1

Damaging Democracy? Security Provision and Turnout in Afghan Elections†

Luke N. Condra, Michael Callen, Radha K. Iyengar, James D. Long, Jacob N. Shapiro‡

This version: February 17, 2016

Abstract In conflict-affected states, election-related violence threatens voter safety and turnout. But when citizens view security providers as corrupt, deploying police to protect voters may reduce citizen participation, undermining the election’s legitimating purpose. We estimate police deployments’ impact on violence and turnout in Afghanistan’s 2010 parliamentary election using data from the universe of polling sites and various household surveys. Locations with similar histories of pre-election violence received different deployments of the Afghan National Police, enabling identification of police’s effects on turnout. Increases in police presence marginally decreased election-week violence near the average polling site, but also decreased voter turnout by an average of 24%. Estimates are robust to a broad range of specifications that control for voter expectations of election day violence, based both on subjective survey responses and objective administrative records. Our results highlight an additional obstacle to using elections to build government legitimacy in weakly institutionalized and conflict-affected states.

† The authors are grateful for the help and support provided by Democracy International, the International Security and Assistance Forces in Afghanistan, and Afghan Ministry of Interior. Eli Berman, James Dobbins, Jim Fearon, Gerard Padro-i-Miquel, Eric Schwab, and participants at the International Relations Colloquium of the University of Wisconsin-Madison and the Order, Conflict, and Violence Workshop of Yale University provided helpful comments. Torben Behmer provided expert research assistance. The U.S. Department of Defense's Minerva Initiative (AFOSR Grant FA9550-09-1-0314) and Democracy International provided generous funding. The authors fully and solely designed and executed the study and any opinions, findings, conclusions are those of the authors alone and do not reflect views of the United States Department of Defense, Democracy International, or institutions with which the authors hold affiliation. ‡ Graduate School of Public & International Affairs, University of Pittsburgh; Kennedy School of Government, Harvard University; RAND Corporation; Department of Political Science, University of Washington; Department of Politics, Princeton University, respectively.

2

“The police are on the front line of public security, law and order, and extending the writ of the government. A police force is critical to helping build a democracy because it has the capacity to generate trust between the government and the people” (Rashid, 2008, 203-204).

In countries affected by violent conflict, newly constituted democratic governments face

challenges in establishing political order and developing legitimacy, understood as the

government’s ability to gain the consent of their population to be governed (Lake, 2009; Levi,

1988). Modern political scientists view elections, and specifically citizens’ participation in voting, as

a core expression of consent and an essential practice to create and maintain a government’s

legitimacy (Brancati and Snyder, 2011; Diamond, 2006; Huntington, 1996; Lindberg, 2006). As a

result, governments and the international community frequently view the introduction of elections

as a necessary institutional benchmark in conflict-affected countries. Indeed, one scholar writes,

“the more elections, the more democratic the regime and society in general” (Lindberg, 2009, 9).

In such cases, donors provide substantial diplomatic and technical assistance for the purpose of

improving electoral processes (Hyde, 2011b; Kelley, 2012a), alongside assistance like peace-

keeping forces and development aid to support the legitimacy of fragile governments (Doyle and

Sambanis, 2006; Fortna, 2008).

Yet, in many cases, elections do not serve to increase government legitimacy. On the contrary,

anti-government elements like insurgents use violence to disrupt elections in order to challenge

and undermine the government by demonstrating a lack of capacity thereby encouraging defection

or vitiatin citizens’ consent.1 Given this, governments rely upon other institutions to help support

the legitimizing effects of elections to make it safe for citizens to vote. This requires that the

government train and enlarge its security forces, such as the police, to protect polling station

1 For example, Weidmann and Callen (2013) note roughly 500 separate election day attacks during

Afghanistan’s 2009 presidential election.

3

workers and voters (Finnemore and Barnett, 1999). The theoretical motivation of increasing such

policing is the intuitive prediction that it will result in a more secure environment to assist the

electoral process.2 This assumes that increasing security deployments necessarily decreases

violence, which in turn encourages citizen participation, with the police serving as a symbol of

peace and order of a legitimate government.

We explore whether or not attempts to improve security services always serve to assist electoral

processes in conflict-prone emerging democracies. Specifically, we examine the conditions under

which the interaction between elections and security provision may erode government legitimacy.

We argue that under some circumstances, increasing policing may actually backfire, and a

government’s attempt to improve security provision may weaken electoral processes. Specifically,

we hypothesize that when citizens view government agencies and security services as corrupt and

predatory, they are less likely to signal consent to the state’s authority through participatory acts like

voting, thereby undermining the government’s use of elections as a means toward increasing its

legitimacy.

To test this, we study how variation in the Afghan government’s assignment of levels of policing

in areas adjacent to and within polling centers affected levels of insurgent violence and voter

turnout in Afghanistan’s 2010 Wolesi Jirga (parliamentary) election. This is a particularly

illustrative case as all entrances to polling centers in Afghanistan were guarded by Afghan security

forces and these forces established multiple police checks in cities and towns. A voter could not

cast a ballot in Afghanistan without some contact with police forces.

We combine data on voting and policing in Afghanistan from numerous sources, including

2 Despite this intuition, the evidence on the economics of crime show mixed results for whether

increased policing reduces violent crime, which we discuss in the next section.

4

administrative records of violent events during the election period, the security classification of

polling centers dictating police deployment, election results, responses from multiple household

surveys to probe Afghan citizens’ views of the government and security services, and primary

qualitative evidence.

However, to test our hypotheses empirically, we must isolate the effects of security force

deployment from the confounding effects of direct violence on voter turnout. We develop a

plausible identification strategy by exploiting unique features of how police were assigned to polling

centers. A committee of government and military officials met in Kabul to sort polling centers into

three security categories (each of which represented a specific police force level designation) for

the 2010 election based on anticipated levels of violence. Officials assigning police to polling

centers had a good sense of the trends in insurgent violence within districts, but did not consider

the dynamics at each individual center within a district in all but the most prominent cases.3 As a

result, polling centers with similar histories of violence were given different levels of policing due to

a range of idiosyncratic factors. Once district traits are taken into account, the local history of

violence around a center does almost nothing to explain its security status. We therefore compare

polling centers that received light deployments to those which received moderate or high

deployments to estimate how extra police affected voting, controlling for violence histories. This

3 Iyengar was present in many of these meetings as an observer and adviser to the Ministry of the

Interior, as well as the ISAF Counterinsurgency Advisory and Assistance Team, in 2010. Our

knowledge of the assignment procedures draws heavily on her experiences, as well as other

coauthors’ extensive interviews with senior officials in the Ministry of the Interior knowledgeable

about the security force deployments process. We provide econometric evidence consistent with

the assessment above later in the article.

