Anger and Support for Punitive Justicein Mexico’s Drug War⇤

Omar García-Ponce† Lauren Young‡ Thomas Zeitzoff§

May 8, 2018

Abstract

Why do civilians affected by violence support vigilante groups? We argue that the anger raised in thewake of violence increases the demand for punitiveness, even at the expense of the rule of law. We test theimplications of this view using three observational and experimental studies with data from an originalsurvey of nearly 1,200 individuals in Western Mexico, where many civilians have been exposed to narco-violence, and vigilante justice. We have three principal findings. First, our observational analysis showsthat individuals who have historically been exposed to violence tend to be angrier and more supportiveof punitive criminal justice policies, as well as policies that enable vigilante groups to punish criminals.Second, both experiments show that citizens that are induced to feel more anger, are more supportive ofharsh punishments, and place less value on the legality of a punishment for morally outrageous crimes.Third, the experimental results suggest that the innocence of a crime’s victim has a stronger effect on angerand moral outrage than the severity of the violence. These results shed light on how the emotional andcognitive reactions of civilians in violence-affected communities may lead to cycles of retributive violencethat undermine the rule of law.

⇤Our deepest thanks to Daniel Hirschel-Burns, Stathis Kalyvas, Beatriz Magaloni, David Shirk, and seminar partici-pants at Essex, USC, UC Davis, MIT, Uppsala, NYU, Yale, and UCSD, for feedback at various stages of this project. Wealso thank Buendía & Laredo for managing the data collection. Isabel Mejía Fontanot and Julio Solís Arce providedexcellent research assistance.

†Assistant Professor, University of California, Davis. ogarciaponce@ucdavis.edu‡Assistant Professor, University of California, Davis. leyou@ucdavis.edu§Assistant Professor, American University. zeitzoff@gmail.com

2

1 Introduction

In 2015, Washington D.C. experienced a sharp 54% increase in homicides.1 In response, the D.C.

Council endorsed a proposal to pay young offenders most at risk of committing firearm crimes

up to $1,000 per month to ‘stay out of trouble.’2 This model, based loosely on similar programs in

Richmond (California), Chicago, and Boston, is credited with causing sharp drops in the rates of

homicide victimization and perpetration among its participants (Davis, May 5, 2016).3 For all its

purported success, the idea of paying violent offenders to not commit crimes does have its share of

detractors. As one victims advocate in Richmond and critic of the program argued,“If I were to find

out that the guy who murdered my twin sons was getting a thousand dollars for a promise (not to

commit crimes)? I mean, how can you trust (them)?”4

It can be hard to garner support for seemingly effective, but less punitive justice policies.

Conversely, harsh justice policies remain quite popular. The Philippines’ President Rodrigo Duterte

won election by a wide margin on a platform pledging to end crime in six months by ignoring

human rights conventions and killing tens of thousands of criminals. Human rights groups have

condemned Duterte for running vigilante death squads that have killed more than 7,000 people as

of March 2017 (Macaraig, May 10, 2016).5 However, political analysts have argued that Duterte’s

popularity actually directly stems from his support for vigilantism stretching back to his alleged

support for death squads when he was mayor of the city of Davao (Kim, July 17, 2015).6

In Central America there is a history of such “Mano Dura” (literally “iron fist”) policies to deal

with violence perpetrated by criminal gangs (Hume, 2007). These policies included the state taking

harsh and punitive measures against suspected gang members, stiffer prison sentences, and tacit

1Based on data from the FBI’s Uniform Crime Reporting (UCR) Program, Washington D.C. experienced 162 homicidesin 2015 (around 24 per 100,000 people), up from 105 (16 per 100,000 people) the year before.

2Most of those identified as ‘at-risk’ had committed crimes in the past with firearms. In addition to not committingfurther crimes, offenders would have to regularly attend behavioral health, education, and as well as job skills training toreceiving payments.

3The D.C. program was ultimately rejected.4See Cowan, Claudia. 2016, August 24. “One California city is paying people not to commit crimes.” Fox News,

Retrieved from: https://goo.gl/7wWny35See Soong, Martin. 2017, March 28. “As Rodrigo Duterte turns 72, a look at his controversial presidency so far.”

CNBC, Retrieved from: https://goo.gl/3ncyKN6See Marshall, Andrew and Manuel Mogato. 2016, May 24. “Philippine death squads very much in business as

Duterte set for presidency.” Reuters, Retrieved from: https://goo.gl/4qZ9Jd

3

and sometimes more overt support for human rights violations against suspected criminals. While

enjoying widespread support (Holland, 2013), many of these Mano Dura policies have had at best

mixed effectiveness,7 and have been linked to state-sanctioned extrajudicial killings (Wolf, 2017).

Why are non-punitive policies to reduce violent crime like the stipend program proposed

in D.C. often unpopular? Why do severe punishments of questionable effectiveness, such as the

extrajudicial killings championed by Duterte, and the Mano Dura policies in Central America, elicit

such strong support? We argue that the popularity of harsh policies is not driven by mispercep-

tions of their effectiveness. When thinking about criminal justice regimes citizens do care about

effectiveness. Yet they also have preferences for punishing criminals in a way that corresponds

to the crimes they have committed. In many cases this preference for punishment outweighs the

demand for other principles that people generally care about in their criminal justice policy, such

as legality and even effectiveness in preventing future violence. In this paper, we argue that how

people make trade-offs between justice and effectiveness changes when they are exposed to heinous

acts of violence that induce anger.

To test whether exposure to violence increases support for punitive and vigilante punishment

of criminals, we draw on cognitive appraisal theory from psychology to theorize that some forms of

violence cause people to feel outraged, which in turn increases perceptions of blame and preferences

for punitiveness in criminal justice policy. We use three separate observational and experimental

tests to elicit these preferences, and to test whether the emotion of anger plays an important role in

the process.

First, we examine whether people who are exposed to higher levels of violence report feeling

anger more frequently, assign higher overall levels of blame to actors involved in the Mexican

drug conflict, and report preferences for more punitive policies than those who are not (Study 1).

Second, we use a survey experiment designed to generate moral outrage by violating community

norms to test whether extreme violence 1) induces higher levels of anger, and 2) causes people

to prefer extrajudicial and harsh punishments (Study 2). Finally, we use a second set of 125

7Many have argued that the policy which has swelled prisons has perversely increased the prison population. SeeDudley, Steven. 2010, November 22. “How ’Mano Dura’ is Strengthening Gangs.” Insight Crime, Retrieved from:http://www.insightcrime.org/investigations/how-mano-dura-is-strengthening-gangs

4

randomly generated scenarios to elicit preferences for criminal justice policy across a wide range of

perpetrators, victims, and types of violence. We again test whether across this broad spectrum of

common types of violence, more severe violence against more innocent victims 1) induces higher

levels of anger, and 2) causes people to prefer extrajudicial and harsh punishments (Study 3).

The combination of these three research designs—all of which were preregistered in advance

of our analysis—enables us to draw conclusions that are based on highly realistic variation, gen-

eralizable to a large population of interest, and causal.8 The first study’s observational design

has the advantage of looking at real variation in the independent variable of interest (exposure to

violence) in a representative sample of Western Mexican citizens, many of whom have been exposed

to violence by narcos (drug traffickers), vigilante groups known as autodefensas, and even govern-

mental security forces.9 Although the estimates from the observational study may not be causal,

the research design enables us to examine whether there is any relationship between emotions

and exposure to violence. The second study, a survey experiment, enables us to test for specific

causal mechanisms and to estimate the effects of anger exposure to outrageous violence through

hypothetical scenarios. Finally, in the third study, we use a factorial experimental design that

provides both causal estimates and an assessment of the generalizability of the experimental results.

Our multi-method research design combines many of the advantages of large-scale observational

research with the precision and causal identification benefits of experiments.

