.0008*** 0.0009*** 0.0009*** (0.0001) (0.0001) (0.0001) (0.0002) Gini 0.361 0.416 4.085*** 4.038*** (1.122) (1.048) (1.439) (1.355) Ln (SCST exp) -‐0.259 -‐0.225 (0.292) (0.365) Ln (UC exp) -‐0.182 -‐0.691*** (0.190) (0.221) Ln (OBC exp) 0.193 0.308 (0.297) (0.361) Time Yes Yes Yes Yes N F-‐stat R-‐squared
828 10.34 0.21
818 9.41 0.2
828 7.74 0.18
818 8.09 0.18
Note: Standard errors in parentheses, clustered at district level. *, **, *** indicate significance at 10%, 5% and 1% respectively. All regressions include district fixed effects, time dummy and constant term. All dependent variables are in logs. Crime rate is crimes per 100000 SC/ST population. Body Crimes are the sum of murder, rape, kidnapping and physical assault. Non-‐body crimes are the sum of dacoity, robbery, arson and other IPC crimes. Other controls included (but not shown) include percent rural, percentage of young males, education dummies, log of expenditure, underemployment rate, Gini and effective number of parties.
24
Table 5 Effect on crime rates against SC/ST after controlling for general crime rates
(1) (2) (3) (4) Crime Rate
(Total crimes) Crime Rate
(Total crimes) Crime Rate (IPC crimes)
Crime Rate (IPC crimes)
Ln (SCST exp/UC exp) 0.247** 0.305** (0.122) (0.151) SCST% -‐0.0851*** -‐0.0928*** -‐0.0978*** -‐0.0991*** (0.0103) (0.00817) (0.0103) (0.0102) SCST%-‐sq 0.0004*** 0.0006*** 0.0006*** 0.0007*** (0.0001) (0.0001) (0.0001) (0.0001) Ln (SCST exp) 0.0299 -‐0.219 (0.184) (0.305) Ln (UC exp) -‐0.322** -‐0.548*** (0.153) (0.185) Ln (OBC exp) -‐0.0361 0.114 (0.178) (0.301) General Crime 0.953*** 0.981*** 0.836*** 0.768*** (0.123) (0.123) (0.167) (0.173) Time Yes Yes Yes Yes N F-‐stat R-‐squared
828 28.54 0.43
818 28.99 0.44
828 12.84 0.27
818 12 0.25
Note: Standard errors in parentheses, clustered at district level. *, **, *** indicate significance at 10%, 5% and 1% respectively. All regressions include district fixed effects, time dummy and constant term. All dependent variables are in logs. Crime rate is crimes per 100000 SC/ST population. IPC crimes are the sum of murder, rape, kidnap, hurt, dacoity, robbery, arson and other IPC crimes. Other controls included (but not shown) include percent rural, percentage of young males, education dummies, log of expenditure, underemployment rate, Gini and effective number of parties. General crimes are (log of) total IPC crimes less total IPC crimes against SC/ST per 100000 non-‐SC/ST population.
25
Table 6a Negative Binomial Regressions using total crimes against SC/ST
(1) (2) (3) (4) Total Crimes Total Crimes Total Crimes Total Crimes Ln (SCST exp /UC exp) 0.263** 0.248** (0.104) (0.102) Ln (SCST exp) 0.212 0.198 (0.164) (0.161) Ln (UC exp) -‐0.265** -‐0.263** (0.122) (0.121) Ln (OBC exp) 0.179 0.140 (0.159) (0.159) General crimes 0.427*** 0.421*** (0.0904) (0.0997) N 824 824 806 806 Note: Standard errors in parentheses *, **, *** indicate significance at 10%, 5% and 1% respectively. Other variables controlled for (but not shown) include SCST%, SCST%-‐sq, percent rural, percentage of young males, education dummies, log of expenditure, underemployment rate, Gini and effective number of parties. SC/ST population is used as scaling variable on right hand side. General crimes are (log of) total IPC crimes less total IPC crimes against SC/ST per 100000 non-‐SC/ST population.
Table 6b Negative Binomial Regressions using IPC crimes against SC/ST
(1) (2) (3) (4) IPC crimes IPC crimes IPC crimes IPC crimes Ln (SCST exp/UC exp) 0.286** 0.289** (0.137) (0.136) Ln (SCST exp) -‐0.0492 -‐0.0735 (0.214) (0.213) Ln (UC exp) -‐0.441*** -‐0.457*** (0.159) (0.159) Ln (OBC exp) 0.290 0.291 (0.195) (0.196) General crimes 0.258** 0.214* (0.106) (0.113) N 820 820 802 802 Note: Standard errors in parentheses *, **, *** indicate significance at 10%, 5% and 1% respectively. IPC crimes are the sum of murder, rape, kidnap, hurt, dacoity, robbery, arson and other IPC crimes. Other variables controlled for (but not shown) include SCST%, SCST%-‐sq, percent rural, percentage of young males, education dummies, log of expenditure, underemployment rate, Gini and effective number of parties. SC/ST population is used as scaling variable on right hand side. General crimes are (log of) total IPC crimes less total IPC crimes against SC/ST per 100000 non-‐SC/ST population.
26
Table 7a Regressions using general crimes as dependent variable (1) (2) General Crimes General Crimes Ln (SCST exp/UC exp) 0.06 (0.0526) Ln (SCST exp) 0.11 (0.0926) Ln (UC exp) -‐0.0191 (0.0595) Ln (OBC exp) -‐0.0032 (0.0991) Time Yes Yes N F-‐stat R-‐squared
828 14.07 0.39
818 13.99 0.37
Note: Standard errors in parentheses, clustered at district level. *, **, *** indicate significance at 10%, 5% and 1% respectively. All regressions include district fixed effects, time dummy and constant term. All dependent variables are in logs. Other controls used but not shown include percent rural, percentage of young males, education dummies, log of expenditure, underemployment rate, Gini and effective number of parties. General crimes are (log of) total IPC crimes less total IPC crimes against SC/ST per 100000 non-‐SC/ST population. Table 7b Negative binomial regressions using general crimes as dependent variable (1) (2) General Crimes General Crimes Ln (SCST exp/UC exp) 0.0452 (0.0372) Ln (SCST exp) 0.0749 (0.0577) Ln (UC exp) -‐0.0247 (0.0421) Ln (OBC exp) -‐0.0113 (0.0580) N 826 808 Note: Standard errors in parentheses *, **, *** indicate significance at 10%, 5% and 1% respectively. Other variables controlled for (but not shown) include SCST%, SCST%-‐sq, percentage rural, percentage of young males, education dummies, log of expenditure, underemployment rate, Gini and effective number of parties. Non-‐SC/ST population is used as scaling variable on right hand side. General crimes are total IPC crimes less total IPC crimes against SC/ST.
27
Appendix 1
Crimes included under the Special and Local Laws (SLL) against Scheduled Castes and Tribes
1) The Protection of Civil Rights Act, 1955
Sections 3 -‐ 7A of the Act define the following as offences if committed on the ground of
“untouchability”:
1) Prevention from entering public worship places, using sacred water resources.
2) Denial of access to any shop, public restaurant, hotel, public entertainment, cremation
ground etc.
3) Refusal of admission to any hospital, dispensary, educational institutions etc.
4) Refusal to sell goods and render services.
5) Molestation, causing injury, insult etc.
6) Compelling a person on the ground of untouchability to do any scavenging or sweeping or to
remove any carcass etc.
2) The Scheduled Castes and Scheduled Tribes (Prevention of Atrocities) Act, 1989
Whoever, not being a member of a Scheduled Caste or a Scheduled Tribe:
1) Forces a member of a Scheduled Caste or a Scheduled Tribe to drink or eat any inedible or
obnoxious substance;
2) Acts with intent to cause injury, insult or annoyance to any member of a Scheduled Caste or
a Scheduled Tribe by dumping excreta, waste matter, carcasses or any other obnoxious
substance in his premises or neighbourhood;
3) Forcibly removes clothes from the person of a member of a Scheduled Caste or a Scheduled
Tribe or parades him naked or with painted face or body or commits any similar act which is
derogatory to human dignity;
4) Wrongfully occupies or cultivates any land owned by, or allotted to, or notified by any
competent authority to be allotted to, a member of a Scheduled Caste or a Scheduled Tribe
or gets the land allotted to him transferred;
5) Wrongfully dispossesses a member of a Scheduled Caste or a Scheduled Tribe from his land
or premises or interferes with the enjoyment of his rights over any land, premises or water;
28
6) Compels or entices a member of a Scheduled Caste or a Scheduled Tribe to do ‘begar’ or
other similar forms of forced or bonded labour other than any compulsory service for public
purposes imposed by Government;
7) Forces or intimidates a member of a Scheduled Caste or a Scheduled Tribe not to vote or
vote for a particular candidate or to vote in a manner other than that provided by law;
8) Institutes false, malicious or vexatious suit or criminal or other proceedings against a
member of a Scheduled Caste or a Scheduled Tribe;
9) Gives any false or frivolous information to any public servant and thereby causes such public
servant to use his lawful power to the injury or annoyance of a member of a Scheduled
Caste or a Scheduled Tribe;
10) Intentionally insults or intimidates with intent to humiliate a member of a Scheduled Caste
or a Scheduled Tribe;
11) Assaults or uses force to any woman belonging to a Scheduled Caste or a Scheduled Tribe
with intent to dishonour or outrage her modesty;
12) Being in a position to dominate the will of a woman belonging to a Scheduled Caste or a
Scheduled Tribe and uses that position to exploit her sexually to which she would not have
otherwise agreed;
13) Corrupts or fouls the water of any spring, reservoir, or any other source ordinarily used by
members of the Scheduled Caste or the Scheduled Tribe so as to render it less fit for the
purpose for which it is ordinarily used;
14) Denies a member of a Scheduled Caste or a Scheduled Tribe any customary rite of passage
to a place of public resort or obstructs such members so as to prevent him for using or
having access to a place of public resort to which other members of public or any section
thereof have a right to use or access to;
15) Forces or causes a member of a Scheduled Caste or a Scheduled Tribe to leave his house,
village, or any other place of residence
1
Homicide and Work: The Impact of Mexico’s Drug War on Labor Market Participation
Ariel BenYishay and Sarah Pearlman*
January 2013
Abstract:
We estimate the impact of the escalation of the drug war in Mexico on the mean hours worked among the general population. We focus on homicides, which have increased dramatically since 2006. To identify the relationship between changes in homicides and hours worked, we exploit the large variation in the trajectory of violence across states and over time. Using panel and instrumental variables regressions, we find that the increase in homicides has negatively impacted labor force activity. An increase in homicides of 10 per 100,000 in a given state is associated with a decline of 0.3 weekly hours worked among the state’s population. For states most impacted by the drug war, in which homicides per 100,000 inhabitants have increased by 30-50 a year, this implies an average decline in hours worked of one and one and a half hours per week. These impacts are larger for the self-employed and are concentrated among the highest income quartiles. This highlights how the costs of crime tend to be unequally born by certain segments of the population.
JEL Codes:
Keywords: Crime, Labor force participation, Mexico
*Contact information: BenYishay, University of New South Wales, School of Economics [email protected], Pearlman, Vassar College, [email protected]
We are grateful to Francisco Ibarra Bravo and participants in the Southern Economic Association Annual Meetings for feedback
2
1 Introduction
In the past five years, Mexico has witnessed a dramatic increase in violent crime. The well-publicized
rise is the result of the drug war, which escalated in late 2006 after newly elected president Felipe
Calderón launched a federal crackdown on drug cartels. The ensuing period has been marked by a
significant rise in violent death, with annual drug related homicides increasing from approximately
2,120 in 2006 to 12,366 in 2011 (Trans-border Institute) and all homicides increasing from 25,780 to
37,375 (National Public Security System). In total, more than 50,000 deaths are attributed to the
conflict (Rios and Shirk 2012). The increases are concentrated geographically, with high intensity
states experiencing increases beyond those seen in some countries officially at war. In a recent
ranking of the most violent cities in the world, Mexican cities claimed five of the top ten spots,
highlighting the relative intensity of the conflict (Citizen’s Council for Public Safety and Criminal
Justice, 2012).
The cost of war between the Mexican state and drug cartels has been high. In addition to
billions of dollars in public funds spent to fight the cartels, casualties have spread beyond members
of the cartels, police and army to the civilian population. The gruesome and public nature of many
of the killings as well as the perception that cartel members can operate with impunity has generated
a high level of fear within the population. Nationally representative victimization surveys show that
the percentage of adults who feel the state in which they live is unsafe rose from 54% in 2004 to
65% in 2009, while the percentage of individuals who feel their work is unsafe rose from 13.7% to
19%1.
This increased fear of victimization can compound the costs of the conflict if it generates
behavioral changes, such a reduction in hours spent working or studying, which are second best.
1 Author’s calculations from the National Survey of Insecurity (ENSI). This is conducted by Mexico’s statistical agency,
INEGI, and the data and documentation are available on their website.
3
Several papers have explored these indirect costs of violent crime. For example, in examining
changes in work trends in the U.S., Hamermesh (1999) finds that higher rates of homicide explain
lower levels of evening and early morning work in large cities and secular declines over time (the
same is not true of non-violent crimes), and that these shifts resulted in losses on the order of 4 to
10 billion dollars. Monteiro and Rocha (2012) find that armed conflicts between drug gangs in Rio
de Janeiro have led to a significant reduction in the educational attainment of children in these areas.
Fernandez, Ibáñez and Peña (2011) find that rural households in Colombia who were victims of
violent crime increase their supply of labor to off-farm activities. Rodriguez and Sanchez (2009) also
look at the armed conflict in Colombia and find it significantly reduces the average years of
schooling. Finally, Barrera and Ibáñez (2004) find that among municipalities in Colombia, in those
where homicide rates are above the national median, school enrollment is significantly lower and
declines as homicide rates increase.
A small number of papers also have investigated the impact of crime on individuals in
Mexico. Braakman (2012) looks at the period from 2002 and 2005, which pre-dates the escalation of
the drug war, and finds that crime victims sleep fewer hours at night and are more likely to take
protective measures against crime, such as changing travel routes and transportation modes. Villoro
and Teruel (2004) examine the impact of homicides in 1997 and estimate losses between 0.03 and
0.6 percent of GDP. Dell (2011) investigates the spillover effects of the federal war on the cartels.
She finds that female labor force participation falls in municipalities where drug traffic and violence
is diverted following government crackdowns on groups in surrounding areas. Finally, nationally
representative victimization surveys in Mexico (ENSI 2009) find that close to half of respondents
stop going out at night as a result of crime, while fifteen percent stop taking public transportation,
eating out and going to events. The changes subsequently may reflect a decline in the willingness of
4
people to work at night or a reduction in work availability due to declining demand for particular
goods and services like restaurants.
We contribute to the literature on the costs of violent crime by investigating the impact of
the drug war in Mexico on the labor force participation of adults. To explore this link we use data
on annual changes in homicides for the years 2007 to 2010 and changes in weekly hours worked
from one year panels on employment and occupation. We begin by estimating the effects on these
annual changes while controlling for time-invariant characteristics and for time-varying individual
and regional level heterogeneity. In particular, we control for economic shocks and changes in the
composition of employment in states over time; factors that may jointly determine changes in
criminal activity and labor market participation. Overall we find a significant negative effect of
changes in homicides on hours spent working. An increase of 10 homicides per 100,000 over a one
year period is associated with an average decline of 0.3 hours worked per week. This constitutes a
decline of roughly one percent. We find larger effects for men than women and even larger
differences by work type. Among salaried workers the response is three times lower than among the
entire population. On the other hand, among self-employed workers the response to changing
homicide rates is one and a half times higher than for the full population. This suggests that in
states most affected by drug war violence, average hours worked by the self-employed have declined
by up to two and a half hours a week. While we do not know if the larger response is due to greater
exposure to violence or an enhanced ability to change work schedules, the results show one
dimension along which the impacts of violent crime may vary.
We next take an instrumental variables approach to further control for regional level
heterogeneity. We instrument for changes in homicides using kilometers of federal toll highways in
a state. The logic behind the instrument is that the increase in violence clearly is linked to the
federal crackdown on drug cartels. The crackdown weakened oligopolistic organizations and
5
increased competition for valuable production and transport routes. This competition has been
most severe over access to land transport routes to the U.S., the largest drug consumer market in the
world and Mexico’s largest trading partner. Federal toll highways are good measures of routes to the
U.S. because the majority of transport of goods and services in Mexico occurs via highway and toll
roads are the highest quality and most rapid highway routes. Furthermore, the majority was built
between 1989 and 1994 or follow previously established routes, which means current economic or
demographic factors do not determine their placement. Regressions of kilometers of toll highways
on homicides show that states with more toll highways register significantly higher homicide
increases than those with fewer routes. Comparisons of other transportation routes find no similar
relationship, while comparisons of economic variables show no similar trend differences. This
confirms that the steeper homicide trajectory experienced by high toll highway states is not simply a
reflection of general transportation access or economic trends.