5

approach isolates plausibly exogenous variation in police deployments, enabling us to estimate

their effect on voter turnout.4 Critically, we also show that the results are unlikely to be an artifact of

ballot stuffing and other kinds of fraud being correlated with police presence.

To preview results, we find that polling centers that receive extra police experience a slight

decrease in insurgent violence around election day, though the effect is not strong statistically.

Those polling centers also experience a marked decrease (about 26% on average) in turnout

relative to similar polling centers that received fewer security forces. We provide primary and

secondary evidence from survey data on attitudes about the government and police corruption to

account for variation in turnout. While the police might have been minimally effective at deterring

insurgent violence on and around election day, the police had a second, deleterious effect on the

electoral process by undermining civilians’ willingness to turn out and vote. We provide multiple

robustness checks and address several alternative explanations for our empirical results in the

Supporting Information.

All told, we find evidence that while increased deployment of security services had small

ameliorative effects on reducing election violence, extra policing seems to have had the unintended

4 The estimated effect of scheduled deployments on turnout remains robust to including measures

based on voters’ subjective forecasts that violence will occur at their local polling center, reducing

concerns that citizens interpreted the presence of police as a sign that violence is likely to happen.

Our measure comes from a survey of citizens in the immediate vicinity of polling centers taken in

the month leading up to the survey. We therefore cannot definitively rule out the argument that

the presence of police caused citizens to strongly update their beliefs about the likelihood of

violence on election day, though this seems unlikely given the robustness of the effect to a broad

set of specifications that control for local violence histories.

6

effect of undermining the government’s legitimacy by discouraging participation in the fundamental

democratic practice of voting. These results stand in stark contrast to the goals of the Afghan

government and the international community in pushing for elections and increased police

deployment to assist the voting process.

A number of important features of this Afghan election suggest it provides a propitious case to

test our hypotheses and develop scope conditions of our argument for comparative analysis. First,

the 2010 parliamentary elections were an important test of the Afghan government’s ability to

independently establish and manage governing institutions to gain citizen compliance and increase

its legitimacy (Berman et al., 2014; Callen and Long, 2015), especially in light of a rising insurgency

and the announced draw-down of international forces. Second, the election took place at a time

when non-state insurgent actors opposing the state like the Taliban had significant potential and

continually threatened to disrupt the elections in many areas of the country (Coburn and Larson,

2014). In response, the Afghan government deployed police to neighborhoods close to polling

centers leading up to the election to protect centers on election day. Last, Afghanistan suffers from

corruption in local and national governing institutions, including the police (Giustozzi and

Isaqzadeh, 2012) . While Afghanistan’s modern political history is obviously distinct in many ways,

we believe that it shares common features of other countries on the path to democratization,

including the desire to hold elections to support government legitimacy where violence and

corruption remain obstacles. Many countries that have moved towards democracy in the

developing world over the last 25 years unfortunately have faced these challenges (Collier, 2009;

Collier and Vicente, 2014), most recently with Arab Spring transitions in Yemen, Libya, Tunisia,

and Egypt.

To our knowledge, we are the first scholars to systematically study the effects of security force

deployment on voting behavior in an emerging democracy. This is due, in part, to the wealth of

7

diverse data we bring to bear in our empirical analysis and an identification strategy that allows us

to better isolate the effects of deployments on turnout. We believe our results make a number of

important contributions to the study of elections and security in emerging democracies.

Theoretically, we explore how variation in views of the government and perceived degree of its

agents’ corruption mediates the process by which security provision affects participation in the

democratic process. In so doing, we account for the conditions under which increased policing

may actually undermine voting, even though intuitively, one would expect increased policing to

help to protect the process. We provide an important caveat to the view that institutional

strengthening among electoral and security sectors necessarily and simultaneously work together.

Whereas any government conducting elections also wants the population to be secure enough to

vote, underlying corruption and predation may cause institutions to have counterproductive

interactions and ultimately undermine a government’s legitimacy.

We also provide insights into a number of related concerns among scholars and policymakers

on elections and security provision in conflict settings. The international community consistently

pushes countries emerging from or engaged in conflict to hold elections as a key benchmark

(Bush, 2015; Hyde, 2011a; Kelley, 2012a; McFaul, 2010).5 But insights from an emergent

5 The international community has supported early elections after conflict in countries as diverse as

South Sudan, Angola, and East Timor. Early elections were an explicit goal of the Bush

Administration’s transition plan in Iraq (Bremer, 2006) and Afghanistan (Rashid, 2008), where the

Obama Administration also lent significant diplomatic and technical support along with its allies

for elections in 2009 and 2010 (Rashid, 2012). In Libya, as the country remained mired in civil

war, Secretary of State Clinton met with rebels to encourage elections even before Muamar

Qaddafi had been captured and killed.