Understanding how exposure to violence and emotions influence attitudes towards justice and

security policy are fundamental questions in political science and psychology. Anger is considered

a core emotion that prepares individuals to take risks and correct perceived wrongs (Frijda, Kuipers

and Ter Schure, 1989). Past research finds that anger makes individuals less risk-averse (Lerner

and Keltner, 2001), more likely to participate in politics (Valentino et al., 2011), and more likely to

support an aggressive foreign policy (Lerner et al., 2003; Skitka et al., 2006). Anger at a perceived

moral transgression may lead to moral outrage or a desire to punish the transgressor (Bastian,

8See EGAP ID #20170504AB for more information.9See Woody, Christopher. 2017, May 23. ”Deadly violence continues to climb in Mexico, where an ascendant

cartel is strengthening its grip on power.” Business Insider, Retrieved from: https://goo.gl/PS4xq1 and Ahmed,Azam. 2017, August 4. “Mexico’s Deadliest Town. Mexico’s Deadliest Year.“ The New York Times, Retrieved from:https://goo.gl/TUdNnQ

5

Denson and Haslam, 2013), and exposure to violence is likely to exacerbate this effect (Zeitzoff,

2014). Additionally, studies have found that anger may be a key mechanism explaining why people

take individually costly actions to enforce norms (Fehr and Gächter, 2002).

While most past studies of anger and costly punishment have looked at non-violent forms

of punishment for social infractions, our study considers a situation in which anger might lead to

a less adaptive form of social sanctioning through punitive and vigilante violence. Our findings

ultimately suggest that civilian anger over certain types of crimes may be an important factor

explaining how direct and indirect victimization can contribute to the escalation of violence and

breakdown of the rule of law. Anger over insecurity, corruption,10 and perceived injustices are

an important factor shaping decisions to support violence against the state and other perceived

transgressors (Brown, Abernethy, Gorsuch and Dueck, 2010; Lotz, Baumert, Schlösser, Gresser and

Fetchenhauer, 2011; Chayes, 2015; Tankebe and Asif, 2016). Our study provides evidence that an

emotional mechanism to explain why many individuals living in contexts with weak institutions

and violence favor extrajudicial, violent punishments.

We use models from psychology and political science on how emotions and exposure to

violence influence decisions to understand when anger is likely to make people support punitive,

and potentially ineffective policies. The insights developed here can inform our understanding

of how emotions affect public opinion not only related to crime, but also foreign policy and

redistribution. Nevertheless, we believe that security policy is a particularly important case to

study for two reasons. First, it is an area of policy where emotions are likely to influence opinion

as violence induces strong emotions, and the risks and benefits of different policy options are

difficult to quantify. Second, security policy is an important issue for many citizens–particularly in

conflict-prone nations–and thus often influences how they vote.

The remainder of the paper is organized as follows: Section 2 provides a literature review

and our theoretical expectations; Section 3 discusses the context of violence and the emergence

of vigilante groups in Mexico’s drug war; Section 4 explains the research design; Sections 5, 6,

and 7 present the results from the three studies; and Section 8 concludes. Additional results and

10See Zechmeister and Zizumbo-Colunga (2013) for a discussion of how corruption influences candidate evaluations.

6

robustness checks can be found in the Appendix.

2 Anger, Exposure to Violence, and Attitudes Towards Justice

How do people form preferences about security and justice policy? Previous research argues

that individuals care both about the ability of security policies to prevent future threats, and

meting out punishments that are appropriately harsh given the crime committed (Carlsmith and

Darley, 2008). Prevention and retribution are related to broader discussions of moralist versus

consequentialist reasoning. Emotions influence how much weight individuals give to these factors

when determining appropriate punishments, as well as how individuals process information. Both

incidental emotions–i.e., those not directly related to the topic at hand, as well as emotions that are

stimulated by the underlying choice affect how this decision is ultimately made.

Early political theorists suggested two rationales for determining punishment: retribution

and prevention (Vidmar and Miller, 1980; Darley, Carlsmith and Robinson, 2000). Retribution is

retrospective, focusing on the perpetrator’s “just deserts” to argue that the punishment should be

proportional to the severity of the crime or how morally outrageous it is (Kant, 1952). If punishments

are determined according to this principle, the severity of the harm and the existence of extenuating

circumstances that mitigate or exacerbate the moral outrage should be strongly related to the

severity of the punishment (Darley, Carlsmith and Robinson, 2000). In contrast, utilitarian legal

scholars argue that “general prevention ought to be the chief end of punishment, as it is its real

justification” (Bentham 1962, qtd. in Carlsmith, Darley and Robinson 2002). While in theory policies

could be both punitive and effective in preventing crime, in practice there is considerable evidence

that the many “get tough” policies are less effective than more “rehabilitative” policies (Farrington,

MacKenzie, Sherman and Welsh, 2003; Chen and Shapiro, 2007; Andrews and Bonta, 2010).

A separate strand of research suggests that citizens also care about equitable treatment in

the process of criminal justice policy (i.e., “procedural justice”) (Lind and Tyler, 1988). Procedural

justice advocates argue that while individuals care about the equality of resources or outcomes

(distributive justice), as well as punishing wrongs (retributive justice), what they really care about is

7

fairness and transparency in judicial outcomes (procedural justice) because punishments derive

their legitimacy from fair processes. This research suggests that people obey the law and prefer

legal punishments to crimes not because of the potential consequences but because they view legal

processes as legitimate (Tyler, 2006).

Yet several studies examining public opinion show that people have preferences for punitive

criminal justice policies that are independent of beliefs that these policies are effective or legal. In

the U.S., Enns (2014) analyzes questions from hundreds of public opinion surveys between 1950

and 2010 to show that attitudes towards punitive rather than preventative policies on crime steadily

rose from the 1960s to the 1990s. Throughout the time series, consistent majorities of Americans

supported the death penalty, reported that the courts were not harsh enough on criminals, and

supported more spending on the police (862). Roberts et al. (2002) coined the term “penal populism”

to capture the concept of criminal justice policies – specifically harshly punitive policies such as the

three strikes laws in the U.S. – that are chosen based on their popularity rather than because they

are effective in reducing crime. Furthermore, as many scholars and others have argued, preferences

in the U.S. over punitive justice policies, and crime more generally, are closely tied to attitudes on

race (Mendelberg, 1997; Gilliam Jr and Iyengar, 2000; Gilliam Jr, Valentino and Beckmann, 2002;

Peffley and Hurwitz, 2007; Mendelberg, 2008).

Political psychologists working in the U.S. have used vignettes to elicit citizens’ preferences

for punishment and test whether those preferences are more in line with a logic of retribution or

prevention. Most of the studies have found strong evidence in favor of a logic of retribution. Darley,

Carlsmith and Robinson (2000) and Carlsmith, Darley and Robinson (2002) present university

students with a series of scenarios that vary in how cases in which someone has perpetrated a

harm ask participants to assess the appropriate punishments. In both studies they find that the

student subjects are highly influenced by the magnitude of the harm, and are minimally affected by

influenced by factors that indicate the potential for the punishment to incapacitate repeat offenders,

or deter future harm. However this literature has rarely traveled outside of the U.S. to areas such

as Latin America, where crime, insecurity, low state capacity, and a lack of trust in the state,11 are

11 Bodea and LeBas (2016) shows that in Nigeria, employment of vigilante groups may erode trust in the state.

8

all the more severe (Arias, 2017). Previous research also has looked at only a handful of scenarios,

raising concerns that the findings might depend on the specific characteristics of the offender or the

crime. Our study extends this literature to the context of narco-trafficking in Mexico, and uses a

design that allows us to vary the kind of crimes, perpetrators, and punishments, to allow hundreds

of variations across individual scenarios (Hainmueller, Hopkins and Yamamoto, 2013).