Results from two-stage least squares estimation provide further evidence that increasing
homicides negatively impact labor force activity. The coefficients on yearly changes in homicides
are negative in all but one case and, for working adults, are three or more times larger than the OLS
coefficients. For salaried adults an increase of 10 homicides per 100,000 over a one year period is
associated with an average decline of 1.2 hours worked per week, while for self-employed workers
the associated decline is one hour per week. These constitute declines between three and five
percent. Similar to the OLS results we find mild differences by gender when work type is not
considered but large differences by work type. Dissimilar to the OLS results, however, we find
higher effects for salaried workers than self-employed ones. Again, we do not know if this is due to
different types of work and exposure to crime or a greater ability to adjust work hours.
This paper also differs from others in that we examine differential responses to the conflict by
income. The ability to avoid crime, and therefore both exposure to and fear of crime, may vary by
6
the resources available to individuals, and several papers provide evidence that crime is unequally
born by households at different income levels. For example, Gaviria and Pagés (1999) examine
victimization data from 17 Latin American countries and find that victims of property crime are
more likely to be upper or middle income. On the other hand, DiTella et. al. (2010) examine a crime
wave in Buenos Aires and find that victimization rates rose much more for poor households than
wealthy ones. They find the increases are largely for crimes for which abatement measures are
expensive, such as burglary, rather than for crimes for which abatement measures are cheap, such as
mugging. Villoro and Teruela (2004) find similar results for Mexico City. Using data from 1999
they find that the impact of crime and crime prevention, particularly for property crime, is much
higher for individuals in the lowest income quintiles. Finally, recent nationally representative
victimization surveys from Mexico (ENSI) suggest that poor households have a reduced ability to
deter crime. Households in the lowest income quintile are five times less likely than those in the
highest income quintiles to invest in high value security items, like alarms and private security
guards, and on average spend 4-5 times less on crime deterrence. We repeat the analysis by income
quartile and find, interestingly, that the impacts are concentrated among the highest income quartile.
While we do not know if this is driven by greater exposure to violence, greater fear of violence, or
an enhanced ability to respond to both by changing work schedules, these results suggest that the
costs of the violent conflict have been born unequally across income levels.
The paper proceeds as follows. In section 2 we describe the data for labor market activity
and crime. In section 3 we outline our empirical strategy and estimate a fixed effect model. In
section 4 we outline an instrumental variables strategy. In section 5 we investigate differential
responses by income. In section 6 we conclude.
7
2 Data
2.1 Homicide Data
To measure the impact of the drug war we use changes in total homicides per 100,000 inhabitants by
state and year. The homicide totals are compiled by the National Public Security System (SESNSP)
using information sent by state police forces. We convert the totals into crimes per 100,000
inhabitants using population data from CONAPO. It is important to note that the totals are for all
reported homicides, not just those linked with drug violence. We limit the scope of violent crime
related to the drug war to homicides for several reasons. First, the reporting rates for homicides are
high, reducing concerns over measurement error that arise with other crimes linked to drug cartels,
such as kidnapping or assault, for which reporting rates are very low. Second, and more
importantly, homicides have been the most visible outgrowth of the conflict. As outlined in Figure
1.A. and Table 1, homicides rose significantly in Mexico from 2005 to 2010. Average homicides per
100,000 inhabitants rose from 23.6 in 2005 to 37.38 in 2010 while the maximum for any state rose
from 47.79 homicides per 100,000 inhabitants to a staggering 127.64.
Table 1 also shows that the drug war has played out unevenly over the country. The yearly
standard deviations close to triple from 2005 to 2010 and the gap between the 25th percentile and the
75th and 90th percentiles widen significantly. Indeed, according to the Trans-border Institute, in
2010, the peak of the drug war, fifty six percent of all drug related killings occurred in just four
states-- Chihuahua, Sinaloa, Tamaulipas and Guerrero (Rios and Shirk 2011). Furthermore, over
seventy percent of the violence was concentrated in just 80 municipalities (out of close to 2500).
Figure 2 demonstrates this uneven trajectory of violence across states. While a handful of states
have experienced sharp increases in homicides, many others have experienced small increases and
some have even experienced a decline. The map also shows that the homicide trajectories lack any
distinct regional trend. High violence states are not concentrated in areas which, ex-ante, one would
8
suspect to face higher levels of drug trafficking, such as the U.S. border, the Pacific or the Gulf
coasts. This shows that geography alone cannot explain why some states have been more affected
by the drug war than others.
Finally, Figure and Table 1 show that homicides did not begin to increase until 2007, after the
federal government launched a crackdown on drug cartels. We detail the crackdown and its role in
increasing violence in Section 4. In the meantime, since our goal is to estimate the impact of
increased homicides stemming from the drug war, our main focus is on the years 2007 to 2010.
2.2 Data on Labor Market Activity
The data on labor market participation, including weekly hours spent on paid work, come from the
Mexican National Survey of Occupation and Employment (ENOE), a rotating labor force survey
conducted by the Mexican Census (INEGI). The ENOE, which began in 20052, follows urban and
rural households for five quarters. From the entire sample in a given year, we restrict attention to
individuals who enter the sample in the first quarter and can be followed for five quarters. This
group constitutes approximately twenty percent of the full sample in any quarter, and given the
rotating nature of the sample, should not differ substantively different from the complete one.
Using the ENOE surveys for the years 2007 to 2010 we create one year panels of adults who
enter the survey during the first quarter of the year and are between the ages of 18 and 65. After
further limiting attention to individuals who were born in the same state in which they reside, the
result is a sample of 94,642 individuals. Given that individuals stay in the sample for a year at most,
this creates a repeated cross-section of one-year panels. Non-participation in the labor force is coded
as zero, such that adults out of the labor force are reported as having zero hours worked. Our
primary emphasis therefore is on the intensive margin of labor force activity, rather than the
2 It replaces the combination of the urban labor force survey (the ENEU) and the urban and rural survey, the ENE.
9
extensive margin. In terms of regional variation, the geographic unit of focus is the state. This is
the finest level of geographic detail we can achieve, as ENOE is not representative at the municipal
level and the regional, time-varying controls are not available at the municipal level. Finally, later in
the paper we investigate differential responses by income. To do this we place individuals into
income quartiles based on their total monthly household income relative to others in their minimum
wage area (defined by INEGI).
Summary statistics on the sample are provided in Table 2. Sixty four percent of the sample
is in the labor force starting at the beginning of the year. Of these, sixty seven percent are engaged
in salaried work while close to thirty percent are self-employed. There are large differences in labor
market outcomes across men and women. While eighty six percent of men enter the sample in the
labor force, only forty five percent of women do. This partially explains the large gaps in hours
worked by gender. On average men work close to forty hours per week, while women work only
seventeen. These differences also are reflected in monthly income, which is significantly higher for
men than women (although household income is not). Considering trajectories of work behavior
over the year, the average change in weekly hours worked is close to zero. Again we see gender
differences. For men the average change is a decline of 0.45 hours per week, while for women it is
an increase of 0.26 hours per week. This gap exists despite the fact that women have higher exit
rates from the labor force than men. On average close to seven percent of men exit the labor force
over the year long period in which they are observed, as compared to ten percent of women. Given
these differences, we explore the possibility that the increase in homicides have differential impacts
by gender.
10
3 Estimation
3.1 Basic Model
We start with a model linking changes in hours spent working to changes in homicide.
isttstitstisteZXicideshours ∆++∆+∆+∆+=∆ λϕγββ ''
10 hom (1)
The outcome variable is the change in weekly hours worked over a one year period (from Q1 in year
t-1 to Q1 in year t) by individual i in state s in year t. This is a function of the one-year change in
homicide rates in state s , changes in individual level characteristics (it
X∆ ), changes in state level
characteristics (st
Z∆ ), year fixed effects, and unobserved changes at the individual and state level (
itse∆ ). The model also implicitly includes controls for time-invariant individual and state
characteristics, as we look at changes by individuals who remain in the same state over a one year
period. As a result fixed individual characteristics, such as skill or risk aversion, that may jointly
determine where people live and their work habits, are netted out. Also netted out are level
differences across states that may jointly determine time use and crime patterns. For example,
certain states may have better institutions or a better educated workforce, leading to lower crime
rates and more employment opportunities.
Given the potential for time varying individual and state level heterogeneity to bias the
coefficient on changes in homicides, we discuss controls for each in detail. To control for time-
varying, individual heterogeneity we include one year change in total household size, change in the
number of children, change in labor force status, and for those in the labor force, change in industry
of work (2 digit code). In addition, to reduce omitted variables bias stemming from individuals who
move to reduce exposure to homicides, we limit the sample to individuals who were born in the
state in which they currently reside. This is the closest we can come to eliminating movers given
the lack of residence history in the survey.
11
To control for regional heterogeneity we include several time varying measures. We include
one year changes in state-level GDP and unemployment rates to capture economic shocks that
might change the returns to work and criminal behavior.. We also consider controls for more
specific changes to the composition of work opportunities within a state over time. To do this we
include changes in the share of output for industries, defined by 2 digit codes, by state and year.
These variables help control for the possibility that over time more labor intensive industries may
differentially locate in states with higher or lower homicide rates. Finally, we include year fixed
effects to capture nation-wide shifts in labor market opportunities and crime. This is particularly
important as the intensification of the federal response to drug trafficking coincides with the
beginning of the Great Recession in the U.S., which severely affected Mexico.
Identification in the model comes from variation within individuals across states and years.
Specifically, we examine how changes in labor market behavior vary by average individuals by state
and year as the homicide rates change across states and time. The large variation in the homicide
trajectories across states over the six year period is key to establishing this relationship. The
identifying assumptions is that after we control for individual and state fixed factors as well as
observable time varying individual and state characteristics, the error term is not correlated with
changes in homicides. In other words, we can identify a relationship between changes in behavior
and crime if 0]|[ =itist
XeE .
We estimate equation (1) using OLS and present the results in table 3. We estimate the
model on the entire sample, and separately for men and women, using population weights and
robust standard errors. To show the importance of controlling for observable individual and
regional heterogeneity, we also present results from a model that contains no controls. Panel A
contains results for all adults, regardless of labor force participation or type of work. Panel B
12
contains results only for adults who have salaried work at the beginning of the year. Panel C
contains results only for adults who are self-employed at the beginning of the year.
Overall the results show a negative relationship between changes in homicides and changes in
work. The coefficients on homicide changes are negative in all but one of the estimations and
significant for all adults and men. In many cases the results from the model that contains the full set
of individual, state and time controls are larger than those from the restricted model, suggesting that
failing to account for individual and regional level heterogeneity leads to an underestimation of the
impact of homicides on labor force behavior. Focusing on the results for the full sample with all
controls (shown in Panel A column two), the results indicate that an increase of 10 homicides per
100,000 over a one year period is associated with an average decline of 0.3 hours worked per week
for all adults. With the sample mean of hours worked of 27.5, this constitutes a decline of roughly
one percent. This coefficient also suggests that in states where the drug war has been most acute--
where homicides per 100,000 inhabitants have increased by 30-50 per year--average hours worked by
adults have declined by 1.0-1.5 hours per week. The results for all adults indicate a slightly larger
response among men than women. An increase of 10 homicides per 100,000 inhabitants is
associated with a decline of 0.45 hours for men, as compared to 0.14 for women.
In examining differential response by labor type, we see that the impact of homicides is much
larger for self-employed workers than salaried ones. The coefficient for salaried workers of both
genders (shown in Panel B column two) is -0.011 and insignificant. This compares with a coefficient
of -0.048 for self-employed workers that is significant at the 10% level (Panel C column two),
despite a much smaller subsample of self-employed respondents. This suggests the response to
changing homicide rates is close to five times as high for the self-employed as for salaried workers.
This large difference may be due to two factors. First, self-employed workers likely have a greater
ability to adjust their work schedules than salaried workers, making it easier to change their labor
13
market activity in response to security concerns. Second, self-employed workers may face greater
exposure to the crime itself; many of these self-employed individuals have less ability to invest in
deterrence and prevention, and may operate their businesses in areas that are more frequented by
criminals. For example, street vendors and drivers of informal buses and taxis are more likely to be
impacted by violent crime while at work than office workers. While we do not know the specific
channel through which the effects operate, the results show that the increase in homicides has led to
significant declines in labor market activity, particularly among the self-employed.
4 Instrumental Variable Analysis
4.1 Empirical Strategy
We now turn to instrumental variables estimation, as we remain concerned that unobserved time-
varying factors may lead to a spurious correlation between changes in homicides and labor market
activity. If state-level homicide changes are correlated with changes in other unobserved state
characteristics that affect changes in hours worked, the coefficients from the fixed effect model
above will be biased. At the same time, changes in labor force participation may lead to changes in
homicide rates, even conditional on a variety of state-level time-varying characteristics.
To eliminate potential bias from unobserved regional heterogeneity and reverse causality, we
instrument for changes in homicide rates. To do so, we focus on changes in the intensity of drug-
related homicides that took place differentially across Mexican states after the federal government
launched a crackdown on drug cartels in December 2006. The crackdown began when newly
elected president Felipe Calderón sent army troops into the state of Michoacán to battle
narcotrafficers. This marked the first time the federal government directly engaged with the cartels,
and the government has since increased its involvement dramatically, deploying thousands of army
14
and federal police officers to various municipalities where cartels have operated. Through the
capture and killing of cartel members and drug seizures, the crackdown achieved a weakening of
oligopolistic organizations, altering the previous equilibrium in which production and transport
routes were divided among rival groups (Rios 2012, Dell 2011). The increase in competition has
come in the form of turf wars, as groups violently vie for the production and transport routes of
their diminished rivals. This competition has been most severe over access to land transport routes
to the U.S., the largest drug consumer market in the world and Mexico’s largest trading partner for
legal goods.
As a result of the crackdown and ensuing intensification in competition, states with more
extensive land transport routes connected to the U.S. have seen sharp rises in cartel-related violence
and homicide rates—whereas states with less extensive routes have not. We therefore consider both
the timing of the crackdown and differences in the extent of federal toll highways across states—
which measure access routes to the U.S. —to construct our instrument for homicide changes.
Specifically, we use the number of kilometers of federal toll highways in each state as our instrument
to estimate the effects of homicide changes in the post-crackdown period3.
Several factors make toll highways suitable instruments for access to transport routes to the
U.S. and therefore increasing homicides. First, the majority of transport of goods and people to the
U.S. from Mexico happens via highways. The North American Transportation Statistics Database
indicates that in 2011, approximately 65% of Mexican exports to the U.S. were transported via
highway, while 82% of Mexican travel to the U.S. occurred via highways and rail (no breakdown is
available, but person transport via rail in Mexico is very low). The reliance on highways for
3 The data on kilometers of highways are from the Annual Statisticals of Secretary of Communications and
Transportation (SCT).
15
transport is higher than in the U.S. and Canada and largely is due to the poor state of Mexico’s
railroads, which only recently have improved under private concessions.
Toll highways themselves are part of a three tier system comprised of federal free highways,
for which no toll is charged, federal toll highways, and state highways. Of the three types, federal
toll highways are of significantly higher quality than the other two. These toll highways were built by
the federal government or under private concession as an alternative to the overused and under-
maintained free federal and state highways, and are the preferred routes for those that can afford the
tolls4. Although toll roads make up only six percent of all highways, they comprise close to fifty
percent of highways with four or more lanes (Secretaría de Comunicaciones y Transportes, 2009).
These toll highways are less congested and usually are the fastest way to travel via road. For these
reasons, toll highways are more likely to capture the transport of valuable goods to the U.S.—
including drug shipments—than do other transport networks. In general, the toll roads follow or
link the most transited and valuable routes, many of which run to the U.S. border. For example, toll
highway segments link into seven crossing points into the U.S. This compares with only one
crossing point into Mexico’s southern neighbor, Guatemala. Figure 3 shows a map of the toll
highway system as well as this map superimposed on the earlier one of homicide increases. The
latter provides some visual confirmation of a relationship between toll highways and the increase in
violence.