8

literature addressing the important question of electoral timing and institutional sequencing in

conflict-prone transitioning countries suggest that these policies may be ill-conceived or poorly

implemented. On one hand, many students of democratization have warned that rushing to

implement elections in countries whose institutions are not equipped to handle a democratic

transition may be counterproductive, producing hybrid or illiberal regimes (Zakaria, 1997) and/or

the resumption of violence (Mansfield and Snyder, 2005; Roessler, 2005; Snyder, 2000). On the

other hand, scholars have argued that these costs do not outweigh the benefits of elections as a

critical means of democratization (Carothers, 2007; Kelley, 2012b). Additionally, beyond elections,

countries like the US continually encourage developing countries to increase training and

deployment of security services, frequently as a part of military assistance to help the US and its

allies pursue strategic security objectives while helping to support nascent regimes. But as Adams

and Sokolsky (2015) note, the US pushes these reforms and provides assistance without taking

seriously enough the problem of corruption and governance within the security sectors of these

countries.6 Our findings do not resolve these debates, but we provide some evidence to suggest

that states transitioning out of conflict and toward democracy should recognize that attempts to

improve security services may not monotonically increase the quality of elections, particularly when

citizens perceive the police as corrupt agents. In such cases, elections are unlikely to foster the

6 Chandrasekaran (2007) and Rashid (2008, 2012) provide extensive evidence about how US

policymakers ignored problems of corruption within the Iraqi and Afghan security sectors while

continuing to encourage increasing deployment numbers. While we focus on the unintended

negative consequences this had for elections, both authors demonstrate how the predatory nature

of the police in both countries has at times produced, rather than resolved, security concerns for

the Iraqi and Afghan governments, their neighbors, and the US and its allies.

9

desired increase in democratic participation, but instead may serve to undermine it. Governments

and the international community may want to re-evaluate and prioritize the degree to which they

work to fight corruption within the police relative to the timing of elections given the perverse

effects that increasing security may have on electoral quality.

Elections, Security, and Corruption in Emerging Democracies

In modern democracies, elections form the critical mechanism to aggregate citizen preferences

and delegate power to representatives (Cox, 1997; Cox and McCubbins, 1986; Przeworski, Stokes,

and Manin, 1999). The strengthening of incipient democratic institutions requires citizen

participation in the voting process to bolster public confidence in elections and government

legitimacy (Norris, 2014). To this end, security forces play a critical role in supporting elections in

countries emerging from, or still engaged in, violent conflict (Binkerhoff, 2007; Leach and

Kingsbury, 2013), because non-state combatants may strategically use violence to deter voting and

undermine support for the government (Berrebi and Klor, 2006; Stedman, 1997).

Prior approaches to the study of elections and security in developing countries have warned of

the dangers of an increased probability of conflict brought about by democratic political

competition (Collier, 2009; Collier and Vicente, 2014; Hafner-Burton, Hyde, and Jablonski, 2014;

Hyde and Marinov, 2012), particularly in countries with histories of ethnic polarization (Horowitz,

1985; Snyder, 2000; Wilkinson, 2004). For these reasons, the question of electoral timing after

periods of conflict has received attention, with evidence to suggest that early elections may increase

the likelihood of violence (Brancati and Snyder, 2013; Höglund, Jarstad, and Kovacs, 2009).

Moreover, past experience with violence could affect whether individuals turn out to vote (Bellows

10

and Miguel, 2008; Blattman, 2009).7 While literature on modern state-building emphasizes the

role that security provision plays in the legitimization of the state and its monopolization of the use

of violence (Fukuyama, 2004; Levi, 1988), scholars have paid less attention to whether and how

the police may directly shape electoral processes by increasing or decreasing the likelihood of

election day violence and turnout at polling centers.

Do elections and policing necessarily support one another in transitioning democracies? On

the surface, it appears intuitive that increased policing should assist electoral processes by providing

protection for polling center workers to conduct the election and voters to feel secure enough to

vote. In Afghanistan, these realities specifically require the deployment of police officers near and

in polling sites. If the police are effective, their presence should decrease violence and increase

voter turnout. Therefore, there is no inherent substitution between improving democracy and

security forces.

However, we argue that these conditions may not always hold in conflict-affected emerging

democracies, and that the interaction of introducing elections and deploying security services may

yield counterproductive effects. First, increased policing may not always cause a decline in

violence. While studies in economics typically find that the deployment of additional police can,

under certain conditions, reduce violence (Di Tella and Schargrodsky, 2004; Draca, Machin, and

7 Although only a few studies look at the effects of violence on the propensity to vote in emerging

democracies, scholars highlight a number of other factors driving turnout in these contexts,

including allegiance to ethnic groups or parties (Horowitz, 1985), commitment to democratic

principles (Bratton, Mattes, and Gyimah-Boadi, 2004), positive incentives from political actors like

vote-buying (Chandra, 2004; Posner, 2005), social pressure (Jung and Long, 2015), and individual-

level characteristics like income (Kasara and Suryanarayan, 2015).

11

Witt, 2011; Klick and Tabarrok, 2005; Levitt, 1997), a series of quasi-experimental results in

criminology do not find a significant impact of policing on decreasing crime (Gottfredson and

Hirschi, 1990; Sherman and Weisburd, 1995). In the context of insurgency, some studies find that

under the right conditions, deploying additional forces can reduce violence (Berman et al., 2013;

Biddle, Friedman, and Shapiro, 2012). But on balance, a priori assumptions about the effects of

policing on election-related violence are mixed.8

Beyond the question of whether increasing policing decreases violence, it is less certain

whether deployments would ultimately affect voter turnout as a second-order effect. Ideally, a

police officer serves as a symbol of the government that fosters peace and order (Serchuk, 2006).

When police act to serve the public good and protect citizens, police deployment around elections

should work to increase turnout, building on increasing levels of political engagement among the

population. In contexts where citizens view the government, and in particular the police, as

ineffective, corrupt, or predatory, we argue that citizens may choose not to participate in elections

overseen by these agents. Illicit activities committed by the state’s agents erode the public’s support

for the government (Rubin, 2007; Torabi and Delesgues, 2007), and therefore citizens are unlikely

to express the consent to be governed by participating in practices like voting for a government

they fear or do not like. If views of corrupt security providers keep voters at home, this

undermines the government’s use of elections as a means of increasing its legitimacy. Indeed, in

the run up to the 2010 parliamentary election, the Afghan Ministry of Interior worried specifically

about how to encourage public trust in the police (Iyengar, 2010).

8 Despite this mixed evidence, we state our first hypothesis H1 in positive terms for the sake of

formal testing.