Behavioral economists have also investigated how and why people choose to punish. They

find that individuals have a preference for fairness, and prefer to punish those who behave in ways

that are perceived as unfair, even if that punishment is personally costly. Camerer and Thaler (1995)

show, for example, that receivers in one-shot ultimatum games will reject very unequal offers, in

effect giving up money to punish the unfair proposer. Carpenter (2007) shows that the demand

for punishment in such situations is relatively inelastic to price or income. Some have argued that

anger rather than cognitive assessments of unfairness is one of the key mechanism that drives

these desires to punish proposers by rejecting the offer (Pillutla and Murnighan, 1996; Srivastava,

Espinoza and Fedorikhin, 2009).

Under what conditions is this preference for punitive policies particularly strong? There

is some evidence that emotions and moral outrage are linked to preferences for punitive justice

policies. Johnson (2009) shows that Americans who say they are angry about crime are more likely

to support punitive crime policies. In an early precursor to our design, Lerner, Goldberg and Tetlock

(1998) shows that anger causes people to make more punitive judgments in hypothetical crime

scenarios. The effect of anger is hypothesized to work by increasing attributions of blame and

simplifying cognitive processes, but is muted when people are accountable for their decisions.

Psychologists have theorized that emotions shape preferences towards punishment. At an

individual level, anger is an approach-oriented emotion that prepares individuals to take action

in order to rectify perceived wrongs or slights (Frijda, 1986; Carver and Harmon-Jones, 2009). Ap-

praisal tendency theory distinguishes anger from other negative emotions by appraisals of certainty,

control, and responsibility. Specifically, it arises from appraisals that someone is responsible for

a negative event, that the decision-maker has individual control, and that the decision-maker is

certain about what happened (Lerner and Keltner, 2000). In turn, experimentally induced anger

9

has been shown to affect a host of appraisals and behaviors that are thought to help the individual

arrive at his desired state, including increasing risk-taking and punitiveness (Lerner and Keltner,

2001; Bastian, Denson and Haslam, 2013).

Several social psychologists have argued that anger plays an integral role in explaining prefer-

ences for punishment because crime violates sacred values and produces moral outrage. Garland

(2012) builds on Durkheim and Swain (2008) to argue that: “(t)he criminal act violates sentiments

and emotions which are deeply ingrained in most members of society – it shocks their healthy

consciences – and this violation calls forth strong psychological reactions, even among those not

directly involved. It provokes a sense of outrage, anger, indignation, and a passionate desire for

vengeance” (30). Indeed, the desire for what are seen as just punishments may invoke the kind of

taboo trade-offs that themselves cause moral outrage (Tetlock, 2003; Ginges et al., 2007).

What leads to these shifts in anger and outrage that increases punitiveness? We argue that

exposure to violence is an important mechanism. Previous research has also shown that exposure

to violence shifts fundamental preferences. Individuals exposed to violence are more altruistic

(Gilligan, Pasquale and Samii, 2014), but this altruism only extends to their neighbors (Voors et al.,

2012), and also more likely to participate in political activities (Blattman, 2009; Bauer et al., 2016).

Individuals exposed to violence are also more risk-taking, but this effect is muted by recollection

of fear and trauma (Callen et al., 2014). Perhaps most relevant to the current research, Bateson

(2012) shows that individuals who are crime victims are more politically active, but also more

willing to support harsh, and punitive justice policies. From the intergroup conflict literature there

is also considerable evidence that exposure to violence drives attitudes that perpetuate conflict

(Canetti-Nisim et al., 2009; Getmansky and Zeitzoff, 2014). However, there is also evidence that

exposure to violence makes people more open to compromise and less likely to support violence

(Lyall, 2009; Hazlett, 2013; Beber, Roessler and Scacco, 2014; Zeitzoff, 2014). This divergence in the

effect of violence across studies could be driven by differences in the type of violence (intergroup

conflict vs. crime vs. terrorism), and whether violence is ongoing and the duration of the conflict or

post-conflict period.

Finally, it should be noted that we are agnostic whether this effect of violence has to be directly

10

experienced by the victim, or whether simply witnessing violence is enough to shift violence. There

is ample evidence that simply being exposed to the violence and impunity that is commonly found

in many developing countries will increase desires for offenders to get their “just desserts” (Bastian,

Denson and Haslam, 2013). Previous research echoes this, showing that a lack of police or state

to effective punish offenders, distrust of police fairness (Tankebe, 2009), along with a history of

extrajudicial violence all increase support for punitive punishment (Messner, Baumer and Rosenfeld,

2006).

To summarize, the existing literature leads us to make several predictions to the data:

• Prediction 1: Exposure to more severe violence increases anger.

• Prediction 2: Exposure to more severe violence increases support for harsher, more punitive punish-

ments.

• Prediction 3: Exposure to more severe violence increases support for extrajudicial punishments.

As discussed in depth in Section 4, we test these hypotheses using both variation in past

exposure to real violence that respondents report, and by asking respondents to evaluate randomly

assigned hypothetical scenarios that describe violent crimes. These predictions as well as the

research design were predigested with EGAP before any analysis.12

3 Violence and Vigilantism in Mexico

3.1 Mexico’s Drug War

Drug-related violence has become the largest threat to security in Mexico, affecting various regions

in the country for more than 10 years. Based on official data from INEGI (the National Institute

of Statistics and Geography), over 200,000 Mexicans have been killed since December 2006, when

former Mexican president Felipe Calderón unleashed a war against organized crime by sending the

12We also preregistered two hypotheses about individual characteristics that would moderate participants’ reactions tothe crime scenarios. Prediction 4: People with more positive attitudes towards vengeance will be even more supportive of harshand extrajudicial punishments for crimes with innocent victims or more severe violence. Prediction 5: People with more exposureto violence will be even more supportive of harsh and extrajudicial punishments for crimes with innocent victims or more severeviolence. In the interest of brevity, we do not present these results in this paper, but rather write them up in a separateanalysis of individual differences and reactions to violent crime.

11

army into the state of Michoacán.13 Mexico’s current President, Enrique Peña Nieto, has adopted a

similar strategy towards combatting drug trafficking organizations. The army and the federal police

remain deployed throughout the Mexican territory aiming at capturing or killing criminal bosses.

On May 24, 2017, Peña Nieto’s administration disclosed that 107 of 122 high-ranking members of

organized crime groups have been either captured or killed.14 However, to date, this kingpin strategy

has not translated into significantly lower levels of violence.15

As shown in Figure 1, despite a modest decline in the national murder rate at the beginning

of Peña Nieto’s administration (2013–2015), violence levels have been trending up again. In fact,

June 2017 has been the most violent month in 20 years, as measured by the number of intentional

homicides.16 It is difficult to estimate how many homicides are strictly related to the so-called drug

war, but leading newspapers and civic organizations in Mexico suggest that drug-related homicides

account for 40% to 50%.17 This includes violence between cartels and the state, intra-cartel violence,

and more recently the emergence of vigilante groups (commonly known as autodefensas).