A second advantage of using toll roads as instruments is that a majority of the system was
either built between 1989 and 1994 and follows previously established highway or rail lines. This
means that more recent factors linked to homicide rates and labor market outcomes largely did not
4 For a review of the highway privatization process in other Latin American countries (Argentina, Colombia and Chile),
see Engel et. al (2003)
16
determine their placement. To further ensure this is the case, we fix our measure of toll highways at
the year 2005, prior to the escalation of the drug war5.
To test if toll highways indeed are related to the increase in homicides due to the escalation of
the drug war, we plot annual homicide rates in states with more extensive federal toll highway
networks (the 12 states above the mean of 231 kilometers) and those with less extensive networks
(the 20 states below the mean). As shown in Figure 4, we observe a sharp increase in these rates in
the former group in 2007, but only a very weak and more gradual increase in the latter group.
Importantly, we observe no similar differences in trends between high and low highway states in
other economic characteristics, such as GDP growth, unemployment, and the extensive margin of
labor force participation. This confirms that states with more toll highways became more violent
during the intensification of the drug war and suggests the higher homicide trajectory for high toll
highways states is not simply a reflection of varying economic trends.
Figure 4 also shows that differences in federal highway networks do not appear to explain pre-
crackdown annual changes in homicides. This further confirms that the increase in violence stems
from the federal crackdown and is not due to other factors. For example, it is believed that Mexico
took over Colombia’s place as the major transport route of drugs into the U.S. starting from at least
1994 (Rios 2012). Second, an increase in political competition, which may have shifted the implicit
agreements between the cartels and various levels of government, also predates the spike in violence
by at least half a decade. In short, cartel-related violence responded most dramatically only after the
federal crackdown began.
Finally, to ensure that toll roads are not picking up general access to transport rather than access
to the U.S. specifically, we examine the relationship between the increase in homicides and other
5 Toll highways do increase over the 2005-2010 time period. Total kilometers rise from 7,409 in 2005 to 8,397 in 2010,
an increase of 13%. However, the rise is not uniform across states, and in approximately half total kilometers remain the
same (Secretaría de Comunicaciones y Transporte, Annual Statisticals).
17
transit routes. We regress changes in homicides on kilometers of all federal highways (which include
toll and free roads) and all highways (federal toll, federal free and state), and compare the results to
those for toll highways alone. The results for toll highways, shown in Panel A of Table 4, show a
clear relationship between toll highways and the post-crackdown increase in violence. Toll highways
are positively and significantly correlated with the level of homicides and changes in the peak years
of the conflict-- 2008, 2009 and 2010—but are not correlated with either prior to the crackdown
taking effect—2005, 2006 and 2007. Meanwhile, no similar patterns are observed for all federal
highways (panel B) and all highways (panel C). There is no significant correlation between either
measure and homicide levels or changes before or after the escalation of the drug war. We therefore
argue that toll highways do a better job of capturing access routes to the U.S. than other forms of
transit.
We thus focus our IV estimation on the 2007-2010 period and use federal toll highways as our
instrument. To do so, we begin by considering the relationship between homicides and federal toll
highway networks in levels:
ℎ�������� = � + ������ℎ��ℎ��� + ������ℎ��ℎ��� ∗ � + ��� +� +� + �
where �captures state-level fixed effects and �captures year-level effects, and � is an error
terms. As homicide rates in toll highway-intensive states appear to have risen consistently since
2007—with the increase approximately linear—we include a linear time trend that varies by the
extent of toll highways in a given state.
Rewriting this equation in differences yields:
Δℎ�������� = � + ������ ��ℎ��ℎ��� + ��Δ� + �� + Δ�
We use this as our baseline first stage. In this specification, we do not rely on the correlation
between federal highways and the levels of homicides in different states (which may be further
correlated with other unobserved factors). Instead, we use only the increasing effect of highways on
18
homicides over time as our exogenous variation by estimating the effect of federal highways on
changes in homicides.
4.2 Estimation and Results
We rewrite the outcome and homicide equations for individual-level estimation as follows:
Δℎ�! �" = # +#�Δℎ�������� + Δ$"� + Δ�% + λt + Δϵ)*+ (4)
Δℎ��������" = � + ������ ��ℎ��ℎ��� + ��Δ� + �� + Δ�" (5)
The identification assumption is that our instrument is uncorrelated with the unobserved
component of changes in hours worked, i.e. that E(Δϵ)*+ ∗ ���� ��ℎ��ℎ���|Δ$", Δ�) = 0.
We estimate this system using two stage least squares. Results from the first stage are presented in
Table 5 and the second stage in Table 6. Similar to the fixed effect model above, we repeat the
exercise for the full sample, separately by gender, and separately by salaried and self-employed
workers.
The first stage results in Table 5 show that the instrument—kilometers of toll highways—is
positively and significantly correlated with annual changes in homicides in all of the estimations.
After controlling for observable time varying factors, toll highways remain significant predictors of
homicide increases. In general, an increase in toll highways of 100 kilometers (close to half of one
standard deviation) is associated with an additional homicide per 100,000 inhabitants. Although
these coefficients are lower than those for yearly increases (Table 4), values for the R-squared, the F-
statistics for the test of weak identification and the Chi-squared statistics for the test of under-
identification are very high. Thus, there is little evidence that the first stage suffers from weak
identification.
The second stage results in Table 6 provide further evidence that homicides negatively impact
labor force activity. The coefficients on the instrumented value for changes in homicides are
19
negative in all but one of the estimations. As expected, however, the standard errors for the IV
estimates are much larger and only one of the coefficients is significant. Nonetheless, it is useful to
interpret the magnitudes of these estimates relative to those of our OLS specification. For example,
the coefficient for all self-employed adults suggests that an increase in homicides per 100,000
inhabitants of 10 leads to an average decline in work hours of approximately 0.3 hours per week—a
decline of approximately 1% that is consistent with our OLS results. Indeed, the IV estimates
suggest that reverse causality and omitted variables are not likely to be major sources of bias in our
OLS estimates for the sample as a whole.
At the same time, our IV results do indicate a difference in the heterogeneous responses to
homicides. We find that salaried women reduce their working hours by roughly two hours in
response to a rise in homicides of 10 per 100,000 residents—a reduction of more than 10% relative
to their baseline hours worked. These results indicate that salaried women in states with the highest
homicide spikes have reduced their working hours by as much as one-third to one-half—a dramatic
response. We cannot identify the mechanism for this reduction, but note that it may be related to
differences in real and psychic costs from violence borne by women in this group, as well as
differences in their ability to afford to forego earnings to avert these costs.
Finally, we also find a large (two hours) but statistically insignificant response among self-
employed men to a homicide rise of 10 per 100,000. The differential response among self-employed
men relative to salaried ones is consistent with our OLS estimates. Indeed, our IV results generally
offer evidence that increasing homicides lead to a non-trivial reduction in work hours that is
consistent with our OLS findings.
20
5 Heterogeneity by Income
Several papers have found that the burden of crime may fall unequally on households at different
points of the income distribution (Gaviria and Pagés 1999, DiTella et. al. 2010, Villoro and Teruel
2004). To investigate the possibility of differential responses to changes in homicides by income, we
return to our base model and estimate this specification on sub-samples of individuals based on their
household income quartile.6 The results are shown in Table 7. They show clear differences in the
response to homicides by income. Interestingly, the results are concentrated in the highest income
quartile. For all adults he estimated coefficient is three to twenty times larger than those for the
other three quartiles. The coefficient for all adults, shown in column 10, implies that an increase of
10 homicides per 100,000 inhabitants is associated with a decline in hours worked per week of 0.5.
This value suggests that for upper income adults in the states most affected by the drug war, hours
worked have declined by one and a half to two and a half hours per week hours per week.
Meanwhile, the response for adults in the lowest income quartile is close to zero while that of adults
in the second lowest one is a decline of 0.2 hours. Similar to the results for self-employment, we do
not know if these results are driven by greater exposure to violence, greater fear of violence, or an
enhanced ability to respond to both by changing work schedules. Nonetheless, the results provide
further evidence that the costs of violent crime are born unequally across income levels.
6 Conclusions
In this paper we estimate the impact of the escalation of the drug war in Mexico on the labor market
outcomes of adults. We focus on the dramatic changes in homicides in a subset of states since 2007
and examine its impacts on changes in hours worked. To identify the relationship between changes
6 Our data are constructed of rolling panels of individuals and thus may reflect compositional changes in labor force
participation over time. However, the correlation between changes in homicide and changes in labor force participation
is weak (-0.006) and thus unlikely to be driving our results.
21
in homicides and hours worked, we exploit the large variation in the trajectory of violence across
states and over time. Using both OLS and instrumental variables regressions, we find that the
increase in homicides has had a decidedly negative impact on labor force activity. An increase in
homicides of 10 per 100,000 is associated with a decline in weekly hours worked of 0.3 hours. For
states most impacted by the drug war, in which homicides per 100,000 inhabitants have increased by
30-50 a year, this implies an average decline in hours worked of one and one and a half hours per
week. These impacts are larger for the self-employed and are concentrated among the highest
income quartiles. This highlights how the costs of crime tend to be unequally borne across the
population.
The findings are consistent with a broad literature on the role of institutions in leading to
both short-run and longer-term economic development. This literature includes evidence on the
effects of institutional deterioration during civil wars and other internal violent conflicts—most
frequently related to control over the state (see Blattman and Miguel 2010 for a review). We
contribute to this field by identifying the impacts of violence related to control over private
resources—the effects of drug cartel violence on the economic behavior of private individuals.
Indeed, we find that even when the risk of directly incurring such violence is less than 1%, private
individuals may adjust their behavior by as much as 30-50%. These results suggest continued
attention on the role of non-state actors in forging institutions that can sustain economic
development.
22
References
Barrera, Felipe and Ana María Ibáñez. 2004. "Does Violence Reduce Investment In Education?: A
Theoretical And Empirical Approach," Documentos CEDE 002382, Universidad de los
Andes- CEDE
Blattman, Christopher and Edward Miguel. 2010. “Civil War.” Journal of Economic Literature 48(1), 3-
57.
Braakman, Nils. 2012. How do individuals deal with victimization and victimization risk?
Longitudinal evidence from Mexico. Journal of Economic Behavior and Organization.
Forthcoming.
Dell, Melissa. 2011. Trafficking Networks and the Mexican Drug War. Mimeograph MIT
Department of Economics
DiTella, Rafael, Sebastian Galiani and Ernesto Schargrodsky. 2010. “Crime distribution and victim
behavior during a crime wave” in The Economics of Crime, NBER Conference Report,
University of Chicago press, DiTella, Edwards, Schargrodsky editors
Engel, Eduardo, Ronald D. Fischer, and Galetovic P. Alexander. 2003. “Privatizing Highways in
Latin America: Fixing What Went Wrong.” Economía, Vol. 4(1), pp 129-164
Fernández, Manuel, Ana Maria Ibáñez and Ximena Peña. 2011. “Adjusting the Labor Supply to
Mitigate Violent Shocks: Evidence from Rural Colombia”, World Bank Working Paper
Series 5684
Gaviria, Alejandro and Carmen Pagés. 1999. “Patterns of crime victimization in Latin American
cities.” Journal of Development Economics, Vol.67, pp 181-203
Hamermesh, Daniel. 1999. “Crime and the Timing of Work.” Journal of Urban Economics, Vol.45,
pp 311-330
Molzahn, Cory, Viridiana Rios and David A. Shirk. 2012. ““Drug Violence in Mexico: Data and
23
Analysis Through 2011.” Trans-Border Institute, Special Report
Monteiro, Joana and Rudi Rocha. 2012. “Drug Battles and School Achievement: Evidence from Rio
de Janeiro’s Favelas.” Working paper
Rios, Viridiana and David A. Shirk. 2011. “Drug Violence in Mexico: Data and Analysis Through
2010”, Trans-Border Institute, Special Report
Rios, Viridiana. 2012. “Why Did Mexico become so violent? A self-reinforcing violent equilibrium
caused by competition and enforcement.” Trends in Organized Crime
Rodriguez, Catherine and Fabio Sanchez. 2009. “Armed Conflict Exposure, Human Capital
Investments and Child Labor: Evidence from Colombia”. Documentos CEDE, No. 5,
Universidad de los Andes- CEDE
Secretaría de Comunicaciones y Transportes. 2005-2010. Anuario Estadístico.