12

We hypothesize that in Afghanistan, increasing police presence is unlikely to encourage

turnout. Significant evidence depicts the Afghan National Police (ANP) as largely corrupt and

predatory (Felbab-Brown, 2013; Giustozzi and Isaqzadeh, 2012). Police in Afghanistan have long

engaged in a wide range of criminal activity (Felbab-Brown, 2013; Rashid, 2008). Police chief posts

are sold to the highest bidder (Rubin, 2007). In a 2010 UN survey, twenty-five percent of

respondents reported having paid at least one bribe to police in the previous year (UNODC, 2010)

(almost surely an undercount), and the police and the justice system were the two sectors perceived

to be the most corrupt in 2006, according to a survey fielded by a prominent Afghan NGO

(Torabi and Delesgues, 2007). Whereas evidence consistently points to corruption in the security

sector in Afghanistan, prior studies have not discussed whether and how this affects voting

behavior. Afghan and ISAF policymakers insisted on significantly increasing police deployments

for this election, despite knowing the challenges perceived corruption posed for the electoral

process.

We formally test the following empirical implications:

H1: Additional police deployments lead to reductions in election-day violence in the vicinity of

polling centers.

H2: Additional police deployments reduce voter turnout.

H3: Negative perceptions of the government and corruption measured through surveys are

greater in places that received additional police deployments.

Data

To study how variation in police deployment affects electoral quality in Afghanistan, we

examine whether police presence increases levels of security at polling centers, measured by

insurgent attacks in the immediate vicinity, directly before and on the day of the election. To do so,

13

we employ data from six different sources: (1) certified voter turnout data from the Afghan

Independent Election Commission (IEC), (2) violence data routinely collected by the NATO-led

International Security Assistance Force (ISAF), (3) levels of policing assigned to these polling

centers, which constitutes the treatment variable, (4) survey data from ISAF’s Afghan National

Quarterly Assessment Report (ANQAR), (5) new survey data that we collected from 5000

individuals across 471 polling center catchment areas in nineteen provinces across all regions, and

(6) survey data from 369 Afghan civilians, virtually all identifying themselves as belonging to

different Pashtun tribes in eighty-four different villages of Maiwand and Arghandab districts of

Kandahar province. Supporting Information (SI) Table 1 provides summary statistics.

Policing. We begin by explaining the process of police allocation to polling centers around

Afghanistan for the 2010 parliamentary election. To select polling centers that warranted higher

levels of police deployment, the Afghan Ministry of Interior, in coordination with ISAF, developed

a three-tier categorization. “Secure” polling centers received no additional police above and

beyond their normal baseline level. “Medium insecurity” polling centers were to receive an extra

deployment of police, and “highly insecure” polling centers were to receive the most extra police in

an effort to provide security around election time. We note that our own interviews with relevant

ISAF officials on the security classification of polling centers is corroborated in public

documentation of this process published after the election was completed (NDI, 2011; FEFA,

2011).

These category designations reflected deliberations within the Ministry of the Interior and

ISAF about the potential for election-day violence and were largely based on the history of

violence around the polling site. A former Ministry of Interior official, who was involved first-hand

in these planning meetings, explained that officials from the National Police Command Center,

ISAF, and Ministry of Interior met at least monthly in early 2010, and reviewed data collection of

14

daily incidents of insurgent attacks across the country, to determine the categorization of polling

centers (Condra, 2014c). These central deliberations also took into account information on

security from provincial-level officials from these institutions (Condra, 2014a, 2014c), including

information on the structural features of the center that affected their desirability as a target, and

proximity to security reinforcements. Some of the decisions regarding levels of police deployment

appeared arbitrary as the discretionary criteria were not always precisely defined. This led to sites

located in areas with similar histories of violence, road access, and even population characteristics

receiving different levels of police. We describe this result in more detail in the next section, but SI

Figure 1 provides a comparison of insurgent violence experienced in high and low police

deployment polling center areas between January 2010 and June 2011. The newly recruited police

were trained by the Ministry of Interior and the European Union Police Mission in Afghanistan

(EUPOL) on how to provide security during the election, on proper election procedures, and on

police conduct (NDI, 2011).

Despite our best and repeated efforts, we have not been able to secure data on actual security

force deployments to individual polling centers around the election because the government will

not release it. While our interviews with multiple officials with first-hand knowledge of the

deployment schedule and process suggest that the security deployment categorization scheme was

adhered to in the field (Condra, 2014a, 2014c), this means that we cannot empirically verify with

full certainty that provincial-level police commanders followed the polling center assignment

protocols that the Ministry of Interior gave them. Nor do we have information about the actual

numbers of security forces assigned across polling centers to quantify the size of the effect in terms

of police forces added.

Lacking the ability to verify actual deployment for all polling centers, our attempt to isolate the

effects of extra policing on turnout, controlling for other sources of variation, therefore adheres to

15

two important research design conventions. First, we argue that our estimates of police

deployments can be interpreted as Intention-to-Treat (ITT) estimates, where perfect compliance

with assignment protocols (deployment) are not easily verified by researchers (Dunning, 2012), a

standard challenge in research that employs experimental or quasi-experimental designs in the field

(e.g., Hyde, 2007). Second, we account for other sources of variation that may confound our

estimates with exhaustive analyses from a wide variety of available data sources (Dunning, 2012).

We develop this argument more fully in the next section.

Voter Turnout. Data on voter turnout comes from the IEC and show the total number of

ballots cast by polling center.9 Candidates run “at-large” within a province and not with respect to

any smaller constituency boundary. Voters cast a single non-transferable vote, although each

province yields multiple members to parliament based on its population size. Therefore, winning

candidates are those who garner the most votes within the province corresponding to the number

of seats allocated to each province (from thirty-three in Kabul to two in Panjshir). The IEC reports

that 3,642,444 votes were cast by the end of polling, or about 40% of the maximum number of

voters (9,203,586).10

Insurgent Violence. As a measure of violence affecting civilians, we use recently declassified

incident reports submitted by ISAF forces and Afghanistan military and police forces that report

combat occurring between ISAF units and insurgents, commonly known as ‘significant activity’ or

SIGACTs. These data, secured and prepared by Shaver and Wright (2016), cover the period

9 “A polling station is the location where a person votes. Each station is part of a polling center. A

center generally is in a geographical location such as a school or mosque. In 2010, the smallest

center had two stations; the largest had 14” (NDI, 2011, 32).