Figure 2 shows the geographic distribution of homicide rates at the municipality level for

2016, based on data from INEGI. While much of the violence has been concentrated in the northern

part of the country, i.e., along the drug-trafficking routes into the U.S., there is substantial spatial

variation across the country. Violence has also intensified in areas that are far from the U.S.-Mexico

border. This is particularly true of Western Mexico, where drug production is concentrated (Dube,

García-Ponce and Thom, 2016). Michoacán and Guerrero, for example, have seen heightened

violence in recent years.18

Scholars and policymakers have pointed to institutionalized corruption and an ineffective

judicial system as key drivers of the violence,19 among several other factors. This consistent with

the theoretical argument that increasing enforcement and sanctions—in this case, militarizing the

country to capture drug lords—can lead to higher crime rates in corrupt and weak governance

13See http://www.pbs.org/wgbh/frontline/article/the-staggering-death-toll-of-mexicos-drug-war/14See http://www.excelsior.com.mx/nacional/2017/05/24/116527215By contrast, recent empirical research has found that arresting or killing leaders of Mexican drug cartels results in

higher levels of violence in the long run (Phillips, 2015).16See http://www.animalpolitico.com/2017/07/nuevo-record-homicidios/17See http://www.animalpolitico.com/2016/10/2016-homicidios-crimen-organizado-aumentaron/18See http://www.insightcrime.org/news-analysis/new-vigilante-groups-in-michoacan-mexico-resemble-predecessors19See http://harvardkennedyschoolreview.com/justice-in-mexico-the-mexican-drug-wars-most-important-change-that-nobody-noticed/

12

environments (Kugler, Verdier and Zenou, 2005).

Figure 1: Homicide rate in Mexico, 1990–2016

Drug War

0

10

20

30

1990 2000 2010Year

Hom

icid

es p

er 1

00,0

00 p

eopl

e

3.2 Vigilantism and Support for Extrajudicial Violence

In early 2013, groups of civilians in the state of Michoacán formed self-defense militias or vigilante

groups (autodefensas) to fight organized crime. In their inception, these militias aimed at kicking

out an exceptionally violent drug cartel known as the “Knights Templars” (Los caballeros templarios).

Violence and extortion had been longstanding problems in Michoacán, particularly in the region

known as the “Hot Land” (Tierra caliente). This motivated civilians to take justice in their own hands.

One of the leaders of the autodefensas explained that the Knights Templar had crossed a line when

they started to kidnap women and children in groups in order to rape them. Others explain that

the Knights Templar started exerting direct control over agricultural production, taking over farms

illegally, displacing owners and exploiting workers.20

Observers have attributed the emergence of these vigilante groups as a semi-popular uprising,

and a reaction to the ineffectiveness of the state to meet out justice and reduce the violence. Many

20See https://www.washingtonpost.com/news/monkey-cage/wp/2014/05/20/how-does-a-drug-cartel-become-a-lime-cartel/?utm_term=.049f2190dc34

13

Figure 2: Homicide rates at the municipality Level, 2016

of these vigilante groups clashed with both drug cartels and state security forces in their attempt to

regain territorial control, and were accused of carrying out lynchings and human rights abuses.21

Based on public opinion polls conducted shortly after the creation of the autodefensas, a majority of

Mexican citizens supported the creation of such groups and perceived them as more effective than

the state security forces.22 On May 2014, after federal authorities arrested one of the autodefensas

leaders, the government offered to incorporate the autodefensas into official public security forces,

which resulted in the newly formed rural police forces. Michoacán’s violence levels did not decrease

as a result of the emergence of vigilante groups. In fact, the murder rate has been increasing sharply

over the past two years.23 More recently, the western state of Colima has experienced a large uptick

of violence based on jockeying of rival carts, with a homicide rate of 82 per 100,000, a four-fold

increase from 2015.24

The interrelationship between violent crime and institutional mistrust shaped the emergence of

vigilantism in Mexico. Several studies have documented that high violent crime has a detrimental

21See https://news.vice.com/article/mexican-authorities-say-they-dont-exist-vigilantes-standing-up-to-the-zetas22See http://www.animalpolitico.com/2014/07/7-de-cada-10-mexicanos-apoyan-a-los-grupos-de-autodefensas/23 See http://www.businessinsider.com/mexico-sinaloa-jalisco-cartel-fighting-violence-in-colima-2017-124See http://www.businessinsider.com/mexico-sinaloa-jalisco-cartel-fighting-violence-in-colima-2017-1

14

effect on political trust, particularly in the case of Latin America, which has been experiencing

trend of escalating criminality (Di Tella, Edwards and Schargrodsky, 2010). For example, citizens

exposed to criminal violence tend to report lower levels of satisfaction with the way democracy

works in their country (Fernandez and Kuenzi, 2010), and specifically lower levels of support for

the justice system (Malone, 2010). With regard to Mexico, there is evidence that perceptions of

insecurity and crime victimization have negatively affected trust in institutions that directly deal

with crime, such as the police and more broadly the judicial system (Blanco, 2013).25 Within this

context, many times citizens turn to vigilante justice. Recent work highlights that perceptions of a

trusting community and an untrustworthy law enforcement jointly influence support for vigilante

justice (Zizumbo-Colunga, 2017). However, vigilante justice often ends up victimizing civilians and

generating more violence. Why do civilians support this behavior? As outlined in the previous

section, we posit that the anger raised in the wake of violence is causally linked to the desire of

punishing perpetrators, even at the expense of the rule of law.

4 Research Design

4.1 Sampling Strategy

Our target population for this study is adults residing in the four states that make up Western

Mexico, namely Colima, Jalisco, Michoacán, and Nayarit. Given our interest in violence, and

in particularly support for vigilantism, half of our sample is chosen from Michoacán, with the

rest selected proportional to population across Colima, Jalisco, and Nayarit. Respondents were

randomly selected using a stratified multistage cluster sampling design. We first took a stratified

random sample of electoral precincts (referred to as primary sampling units or PSUs going forward),

then selected blocks or clusters of households (secondary sampling units or SSUs), then selected

individual households, and finally selected individuals within households for surveying. A full

description of the sampling strategy can be found in the Appendix.

25Furthermore, studies have shown that electoral turnout is lower in the most violent regions of the country (Carreras,Forthcoming), and that support for the national incumbent party varies inversely with prevailing levels of violence (Ley,2013).

15

Our sampling strategy takes into account variation in autodefensas group presence, violence

levels, and degree of urbanization. Stratifying along these variables ensures that within each

geographic subregion (Michoacán versus the rest of Western Mexico), we have variation in key

variables of interest such as exposure to violence and vigilante groups.26 After stratifying based on

these characteristics, we sampled PSUs proportional to the size of their populations.

Figure 3 shows monthly homicide rates in each of the four states of western Mexico between

January 2014 and June 2017. The blue line denotes the date in which our survey experiment was

conducted.

Figure 3: Recent homicide rates in Western Mexico

CO

LIMA

JALISCO

MIC

HO

ACAN

NAYAR

IT

2014 2015 2016 2017

0

50

100

150

0

50

100

150

0

50

100

150

0

50

100

150

Year month

Hom

icid

es p

er 1

00,0

00 p

eopl

e (a

nnua

lized

rate

)

We generated five random samples so that surveyors could replace PSUs in case it was

needed. The interviewing mode was based on face-to-face interviews with structured questionnaires

administered by trained interviewers through a personal electronic device (tablet).

Within each PSU, surveyors used maps of the PSUs and a random number generator to select

blocks or clusters of households proportional to their size. Within each of those blocks or household

clusters, surveyors used a systematic random sampling method to select households. Blocks were26Because the vast majority of localities outside of Michoácan do not have known vigilante group presence, we only

stratified on vigilante group presence in Michoácan.

16

covered starting by the northeast corner, walking clockwise, and skipping households in regular

intervals of three units. Once a household was selected, they then created household rosters using

a program on their tablet. The program randomly selected the gender of the respondent to be

surveyed in each household,27 and then selected which of the household members of that gender

would be invited to participate. If the respondent was not available or declined, we made one

replacement within the household and then replaced the household with its nearest neighbor.

4.2 Descriptive Statistics

This strategy produced a sample with considerable variation in terms of exposure to violence

and criminal justice preferences. Demographic summary statistics are presented in the Appendix

Table. Half of the respondents in our sample are in Michoácan, 40% are in Jalisco (due to its large

population), 6% are in Nayarit, and 4% in Colima.