Villoro, Renata and Graciela Teruel. 2004. “The social Costs of Crime in Mexico City and
Suburban Areas.” Estudios Económicos. Vol. 19 (1), pp 3-44
24
Figure 1: Change in Total Homicides
Total 2005= 25,780
Total 2011= 37,375
250
00
30
00
03
50
00
400
00
2004 2006 2008 2010 2012Source:SESNSP
Years 2005-2011
Total Homicides by Year, Mexico
25
Figure 2: Geographic Distribution of 6 Year Change in Homicides
Change in homicides per 100,000 inhabitants75 to 10050 to 7520 to 500 to 20-20 to 0
Source: SESNSP
Years 2005-2010
Change in Homicides per 100,000 Inhabitants
26
Figure 3: Federal Toll Highway System
Federal Toll Highways
Change in homicides per 100,000 inhabitants75 to 10050 to 7520 to 500 to 20-20 to 0
Source: SESNSP & Secretaria de Comunicaciones y Transporte
Years 2005-2010
Change in Homicides per 100,000 Inhabitants and Kilometers Federal Toll Highways
27
Figure 4: Falsification Tests
20
30
40
50
2005 2006 2007 2008 2009 2010
States above mean toll highways
States below mean toll highways
Average Homicides per 100,000
-.05
0.0
5
2005 2006 2007 2008 2009 2010
States above mean toll highways
States below mean toll highways
Average GDP Growth
33.5
44.5
55.5
2005 2006 2007 2008 2009 2010
States above mean toll highways
States below mean toll highways
Average Unemployment Rate
.57
.575
.58
.585
.59
.595
2005 2006 2007 2008 2009 2010
States above mean toll highways
States below mean toll highways
Average Labor Force Participation Rate
28
Table 1: Summary Statistics Homicides
Homicides per All
100,000 inhabitants Years 2005 2006 2007 2008 2009 2010
Mean 28.72 23.60 24.59 25.09 29.16 32.46 37.38
Standard deviation 18.46 9.79 10.98 10.14 16.55 23.74 28.16
Minimum 4.19 4.19 12.41 14.22 14.04 10.51 8.69
Maximim 127.64 47.79 53.33 55.04 77.11 107.06 127.64
25th percentile 17.09 17.27 16.19 17.63 16.80 16.59 19.79
50th percentile 22.09 19.96 20.58 20.69 22.00 23.36 26.67
75th percentile 35.10 32.65 32.96 34.62 37.47 38.57 47.30
90th percentile 48.83 37.12 40.73 37.18 47.91 63.05 65.66
Observations 192 32 32 32 32 32 32
Source: SESNSP (National Public Security System)
Separately by Year
29
Table 2: Summary Statistics Labor Force Survey Data
Adults age 18-65 20005-2010
All Adults Men Women All Adults
(1) (2) (3) (4)
Age 38.23 38.25 38.21 38.18
(12.56) (12.65) (12.49) (12.57)
Education:
Primary 43.5% 41.3% 45.4% 44.7%
Secondary 30.5% 29.7% 31.2% 30.0%
Tertiary 25.9% 28.9% 23.3% 25.2%
Household size 4.70 4.68 4.72 4.73
(all indivinduals) (2.02) (1.99) (2.05) (2.03)
Children in households 0.99 0.99 0.99 1.00
(1.04) (1.04) (1.04) (1.05)
Monthly income 2509.57 3862.64 1419.64 2426.51
(pesos) (4281.67) (5175.04) (2977.42) (4609.55)
Household monthly 6511.00 6766.26 6290.28 6407.85
income (pesos) (8096.99) (8075.69) (8108.96) (8458.72)
At beginning of year:
In labor force 64.1% 85.9% 45.3% 64.1%
Unemployed 2.8% 3.6% 2.0% 2.6%
Of those in labor force:
Self employed 29.2% 30.7% 27.0% 29.5%
Salaried work 65.4% 66.3% 63.9% 65.0%
Weekly hours worked 27.27 39.37 16.85 27.31
(24.50) (21.72) (21.83) (24.52)
Change in weekly -0.07 -0.45 0.26 -0.05
hours worked (20.81) (22.43) (19.29) (20.59)
Over one year period:
Exit labor force 8.7% 7.0% 10.2% 13.3%
Change industry 12.6% 18.7% 7.4% 19.8%
Observations 94642 43759 50883 141054
Averages weighted by population weights. Standard errors
in parentheses. Source of data: ENOE
Years 2007-2010
30
Table 3: Change in Hours Worked, OLS Results
Outcome Variable= Change Years:2005-2010
in weekly hours worked
PANEL A: All Adults All Adults
(1) (2) (3) (4) (5) (6) (7)
Change Homicides -0.022* -0.030** -0.018 -0.045** -0.025 -0.014 -0.031***
(0.012) (0.012) (0.018) (0.019) (0.015) (0.015) (0.011)
Year fixed effects No Yes No Yes No Yes Yes
Individual time controls No Yes No Yes No Yes Yes
State time controls No Yes No Yes No Yes Yes
Observations 94,116 94,116 43,408 43,408 50,708 50,708 140,281
R-squared 0.000 0.001 0.000 0.002 0.000 0.002 0.001
PANEL B: Salaried Only
(1) (2) (3) (4) (5) (6) (7)
Change Homicides -0.004 -0.011 -0.020 -0.008 0.025 -0.011 -0.007
(0.016) (0.012) (0.021) (0.016) (0.027) (0.018) (0.011)
Year fixed effects No Yes No Yes No Yes Yes
Individual time controls No Yes No Yes No Yes Yes
State time controls No Yes No Yes No Yes Yes
Observations 41,772 41,772 25,226 25,226 16,546 16,546 61,774
R-squared 0.000 0.406 0.000 0.350 0.000 0.509 0.402
PANEL C: Self Employed
Only (1) (2) (3) (4) (5) (6) (7)
Change Homicides -0.044 -0.048* -0.061 -0.057* -0.018 -0.023 -0.033
(0.032) (0.028) (0.037) (0.034) (0.062) (0.047) (0.025)
Year fixed effects No Yes No Yes No Yes Yes
Individual time controls No Yes No Yes No Yes Yes
State time controls No Yes No Yes No Yes Yes
Observations 16,794 16,794 10,713 10,713 6,081 6,081 25,196
R-squared 0.000 0.239 0.001 0.230 0.000 0.277 0.241
*** p<0.01, ** p<0.05, * p<0.1
State time controls include changes in GDP, unemployment and the total share of production by industry
change in children, change in labor force status, and change in industry for those in labor force
All Adults Men Women
Robust standard errors in parentheses. Estimated using survey weights
All Adults Men Women
Individual time controls incldue change in household size,
Years: 2007-2010
All Adults
All Adults
WomenMenAll Adults
31
Table 4: Transport Routes and Homicides
Total Change
Year 2005 2006 2007 2008 2009 2010 2005 2006 2007 2008 2009 2010 2005-2010
Panel A: Federal Toll Highways
Kilometers Toll 0.009 0.014 0.010 0.034** 0.052** 0.061** 0.000 0.005 -0.004 0.024** 0.019** 0.009 0.043**
Highways (0.009) (0.010) (0.009) (0.014) (0.020) (0.024) (0.004) (0.005) (0.005) (0.010) (0.008) (0.008) (0.019)
R-squared 0.032 0.060 0.034 0.155 0.183 0.179 0.000 0.031 0.023 0.160 0.168 0.040 0.154
Panel B: All Federal Highways
Kilometers Federal 0.002 0.002 0.002 0.005 0.008 0.009 0.001 -0.000 0.000 0.003 0.003* 0.000 0.006
Highways (0.002) (0.002) (0.002) (0.003) (0.005) (0.006) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.005)
R-squared 0.028 0.020 0.025 0.068 0.086 0.064 0.010 0.000 0.000 0.054 0.088 0.000 0.062
Panel C: All Highways
Kilometers All 0.000* 0.000 0.000 0.000 0.001 0.001 0.000 0.000 -0.000 0.000 0.000 0.000 0.000
Highways (0.000) (0.000) (0.000) (0.000) (0.001) (0.001) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.001)
R-squared 0.097 0.085 0.067 0.028 0.027 0.023 0.002 0.001 0.011 0.000 0.016 0.002 0.002
Observations 32 32 32 32 32 32 32 32 32 32 32 32 32
Homicides per 100,000, Levels Homicides per 100,000, Annual Changes
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
Highway values as of 2005. Source: Secretaria de Comunicaciones y Transporte
32
Table 5: Instrumental Variable First Stage Results
Outome variable=Change in Homicides All Adults Men Women
(1) (2) (3)
Toll highway kilometers 0.008*** 0.008*** 0.008***
(0.000) (0.000) (0.000)
Observations 94,116 43,408 50,708
R-squared 0.285 0.283 0.288
Kleibergen-Paap F statistic for weak identification 665.1 306.1 360.1
Kleibergen-Paap Chi squared statistic for underidentification 636.5 296.4 341.8
PANEL B: Salaried Workers Only All Adults Men Women
(1) (2) (3)
Toll highway kilometers 0.008*** 0.007*** 0.009***
(0.000) (0.001) (0.001)
Observations 41,772 25,226 16,546
R-squared 0.288 0.286 0.297
Kleibergen-Paap F statistic for weak identification 249.4 134.8 119.0
Kleibergen-Paap Chi squared statistic for underidentification 238.9 130.6 112.3
PANEL C: Self -Employed Workers Only All Adults Men Women
(1) (2) (3)
Toll highway kilometers 0.007*** 0.008*** 0.004***
(0.001) (0.001) (0.001)
Observations 16,794 10,713 6,081
R-squared 0.260 0.277 0.258
Kleibergen-Paap F statistic for weak identification 126.6 115.9 16.06
Kleibergen-Paap Chi squared statistic for underidentification 123.5 113.9 16.10
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
Individual time controls incldue change in household size, change in children, in labor force status,
and the total share of production by industry. Years 2007-2010
and in industry for those in labor force. State time controls include changes in GDP, unemployment
33
Table 6: Instrumental Variable, Second Stage Results
Outcome variable: Change in hours worked
PANEL A: All Adults All Adults Men Women
(1) (2) (3)
Change in homicides -0.031 -0.032 -0.033
(0.078) (0.123) (0.100)
Observations 94,116 43,408 50,708
R-squared 0.001 0.002 0.002
PANEL B: Salaried Workers All Adults Men Women
(1) (2) (3)
Change in homicides -0.125 -0.075 -0.212*
(0.088) (0.129) (0.109)
Observations 41,772 25,226 16,546
R-squared 0.404 0.350 0.503
PANEL C: Self-Employed All Adults Men Women
Workers (1) (2) (3)
Change in homicides -0.099 -0.204 0.355
(0.195) (0.177) (0.743)
Observations 16,794 10,713 6,081
R-squared 0.239 0.228 0.267
Robsut standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
Individual time controls include change in household size, children, labor force status,
and the total share of production by industry. Year fixed effects included. Sample years 2007-2010
and industry for those in labor force. State time controls include changes in GDP, unemployment
34
Table 7: Heterogeneity by Income, OLS Results
Fixed Effect Results
Change, hours worked All Adults Men Women All Adults Men Women All Adults Men Women All Adults Men Women
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
Change Homicides 0.002 -0.039 0.026 -0.019 -0.004 -0.025 -0.001 -0.043 0.045 -0.052*** -0.035 -0.061**
(0.023) (0.041) (0.027) (0.025) (0.040) (0.032) (0.023) (0.034) (0.030) (0.020) (0.030) (0.026)
Year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Individual time controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
State time controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Observations 20,616 8,869 11,747 23,816 10,819 12,997 24,900 11,755 13,145 24,784 11,965 12,819
R-squared 0.053 0.069 0.048 0.002 0.007 0.011 0.013 0.033 0.010 0.053 0.082 0.044
Robust standard errors in parentheses
-0.15
-2.5
*** p<0.01, ** p<0.05, * p<0.1. Estimated using survey weights.
Income Quartile 4Income Quartile 3Income Quartile 1 Income Quartile 2
0
Living on the Edge: Youth Entry, Career
and Exit in Drug-Selling Gangs
Leandro Carvalho†
and
Rodrigo R. Soares‡
January 2013
Abstract
We use data from a unique survey of members of drug-trafficking gangs in favelas (slums) of Rio de Janeiro, Brazil, to characterize drug-trafficking jobs and study the selection into gangs, analyzing what distinguishes gang-members from other youth living in favelas. We also estimate wage regressions for gang-members and examine their career path: age at entry, progression within the gangs’ hierarchy, and short- to medium-term outcomes. Individuals from lower socioeconomic background and with no religious affiliation have higher probability of joining a gang, while those with problems at school and early use of drugs join the gang at younger ages. Wages within the gang do not depend on education, but are increasing with experience and involvement in gang-related violence. The two-year mortality rate in the sample of gang-members reaches 20%, with the probability of death increasing with initial involvement in gang violence and with personality traits associated with unruliness.
Keywords: crime, youth, gangs, drugs, trafficking, Brazil JEL codes: J4, K42, O15, O17
The authors wish to thank Guilherme Hirata for outstanding research assistance. This paper benefited from comments from Lars Lefgren, Giovanni Mastrobuoni, Yaw Nyarko, and seminar participants at RAND Corporation, the 2011 World Bank’s ABCDE Conference (Paris), and the LACEA 2012 Annual Meeting (Lima). Contact information: carvalho at rand.org and soares at econ.puc-rio.br. † RAND Corporation ‡ Pontifical Catholic University of Rio de Janeiro and IZA
1
1. Introduction
Youth account for a disproportionally high fraction of the perpetrators and victims of violence
(see, for example, Levitt and Lochner, 2001 and Soares, 2006). Part of this involvement with violence is
associated with membership to criminal groups, as exemplified by the Crips and Bloods in the 1980s’
Los Angeles, the pandillas and maras in several Latin American countries and throughout the US
penitentiary system, and the drug-trafficking gangs in the favelas (slums) of Rio de Janeiro. Information
on the activities and organization of these groups and on the individuals involved with them is extremely
rare. In the particular case of Brazil, a country with high violence and strong presence of drug-
trafficking gangs in poor areas of virtually every major urban center, very little is known about the way
they function.
A greater understanding of how criminal organizations attract the youth and how they operate is
paramount in designing an effective strategy to fight crime and curb violence, but the difficulty in
“getting inside” these criminal organizations has been a major obstacle. The handful of studies that have
been able to overcome this obstacle have made important contributions. With a more historical
perspective, Reuter (1983) and Gambetta (1993) studied the organization and functioning of the Italian
Mafia, while Leeson (2007) discussed the governance rules among 18th century pirates. Directly related
to this paper, there is also a considerable literature on the contemporaneous phenomenon of urban gangs,
most of a descriptive nature and with an ethnographic approach. Examples include Moore (1990), who
discussed the role of gangs in violence and the drug trade, Levitt and Ventakesh (2000 and 2001), who
analyzed the financial organization and history of a drug-selling gang and the outcomes of a cohort of
youth growing up in the Chicago projects, Dowdney (2003), who described the structure and social
norms of gangs in favelas of Rio de Janeiro, and Rubio (2007), who coordinated an impressive effort to
survey the perceptions and involvement of youth with pandillas and maras in Guatemala, Honduras,
Nicaragua, and Panama. Still, little is known about the selection of youth into these gangs, the
occupational structure of these organizations, and the typical “careers” of gang-members.
This paper uses a unique dataset to help fill in this gap. In 2004, a Brazilian NGO, Observatório
de Favelas, interviewed 230 individuals who worked for drug-selling gangs in 34 favelas of Rio de
Janeiro. 1 The survey collected detailed information on characteristics of both gang-members
(demographics, socioeconomic background, etc.) and their jobs (wages, hours, occupation, involvement
with violence, etc.). Gang-members were between 11 and 24 years old when first interviewed from June 1 Observatório de Favelas translates as Slum Observatory. Favela is used in the remainder of the paper to refer to Brazilian slums. The Observatório de Favelas (www.observatoriodefavelas.org.br) is an NGO founded by individuals originally from poor communities in Rio de Janeiro, focused on research and policy action to promote knowledge, public debate, and policy proposals on issues relevant to the favelas and other urban phenomena.
2
to August of 2004. After this baseline interview, interviewers attempted – with limited success – to
follow individuals monthly for the four subsequent months. In addition, death records were collected for
the two-year period following the baseline interview. The data allow us to draw an unprecedented
picture of the criminal entry, career, and exit among gang members. Despite limitations associated with
the non-random nature of the sample, the data provide an insight into drug-trafficking gangs that
represents an important progress in our understanding of the way these groups function.
This paper makes four main contributions. First, we describe in detail the characteristics of the
drug-trafficking jobs that existed in the favelas of Rio de Janeiro in 2004. We document that gang-
members earned on average $300 per month, only 23% more than other youth from the favelas, and
worked typically more than 10 hours a day. There were large risks associated with these jobs. At the
time of the first interview, more than half of the sample had participated in armed confrontations with
rival gangs and roughly two-thirds had participated in gun fights with the police. At the end of two
years, 20% of the initial sample had died. We also document how job characteristics vary according to
occupation within the gang. We show, for example, that the risks are even larger for members higher up
in the drug-trafficking hierarchy. Members at the top of the hierarchy earned 90% more than members in
entry-level occupations, but were also 10 percentage points more likely to die within two years.
Second, we investigate who the young men who voluntarily join drug-trafficking gangs are. We
combine data from the 2000 Brazilian Census with the survey of gang-members and use the procedure
suggested by Lancaster and Imbens (1996) for contaminated samples. This allows us to estimate a model
of selection into the criminal sector to investigate what distinguishes gang-members from other young
men living in the favelas of Rio de Janeiro that chose not to join a gang. This is one of the first estimates
of determinants of participation in criminal activities available for Latin America. We find that younger
individuals, from lower socioeconomic background (black, illiterate, and from poorer families) and with
no religious affiliation are more likely to join drug-trafficking gangs. For example, blacks are between 6
and 17 percentage points more likely to join a drug-trafficking, while the same number for illiterates is
between 6 and 20 percentage points.
The third contribution of the paper is to analyze the determinants of wages in drug-trafficking
jobs. We present what we believe to be the first set of Mincerian regressions for individuals employed
by criminal organizations. We find that there are no returns to schooling within the gangs, but there are
strong returns to experience, bravery, and loyalty. Each additional year working for a gang is associated
with a 10% increase in wages, while participation in an armed conflict increases wages by 5% and a
punishment for failing to comply with gang rules reduces wages by 17%.
3
Finally, the paper also examines the “career path” of criminals: age at entry, progression within
the gang hierarchy, and short to medium term outcomes. We find that troubled kids who have problems
at school and start using drugs early on are at greater risk of being recruited at younger ages. In line with
the evidence from the wage equations, we also document that position within the gang hierarchy is
positively correlated with experience and with participation in gun fights. Finally, we present evidence
that gang-members with weaker attachment to the gang and better outside opportunities are more likely
to quit the gang. For those who choose to continue in the gang, prospects are bleak. Each additional
experience of gun fight at the time of the initial interview is associated with an increase of 2 percentage
points in the probability of death in the following two years. Individuals with personality traits
associated with aggressiveness and lack of control are also more likely to die.
Our paper speaks to several streams of literature. There is a large literature on the relationship
between human capital and crime, with an almost exclusive US focus. Part of this literature analyzes the
short-run effects of schooling, through incapacitation of children and teenagers, on the incidence of
crime (Snyder and Sickmund, 1999, Jacob and Lefgren, 2003, Gottfredson and Soulé, 2005, Luallen,
2005), while other papers evaluate the long-run effects of education on criminal involvement through
better legal market opportunities and possibly changed preferences (Lochner and Moretti, 2004,
Deming, 2011, Lochner, 2010, Machin et al, 2010). Other relevant work includes studies on criminal
careers (see review on Blumstein et al, 1986) and on violence in connection with drug-trafficking (see
papers in De La Rosa et al, 1990). In the case of Brazil, NEPAD & CLAVES (2000), Neto et al (2001),
Sousa and Urani (2002), Dowdney (2003), and Observatório de Favelas (2006) present ethnographic
analyses of drug-trafficking gangs in the favelas of Rio de Janeiro that can be seen as complementary to
the current study.2
As mentioned before, our work is most closely related to the literature on the structure and role
of criminal organizations (Reuter, 1983, Moore, 1990, Gambetta, 1993, Levitt and Venkatesh, 2000,
Leeson, 2007, and Rubio, 2007). The contribution of the current study is to focus on individual gang-
members and their careers. We have a unique dataset with a wealth of individual characteristics that
allows us to paint a picture of the background and career of gang-members that, up to now, had been
impossible. Among other things, we present what we believe to be the first set of results related to wage
and occupational determinants within criminal organizations.