10 Afghan Analysis Network. Available from: http://aan-afghanistan.com/index.asp?id=1066.

16

January 2002 through January 2015. The date, time, and georeferenced location of each incident is

provided. In this article, we use the subset of that dataset of insurgent attacks on ISAF and Afghan

forces for the period March through December 2010 (29,324 incidents). These data include all

records of the following attack types: direct fire; indirect fire; improvised explosive device; surface-

to-air fire; mine strike; unexploded ordnance; assault; assassination; arson; and sabotage. We

analyze the impact of police deployment on the sum of all these categories. We create a count of

incidents within a radius of one kilometer for each polling center to isolate the impact of police in

the immediate vicinity of the location they deployed to protect.11 While the SIGACTs data do not

measure all violence that civilians experience and that would plausibly affect their behavior and

attitudes, it is the most complete and comprehensive dataset measuring violence during this time

period. 12 The summary statistics of outcomes and descriptive variables at polling centers are in SI

Table 1.

11 The Afghan Government mandated a force laydown of one kilometer around polling centers for

the ANP. Long was privy to information on force laydown protocols in his capacity as an

accredited observer for the 2010 election. Our interviews with officials confirm this force laydown

plan (Condra, 2014a, 2014b, 2014c), as does independent election reports (FEFA, 2011, 46).

Results are robust to using a 2 km-buffer instead.

12 Other studies have demonstrated that civilians bear considerable risk of being killed or wounded

during the course of SIGACTs (Condra and Shapiro, 2012). As a measure of electoral violence,

SIGACTs are highly preferable to other possible measures primarily because others (e.g.,

UNAMA, National Democratic Institute) do not provide information at sufficient temporal or

geographic specificity to enable matching incidents to polling center locations.

17

Surveys. We use three different survey data sets related to Afghans’ opinions about police and

corruption. The first dataset comes from ISAF’s ANQAR survey waves (Berman et al., 2011;

Blair, Imai, and Lyall, 2014). The samples for the surveys are a nation-wide poll of Afghans aged

eighteen or older, and UN and World Food Program population statistics (region, province, and

district-level) are used to draw the sample. We use waves 7 (9,191 respondents) and 8 (10,388

respondents), fielded in March 2010 and May/June 2010.

The second dataset comes from two surveys that we designed and administered of households

living in the immediate vicinity of polling sites. We fielded the first baseline survey in August 2010

and the endline survey in December 2010. Our baseline sample comprises 450 polling centers in

nineteen of thirty-four provincial capitals in Afghanistan and our endline comprises 471 polling

centers (7.8% of polling centers operating on election-day), matching the baseline sample with

twenty-one polling centers added in Kabul after receiving additional funding. We selected our

sample of 471 polling centers by identifying polling centers scheduled to open on election day and

deemed secure by ISAF and ANP for the safety of our field staff. The baseline survey contained

2,904 respondents and the endline, 3,100 respondents.13

The third dataset is geocoded survey responses from 369 Afghan civilians between August and

October 2010 in eighty-four villages of Maiwand and Arghandab districts of Kandahar province

collected by a commercial entity with experience in marketing and psychological data collection.

The data record general demographic information and answers to questions about perceptions of

13 To obtain a representative sample of respondents living near polling sites, enumerators

employed a random walk pattern starting at the polling site, with random selection of every fourth

house or structure. Respondents within households were randomly selected using the Kish grid

method.

18

the Taliban, the ANA, the ANP, and ISAF. Villages were selected within key terrain districts with

between five and twenty individuals surveyed, depending on population density.

Estimating Police’s Effect on Turnout and Insurgent Violence

In this section, we estimate the effect of assigned police deployment on both insurgent violence

and voter turnout at polling centers across Afghanistan. As discussed above, these estimates should

be considered ITT and are indicative of the potential effect of increased police forces. We begin

by showing that there is a random component of the allocation of police deployment that can be

exploited to identify the deployment’s effect on our two outcomes of interest; that is, the polling

center area’s history of violence does not fully predict the level of police deployment. This is

important for establishing a source of plausibly exogenous variation which allows us to disentangle

actual violence on and around the election from the history of violence which informed police

deployments.

Having established this empirically, we proceed to estimate the effects of interest using linear

regression models and discuss those results. In the Supporting Information, we address a set of

factors that could plausibly affect both police deployment and turnout, rendering our main results

spurious. The results from these tests of competing explanations substantiate our claim of

exogenous variation in treatment assignment and strengthen confidence in our results.

To begin, within districts the security classification assigned to a polling center is not a function

of previous trends in insurgent-ISAF violence, which is the most readily available indicator that

citizens would have of risk on election day. Our approach shows that neither the trends in violence

nor the levels of contestation in the vicinity of a polling center predict its security status. To

measure the trends in violence we use four lags of weekly violence. For levels of contestation, we

use a cubic polynomial in average violence over the first part of the 2010 fighting season. To

19

measure time-invariant regional factors we use province or district fixed effects. We estimate the

models only on the subset of 1,823 polling centers that were not obviously fraudulent (described

below) and that were in one of the 251 districts that had at least one medium- or high-security

polling center.

In our full model we estimate the following using Ordinary Least Squares:

𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖 = 𝛽𝛽1�𝑉𝑉𝑖𝑖,𝑡𝑡−1� + 𝛽𝛽2�𝑉𝑉𝑖𝑖,𝑡𝑡−2� + 𝛽𝛽3�𝑉𝑉𝑖𝑖,𝑡𝑡−3� + 𝛽𝛽4�𝑉𝑉𝑖𝑖,𝑡𝑡−4� + 𝛾𝛾1�𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖 � + 𝛾𝛾2(𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖2) +

𝛾𝛾3(𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖3) + 𝑑𝑑𝑖𝑖 + 𝜇𝜇𝑖𝑖,𝑡𝑡,

where, 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖 is the security classification assigned to the polling center (low, medium, or high),

the 𝑉𝑉𝑖𝑖,𝑡𝑡−𝑘𝑘 are four lags of insurgent violence (1 to 4 weeks prior to the election), and 𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖 is the

average weekly violence in the earlier part of the 2010 fighting season (weeks t-5 to t-22, April

through July), and 𝑑𝑑𝑖𝑖 is a district fixed effect (or province in some specifications). We report robust

standard errors clustered at the district level because that is the geographic level at which the ANP

and Taliban operational command structures are typically organized.