In terms of exposure to violence and armed groups, our data shows considerable variation

across states that is generally in line with qualitative and journalistic accounts. We first examine

variation in exposure to armed groups by state. Figure 4 plots the proportion of respondents by

state who say that it’s “very likely” or “sure” that four different types of armed groups were active

in a municipality like theirs in the past year.28

The survey responses show that the distribution of armed groups varies in substantively

important ways across the four states. In most of the states, respondents estimate that the narcos

and state actors deployed to fight them (the police and the army) are much more likely to be active

than autodefensas. Michoácan is a notable exception, where about 20% of respondents say that it’s

very likely or sure that the autodefensas are active in a similar locality. Michoácan similarly has the

lowest level of estimated army presence. Around 40-50% of respondents across all four states find it

likely that the police would have been active, while the perceived recent presence of the army is

relatively high in Colima and Nayarit.

Next we examine the estimated incidence of five major types of violence. Because we assessed

27We sampled men with a 60% probability in order to produce a sample with better gender balance because men weremore likely to be unavailable to participate in the survey.

28We intentionally asked these questions indirectly due to the sensitive nature of directly asking respondents to reporton various actors in their municipality.

17

Figure 4: Estimated presence of armed groups by state

Michoacan Nayarit

Colima Jalisco

Police Ar

my

Narcos

Auto

Police Ar

my

Narcos

Auto

0.00.10.20.30.40.5

0.00.10.20.30.40.5

Group

Proportion

that asking directly about personal exposure to major violence in this region would put both subjects

and surveyors at an unjustifiable level of risk, we proxy for personal exposure to violence with

questions that ask respondents to estimate how likely it is that someone in their community has

experienced different types of violence in the past 30 days. We selected these five forms of violence

based on past applications of the Harvard Trauma Questionnaire and crime statistics for the four

states included in our study (O’connor, Vizcaino and Benavides, 2015). To validate these measures

as proxies for personal exposure, we also measure a subset of less sensitive types of violence directly

by asking people if they have experienced them in the past year. As Appendix D shows, the direct

questions are strongly predictive of responses to the indirect questions, even after including fixed

effects for the 135 precincts (PSUs) that we selected for the study. This suggests that respondents

are drawing on their personal experiences to answer the indirect questions about violence, and that

these measures can indeed be used as proxies for personal exposure.

In Figure 5 we plot the distribution of the estimated likelihood that someone in the respondent’s

community has experienced each type of violence in the past thirty days. Each dot represents the

average response for a PSU based on approximately ten individual respondents, while the boxes

18

represent the 25th, 50th, and 75th quartiles of the PSU averages by state.

Figure 5: Estimated incidence of severe violence by state

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●●

●

●●

●●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●●

● ●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●●

●

●●

●

●

●

●

●

●

● ●

●

●

●

●●

●

●●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

● ●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●●

●

●●

●●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

Not at all

A little bit

Somewhat

Very

Abdu

ction

Assa

ult

Extor

tion

Protec

tion

Weap

on

Violence

Like

lihoo

d

State ● ● ● ●Colima Jalisco Michoacan Nayarit

Figure 5 shows that there is considerable variation in the estimated exposure to violence both

within and across states. Respondents in localities in Nayarit generally estimate the lowest incidence

of violence, on average finding it just a little bit or not at all likely that someone in their community

has experienced all of the five types of violence in the past month. Colima has the highest averages,

although Michoácan has the most individual PSUs with the highest estimated prevalence of most

types of violence. Finally, the data from all the states suggests that extortion is the most common

type of violence that people experience. The extremely high incidence of extortion is also supported

by our direct measure of exposure to this type of violence: 14% of our respondents say that they

have personally experienced extortion in the context of the drug war or drug trafficking. The high

prevalence of extortion is also in line with other data sources: according to the 2016 National Crime

Victimization Survey (ENVIPE), extortions have been increasing over the past years throughout the

country, representing 24% of all cases of self-reported crime victimization, and ranking as the most

19

common crime in the four states of Western Mexico.29

5 Study 1: Is Violence Correlated with Support for Harsh, Pro-vigilante

Policies?

In the first study, we test whether exposure to violence is correlated with anger, attributions of

blame, and punitiveness. This provides an initial observational test of whether the kinds of violence

to which Western Mexicans are regularly exposed to are related to stronger preferences for harsh and

vigilante punishments, exhibit more of the proposed psychological underpinnings of punitiveness

including more punitiveness, stronger attributions of blame, and more anger. If we find the expected

relationships–exposure to violence increases punitiveness and anger–then this would suggest a

mechanism for how violence corrodes support for the rule of law.

Our main measure of exposure to violence, the key independent variable in Study 1, is a

standardized additive index of five different types of violence (abduction, extortion, paying for

protection, being threatened with a weapon, and assault) discussed in depth in the previous

section. We examine the correlation between this measure of exposure to violence and five different

outcomes, including two indices of policy preferences and three psychological variables that our

theory predicts should mediate the relationship between violence and support for harsh, vigilante

criminal justice policy. Before we present tests of our hypotheses about the relationship between

violence and these outcome variables, we present a short descriptive analysis of attitudes towards

criminal justice policy.

First, we present a description of respondent’s attitudes on the six policy questions that underly

our two indices of support for vigilantism and support for harsh criminal justice policy. The four

questions that measure support for vigilantism asked respondents the extent to which they support

or oppose non-governmental armed groups, the autodefensas, lynching a criminal rather than

releasing him on a technicality, and formally legalizing the autodefensas. The two questions that

measure support for harsh punishments ask about whether the respondent supports reinstating

29See http://www.inegi.org.mx/saladeprensa/boletines/2016/especiales/especiales2016_09_04.pdf.

20

the death penalty, and whether they would oppose a proposal to pay narcotraffickers to stop

participating in violence. Figure 6 presents the proportion of our respondents who agree with each

of these six policy proposals.

Figure 6: Support for pro-vigilante and harsh criminal justice policy

Prefer Lynching to Release on Technicality

Support Non−Govt Armed Groups Support Autodefensas

Support Death Penalty Oppose Paying Narcos

Stron

gly Di

sagre

e

Disag

ree

Neith

erAg

ree

Stron

gly Ag

ree

Stron

gly Di

sagre

e

Disag

ree

Neith

erAg

ree

Stron

gly Ag

ree

0.0

0.2

0.4

0.6

0.0

0.2

0.4

0.6

0.0

0.2

0.4

0.6

Response

Prop

ortio

n

Beginning with the two questions that measure support for harshly punitive criminal justice

policies in the top panels, we find fairly strong support for punitive policies. 36% of respondents

support bringing back the death penalty. In contrast to the other policy questions, where opinion

is fairly widely distributed across both sides of the policy, a large majority (87%) of Western

Mexicans are opposed to a policy that would financially reward narcotraffickers for abstaining from

violence. Generally, the bottom three panels of Figure 6 show that respondents are fairly mixed in

21

terms of their support for vigilante solutions to drug violence. Across the two questions that ask

directly about support for non-governmental armed groups or autodefensas, a sizable minority of

respondents say that they support or strongly support these groups (⇡34-36%). About 28% would

prefer that a criminal be lynched than released legally on a technicality. Surprisingly given it’s

reputation as the “tierra caliente”, Michoacan residents exhibit relatively low levels of support for

these policies in our sample. In general, the highest level of support for pro-vigilante and punitive

policies is in Colima.

We now test whether people who are exposed to more narcotrafficking violence are generally

more likely to prefer criminal justice policies that enable vigilante groups and mete out harsher

punishments.30 This analysis is based on two mean effects indices built from the five distinct survey

questions presented above. The Harsh Punishment Index is based on two questions measuring

the extent to which the respondent supports bringing back the death penalty and opposes paying

criminals to stop perpetrating crimes. The Pro-Vigilante Index is based on three questions measuring

the extent to which the respondent believes the autodefensas are necessary, believes that armed groups

outside of the army and police are necessary, and prefers to see a criminal lynched than go free

on a technicality.31 The full text of all the measures used to construct these indices is included in

Appendix B.1.

