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: gkhanna@umich.edu and lvzimmer@umich.edu. 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:smriti@econdse.org.   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 a.benyishay@unsw.edu.au, Pearlman, Vassar College, sapearlman@vassar.edu

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