If previous violence predicted treatment assignment, we would expect to see this reflected in

the results of these models, shown in SI Table 2. The lags of violence and cubic polynomial in

fighting season averages generally do not predict treatment assignment very well (column 1),

explaining less than 1% of the variance in security classifications. In column 2, we add province

fixed effects to the model, which improves the model fit marginally, bringing the r2 up to 0.07.

Adding district fixed effects in column 3 brings the explained variance up to 32%. In the final

20

model (column 4), we exclude the weekly lags of violence and the cubic polynomial in average

previous violence remains jointly significant.14

While police deployment was designed to account for the violence previously experienced in

an area, the evidence shows that violence in the vicinity of polling centers does not do a good job

of explaining variation in treatment. As we noted, this is in part due to the police assignments being

based on provincial and higher level data while we can rely on much more local variation in

violence levels. We therefore proceed to estimate the effect of policing on turnout and violence

levels around election time, controlling for the previous history of violence at each polling center.

Police’s Effect on Violence. Our analysis of police deployment’s effect on violence and

turnout is limited to a subsample of the full 5,524 polling centers in operation in the election. For

inclusion, the polling center must meet three criteria. First, we omit all 1,324 centers that reported

turnout in excess of an average of 590 votes across the polling sites within that center. The reason

for this is to avoid including in our analysis polling centers where electoral fraud occurred. Polling

sites within each center were set up to have no more than 600 ballots cast, so any polling center

whose average site was close to that level is suspect and the IEC used this threshold in their

decisions to nullify results (DI, 2011, 33).15 If the average turnout across the polling sites is that

14 F - tests for joint significance on violence lags show that we can reject the null hypothesis that the

lags are jointly zero in these specifications.

15 In the 2009 presidential election, all ballot boxes containing 600 or more votes were recounted,

many of which showed physical evidence of manipulation (Weidmann and Callen, 2013). Callen

and Long (2015) discuss the mechanics of electoral fraud in the 2010 election and why 590 votes

or more per station is a likely indicator of fraud, especially those dedicated for female voters who

were much less likely to turn out than men.

21

high, fraud in the form of artificial ballot-stuffing or inflated vote totals in the tallying process likely

occurred. All results reported below are substantively stronger when we drop these polling centers

from the analysis, but the statistical significance of the estimated coefficients are largely unchanged.

This is consistent with the possibility that overt fraud is positively correlated with police presence.

Second, we exclude all 1,502 non-fraudulent polling centers located in districts that do not

have at least one polling center with a medium or high security provision classification. In such

districts there is no variation in treatment and our identification strategy relies on within-district

variation in security classifications.

Finally, to enable the difference-in-differences estimation described below, we exclude the

1,518 centers that were operational in 2010, but not in 2009 (of which 152 meet our other criteria).

All results on 2010 turnout are robust to including these polling centers.16 After this pruning, we do

our analysis on 1,823 out of a possible 2,031 polling centers that have security classification

designations from the Ministry of Interior in 2010.

To estimate the impact of security status on violence we estimate security classification on a

number of measures of changes in violence. For all regressions, we measure police deployments

two ways: first as a binary variable which takes a value of ‘1’ if the polling center was classified to

receive any additional police deployment; and second, we include dummy variables for polling

centers that received ‘medium’ or ‘high’ security provision classifications. That latter specification is

estimated as the following using OLS:

16 The results in SI Table 2 – predicting treatment assignment with past violence – are substantively

and statistically very similar if we include polling centers that were not operational in 2009,

indicating that there is no bias resulting from dropping these polling centers in our core turnout

and violence results presented below.

22

Δ𝑉𝑉𝑖𝑖 = 𝛼𝛼1(𝑀𝑀𝑖𝑖) + 𝛼𝛼2(𝐻𝐻𝑖𝑖) + 𝛽𝛽1�𝑉𝑉𝑖𝑖,𝑡𝑡−1� + 𝛽𝛽2�𝑉𝑉𝑖𝑖,𝑡𝑡−2� + 𝛽𝛽3�𝑉𝑉𝑖𝑖,𝑡𝑡−3� + 𝛽𝛽4�𝑉𝑉𝑖𝑖,𝑡𝑡−4� + 𝛾𝛾1�𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖 � +

𝛾𝛾2(𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖2) + 𝛾𝛾3(𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖3) + 𝑑𝑑𝑖𝑖 + 𝜇𝜇𝑖𝑖,𝑡𝑡,

where Δ𝑉𝑉𝑖𝑖 represents a difference in violence within one kilometer of the polling center as

described below, 𝑀𝑀𝑖𝑖 and 𝐻𝐻𝑖𝑖 are dummy variables for whether a polling center was assigned a

medium or high security provision classification, the 𝑉𝑉𝑖𝑖,𝑡𝑡−𝑘𝑘 are lags of weekly violence prior to the

last period of the difference, and the 𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖 are prior fighting season violence as before.

The left-hand side variable across the specifications in SI Table 3 is a series of differences in

violence within one kilometer of the polling center: election week versus the week prior to election

(columns 1-2), election week versus the average violence in the four-week run up to the election

(columns 3-4), the average violence during the four weeks after the election versus the average

violence during the four weeks prior to the election (columns 5-6), and the average violence during

the eight weeks after the election versus the average violence during the eight weeks prior to the

election (columns 7-8). All models include district fixed effects and robust standard errors are

clustered at the district level.

As the results show, there is not a strong discernible effect of security status of a polling center

on the change in violence experienced at polling centers before and during/after the election.