For each index we estimate the relationship between exposure to violence and policy pref-

erences using OLS. For each outcome of interest, we estimate a specification without any control

variables, with individual-level controls, and with PSU fixed effects. The individual controls include

gender, education, an assets index, age, marital status, and employment of the household head.

We selected these demographic characteristics because they are both likely to explain variation in

exposure to violence, and because they are the kind of slow-changing demographic characteristics

that are unlikely to be affected by post-treatment bias. We cluster standard errors by locality to take

30This analysis was not preregistered and so should be interpreted as exploratory.31We also asked respondents whether they thought the autodefensas should be legalized and whether they thought El

Chapo (a notorious drug trafficker extradited to the US shortly before our survey) should be tried in the US. We do notinclude these questions in either index because they are not unambiguously related to either of the underlying attitudeswe are measuring. Preferences for legalizing the autodefensas, for example, could be driven by a desire for all criminaljustice to be carried out under the auspices of the legal system or by strong support for the autodefensas. People whooppose El Chapo’s extradition could do so because they think he’s less likely to be harshly punished in the US or becausethey think he should be tried by the citizens whom he most harmed, among other reasons.

22

into account the fact that violence exposure is likely correlated across residents at the local level.32

Table 1 presents the results of this analysis.

Table 1: Exposure to violence and criminal justice policy preferences

Dependent variable:Harsh Punishment Index Pro-Vigilante Index(1) (2) (3) (4) (5) (6)

Violence Index 0.07⇤⇤⇤ 0.06⇤⇤⇤ 0.07⇤⇤⇤ 0.08⇤⇤⇤ 0.07⇤⇤⇤ 0.05⇤(0.02) (0.02) (0.03) (0.03) (0.03) (0.03)

Female 0.0004 0.02 �0.04 �0.06(0.05) (0.05) (0.05) (0.04)

Education 0.05⇤⇤⇤ 0.05⇤⇤⇤ �0.02 �0.02(0.01) (0.01) (0.01) (0.02)

Assets Index 0.07⇤⇤⇤ 0.06⇤⇤ 0.0004 �0.005(0.02) (0.03) (0.03) (0.03)

Age �0.002⇤ �0.003⇤ �0.004⇤⇤⇤ �0.01⇤⇤⇤(0.001) (0.002) (0.002) (0.002)

Married 0.06 0.06 �0.02 �0.01(0.04) (0.05) (0.05) (0.06)

Employed �0.04 �0.07 �0.02 �0.01(0.05) (0.05) (0.05) (0.06)

Constant 0.02 �0.07 �0.07 0.02 0.29⇤⇤ 0.11(0.02) (0.09) (0.11) (0.03) (0.12) (0.12)

PSU FEs X XNumber of PSUs 134 134Observations 1,181 1,149 1,149 1,185 1,152 1,152R2 0.01 0.05 0.17 0.01 0.02 0.21⇤p

23

all specifications. Disaggregated results presented in Appendix E.2 show that while these effects do

vary across sub-indicator and are not significant in all cases, their signs are generally consistent.33

Substantively, some of the shifts in policy preferences are important. People who are exposed

to above-average levels of narcotrafficking violence, for example, are 10 percentage points more

likely to support bringing back the death penalty, an increase of 30% over the low-violence group.

High-violence respondents are also ten percentage points more likely to prefer that criminals are

lynching in the town square rather than released from jail on a technicality, a 37% increase over the

low-violence group.

Our proposed explanation for the link between exposure to violence and policy preferences is

that violence leads to changes in the psychological states of those affected, which in turn leads to

changes in their policy preferences. We next test whether past exposure to narcotrafficking violence

is associated with psychological factors that should increase support for punitive and vigilante

justice policies: higher levels of anger, stronger attributions of blame, and a stronger general desire

to punish. Following the appraisal tendency framework, we view blame attributions as a potential

mechanism linking anger to punitiveness (Lerner, Goldberg and Tetlock, 1998).

Before presenting our analysis of the relationship between exposure to violence and these

psychological variables, we briefly discuss our measurement strategy and present a descriptive

analysis. To measure attributions of blame, we asked people for their opinions about how responsi-

ble different actors are for the drug-related violence in Mexico, including the narcotraffickers, local

and federal police, army, autodefensas, and politicians. Figure 7 plots the distribution of blame by

actor across PSUs. Each dot represents the average response for a PSU based on approximately

ten individual respondents, while the boxes represent the 25th, 50th, and 75th quartiles of the PSU

averages by state.

Interestingly, our respondents on average place the most blame for drug trafficking violence

on the local police and politicians. Narcos, autodefensas, and the federal police are seen as less

33The results on the harsh punishment index are driven primarily by the relationship between violence and supportfor the death penalty. The null result on the question asking about opposition to paying narcos to abstain from violencecould be due to the low variation on this policy, which 87% of respondents oppose. There is a positive and significantrelationship between violence and support for two out of the three pro-vigilante policy questions. Interestingly, the onenull result is on support for the actual autodefensas.

24

Figure 7: Blame for different actors across PSUs

●

●

●

●●

●

●

●

●

●●

●

●

●

●

●

●●

●

●●

●

●

●

●●

●●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

● ●

●

●

●●

●●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●●

●●

●

●

●

●

●

●●

●

●●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

Not at all

A little bit

Somewhat

Very

Army

Autod

efens

as

Fede

ral Po

lice

Loca

l Polic

e

Narco

s

Politic

ians

Actor

Blam

e

responsible, while the army falls in between these two poles. Blame for different actors is consistently

positively correlated (r = [0.13,0.44]), suggesting that blaming all of the groups collectively (i.e.,

blameworthiness) is an important underlying attitude.

We use a similar strategy to measure general punitiveness by asking respondents how much

they would like to see the same six actors be punished for the violence affecting Mexico on a

four-point scale. To measure respondents’ general levels of anger, we ask them to report how often

they felt angry in the last 30 days on a four-point frequency scale. Descriptive analyses of both of

these variables are presented in Appendix E.1.

Table 2 shows that past exposure to violence is strongly and positively associated with the

propensity to feel anger, attributions of blame, and general punitiveness. Based on the most con-

servative specifications with PSU fixed effects, a one standard deviation increase in exposure to

violence is associated with a 0.16 standard deviation increase in trait anger, a 0.1 standard deviation

increase in blame attributions, and a 0.09 standard deviation increase in general punitiveness. Sup-

plemental results in Appendix E show that the increases in blame and punitiveness are consistent

across all six of the sub-indicators that make up the indices. In addition, we test for associations be-

25

Table 2: Exposure to violence and psychological outcomes: emotions and attributions

Dependent variable:Anger Blame Index Punitiveness Index

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Violence Index 0.14⇤⇤⇤ 0.14⇤⇤⇤ 0.16⇤⇤⇤ 0.10⇤⇤⇤ 0.09⇤⇤⇤ 0.10⇤⇤⇤ 0.11⇤⇤⇤ 0.09⇤⇤⇤ 0.09⇤⇤⇤(0.03) (0.03) (0.03) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02)

Female 0.03 0.05 �0.02 �0.01 0.01 0.01(0.05) (0.06) (0.04) (0.04) (0.04) (0.04)

Education �0.01 �0.01 �0.002 �0.01 0.02⇤ 0.01(0.02) (0.02) (0.01) (0.01) (0.01) (0.01)

Assets Index �0.01 �0.01 0.05⇤⇤ 0.05⇤⇤ 0.07⇤⇤ 0.05⇤(0.03) (0.03) (0.03) (0.03) (0.03) (0.03)

Age �0.003⇤⇤ �0.003 �0.004⇤⇤ �0.004⇤⇤ �0.005⇤⇤⇤ �0.01⇤⇤⇤(0.001) (0.002) (0.002) (0.002) (0.001) (0.002)

Married �0.04 �0.05 0.05 0.07 0.03 0.06(0.06) (0.07) (0.04) (0.05) (0.04) (0.05)

Employed 0.01 0.04 0.05 0.07 0.02 0.01(0.06) (0.07) (0.04) (0.04) (0.04) (0.04)

Constant 0.01 0.17 0.34⇤⇤ 0.02 0.13 0.22⇤⇤ 0.04 0.12 0.46⇤⇤⇤(0.03) (0.11) (0.14) (0.02) (0.10) (0.11) (0.03) (0.08) (0.11)

PSU FEs X X XNumber of PSUs 134 134 134Observations 1,183 1,150 1,150 1,177 1,147 1,147 1,176 1,145 1,145R2 0.03 0.04 0.15 0.03 0.05 0.22 0.03 0.07 0.23⇤p

26

justice attitudes. In the next sections, we turn to experimental research designs that use randomly-

assigned variation in exposure to hypothetical forms of violence to address some of the endogeneity

concerns and home in on causal mechanisms.