The remainder of the paper is structured as follows. Section 2 presents the data and describes the
characteristics of drug-trafficking jobs and of gang-members. Section 3 introduces a selection model,
2 The topic of this paper also bears some relationship with the literature on child combatants in contexts of civil war, such as exemplified by Blattman and Annan (2008 and 2010).
4
which we estimate by combining data from the gang-members survey with the 2000 Brazilian Census.
In section 3, we estimate Mincerian equations for drug-trafficking jobs. In section 4, we examine the
“career path” of criminals, analyzing their entry, progression within the gang hierarchy, and short to
medium term outcomes. Section 5 concludes the paper.
2. Data and Descriptive Statistics
Rio de Janeiro is an important transit point for cocaine exports to Europe and South Africa
(UNODC, 2012) and a regional distribution point for cocaine and marijuana. The city also has an active
retail market controlled by drug-trafficking gangs operating in some of its 600 favelas.3 The gangs are
armed groups formed by young men living in the favelas in which the retail markets operate.
This study relies on unique data collected by the Brazilian NGO Observatório de Favelas (OF)
on 230 individuals who worked for the drug-trafficking business in favelas of Rio de Janeiro, Brazil,
when first interviewed between June and August of 2004.4 The baseline interviews collected detailed
information on demographics, family background, and criminal activities. The NGO attempted to follow
individuals for a period of four months after the initial interview. Death records covering the two years
subsequent to the baseline interview were also collected. Given the difficulty in reaching the population
of interest, the study adopted a convenience sampling scheme, where the selection of interviewees was
based primarily on the connections that interviewers had to the drug-trafficking network. To conduct the
survey, the NGO selected 10 interviewers (5 men and 5 women) who had had some previous
relationship to drug gangs.5 Some of them had worked for a criminal organization before, while others
were in contact with individuals employed by the drug-trafficking business. Representativeness was also
taken into consideration when selecting the interviewers: the 230 interviewees worked in 34 different
favelas, geographically distributed across the city. Each interviewer was supposed to survey between 20
and 25 individuals.
In our analysis, we use several variables constructed using data from the OF survey. These
convey the information listed below and can be broadly classified into the following categories:
- individual demographic characteristics: age, race, years of schooling, and illiteracy;
3According to the 2010 Brazilian Census, approximately 1.4 million people, 22% of the city’s population, live in favelas. Depending on the criterion used to define a favela, it is estimated that there are either 868 or 600 favelas in the city of Rio de Janeiro. 4 The specific timing was the following: 7 interviewers conducted 157 interviews in June 2004; 2 other interviewers conducted 53 interviews in July; and 1 interviewer conducted 20 interviews in August. 5 All interviewers with the exception of one had completed high school, 3 were going to college at the time of the survey, and one had graduated from college.
5
- family background: whether individual was raised in a female single-headed household, number of
siblings, ownership of real estate, and occurrence of domestic violence;
- drug-trafficking jobs characteristics: age at entry, years of experience, monthly wage, and
occupation;
- history of violence and attachment to the gang: number of previous participations in armed
confrontations with rivals or with the police, possession of weapon during work time, and whether
individual had previously stopped working for the drug-trafficking gang;
- individual personality and behavioral traits: whether individual was perceived as unruly by his
family, age at which started using drugs, and whether individual was religious.
The interpretation of each of these variables is discussed when they are introduced in our
empirical analysis. For purposes of comparison with the average population of the city and to analyze
the selection into the gang, we also use individual level variables from the 2000 Brazilian Census: race,
age, illiteracy, real estate ownership by the family, and religiosity. We choose the 2000 Census because
it corresponds to the closest point in time for which we have data on households living in the favelas.6
We start in this section by describing the characteristics of gang-members and the occupational
features associated with drug-trafficking jobs. Given the scarcity of information on the subject, we
believe this descriptive characterization has enough value to warrant a detailed discussion.
2.1. Characteristics of Individuals Employed in Drug-Trafficking Jobs
Almost all gang-members interviewed were males (225, or 98%). They were between 11 and 24
years-old at the time of the baseline interview (average age of 16, and 67% between 16 and 18), with 9%
married, 28% with kids, but 67% still living with a parent. Only 37% had been raised by both parents
and the father had been involved in the upbringing in fewer than 40% of the cases (23% of the 230
reported that their fathers were deceased, while more than 10% did not know whether the father was still
alive). The mother had been involved in the upbringing of 81% of them. They had on average 2.7
siblings, with 22% having 4 siblings or more. Drug problems in the family, as well as domestic violence,
were relatively common: 30% reported having a sibling with drug problems, 10% having an addicted
father, 7% an addicted mother, and 23% reported having been a victim of domestic violence.
Despite the fact that 88% were below 18 years-of-age, only 10% were attending school.
Schooling outcomes at the time of the interview can be summarized as follows: 15% dropped out before
completing elementary school (without retention, a child in Brazil should complete elementary school
6 There were no major economic changes in Brazil between 2000 and 2004, period which predates the more recent acceleration in growth experienced by the country. According to the Penn World Tables, for example, income per capita in Brazil between 2000 and 2004 grew by 4% only (total growth accumulated over the period).
6
by age 11); 70% finished elementary school but did not graduate from middle school; 10% graduated
from middle school; and only 5% graduated from high school. Among the drop-outs, 57% had left
school before turning 15 and 35% had dropped out between ages 15 and 16. Roughly 60% reported
having had a previous legitimate job before working for the gang. One reason why these individuals
may have dropped out of school at early ages is because they started working for the drug business also
very young: 8% started before age 13, 58% started between 13 and 15, and less than 2% started after 18.
2.2. Characteristics of Drug-Trafficking Jobs
Table 1 presents the characteristics associated with drug-trafficking jobs in our sample. Average
monthly wages were around US$ 300.7 Individuals were very young, on average around 17, and worked
typically more than 10 hours a day, with roughly one day-off every two weeks. The vast majority of
individuals carried a gun at work, but some only sporadically.
A large fraction of them had trouble with the police: 28.5% had gone through juvenile detention
and 53% had been arrested at least once. The majority reported having been victims of police violence
and extortion. There was also a history of direct involvement with violence. Many took part in gun
fights both with the police and with rival gangs, 18% had committed at least one homicide, and 23.9%
had been wounded in shoot-outs; 24% had carried out a castigo, i.e., a physical punishment that is
applied to residents of the favela – including gang-members – for breaking the rules imposed by the
gang; and 22.3% reported having been a victim of physical punishment.8
Maybe the most striking information in Table 1 is the mortality rate of those employed in drug-
trafficking jobs. Two years after the initial interview, 45 individuals, or roughly 20% of the sample, had
died; 40 of them had been killed.9 To put it in perspective, this number is higher than the mortality rates
of military personnel in severe conflict areas. For example, less than 1% of the American military
personnel fighting in the Vietnam War were killed (Leland and Oboroceanu 2010).10 Death rates are
typically higher, however, in settings involving youth at risk. Levitt and Venkatesh (2001) follow a
cohort of men coming of age in a Chicago housing project in 1991 and record a 10-year mortality rate
above 10% for individuals aged between 17 and 26 at the time of the first interview, in a sample that
includes both gang-members and others. In a previous work, Levitt and Venkatesh (2000) estimate the
7 The information on earnings is provided in brackets defined in terms of multiples of the monthly minimum wage, while the information on hours is provided in brackets of daily hours. We translate these variables into scalars by choosing the mid-point in each bracket, and an equivalent distance above the last bracket. We use an exchange rate of 2 R$/US$. 8 Punishments may involve being shot in the hands or feet, beatings or death (Dowdney 2003). A gang member may suffer punishment for misappropriation of gang money or drugs or for committing a theft in the favela (Souza e Silva et al 2006). 9 Of the remaining five, two died of overdose, two were ran-over during armed robberies, and one died in a car accident. 10 For example, according to Leland and Oboroceanu (2010), the total casualty rate – deaths plus injuries – among American military personnel during the Vietnam war was 4.1%; death rate alone was below 1%.
7
four-year mortality rate for members of a Chicago drug-selling gang to be of the order of 28%, with the
average number of non-fatal injuries over the same period reaching 2.4 per gang-member, and the
average number of arrests being close to 6. Also in their setting, gang activity is an extremely risky
business.
All these patterns point to the necessity of better understanding the decisions of youth to join
drug-trafficking gangs in different contexts. The evidence available, both in Brazil and in the US, make
it hard to rationalize these decisions from a purely monetary perspective. Though issues of social status
within the local communities may be important, it is also possible that behavioral aspects associated
with lack of self-control and time-inconsistency may play some role. In fact, Lee and McCrary (2005)
present evidence on criminals’ behavior consistent with the existence of hyperbolic discounting.
The last four columns in Table 1 present the same descriptive statistics by occupation in the
drug-trafficking business. Dowdney (2003) argues that in each favela the drug gang operates as a strict
hierarchical structure with well-defined occupations and that such structure is repeated very similarly in
all favelas. Following the structure outlined in Dowdney (2003), we categorize our data into four
occupational levels: (a) occupation 1: look-outs (observe the area to warn of a police raid or a rival gang
invasion) and local transporters (move small quantities of drugs within a favela); (b) occupation 2:
street-sellers (sell drugs directly to consumers) and wrappers (handle and wrap the drug before sale to
consumers); (c) occupation 3: soldiers (responsible for local security and also for the main role in armed
confrontations); and (d) occupation 4: managers (depending on the level, accountable for the entire
operation in a favela or for the market for one specific drug). The distribution across occupations in our
sample is as follows: 20.2% were look-outs and transporters; 42.6% were street-sellers and wrappers;
24.8% were soldiers; and only 12.4% were managers.
Our data provide support for the hierarchic structure proposed by Dowdney (2003). Managers
are older (17.7 years old), have longer tenure in the gang (2.6 years of experience) and earn higher
wages than street-sellers, wrappers and soldiers, which, in turn, are older (16.8 years old for both
occupations), have greater experience (2.1 for street-sellers and 1.8 for soldiers) and earn more than
look-outs and local transporters (16 years old with 1.5 years of experience). Thus, look-outs and local
transporters seem to be the entry level occupations in the gang, while managers sit at the top of the
hierarchy. The ranking between occupations 2 and 3 is less clear. The wage, age, and tenure of the two
occupations are similar.
Table 1 shows that involvement with violence increases as we move in the occupational structure
from look-outs and transporters to street-sellers and wrappers, then to soldiers and to managers. While
8
less than 30% of those in lower level occupations claim to carry a gun at work on a daily basis, 78% of
managers report working armed. Similarly, soldiers and managers are more likely to have participated in
gun fights with police and rival gangs, to have carried out punishments, to have committed homicide, to
have been wounded in combat, and to have been killed in the two years following the initial interview.
Finally, soldiers and managers are less likely to have been punished for bad behavior, consistent with the
idea that these positions reflect a higher status within the gang and, therefore, require a history of
compliance with gang’s rules.
The high risks – of being arrested and/or victim of police violence, of being wounded in armed
conflicts with the police and with rival gangs and, most important, of being killed – associated with the
drug-trafficking jobs beg the question of why these young men join the drug-trafficking ranks in the first
place. Dowdney (2003) argues that children and adolescents are not forced or coerced to join a gang.
Consistent with this view, 39% of the sample reported having voluntarily stopped working for the drug
business at some point in the past. Roughly half of the sample mentioned they had been introduced to
the gang by a friend or a family member.
When asked what had led them to join the drug-trafficking gang, “to make a lot of money” and
“to financially support my family” were presented as the main reasons. These were also presented as the
main reasons for why they remained in the gang (together with “connection to the drug gang-
members”). Finally, when pressed to think about reasons that would make them quit, “earn a lot of
money doing something else” and “find a legitimate job” were the main answers for, respectively, 50%
and 30% of the respondents (not shown in the table).
2.3. Comparing Individuals Employed in Drug-Trafficking Jobs to Young Men Living in Favelas
In Table 2, we compare individuals in the sample of gang-members to young men residing in the
favelas of Rio de Janeiro. The Brazilian Census classifies favela areas as “subnormal urban
agglomerates,” allowing identification of the households located in favelas of a given city. We use the
2000 Census files to construct descriptive statistics for the population between 10 and 25 years-of-age in
favela areas of Rio de Janeiro.
We concentrate the comparison and the remainder of our analysis on males because they
correspond to 98% of our sample of gang-members, which is consistent with evidence that victims and
perpetrators of violence in Brazil are predominantly young males (see, for example, Cerqueira and
Soares, 2012). Table 2 presents the set of variables that can be compared across the Census and the
9
gang-members datasets: individual characteristics related to race, religion, marital status, age, education,
labor market status, and earnings.11
Gang-members earned on average 23% more than other young men living in the favelas.12 The
counterfactual wages of gang-members may be somewhat lower than the wage of the average young
man living in a favela, given that gang-members are more likely to be black, illiterate, and younger,
characteristics that are penalized in the legal labor market. Finally, the 12% unemployment rate among
young men living in favelas may also increase the attractiveness of drug-trafficking jobs. Gang-members
also have characteristics indicating a higher probability of being at risk: they are 6 percentage points
more likely to be illiterate, 44 percentage points less likely to be attending school, 5 percentage points
more likely to be married at such young ages, and 19 percentage points less likely to be religious.
We proceed next to compare the socioeconomic status (henceforth, SES) of gang-members’
families to the SES of families of young men living in favelas. One complicating factor for such
comparison is that in the Census we can only observe the parents’ SES if the young man still lived with
his parents, which is a selected sample of young men living in favelas. For this reason, we use as
comparison group men between ages 25 and 65 and women between ages 25 ages 60 living in favelas,
who would have been old enough to be parents of the young men surveyed in the OF sample (see
Appendix Table A1).
The limited evidence we have suggests that gang-members come from lower SES families. The
comparison indicates that the mothers of gang-members had on average 0.8 less years of schooling than
women aged 25-60 living in favelas. The income of families living in favelas (in which either the head
or the head’s spouse was a man aged 25-65 or a woman aged 25-65) was on average 26% higher than
the parental income of gang-members. Gang-members also came from larger families with 1.4 more
children and were 15 percentage points more likely to have been raised by a single mother.
These results should be interpreted with caution given the limitations of the data. First, a large
fraction of individuals interviewed in the OF sample did not know their parents’ income or schooling:
36% failed to report their mother’s education and 29% did not answer about their parents’ income.
Another issue concerns the identity of the survey respondent: while in the OF survey gang-members
were reporting the schooling and the income of their parents, the household head or the head’s spouse
may have reported this information when interviewed by the Census. Finally, some of these
characteristics may not be directly comparable. For example, in the OF survey respondents were asked a
11Values from the 2000 Census are deflated to June 2004 and then converted into US dollars. 12 The monthly minimum wage at the time was 260 Brazilian Reais, or roughly US$ 130.
10
retrospective question about whether they had been raised by their mother only while in the Census we
can identify the fraction of families in which the head is a single female.
In the following sections, we investigate in more detail the relationship between individual
characteristics and criminal careers. First, we study the issue of selection into the criminal career and
investigate the determinants of drug-trafficking wages. We then proceed to analyze the determinants of
the typical career – entry, progression, and exit – in the gang.
3. Selection and Wages in Drug-Trafficking Jobs
3.1. Model and Empirical Strategy
All the young men living in a favela in which a drug-trafficking gang is active face the same
decision: whether to join the gang or not. The members of a gang are recruited from young men living in
the favela in which the gang operates and to all accounts any young man can join the drug-trafficking
ranks. If so, why did some young men decide to work for the drug-trafficking business while some of
their peers did not? Are their families different in any peculiar way? Are they distinguishable from other
youth? Would it possible to tell, beforehand, who are the children most likely to follow a criminal path?
We tackle this question by developing a model of selection into drug-trafficking jobs that makes use of
data from the OF survey and from the 2000 Brazilian Census.
The analysis of the decision to participate into crime gains additional relevance in light of the
evidence from the US. Blumstein et al (1986), for example, document a great heterogeneity in
participation into crime across demographic groups (extensive margin decision across gender, age, race,
and socioeconomic background), but a striking homogeneity in the level of criminal activity for those
who do participate (intensive margin). In other words, to understand the dynamics and the determination
of criminal involvement, the extensive margin (entry and exit of individuals) seems to be much more
relevant than the intensive margin (level of criminal activity of an individual involved with crime).