Police’s Effect on Turnout. To assess the effect of deployment of extra police on turnout

recorded at polling centers we estimate a series of regressions like:

Δ𝑇𝑇𝑖𝑖 = 𝛼𝛼1(𝑀𝑀𝑖𝑖) + 𝛼𝛼2(𝐻𝐻𝑖𝑖) + 𝛽𝛽1�𝑉𝑉𝑖𝑖,𝑡𝑡−1� + 𝛽𝛽2�𝑉𝑉𝑖𝑖,𝑡𝑡−2� + 𝛽𝛽3�𝑉𝑉𝑖𝑖,𝑡𝑡−3� + 𝛽𝛽4�𝑉𝑉𝑖𝑖,𝑡𝑡−4� + 𝛾𝛾1�𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖 � +

𝛾𝛾2(𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖2) + 𝛾𝛾3(𝑉𝑉𝑉𝑉𝑉𝑉𝑃𝑃�������𝑖𝑖3) + 𝑑𝑑𝑖𝑖 + 𝜇𝜇𝑖𝑖,𝑡𝑡,

23

where the only difference from the violence regressions is that we look at the differences in turnout

since 2009.

Table 1A reports results using the binary measure of police, with the least secure polling

centers receiving a value of ‘1’ on the police variable (i.e., high or medium deployment) and the

most secure category receiving a value of ‘0’ on the variable (i.e., low deployment). Table 1B

reports results of models that include dummy variables for the medium and high security

classifications (high/medium/low deployment) instead, with the most secure classification as the

omitted category.

We measure turnout at the polling center level, 𝑇𝑇𝑖𝑖, in three different ways. In columns 1-3 of

both panels, we use the reported raw turnout in the 2010 election. In column 1, we control for

predicted levels of violence at polling centers, generated from a linear regression of election day

violence on four (week) lags of violence, violence in the previous five months, and the squared and

cubic levels of violence. Column 2 includes the residual of that regression as a control. Finally, in

column 3, we control for previous levels of violence in the area. In both panels, there is an

estimated negative effect of having a higher security classification (medium or high police) on raw

turnout.

This relationship is confirmed when we use the difference in turnout between the 2009 and

2010 elections as the dependent variable (columns 4-6).17 In this specification, a polling center that

17 The election held in 2005 might seem like a more appropriate comparison case for this analysis,

given that it, too, was parliamentary, while the 2009 election was presidential. However, no polling

station level data was released for the 2005 parliamentary election, which renders comparison

impossible. But on comparability, the security situation changed in many parts of the country

between 2005 and 2010 (Coburn and Larson, 2014; Vendrell, 2011), and insurgents’ strategic use

24

is designated to receive more police (i.e., medium or high) experiences a statistically and

substantively significant decrease in turnout relative to the previous year’s election, an estimated

24% change. Note that the point estimate of the treatment effect on turnout hardly changes across

specifications, whether turnout is analyzed in the cross-section or is differenced. As expected, the

precision of the differenced estimate decreases because of added noise and reduced power relative

to the cross-sectional equation. Columns 7-9 in both panels show the estimated effect when we use

the difference in standardized turnout numbers (such that the variable has a mean of 0 and

standard deviation of 1) within each election as our turnout measure. This ensures that the effect is

not driven by different distributions across the elections, allowing for maximum comparability of

turnout results. Similarly, polling centers that receive more police experience a decrease in turnout.

Taken together, these results are striking. We make use of the fact that previous levels of

violence do not predict security classification of polling centers, and that after controlling for

previous levels of violence, the majority of variation in treatment assignment is left unexplained.

This allows for the identification of the effect of extra police on turnout and violence. While extra

police may have had the effect of slightly decreasing the amount of violence at their polling centers

relative to others with less police, we see some evidence that the police rollout had an unintended

and deleterious effect on turnout. We refer the reader to this article’s Supporting Information,

which provides numerous robustness checks on our main results, as well as tests of alternate

explanations for them.

of violence in the context of elections also changed over that period (Giustozzi, 2008). For these

reasons, the 2009 election is more suitable for comparison because of the need to account for the

effect violence may have had on voting behavior.

25

We interpret these results with some caution, however. As noted, the lack of data on actual

security force deployments raises two important issues for our results. First, we cannot definitely

rule out the possibility that the data on security classifications of polling centers does not perfectly

represent security force deployment around the election. If, for example, provincial-level police

chiefs did not follow the classification protocol, our attempt to identify the effect of policing on

violence and turnout is flawed, and the results cannot be interpreted as a reliable estimate of the

relationship under study. We note that our first-hand observation of the treatment assignment

process and the Afghan government’s planning of the rollout, as well as interviews with officials

who had responsibility for the rollout, give us confidence in the validity of our assumptions. In

addition, these differences can at worst be interpreted as ITT estimates in which the assignment of

higher police levels is associated with lower violence. Second, we cannot relate the security

classification designation of polling centers to actual levels of increased policing because we do not

have those data. The magnitude of our ITT estimate, however, suggests that the protocols were

followed.

Explaining the Results

Citizens and observers frequently view the government, and in particular the police, as corrupt in

Afghanistan. We advance a plausible explanation of the results that voters stayed away from polling

centers in 2010 with higher levels of police deployment in order to avoid interaction with the

police, because every voter would have confronted multiple police officers in the process of voting

due to security deployments on election day. To further develop our argument, we first present

descriptive data from multiple sources attesting to the variation in these attitudes in the population

and then provide survey evidence.

26

First, the competence of the police force is heavily criticized and the government poorly

manages this institution. Attrition and desertion rates are high, and the Ministry of the Interior has

been unable to attract high quality, educated recruits (Giustozzi and Isaqzadeh, 2012). In 2011, the

Minister admitted that 90% of the force was illiterate. Data from the survey enumerated in

Kandahar and almost exclusively to Pashtuns also attests to this problem: 72% of respondents

agreed a little or completely with the statement, “ANP officers in my area are illiterate” (SI Table

4). Drug addiction is also a problem, as is involvement in the narcotics trade (Giustozzi and

Isaqzadeh, 2012).