6 Study 2: Do Crimes that Inspire Outrage Increase Support for Harsh,

Vigilante Policies?

In Study 2 we use an experiment to test whether outrage is driving the relationship between

exposure to violence and preferences for harsh, vigilante criminal justice policy. This eliminates the

possibility that a confounding factor (i.e., maybe angry people are more likely to report exposure

to violence and prefer vigilante justice) might bias our estimate of the relationship between anger,

violence, and political preferences. Therefore, we directly test how individuals respond to morally

outrageous violence. While the experimental design comes at the cost of less external validity

because we rely more heavily on hypothetical scenarios and respondents’ self-reports of what

they would do and feel in such situations, these advantages make it a strong complement to the

observational methods in Study 1.

Specifically, we use three survey experiments designed to manipulate the level of moral outrage

that respondents feel in response to a crime by violating moral tenets. We then ask respondents to

report how they would react if such a crime occurred, including what emotions they would feel

and how they would evaluate two different potential punishments: one that is clearly very harsh

and extra-judicial (Outcome B), and another that is legal and less severe (Outcome A). The full text

of the crime scenarios and two potential punishments are presented in Table 3. The sections of the

scenarios that are randomized are italicized, and the “moral outrage” version of the scenario is also

bolded.

The three scenarios violate various commonly-held moral tenets, and in two of the three

scenarios the crime is violent. In all of the scenarios, the victims are presented as innocents, but

this is particularly strong in Scenarios 1 and 2 where the victims in the moral outrage version are

children. In Scenario 3, the crime evokes the idea of “extra-lethal” violence (Fujii, 2013) that is

27

Table 3: Crime scenarios and punishment options in Study 2

Scenario 1 Scenario 2 Scenario 3

Scen

ario Imagine a situation in which anarco gang controls the town.

They control the drug trade, andthey also are notorious for abus-ing and exploiting the local popu-lation / children under the age of10.

Imagine a situation in which acorrupt politician is in charge ofa large city. He does political fa-vors for his friends and powerfulpeople, and steals money fromgovernment contracts / a hospitalfor disabled children.

Imagine that a narco abducts asmall business owner becausehe won’t pay them part of hisprofits. A week later, the busi-ness owner’s body is foundoutside town, and he has beenshot to death / beheaded and hisbody shows signs of torture.

Out

com

es

A: The narco gang members arearrested and put on trial for theircrimes.

A: The politician is arrested andput on trial for corruption.

A: The narco is arrested andput on trial.

B: The narco gang members arekilled by locals in the townsquare.

B: Local citizens attack themayor and burn his house down.

B: The narco is killed by au-todefensas.

particularly performative and brutal. On the other hand, a number of factors are held constant

between the moral outrage and control versions of the scenarios that might influence the perceived

effectiveness and justice of punishment, including the perpetrator’s identity, the likely motivation,

and the amount of harm. These factors are held constant such that the treatment version of the

scenarios isolates the effect of moral outrage as cleanly as possible. The two possible outcomes are

designed to contrast the punishment that is clearly legal (“arrested and put on trial”) with a range

of vigilante solutions: spontaneous lynching, retributive destruction of property, or violence by an

organized vigilante group.

All respondents were asked to evaluate all three scenarios, and we present out main analysis

based on a stacked dataset of all responses to all three scenarios to avoid multiple comparisons

concerns. The order of the three scenarios was randomized. We cluster standard errors at the

individual level to take this stacked structure into account. This analysis was pre-registered with

EGAP and we do not deviate from the pre-analysis plan.

In this experiment we have four main outcomes of interest. First, we test whether respondents

say that the scenarios would make them angry. Because the experiment is designed to induce anger,

we consider this a manipulation check. We also measure another emotion that could be plausibly

28

induced by the moral outrage versions of the scenarios, fear. As specified in our pre-analysis plan,

we consider the experiment to have passed the manipulation check if participants report that the

outrage scenario would make them feel significant levels of anger and have little effect on fear.

Figure 8 plots the coefficients from an analysis of the effects of the three pooled treatments on how

angry and afraid respondents say they would feel if the hypothetical crime scenario occurred in

their community.

Figure 8: Effect of outrage scenarios on hypothetical anger and fear

●

●

−0.1

0.0

0.1

0.2

0.3

Anger FearEmotion

Stan

dard

ized

Trea

tmen

t Effe

ct

Outcome●

●

AngerFear

The treatments overall had a large, statistically significant positive effect on how angry respon-

dents thought they would be if the crime occurred in their community. They had no detectable

effect on how afraid respondents would be. Appendix Figure F.1 in shows that this effect is driven

by Scenarios 1 and 2, with a much larger effect in Scenario 1, while the outrage version of Scenario 3

had no effect on either anger or fear. We’ll come back to this heterogeneity in the effectiveness of

the scenarios in manipulating outrage when discussing the interpretation of the substantive results,

and in Study 3, where we use an experimental design to test more systematically for the types of

violence that makes people most outraged.

Next, we turn to the substantive outcomes of interest, which measure the respondents’ prefer-

ences over and perceptions of the two punishments for the crime scenario. First, we test whether

respondents are more likely to prefer the vigilante outcome if they are presented with the outra-

29

geous crime. Second, we examine two perceptions that might underlie such a preference shift:

the perception that the vigilante solution is more effective in preventing future violence, and the

perception that it is more just. The bars in Figure 9 presents the proportion of the respondents by

state and scenario who said that they preferred the vigilante outcome, aggregated across both the

“outrageous” version of the crime and the less offensive version and with 95% confidence intervals

in black.

Figure 9: Proportion of respondents who prefer the vigilante outcome

Michoacan Nayarit

Colima Jalisco

Child

Abus

eGr

aft

Mutilia

tion

Child

Abus

eGr

aft

Mutilia

tion

0.0

0.1

0.2

0.3

0.4

0.0

0.1

0.2

0.3

0.4

Scenario

Prop

ortio

n

Figure 9 shows first that large minorities of respondents prefer the harsh, extrajudicial pun-

ishment in these scenarios. Overall, 17% of respondents prefer the vigilante solution in the child

abuse scenario, 10% in the graft scenario, and 19% in the mutilation scenario. Second, there are

interesting differences across states. More respondents in Colima prefer the vigilante outcome in

two out of three of the scenarios, although these differences should be interpreted with caution due

to the small Colima sample. More surprisingly given its reputation, Michoacan actually has low

levels of support for the harsh, vigilante outcomes relative to the other states.

Our primary analysis of the effect of the outrageous versions of the treatments is based on a

Linear Probability Model (OLS), but in Appendix F we show robustness to a logit model that takes

into account the binary nature of the three dependent variables of interest. Table 4 presents the

30

results of the analysis.