Suppose there are two sectors that workers can choose: an illegal sector denoted by the subscript
I and a legal sector denoted by the subscript L. Earnings in the illegal sector are determined according to
, , , (1)
where is a vector of characteristics of individual j and , ~ 0, . Similarly, earnings in the legal
sector are determined according to
, , , (2)
11
where , ~ 0, . and are the returns to individual characteristics in the illegal and legal
sectors, respectively. In principle, some entries of and may be zero, such that some characteristics
are valued in one sector but not in the other. Worker j will choose to work in sector I if illegal earnings
are sufficiently higher than legal earnings to compensate for potential negative characteristics associated
with illegal jobs:
, , , (3)
where indicates the compensating differential related to the risks associated with illegal jobs and the
moral conflicts that some may face when working in illegal occupations. The weight attributed to the
risk and moral dimensions of illegal activities may depend on individual characteristics, so that we
write:
, (4)
where indicates a set of demographic characteristics and a non-observable component with
~ 0, . In principle, may contain some of the same elements of and other components
(variables related to risk aversion and moral views on crime, for example, could enter but not ).
Under these assumptions, the probability that worker j works in the illegal sector is given by:
1| , ; , Φ , (5)
where is an indicator variable that is equal to 1 if individual j is employed in the illegal sector and 0
otherwise, and .
One challenge to the estimation of the participation equation (5) is that we do not have a dataset
in which we observe simultaneously individual characteristics and whether the individual is employed in
the legal or illegal sectors. Instead, we have two independent samples. The first is drawn from the
subpopulation who chose to work in the illegal sector (i.e., 1). These data come from the OF survey
that interviewed young men employed in the drug business. The second sample is a random sample of
the population of young men living in favelas for whom only covariates are observed – that is, we do not
observe whether these workers are employed in the illegal or in the legal sector. These data come from
the 2000 Brazilian census. We use the variable to distinguish the two samples: is equal to 1 if the
data on individual j come from the OF survey and 0 if they come from the Census. Let be the number
of observations in the OF sample, the sample size of the Census sample, and q denote the fraction of
the overall population employed in the illegal sector.
12
Lancaster and Imbens (1996) propose a generalized method of moments (GMM) estimator for
equation (5) when is known. They assume that the relative sizes of the two samples are determined by
a sequence of Bernoulli trials with unknown parameter h and consider the following moments:
, , , , , ,Φ
1 Φ Φ, 6
, , , , , ,1 Φ
1 Φ, 7
, , , , , ,Φ
1 Φ. 8
They show that , and , which equate the sample moments to zero, are consistent estimators of
, and h:
, , argmin , , . (9)
We estimate the participation equation (5) through GMM using their proposed estimator,
providing different estimates for different values of . Here we give some intuition for their estimator by
looking at the simplest case possible, when there is only one dichotomous independent variable.
Suppose we are interested in identifying 1| 1 . Notice that 1 , 1|
1 and 1 are all known. From Bayes’ rule, we can estimate:
1| 11| 1 1
1. 10
The second goal of our empirical analysis is to estimate the wage equation (1) for the illegal
sector. Because youth who self-select into the illegal sector may be very different from the population at
large, ordinary least squares are potentially biased. To correct for self-selection, we use Heckman’s two-
step method. After estimating the participation equation (4), we estimate the following modified wage
equation:
, , , (1’)
13
where
is the Inverse Mills’ Ratio calculated using the estimated coefficients
and obtained from the estimation of the participation equation.
3.2. Selection into Drug-Trafficking
Table 3 presents the results associated with the participation equation (5). To the best of our
knowledge, this is one of the first set of estimates on the determinants of participation in criminal gangs
available for Latin America. GMM estimates are reported in Panel A while Panel B shows the marginal
effects. The table shows results for different values of q – i.e., for different assumptions about which
fraction of young men living in favelas was working for the drug-trafficking business, namely 5%, 10%
and 15%. Dowdney (2003) estimates that 1% of residents of Rio de Janeiro’s favelas are employed by
the drug-trafficking business. According to the 2000 Brazilian Census, the population of men between
10-25 corresponded to 15.4% of the total population living in Rio de Janeiro’s favelas. These two
figures together suggest that 6.5% of men ages 10-25 living in favelas were members of drug-trafficking
gangs. This is a relatively small number compared, for example, to the Chicago neighborhood analyzed
by Levitt and Venkatesh (2000), where a quarter of the males between 16 and 22 were estimated to work
as foot-soldiers for the gang. Rubio (2007), looking at Central American countries, reports a
participation rate in pandillas among individuals of lower socioeconomic background of 5% for
individuals enrolled in school and of 17% for those who dropped out.
We focus our attention on Panel B for ease of interpretation. The first specification includes only
a dummy for black and age. In the second specification, we add a dummy for illiteracy. The third
specification includes in addition a dummy for whether the family owned the house in which the gang-
member lived, a proxy for the socioeconomic status of the individual’s family. Finally, in the last
specification, we add a dummy for whether the individual was not religious. It is apparent from the table
that the magnitudes of the estimates vary with q, but the results are qualitatively the same irrespective of
the assumption on the fraction of gang-members.
The table indicates that individuals with fewer opportunities in the legal labor market are more
likely to select into the illegal sector: black, young, and illiterate men are the ones at highest risk of
working for the drug-trafficking gang. Depending on the assumption about q, blacks are from 6 to 17
percentage points more likely to work for the gang, while the illiterate are 6 to 20 percentage points
more likely. Consistently with this pattern, young men from poorer socioeconomic backgrounds
(proxied by ownership of the house) are also more likely to work for the drug-trafficking business.
Finally, young men that reported having no religion were from 4 to 13 percentage points more likely to
be part of a drug-trafficking gang. Overall, the qualitative pattern of correlation between socioeconomic
14
background and participation into crime is broadly consistent with the evidence available from other
settings, such as that presented by Blumstein et al (1986) and Rubio (2007).13
3.3. Wages
In this section, we investigate the determinants of criminal productivity within the gang. We
proceed with two exercises. First, we follow the model from section 3.1 and estimate a Mincerian
equation for the illegal sector, accounting for the possibility of selection into the gang. We run a
regression of the natural logarithm of wages on the first four variables from Table 3 – namely race, age,
illiteracy, and ownership of real estate by the family – while controlling for the Inverse Mills’ Ratio
implied by the most complete specification in Table 3. The variable excluded from the wage regression
is a dummy for being religious, implying that the identifying assumption is that religion affects the
probability that the individual joins the gang, but, conditional on joining the gang, religion does not
affect wages. While this analysis takes into account selection, it restricts us to using information that are
available both in the OF and in the Census datasets.
In our second exercise, we depart from the model outlined in section 3.1 to further explore the
wealth information available in the OF survey. We begin by running specifications similar to those
typically used in labor economics, regressing the natural logarithm of wages on race, experience, years
of schooling and age at which one had started working for the drug gang. We also include in this
benchmark specification a set of controls for family background, namely whether the individual was
raised by a single mother, whether the family owned the house where he lived, and number of siblings.
We then extend this traditional specification by including variables that may be particularly important
for productivity and career advancement in crime: a dummy indicating whether the individual had a
physical disability, a variable indicating the number of times that the individual had participated in
armed confrontations with rival gangs or with the police, and another dummy variable indicating
whether he had ever been physically punished for breaking the gang’s rules. Finally, we include
occupational dummies indicating different positions within the gang hierarchy. Again, to the best of our
knowledge, these two sets of results are the first estimates of Mincerian regressions for individuals
employed by criminal organizations.
Table 4 presents the results of the first exercise, in which we keep only the variables used to
estimate the selection equation. The table contains four panels, each representing a different assumption
concerning the selection correction. In the first panel, there is no correction for selection, while in the
remainder three panels we include the Inverse Mills’ Ratio corresponding to the three models estimated
13 Even though Rubio (2007) documents a significant correlation with socioeconomic only for individuals with lower educational levels.
15
in Table 3, respectively: q = 5%, q = 10%, and q = 15%. In each panel, the first column includes only
race and age as explanatory variables; the second column includes in addition a dummy for illiteracy;
and the third column includes also a dummy indicating ownership of the house by the family.
Qualitative results across the four panels are identical: age and illiteracy are positively correlated
with wages, while race and ownership of real state by the family are not. Wages exhibit a very steep
profile within the gang, with a roughly 10% increase associated with each additional year. Surprisingly,
illiteracy also appears positively correlated with wages. Since we do not control for actual experience
within the gang in these specifications, it is possible that, conditional on age, illiteracy is correlated with
a longer tenure in the drug-trafficking business (indicating earlier drop-out from school and entry into
the gang). If anything, education seems to reduce the attractiveness of illegal occupations.
The other important result to come out of Table 4 is that selection into the drug-trafficking gang
does not seem to affect much the estimated coefficients in the wage regressions. The coefficients on age
are almost identical across the four panels, while those on illiteracy increase by roughly 20% when the
Inverse Mills’ Ratio is included, but remain unchanged irrespectively of the assumption on q. Given
these results, it looks as if selection into crime is not a serious enough hindrance to the estimation of
wage regressions for the illegal sector, even more so if one controls for a broad set of characteristics.
With these results in mind, we move to Table 5 in which we estimate Mincerian regressions
ignoring selection and exploring in more detail the wealth of information from the OF survey. The first
column includes only the individual characteristics and family background variables described before.
Columns 2 to 4 include, in sequence, a dummy for physical disability, the number of times the
individual participated in armed confrontations, and a dummy indicating whether he had been physically
punished for breaking gang’s rules. Column 5 includes all these three variables simultaneously. Column
6 goes back to the benchmark specification, but now including dummies for occupations, and column 7
includes all independent variables simultaneously.
The first result to come out of Table 5 is that there are no returns to education within the gang.
The estimated coefficients are very close to zero and are precisely estimated. There is also no penalty for
being black and no returns to family background (being raised by a single mother appears as borderline
significant in some specifications, but this significance disappears as other controls are included). There
are, however, high returns to experience. In the first specifications, each additional year working for the
gang is associated with a 10% increase in wages. Age at entry into the gang, in turn, is also positive and
large in magnitude: conditional on experience within the gang, having entered later is associated with
16
higher wages. This may reflect the fact that there is a negative selection on unobservables for individuals
who enter very young into the gang.
Columns 2 to 5 show which individual characteristics the gang rewards. The lack of a physical
disability, a proxy for physical prowess, is associated with a wage premium of 37%. The number of
armed confrontations in which the individual had participated, which supposedly reflects both combat
experience as well as the skills required to have survived, is also associated with higher wages. Each
additional conflict is associated with a 5% increase in wages. Finally, obedience to the gang’s rules is
also rewarded. Members who reported having been physically punished at least once for failing to
comply with such rules experienced a 17% wage penalty.
In the last two columns of Table 5 we include dummies for occupation within the drug-
trafficking hierarchy. The results confirm the structure discussed above. Managers earn higher wages
than street-sellers, wrappers and soldiers, which, in turn, earn higher wages than lookouts and
transporters. Conditional on observables, managers earn typically 35% more than members in entry-
level occupations. We cannot reject the hypothesis that street-sellers and wrappers are paid as much as
soldiers. When the occupational dummies are included in the benchmark specification, the coefficient on
experience is reduced by almost 50% while the one on age at entry is reduced by 30%, suggesting that
the previous results were partly capturing wage increases associated with upward movements within the
rankings of the criminal organizations. This is reassuring in that the estimated equations do seem to
capture relevant measures of productivity within drug-trafficking gangs.
4. Careers in Drug-Trafficking Jobs
4.1. Age at Entry
In this section, we use information on the age at which individuals joined the gang (henceforth,
age at entry) to investigate which characteristics predict early entry. Early entry into crime is supposedly
associated with a lower accumulation of human capital (that would be rewarded in the legal labor
market) and consequently with higher chances of remaining in the criminal path. Chen et al (2005), for
example, analyze the long-term criminal outcomes for youth that were apprehended by the police or
brought before the Children’s Court in Australia. They show that early initial contact with the justice
system is correlated with a high probability of later involvement in crime. Therefore, age at entry is
likely to be correlated with a more consistent long-term involvement with gang activity.
We estimate OLS regressions of age at entry on the following groups of variables:
socioeconomic background (proxied by whether the gang-member had been raised by a single mother,
17
the number of siblings he had, and whether the family owned the house in which he lived), home
environment (proxied by occurrence of domestic violence while growing-up), schooling (a dummy for
illiteracy), personality traits (whether individual was perceived by his family as unruly), and history of
drug-use (whether he had started using drugs before age 13).14 We choose to use illiteracy as a measure
of schooling because years of schooling may be endogenous to age at entry – i.e., joining the gang may
lead members to drop out of school. Illiteracy status, on the other hand, should have been already
determined by the time individuals decide to join the gang. Similarly, we choose the other family and
individual related characteristics to reflect dimensions that, in principle, would be predetermined at the
time of this decision.
The results are presented in Table 6. The two main correlates of age at entry are illiteracy and
history of drug-use. Illiterate individuals typically joined the gang 1.3 years before literate individuals,
suggesting that early school drop-outs join the gang at younger ages (there are very few – if any –
employment opportunities in the legal labor market for individuals younger than 14-15). Individuals
who started using drugs before 13 typically entered the gang 1 year before other individuals. Of the
other estimated coefficients, none is close to being statistically significant. These results suggest that
troubled kids who have problems at school and start using drugs early on are at greater risk of being
recruited at younger ages. Inciardi (1990), analyzing a population of adolescent offenders in Dade
County (Miami), also highlights a strong relationship between drug use, entry into crime, and
involvement with violence.
There are two limitations in this exercise. First, because we observe individuals years after initial
entry, this sample is a selected sample of individuals who had not been killed and had not quit the drug
gang between entry and the moment of the interview. Second, because the variables in the survey are
measured after entry, they may be endogenous to age of entry. As mentioned before, to minimize the
latter potential problem, we restricted the analysis to independent variables that would have been
determined before age 13 and remained constant throughout the individual’s lifetime.
4.2. Occupational Progression
Next, we consider which characteristics determine an individual’s occupation within the drug-
trafficking hierarchy. In order to conduct this analysis, we estimate a multinomial logit with outcomes
corresponding to the four different occupational categories listed before: lookouts and local transporters;
street-sellers and wrappers; soldiers; and managers. For ease of exposition, we call these occupational
14 Qualitative and quantitative results are very similar if we estimated the model as an ordered logit with the independent variable indicating age brackets (less than 10, from 10 to 12, from 13 to 15, from 16 to 18, and above 18). We show the OLS results because they are easier to interpret.
18
categories from now on, respectively, entry, mid-level, soldier, and manager. Given the evidence from
the wage equations, we estimate two specifications. The first one includes demographic and family
background variables, namely race, experience, age at entry, schooling, a dummy for being raised by a
single mother, ownership of real estate, number of siblings, and a dummy for physical disability. The
second specification includes also variables proxying for performance within the gang: participation in
armed conflicts and physical punishment for breaking the gang’s rules.
Table 7 presents the marginal effects from the multinomial logit estimation for the probability
that the individual was employed in each of the 4 different occupational categories. As expected,
individuals with more experience are more likely to be in the higher-ranking occupations. The main
result in Table 7 is that there is a positive association between participation in armed confrontations and
ranking: individuals who had participated in armed confrontations were less likely to be employed in
lower level occupations and more likely to be soldiers and managers. Individuals with one additional
previous conflict experience were roughly 3 percentage points less likely to be in entry and mid-level
positions and 2 percentage points more likely to be managers. There are two possible interpretations of
this result. One is that individuals who participate in armed confrontations, either to defend the gang’s
turf or to invade rivals’ territories, are promoted for showing commitment to the gang’s core activities
and willingness to engage in risk. Another possible explanation is that a greater involvement with
violent activities is one of the duties of higher ranking occupations. Interestingly, having being
physically punished for failing to comply with the gang’s rules is associated with a lower probability of
advancing to higher occupations (although the coefficient is imprecisely estimated and thus not
significant). Age at entry is positively associated with ranking, indicating that individuals who join the
gang at older ages may skip the entry-level positions and already start in higher occupations. Members
with a physical disability were less likely to be employed in entry level occupations – look-outs and
transporters – because these occupations require mobility.
Overall, the results related to the determinants of occupation within the gang’s hierarchy fall in
line with the previous results from the wage equations. This is reassuring in that we seem to be
identifying indeed dimensions that are deemed relevant for productivity within the gang.