Table 4: Effect of outrage scenarios on preferences over and perceptions of harsh, vigilante punish-ments

Dependent variable:Vigilante Preferred Vigilante More Just Vigilante More Effective

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

Outrage Treatment 0.01 0.02 0.01 0.01 0.03⇤⇤ 0.03⇤⇤(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)

Violence Index 0.03⇤⇤⇤ 0.05⇤⇤⇤ 0.05⇤⇤⇤(0.01) (0.01) (0.01)

Female �0.05⇤⇤⇤ �0.02 �0.01(0.02) (0.02) (0.02)

Education �0.02⇤⇤⇤ �0.02⇤⇤⇤ �0.01⇤(0.01) (0.01) (0.01)

Assets Index 0.01 0.01 �0.001(0.01) (0.01) (0.01)

Age �0.002⇤⇤⇤ �0.001⇤⇤ �0.002⇤⇤⇤(0.001) (0.001) (0.001)

Married �0.01 �0.01 0.02(0.02) (0.02) (0.02)

Employed �0.01 0.002 �0.002(0.02) (0.02) (0.02)

Constant 0.15⇤⇤⇤ 0.25⇤⇤⇤ 0.16⇤⇤⇤ 0.25⇤⇤⇤ 0.18⇤⇤⇤ 0.28⇤⇤(0.01) (0.08) (0.01) (0.09) (0.01) (0.13)

PSU FEs X X XObservations 3,615 3,456 3,615 3,456 3,615 3,456R2 0.0004 0.10 0.0003 0.10 0.001 0.10⇤p

31

irrational or extremely dedicated to their efforts.

Given that not all of the three scenarios passed the manipulation check, we also turn to a

disaggregated analysis to help interpret these results. Figure 10 shows that Scenario 1, in which a

group of narcos is abusing young children in the outrageous version of the crime and the narcos

are lynched in the town square in the vigilante outcome, caused significant increases in all three

substantive outcomes. In that scenario, respondents are 6-7 percentage points more likely to prefer

the vigilante solution and to find it more just and more effective.

Figure 10: Effect of disaggregated outrage scenarios on preferences and attitudes towards vigalin-tism

●

●

●

●

●

●

●

●

●

−0.05

0.00

0.05

0.10

AbusingChildren

Stealingfrom Disabled

Decapitation

Scenario

Stan

dard

ized

Trea

tmen

t Effe

ct

Outcome●

●

●

PreferredMore JustMore Effective

Taken together, the main and disaggregated results provide tentative evidence that violence

that causes moral outrage causes an increase in support for harsh, vigilante punishments, under

some circumstances. The statistically significant result on perceptions of effectiveness in the pooled

analysis was driven in part by a slightly higher level of perceived effectiveness of the vigilante

solution in the third scenario, which did not pass our manipulation check. This suggests that we

should not conclude that moral outrage causes increases in perceived effectiveness of vigilantism

32

but not perceived justice or overall preferences for vigilante solutions. The disaggregated results,

on the other hand, suggest that some scenarios, particularly those that involve the violence against

innocents, do cause people to prefer harsh, vigilante solutions, and to perceive them as both more

just and more effective.

These three scenarios were designed to use specific language to maximize the amount of

outrage that respondents would feel. However, because a number of factors change across the three,

they do not allow us to definitively pinpoint the elements of a crime that make citizens outraged.

In the next study, we turn to an experimental research design that uses a much larger range of

variation in crime scenarios to do just that.

7 Study 3: Which Crimes Increase Support for Harsh, Vigilante Poli-

cies?

In Study 3, we test whether the severity of a crime and innocence of the victim are general properties

that make people more likely to support a punitive or vigilante solution. We generate a survey

experiment with 125 unique scenarios with different perpetrators, victims, and crimes that represent

common crimes that people in Western Mexico face.

The survey enumerators read a script to the participants during the interview that described

this randomly generated crime. Each element was randomly assigned independently and had an

equal probability of appearing. Table 5, with randomized segments in bold:

Table 5: Crime scenarios in Study 3

Imagine that a grandmother / student / local small business owner /soldier / narco has been robbed / extorted for money / tortured / disap-peared / killed in your community by a narco / autodefensas member/ local police officer / federal police officer / soldier.

In order to test hypotheses about the general elements of crimes that citizens find outrageous,

we code the individual crime scenarios along two distinct dimensions:

• Innocence of Victim: can take a value of -1 (narco), 0 (soldier), or 1 (grandmother, small

33

business owner, student)

• Severity of Crime: can take a value of -1 (robbed), 0 (extorted for money, tortured), or 1

(disappeared, killed)

In addition, although we did not have specific hypotheses about how the identity of the

perpetrator would affect outrage or preferences for harsh, vigilante punishments, we also code our

five perpetrator categories according to whether they are members of state security forces or not.

We use reactions to these crime scenarios to test first whether crimes that involve more severe

violence and have more innocent victims are more likely to induce anger, but not fear (Prediction

1). To measure the outcomes for this hypothesis, we asked participants to assess how angry and

afraid they would be on a five-point scale if a crime like the one described happened in their locality.

Second, we test whether participants prefer harsher punishments for crimes that involve more

severe violence and have more innocent victims (Prediction 2). In order to measure this outcome,

we asked participants to choose from a list of possible punishments the one they would be most

satisfied with for the perpetrators of the crime described. We then coded the punishments by

severity according to logical criteria that we specified in advance, so that the outcome variable

Severity of Punishment can take a value of 0 (no punishment), 1 (beaten, one year of jail), 2 (ten years

of jail), or 3 (death penalty, lynched, shot). Similarly, in order to test Prediction 3 that participants

would be more likely to prefer extrajudicial punishments for perpetrators of more violent crimes

against more innocent victims, we code the same preferences according to whether they are legal

or extrajudicial punishments: in this case, the variable takes a value of 1 for punishments that are

clearly extrajudicial, and zero otherwise. As a secondary test of how participants are choosing

punishments, we also directly asked them to rank three principles that are commonly invoked to

justify punishing criminals: punitiveness, effectiveness in preventing future crimes, and legality. As

per Predictions 2 and 3, we expect that punitiveness will increase and legality will decrease in this

ranking for crimes involving more severe violence and innocent victims.

We carry out our main hypothesis tests using OLS. Appendix G also presents coefficients

estimated using an ordered logit. In order to calculate an estimate of the treatment effects that is

as close as possible to the effect of violence that this population is exposed to in the real world,

34

we weight each scenario by how likely respondents found it to be. Appendix G.1 provides more

information on how the weights were calculated and shows that our results are robust to those

without weights.

Table 6 presents our tests of whether crimes with more innocent victims and more severe

violence induce more anger.

Table 6: Characteristics of scenarios that induce anger and fear

Dependent variable:Anger Fear

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

Innocence of Victim 0.744⇤⇤⇤ 0.741⇤⇤⇤ 0.726⇤⇤⇤ 0.234⇤⇤⇤ 0.219⇤⇤⇤ 0.218⇤⇤⇤(0.035) (0.038) (0.039) (0.042) (0.044) (0.045)

Severity of Violence �0.069⇤ �0.051 �0.088⇤ 0.044 0.061 0.057(0.037) (0.039) (0.046) (0.044) (0.045) (0.053)

Government Perpetrator 0.159⇤⇤⇤ 0.142⇤⇤ 0.139⇤⇤ 0.026 0.051 0.050(0.055) (0.060) (0.060) (0.066) (0.069) (0.069)

Violence Index 0.058⇤ 0.059⇤ 0.049 0.050(0.031) (0.031) (0.036) (0.036)

Female 0.056 0.056 0.528⇤⇤⇤ 0.529⇤⇤⇤(0.060) (0.060) (0.069) (0.069)

Education �0.024 �0.024