4.3. Exit
In this section, we analyze the determinants of exit (temporary or permanent) from the gang
using longitudinal data from monthly follow-up surveys that tracked the gang-members over a four-
month period after the initial interview. Even though the data are noisy, they provide a unique
opportunity to study what drives gang-members to voluntarily disassociate from drug-trafficking. We
19
have data for only 189 gang-members out of the 225 that were initially interviewed and the attrition rate
is high (in each wave, on average, 50 respondents could not be found; in addition, 19 were killed and 1
arrested during this period),
In the analysis, we estimate a conventional logistic hazard model (Allison, 1982, and Efron,
1988). The unit of observation is individual i in wave t if he was known to be at risk (so individuals who
were not interviewed in wave t – either because they were in prison, had been murdered, or could not be
located – are not part of the sample). For simplicity, we assume that exiting the gang is an absorbing
state, such that an individual who quit in wave t is not at risk in subsequent waves. The dependent
variable, , , is 1 if individual i exited the drug-gang in wave t and 0 if individual i was still drug-
trafficking in wave t. We run logit regressions, including dummies for each wave.
We start from a specification including as independent variables demographic and family
characteristics (race, experience, education, whether individual was raised by a single mother, whether
the family owned the house where he lived, and number of siblings). Following, to capture the
individual’s attachment to the gang, we include a dummy indicating whether he had a legal job before
joining the gang, and another dummy indicating whether he had previously quit the gang voluntarily.
Finally, we include variables capturing his exposure to violence inside the gang (number of gun fights
with police and rivals) and personality traits (a dummy indicating whether the individual was seen as
unruly by his family). The main goal of this sequence of specification is to capture, in addition to
standard demographic and family related variables, dimensions that would reveal the individual’s
connection with the gang and his history inside it.
The results shown in Table 8 suggest that gang-members with weaker attachment to the gang and
better outside opportunities were more likely to quit. Members who reported in the baseline interview
that they had exited the gang before were roughly 10 percentage points more likely to quit again. Gang-
members with more years of experience, who supposedly had developed stronger ties, were less likely to
quit. Each year of experience reduces the probability of exit by 2 percentage points. Together, these
results indicate that some members of the gang had a weaker attachment to it, working on and off for the
gang. Members who had a legal job before joining the gang, and who supposedly could more easily find
another job, were 4 percentage points more likely to quit (although the result is not statistically
significant). Finally, members more exposed to risk – as proxied by the number of armed confrontations
with rival gangs and with the police – were more likely to quit, with each participation in an armed
confrontation being correlated with an increase of 1 percentage point in the probability of exiting.
20
The results suggest that weaker initial attachment to the gang increases the probability of later
exit. At the same time, individuals seem to be aware of the risks that they are exposed to when working
for the gang: conditional on other observable characteristics, previous involvement with violence does
increase the probability of later exit. In combination, these open up the possibility for better outside
options for gang-members – maybe through targeted reintegration programs – to have potentially large
effects on the involvement of youth with gang violence.
4.4. Mortality after Two Years
In Table 9, we examine which individual characteristics predict the likelihood of death within
two years of the baseline interview. These data were collected either directly or indirectly, through
relatives and friends. We estimate a linear probability model by OLS following the same sequence of
specifications from Table 8 (marginal effects from a logit specification are almost identical).
The results show – perhaps not surprisingly – that members who were more involved with the
violent activities of the gang at the time of the baseline interview, proxied by the number of armed
confrontations, were more likely to die in the following two years. Each additional confrontation is
associated with a 2 percentage-point increase in the probability of death. Consistent with this result,
members who reported in the baseline interview that they had temporarily stopped working for the gang
before were 8 percentage points less likely to die. The standard errors are, however, imprecisely
estimated and the coefficient is not statistically different from zero.
The risk of death is also predicted by whether the individual was perceived as unruly by his
family. Unruly individuals were 16 percentage points more likely to be dead at the end of the two-year
period. This result is similar to that obtained by Chandler et al (2011). The authors analyzed data from
Chicago public schools and found that children with disciplinary and behavioral problems had a higher
risk of being shot within a given year. Blumstein et al (1986) also document a relationship between
personality traits associated with lack of discipline and parental control and early involvement with
violence. Extreme negative outcomes, therefore, seem to be associated to a great extent with recurrent
involvement with violence and behavioral problems.
None of the other variables, including proxies for family background, predict mortality. The only
exception is the coefficient on being raised by a single mother, which is large and marginally significant.
If taken at face value, the point estimate implies that gang-members that had been raised by a single
mother were 9 percentage points more likely to end up killed.
21
5. Concluding Remarks
This paper uses a unique dataset of individuals employed in drug-trafficking gangs operating in
favelas of Rio de Janeiro to draw an unprecedented picture of the criminal entry, career, and exit
alternatives among gang-members. We explore a survey conducted by Observatório de Favelas, a
Brazilian NGO, that collected detailed information on demographics, family background, and criminal
activities for individuals involved with drug-trafficking gangs. We document that gang-members worked
long hours and earned little more than other youth from the favelas working in the legal sector. Still,
there were large risks associated with gang-membership: at the time of the first interview, more than
two-thirds of the sample had participated in gun fights and, after two years, 20% were dead. The results
also show that younger individuals, from lower socioeconomic background (black, illiterate, and from
poorer families) and with no religious affiliation were more likely to be gang-members. Education was
not rewarded within the gang, but experience, displays of bravery, and loyalty were. We find that kids
who dropped out of school early on and who had early problems with drugs were more likely to join the
gang also at earlier ages. Finally, we present evidence that gang-members with weaker attachment to the
gang and better outside opportunities were more likely to quit the gang; those with more previous
involvement with violence and personality traits associated with aggressiveness and lack of control, on
the other hand, were more likely to die within two years. The provision of education and social support
for troubled children may therefore constitute a good combination of policies to delay entry into crime
and to reduce the probability of extreme negative outcomes for those who enter crime. Still, the results
presented here should be taken only as a first step in the direction of understanding the determinants of
youth involvement with criminal gangs. A main puzzle that remains is the motivation behind youth
involvement with gang activities: monetary returns are relatively low and risks are extremely high. A
departure from the more strict economic analysis of crime, focused simply on monetary returns, seems
to be needed in order to advance the knowledge in the area.
22
References
Allison, P.D. (1982). Discrete-time methods for the analysis of event histories. In: Leinhardt, S. (Ed.), Sociological Methodology. Josey-Bass, San Francisco, CA.
Blattman, Christopher and Jeannie Annan (2008). Child Combatants in Northern Uganda: Reintegration Myths and Realities. In: Robert Muggah (ed.). Security and Post-Conflict Reconstruction: Dealing with Fighters in the Aftermath of War, Routledge, 103-126.
Blattman, Christopher and Jeannie Annan (2010). The Consequences of Child Soldering. Review of Economics and Statistics, v92, n4, 882-898.
Blumstein, Alfred, Jacqueline Cohen, Jeffrey A. Roth, and Christy A. Visher (1986). Criminal Careers and “Career Criminals.” National Academy Press, Washington DC, 489p.
Cerqueira, Daniel and Rodrigo R. Soares (2012). “The Welfare Cost of Homicides in Brazil: Accounting for Heterogeneity in the Willingness to Pay for Mortality Reductions.” PUC-Rio Working Paper n600.
Chandler, Dana, Steven D. Levitt, and John A. List (2011). Predicting and Preventing Shootings among At-Risk Youth. American Economic Review Papers and Proceedings, 101(3), 288-92.
Chen, Shuling, Tania Matruglio, Don Weatherburn, and Jiuzhao Hua (2005). The Transition from Juvenile to Adult Criminal Careers. Crime and Justice Bulletin – Contemporary Issues in Crime and Justice, n86, 1-12.
Chioda, Laura, João Manoel De Mello, and Rodrigo Soares (2010). “Does Education Reduce Crime and Violence? Evidence from Bolsa Família and Extending the School Day in São Paulo.” Unpublished manuscript, PUC-Rio.
De La Rosa, Mario, Elizabeth Y. Lambert, and Bernard Gropper (1990). Drugs and Violence: Causes, Correlates, and Consequences. NIDA Research Monograph 103, US Department of Health and Public Services, Rockville, 293p.
Deming, David (2011). Better Schools, Less Crime? Quarterly Journal of Economics, forthcoming.
Dowdney, Luke (2003). Crianças do Tráfico: Um Estudo de Caso de Crianças em Violência Armada Organizada no Rio De Janeiro. Editora Sete Letras, Rio de Janeiro, 211p.
Efron, B. (1988). Logistic regression, survival analysis, and the Kaplan–Meier curve. Journal of the American Statistical Association 83, 414–425.
Gottfredson, Denise C. and David A. Soulé (2005). The Timing of Property Crime, Violent Crime, and Substance Use among Juveniles. Journal of Research in Crime and Delinquency, 42(110), 110-120.
Iniciardi, James A. (1990). The Crack-Violence Connection within a Population of Hard-Core Adolescent Offenders. In: Mario De La Rosa, Elizabeth Y. Lambert, and Bernard Gropper. Drugs and Violence: Causes, Correlates, and Consequences. NIDA Research Monograph 103, US Department of Health and Public Services, Rockville, 160-176.
Jacob, Brian A. and Lars Lefgren (2003). Are Idle Hands the Devil’s Workshop? Incapacitation, Concentration, and Juvenile Crime. American Economic Review, 93(5), 1560-1577.
Lancaster, Tony and Guido Imbens (1996). Case-control Studies with Contaminated Controls. Journal of Econometrics, v71, n1-2, 145-16.
Lee, David S. amd Justin McCrary (2005). “Crime, Punishment, and Myopia.” NBER Working Paper 11491.
23
Leland, Anne and Mari-Jana Oboroceanu (2010). “American War and Military Operations Casualties: Lists and Statistics.” Congressional Research Service Report for Congress 7-5700, RL32492. Washington, DC.
Leeson, Peter T. (2007). An-arrg-chy: The Law and Economics of Pirate Organization. Journal of Political Economy, v115, n6, 1049-1094.
Levitt, Steven D. and Lance Lochner (2001). The Determinants of Juvenile Crime. In: Jonathan Gruber (ed). Risky Behavior among Youths: An Economic Analysis. University of Chicago Press, Chicago, 327-373.
Levitt, Steven D. and Sudhir Alladi Venkatesh (2000). An Economic Analysis of a Drug-Selling Gang's Finances. Quarterly Journal of Economics, v115, n3, 755-789.
Levitt, Steven D. and Sudhir Alladi Venkatesh (2001). Growing Up in the Projects: The Economic Lives of a Cohort of Men Who Came of Age in Chicago Public Housing. American Economic Review Papers and Proceedings, v91, n2, 79-84.
Lochner, Lance and Enrico Moretti (2004). The Effect of Education on Crime: Evidence from Prison Inmates, Arrests, and Self-Reports. American Economic Review, 94(1), 155-189.
Lochner, Lance (2010). “Education Policy and Crime.” NBER Working Paper 15894.
Luallen, Jeremy (2006). School’s out. . . forever: A study of juvenile crime, at-risk youths and teacher strikes. Journal of Urban Economics, 59(1), 75–103.
Machin, Stephen, Olivier Marie, and Sunčica Vujić (2010). “The Crime Reducing Effect of Education.” IZA Discussion Paper n5000.
Moore, Joan (1990). Gangs, Drugs, and Violence. In: Mario De La Rosa, Elizabeth Y. Lambert, and Bernard Gropper. Drugs and Violence: Causes, Correlates, and Consequences. NIDA Research Monograph 103, US Department of Health and Public Services, Rockville, 160-176.
NEPAD & CLAVES (2000). “Estudo Global Sobre O Mercado Ilegal de Drogas no Rio de Janeiro.” Unpublished Manuscript, UERJ / FIOCRUZ.
Neto, Octávio Cruz, Marcelo Rasga Moreira and Luiz Fernando Mazzei Sucena (2001). Nem Soldados Nem Inocentes. Rio de Janeiro, Editora Fiocruz, 200p.
Observatório de Favelas (2006). “Caminhada de Crianças, Adolescentes e Jovens na Rede do Tráfico de Drogas no Varejo do Rio de Janeiro, 2004-2006 – Sumário Executivo.” Unpublished Manuscript, Observatório de Favelas.
Paula, Liana and Renato Lima (2008). Violência e Juventude: O Sistema Brasileiro de Atendimento Socioeducativo. Segurança Pública, n4, 71-82.
Rubio, Mauricio (2007). De la Pandilla a la Mara: Pobreza, Educación, Mujeres y Violencia Juvenil. Publicaciones Universidad Externado de Colombia, 579p.
Silva, Jailson de Souza and André Urani (2002). Brazil, Children in Drug-trafficking: A Rapid Assessment. International Labour Organisation (IPEC), Geneva, 108p.
Snyder, Howard N. and Melissa Sickmund (1999). Juvenile Offenders and Victims: 1999 National Report. US Department of Justice Programs, Office of Juvenile Justice and Delinquency Prevention, Washington DC.
Soares, Rodrigo R. (2006). The Welfare Cost of Violence across Countries. Journal of Health Economics, 25(5), September 2006, 821-846.
24
UNODC (2012). World Drug Report 2012. United Nations Office on Drugs and Crime, Vienna.
All Watchers and Transporters
Street-sellers and Wrappers
Soldiers Managers
(100%) (20.2%) (42.6%) (24.8%) (12.4%)Monthly Earnings $308 $245 $299 $292 $467Work Hours (per work day) 10.8 10.8 10.9 10.5 11.4Number of Days Off (per week) 0.6 0.4 0.5 0.8 0.5Works Armed?No 10% 23% 5% 6% 4%Only We Needed 18% 18% 22% 17% 11%Sporadically 26% 30% 34% 17% 7%Everyday 47% 30% 39% 61% 78%Apprehension and Police ExtorsionEver in Juvenile Detention? 29% 23% 29% 28% 41%Ever Arrested? 53% 45% 49% 54% 70%Ever Victim of Police Extorsion? 54% 30% 59% 52% 81%History of ViolenceEver Victim of Police Violence? 73% 61% 78% 70% 89%Ever Gun Fight with Rival Gang? 53% 23% 45% 70% 96%Ever Gun Fight with Police? 68% 36% 63% 85% 100%Ever Perpetrator of Physical Punishment? 34% 23% 32% 45% 48%Ever Comitted Homicide? 18% 2% 19% 27% 27%Ever Wounded in Combat? 24% 7% 14% 37% 56%Ever Victim of Physical Punishment? 22% 27% 27% 20% 11%Killed within 2 Years 20% 18% 18% 20% 30%
Observations 225 44 93 54 27
Table 1: Characteristics of Drug Trafficking Jobs
Occupation in Drug Traffic Hierarchy
Note: The table reports summary statistics for the 225 young men interviewed in the Observatorio de Favelas survey. The first column shows statistics for the entire sample, while the last four columns report statistics by occupation. Following Dowdney (2003), we classify respondents into four occupation categories: (1) "Watchers" who observe the area to warn in case of police presence and "Transporters" who move small quantitities of drugs within a favela; (2) "Street-sellers" who sell directly to consumers and "Wrappers" who handle and wrap the durg before sale to consumers; (3) "Soldiers" who are responsible for local security and who play a major role in armed confrotations; and (4) "Managers" who may be responsible for the entire operation in a favela or be in charge of the market for one specific drug. There were 7 respondents who we did not categorize into any of the four occupation categories because their job assignments did not fall into any of the specified categories.
OF Survey Brazilian CensusMen 10-25 Living
Men in Rio's FavelasAge 16.7 17.5RaceWhite 29% 38%Black 27% 14%Mixed-Race 37% 46%Other 7% 0%ReligionNo Religion 43% 24%Catholic 39% 54%Evangelical 16% 19%Other 2% 2%Marital StatusSingle 91% 96%Married 9% 4%Illiterate 9% 4%Currently Attends School? 10% 54%Years of Schooling 5.4 5.5Employment StatusUnemployed - 12%Employed - 40%Not in Labor Force - 48%Monthly Earnings 316 256Less than minimum wage 8% 21%Between 1 and 3 times the Minimum Wage 68% 67%Between 3 and 5 times the Minimum Wage 19% 9%More than 5 times the minimum wage 4% 2%Family Owns House 73% 83%
Observations 225 17,175
Table 2: Characteristics of Gang-Members and Young Men Living in Favelas
Note: The table reports means, separately for young men employed in the drug traffic business interviewed in the Observatorio de Favelas survey (column 1) and young men ages 10-25 living in households located in favelas in the city of Rio that were interviewed in the 2000 Brazilian Census (column 2).
Black 0.44 0.42 0.43 0.33 0.54 0.51 0.51 0.38 0.63 0.59 0.58 0.43[0.096]*** [0.083]*** [0.081]*** [0.070]*** [0.096]*** [0.102]*** [0.099]*** [0.087]*** [0.129]*** [0.120]*** [0.116]*** [0.102]***
Age -0.02 -0.02 -0.03 -0.03 -0.03 -0.03 -0.03 -0.03 -0.03 -0.03 -0.04 -0.04[0.004]*** [0.003]*** [0.004]*** [0.004]*** [0.005]*** [0.004]*** [0.005]*** [0.005]*** [0.006]*** [0.005]*** [0.006]*** [0.006]***
Illiterate 0.46 0.41 0.39 0.56 0.50 0.46 0.68 0.57 0.51[0.140]*** [0.132]*** [0.140]*** [0.176]*** [0.164]*** [0.174]*** [0.214]*** [0.194]*** [0.206]**
Owns House -0.29 -0.26 -0.34 -0.29 -0.38 -0.32[0.067]*** [0.071]*** [0.082]*** [0.086]*** [0.096]*** [0.101]***
No Religion 0.38 0.45 0.50[0.064]*** [0.078]*** [0.090]***
h 0.010 0.013 0.013 0.013 0.010 0.013 0.013 0.013 0.011 0.013 0.013 0.013[0.001]*** [0.001]*** [0.001]*** [0.001]*** [0.0011]*** [0.001]*** [0.001]*** [0.001]*** [0.001]*** [0.001]*** [0.001]*** [0.001]***
Constant -1.73 -1.76 -1.53 -1.64 -1.39 -1.42 -1.14 -1.28 -1.16 -1.19 -0.88 -1.03[0.025]*** [0.023]*** [0.057]*** [0.069]*** [0.024]*** [0.027]*** [0.071]*** [0.084]*** [0.030]*** [0.030]*** [0.084]*** [0.096]***
Black 0.06 0.05 0.05 0.04 0.11 0.11 0.11 0.07 0.17 0.16 0.16 0.11[0.014]*** [0.012]*** [0.012]*** [0.010]*** [0.024]*** [0.025]*** [0.024]*** [0.020]*** [0.041]*** [0.037]*** [0.036]*** [0.030]***
Age -0.002 -0.002 -0.003 -0.003 -0.004 -0.005 -0.006 -0.005 -0.007 -0.007 -0.009 -0.009[0.0004]*** [0.0003]*** [0.0003]*** [0.0003]*** [0.0008]*** [0.0007]*** [0.0007]*** [0.0007]*** [0.0013]*** [0.0011]*** [0.0012]*** [0.0012]***
Illiterate 0.06 0.05 0.05 0.13 0.11 0.10 0.20 0.16 0.14[0.025]** [0.022]** [0.022]** [0.050]** [0.044]** [0.045]** [0.076]*** [0.066]** [0.067]**
Owns House -0.03 -0.03 -0.06 -0.05 -0.09 -0.08[0.009]*** [0.009]*** [0.017]*** [0.018]*** [0.027]*** [0.027]***
No Religion 0.04 0.08 0.13[0.008]*** [0.016]*** [0.025]***
Observations 17,399 17,399 17,399 17,399 17,399 17,399 17,399 17,399 17,399 17,399 17,399 17,399Note: The table reports results of the Maximum Likelihood estimation of participation in the illegal sector using the estimator proposed by Lancaster and Imbens (1996). Panel A reports the results from the MLE estimation. Panel B reports the marginal effects. The three different set of results correspond to different assumptions about which fraction of young men 10-25 living in favelas were employed by the drug business (q). The sample is formed by young men 10-25 living in favelas of Rio surveyed by the 2000 Brazilian Census and 225 young men employed by the drug traffic business interviewed in the OF survey. Robust standard errors in brackets.
Fraction young men drug-trafficking = 5% Fraction young men drug-trafficking = 15%
Dependent Variable: Employed by Drug Traffic Business
Fraction young men drug-trafficking = 10%
Table 3: Participation in the Illegal Sector
Panel A: Participation Equation (GMM)Fraction young men drug-trafficking = 5% Fraction young men drug-trafficking = 15%
Panel B: Marginal Effects
Fraction young men drug-trafficking = 10%
Black 0.03 0.03 0.03 0.01 0.10 0.09[0.092] [0.092] [0.092] [0.109] [0.115] [0.124]
Age 0.10 0.10 0.10 0.10 0.09 0.09[0.020]*** [0.019]*** [0.019]*** [0.021]*** [0.021]*** [0.022]***
Illiterate 0.25 0.26 0.32 0.31[0.109]** [0.113]** [0.138]** [0.140]**
Owns House in which Lives 0.05 0.01[0.092] [0.109]
Mills Ratio -0.05 0.18 0.16[0.174] [0.212] [0.251]
Constant 5.63 5.61 5.57 5.75 5.23 5.25[0.045]*** [0.047]*** [0.085]*** [0.362]*** [0.450]*** [0.493]***
Observations 221 221 221 221 221 221
Black 0.01 0.10 0.09 0.01 0.10 0.09[0.109] [0.115] [0.123] [0.109] [0.115] [0.123]
Age 0.10 0.09 0.09 0.10 0.09 0.09[0.021]*** [0.021]*** [0.023]*** [0.021]*** [0.022]*** [0.023]***
Illiterate 0.32 0.32 0.32 0.32[0.138]** [0.140]** [0.137]** [0.139]**
Owns House in which Lives 0.01 0.01[0.108] [0.108]
No Religion
Mills Ratio -0.05 0.16 0.15 -0.04 0.16 0.14[0.157] [0.192] [0.225] [0.150] [0.179] [0.209]
Constant 5.72 5.31 5.33 5.70 5.35 5.36[0.279]*** [0.350]*** [0.373]*** [0.233]*** [0.292]*** [0.306]***
Observations 221 221 221 221 221 221Note : The table reports results from OLS regressions in which the dependent variable is log of drug-trafficking wages. The first set of results do not control for selection. All the other results control for the Inverse Mills' Ratio implied by the most complete specification in Table 3. The Mills' Ratio depends on the assumption about which fraction of young men were employed by the drug traffic business (q), which we assume to be either 5%, 10% or 15%.
Table 4: Wages in the Illegal Sector
Fraction young men drug-trafficking = 5%
Fraction young men drug-trafficking = 15%
No Correction for Self-Selection
Dependent Variable: Log Wages (OLS)
Fraction young men drug-trafficking = 10%
(1) (2) (3) (4) (5) (6) (7)Black -0.02 -0.02 -0.04 -0.02 -0.05 -0.04 -0.07
[0.096] [0.095] [0.092] [0.096] [0.091] [0.092] [0.090]Experience 0.10 0.10 0.08 0.10 0.09 0.05 0.06
[0.027]*** [0.027]*** [0.026]*** [0.027]*** [0.025]*** [0.023]** [0.023]**Age at Entry 0.09 0.09 0.09 0.10 0.09 0.07 0.07
[0.028]*** [0.028]*** [0.027]*** [0.028]*** [0.026]*** [0.025]** [0.024]***Education 0.00 0.00 0.00 -0.01 -0.01 -0.01 -0.01
[0.020] [0.020] [0.019] [0.020] [0.019] [0.019] [0.018]Single Mother 0.17 0.15 0.15 0.17 0.13 0.12 0.11
[0.086]* [0.086]* [0.082]* [0.085]** [0.082] [0.082] [0.080]Owns House in which Lives 0.03 0.03 0.04 0.02 0.05 -0.03 -0.02
[0.095] [0.095] [0.092] [0.095] [0.092] [0.089] [0.086]Number of Siblings -0.01 -0.01 -0.02 -0.01 -0.02 -0.01 -0.02
[0.025] [0.025] [0.025] [0.025] [0.024] [0.023] [0.023]Physical Disability -0.23 -0.37 -0.22
[0.181] [0.183]** [0.142]# Gun Fights 0.05 0.05 0.04
[0.011]*** [0.010]*** [0.011]***Ever Victim of Physical Punishment? -0.18 -0.18 -0.17
[0.099]* [0.094]* [0.093]*OccupationMid 0.21 0.13
[0.112]* [0.114]Soldier 0.24 0.07
[0.111]** [0.116]Manager 0.63 0.35
[0.126]*** [0.149]**Constant 3.96 3.96 3.94 3.99 3.95 4.35 4.29
[0.420]*** [0.417]*** [0.398]*** [0.420]*** [0.386]*** [0.380]*** [0.360]***
Observations 220 220 220 219 219 212 211
Dependent Variable: Log Wages (OLS)
Table 5: Wages in the Illegal Sector
Note : The table reports results from OLS regressions in which the dependent variable is the log of drug-trafficking wages. Robust standard errors between brackets.
(1) (2) (3) (4) (5) (6)Single Mother 0.32 0.30
[0.249] [0.265]Owns House in which Lives 0.17 0.01
[0.317] [0.316]Number of Siblings -0.13 -0.05
[0.086] [0.096]Illiterate -1.46 -1.26
[0.604]** [0.615]**Started Using Drugs Before 13 -1.09 -1.11
[0.305]** [0.303]**Domestic Violence -0.04 0.04
[0.314] [0.307]Unruly -0.01 0.09
[0.300] [0.312]Constant 14.94 14.99 15.12 14.88 14.83 15.32
[0.363]*** [0.121]*** [0.135]*** [0.160]*** [0.145]*** [0.400]***
Observations 224 223 224 204 224 203
Dependent Variable: Age at Entry (OLS)
Table 6: Age at Entry
Note : The table reports results from OLS regressions in which the dependent variable is age at which individual started working for the drug-traffic business. Age at entry was reported in brackets. We converted the answers into scalars by choosing the mid-point in each bracket, and 19 for the upper bracket (above 18 years old). Robust standard errors are reported between brackets.
Entry Mid Soldier Manager Entry Mid Soldier Manager(1) (2) (3) (4) (5) (6) (7) (8)
Black 0.00 0.05 -0.04 0.00 0.01 0.09 -0.08 -0.02[0.038] [0.080] [0.071] [0.055] [0.028] [0.084] [0.080] [0.019]
Experience -0.03 0.00 -0.01 0.04 -0.01 0.02 -0.02 0.01[0.012]** [0.027] [0.029] [0.015]** [0.009] [0.030] [0.031] [0.005]**
Age at Entry -0.02 -0.02 0.01 0.03 -0.01 -0.02 0.01 0.02[0.012] [0.026] [0.025] [0.016]** [0.009] [0.027] [0.027] [0.008]**
Education -0.01 0.01 0.00 0.00 -0.01 0.01 0.00 -0.01[0.009] [0.019] [0.018] [0.013] [0.007] [0.020] [0.019] [0.005]
Single Mother -0.06 0.03 0.00 0.02 -0.04 0.03 0.00 0.01[0.034]* [0.074] [0.066] [0.051] [0.025] [0.079] [0.074] [0.021]
Owns House in which Lives 0.01 -0.05 0.03 0.02 0.01 -0.06 0.04 0.01[0.036] [0.081] [0.075] [0.052] [0.027] [0.087] [0.085] [0.018]
Number of Siblings -0.01 0.00 0.01 -0.01 0.00 0.00 0.00 0.00[0.011] [0.024] [0.022] [0.016] [0.008] [0.026] [0.025] [0.005]
Physical Disability -0.20 0.06 -0.08 0.21 -0.13 0.15 -0.10 0.08[0.030]*** [0.162] [0.124] [0.142] [0.036]*** [0.150] [0.138] [0.055]
# Gun Fights -0.02 -0.03 0.03 0.02[0.004]***[0.012]**[0.011]***[0.006]***
Ever Victim of Physical Punishment? 0.03 0.08 -0.08 -0.03[0.032] [0.087] [0.082] [0.020]
Observations 213 213 213 213 212 212 212 212Note : Robust standard errors between brackets. The table presents marginal effects from a multinomial logit, where the dependent variable is occupation in the drug traffic hierarchy. The occupation categories are: (1) entry-level (watchers and transporters); (2) mid-level (street-sellers and wrappers); (3) soldiers; and (4) managers.
Table 7. Occupation in the Drug Traffic Hierarchy
Dependent Variable: Occupation (Multinomial Logit - Marginal Effects)
(1) (2) (3) (4) (5)Black 0.07 0.05 0.06 0.07 0.05
[0.038]* [0.034] [0.037]* [0.038]* [0.034]Experience -0.02 -0.02 -0.02 -0.02 -0.02
[0.008]** [0.008]** [0.008]*** [0.009]** [0.008]**Education -0.01 -0.01 -0.01 -0.01 -0.01
[0.006] [0.006] [0.006] [0.006] [0.006]Single Mother -0.02 -0.04 -0.02 -0.02 -0.03
[0.029] [0.026] [0.028] [0.030] [0.026]Owns House in which Lives 0.03 0.03 0.02 0.03 0.03
[0.032] [0.029] [0.033] [0.033] [0.029]Number of Siblings 0.01 0.00 0.01 0.01 0.00
[0.008] [0.008] [0.008] [0.009] [0.008]Have Had Other Job 0.04 0.04
[0.030] [0.029]Have (Temporarily) Quit Before 0.10 0.09
[0.036]*** [0.035]***# Gun Fights 0.01 0.01
[0.004]* [0.003]*Unruly 0.04 0.04
[0.045] [0.044]
Observations 349 348 349 349 348
Dependent Variable: Quit (Logistic Hazard Model - Marginal Effects)
Table 8: Quit
Note : Robust standard errors between brackets. The table presents marginal effects from a logistic hazard model. The unit of observation is individual i in wave t if he was known to be at risk (so individuals who were not interviewed in wave t – either because they were in prison, had been murdered, or could not be located – are not part of the sample). We assume that exiting the gang is an absorbing state, such that an individual who quit in wave t is not at risk in subsequent waves. The dependent variable, Y_i,j is 1 if individual i exited the drug-gang in wave t and 0 if individual i was still drug-trafficking in wave t. We run logit regressions, including dummies for each wave.
(1) (2) (3) (4) (5)Black -0.03 -0.03 -0.04 -0.03 -0.04
[0.059] [0.059] [0.060] [0.058] [0.058]Experience 0.01 0.02 0.01 0.02 0.02
[0.017] [0.017] [0.017] [0.017] [0.017]Education -0.02 -0.02 -0.02 -0.01 -0.02
[0.013] [0.013] [0.012] [0.013] [0.012]Single Mother 0.09 0.09 0.08 0.10 0.09
[0.058] [0.058] [0.058] [0.057]* [0.057]Owns House in which Lives 0.00 0.00 0.01 0.00 0.00
[0.064] [0.066] [0.064] [0.065] [0.066]Number of Siblings 0.01 0.02 0.01 0.01 0.02
[0.018] [0.019] [0.018] [0.017] [0.018]Have Had Other Job 0.02 0.01
[0.056] [0.054]Have (Temporarily) Quit Before -0.08 -0.09
[0.058] [0.058]# Gun Fights 0.02 0.02
[0.008]** [0.008]**Unruly 0.16 0.17
[0.075]** [0.073]**Constant 0.19 0.20 0.15 0.14 0.11
[0.091]** [0.096]** [0.091]* [0.096] [0.100]
Observations 224 223 224 224 223
Table 9: Killed Within 2 Years
Dependent Variable: Killed (OLS)
Note : Robust standard errors between brackets. The table presents results from an OLS regression in which the dependent variable is 1 if gang-member was dead two years after baseline interview and 0 otherwise.
Father's Education 5.5 5.5 Education of Men Ages 25-65Mother's Education 4.8 5.6 Education of Women Ages 25-60
Parental Income $310 $392 Income Head + SpouseFraction Raised by Single Mother 37% 22% Fraction of Single-Female-Headed Families
Number of Siblings 2.7 2.3 # of Alive Children of Women Ages 25-60
Father's Education Missing 58%Mother's Education Missing 36%
Parental Income Missing 29%
Observations 225 24,396 men / 25,484 women / 26,514 families
Appendix Table A1: Family Background
2000 Brazilian Census
Note: The table reports means of family background characteristics for the OF sample and the average education and income of men ages 25-65 and women ages 25-60 living in favelas in the city of Rio who were surveyed in the 2000 Brazilian Census. The women and men in these age groups were old enough to be parents of our OF sample who were between 11-24 years old. The table also reports the income and the fraction of single-female-headed families among families in which either the head or the head's spouse was a man ages 25-65 or a woman ages 25-60.
OF Sample Residents of Favelas in City of Rio
Top Related