Second, aside from incompetence, there is considerable evidence that the police treat civilians

poorly, in a variety of ways (Giustozzi, 2008). Giustozzi and Isaqzadeh (2012) document a litany of

corruption allegations from witnesses throughout Afghanistan. The authors interviewed truck

drivers who reported being regularly asked for bribes on the primary ring road connecting

Afghanistan’s major cities. A 2010 survey focusing on police indiscipline uncovered the practice of

guns being taken from recruits before going off duty because recruits were using them to rob

civilians. Police also frequently steal from civilians (Giustozzi and Isaqzadeh, 2012). The Kandahar

survey (SI Table 4) revealed that 73% of respondents disagreed a little or a lot to the statement,

“ANP officers treat members of the local community with respect”; similarly, 74% disagreed with

the statement, “ANP officers are well respected by local people.” A full 62% of Kandahar

respondents agreed with the statement, “ANP officers in my area sometimes beat people up.”

Police reportedly have engaged in rape, torture and extrajudicial executions of civilians (Giustozzi,

2008, 176).

There is further evidence in public opinion surveys that Pashtuns had a particularly challenging

relationship with the ANP, who are commonly thought to be dominated by Tajiks and Uzbeks,

ahead of the 2010 election. Of 19,579 respondents in the two ANQAR waves of quarterly public

27

opinion surveys immediately before the election (SI Table 4), 31% of Pashtuns reported seeing the

police engage in corrupt acts while only 17% of Tajiks and 9% of Uzbeks reported the same.

These are not simply reflections of the geographic dispersion of corruption and ethnicity. Pashtun

respondents are significantly more likely than others to report having seen corruption even when

adding province fixed-effects to a range of regression models in the ANQAR data. When

respondents to the Kandahar survey were asked to state their level of agreement with the

statement, “most ANP officers are corrupt,” 64% agreed, either a little or completely. To the

statement, “ANP officers put the interests of their community before their own interests,” 70%

disagreed.

To further support our argument, we use responses from surveys taken before the election to

investigate variation across polling centers in attitudes toward the ANP and the Afghan National

Army (ANA) as a function of security classification of the polling centers. Respondents were

sampled in the immediate vicinity around the polling centers, so are located in the same area as the

insurgent violence that we include in our models. Our survey asked whether people felt that the

presence of the ANP and ANA would alter the safety of their polling center.18 We regress

responses to this question (-1 = Less Safe; 0 = No Difference; 1 = Safer) on the security

classification of the polling center and measures of insurgent violence in the five months prior to

the election (regression results shown in SI Table 5). While the estimates are somewhat sensitive to

the inclusion of province and district fixed effects, in general, the evidence suggests that in places

that were designated to receive more police, respondents felt the ANP made the area less safe.

The intensity of concern with ANP presence is stronger in places where more police were

18 “In your opinion, does the presence of the army [police] near your polling centre make it: Safer,

Less Safe, No Difference?”

28

deployed. This is not simply an artifact of the security classifications leading to generalized

disaffection with the state, as we see no evidence of this in the case of the ANA. While one would

not hold up the ANA as a model of professionalism, it is generally seen as a less frequent violator

of law and human rights than the ANP, perhaps in part because the ANP grew out of local Afghan

militias, whereas the ANA was more carefully built.19 We have few observations in our survey from

polling centers designated to receive “medium” or “high” police deployment, so we do not place

too much weight on the results. But they are consistent with our argument that it is the police in

particular that people want to avoid, not just any security provider.

We present more evidence in support of our argument in SI Table 6, which shows how police

deployments around the election affected variation in responses to questions designed to evaluate

people’s views of the police and the government more generally. An observable implication of our

theory is that negative views of the police should correlate with other proxies related to views of the

government (as the police are the main representatives of the government visible at the elections):

appropriate authority for dispute resolution, importance of paying taxes, and performance of the

central government (columns 3-5). Presence of the police should be less likely to correlate with

more general opinions about Afghanistan’s regime type and satisfaction with Afghan democracy

(columns 1-2).

19 Giustozzi (2008) observes that “the new national army created from May 2002 was a substantially

different force from both the various militias and the police. It was in fact the only Afghan security

force to be created from scratch on a professional basis. It was widely touted as one of the few

success stories of post-2001 Afghanistan, and compared to abject failures such as the formation of

a police force with the name it was indeed a success.”

29

Panel A regresses the pre-election (baseline) mean response value to these questions at the

polling center-level on security classification, as well as previous levels of violence, showing the

security classifications were not conditionally correlated with views before forces were deployed.

Panel B provides regression results from post-election (endline) responses as the dependent

variable, showing a general reduction in three of five questions for places that had a medium or

high security designation. In Panel C, we take the difference between post-election and pre-election

responses as the dependent variable and regress that change on security classification and previous

violence.

Focusing on the differenced results in Panel C, we see that attitudes about democracy do not

change much with security status (columns 1-2).20 We see substantial negative movements,

however, in responses to specific questions that should be affected by interaction with a corrupt

police force. First, the percentage of respondents who would trust either district officials or the

police in a dispute decreases pre- to post-election in areas designated to receive more police

(column 3).21 Second, respondents in areas receiving additional security forces become less likely

to view paying taxes as important (column 4), though this is not a statistically strong result.22 Third,

respondents’ rating of how well the central government is doing its job decreases (column 5).23

These results support what we label a corruption mechanism linking security force deployment to

reduced turnout, consistent with our theory that focuses on the police as a symbol of legitimate

20 “In your opinion, is Afghanistan a democracy?”; “Overall, how satisfied are you with the way

democracy works?”

21 “If you had a dispute with a neighbor, who would you trust to settle it?”

22 “How important is it to pay taxes?”

23 “Does the central government do an [excellent/good/just fair/poor] job?”

30

government and law and order (Rashid, 2008). Our argument is that citizens’ prior experience with

the police’s corrupt and predatory behavior caused them to avoid polls altogether where more

police were present, to avoid potentially injurious interaction with police.24

Finally, a plausible alternative to our corruption mechanism is what might be called a signaling

mechanism: the increased pre-election presence of police around a polling center might ‘signal’ to

citizens that they should expect heightened violence on election day, causing them to stay home.

Given that the Taliban warned of violence on and around election day in an effort to deter voting

(Farmer, 2010; NDI, 2011; Adler, 2002), this represents a threat to our argument. In their

ethnographic study of the 2010 Afghan elections, Coburn and Larson (2014) emphasize the

24 Our results do not preclude the possibility that conditional on the decision to turn out, some

voters cast ballots motivated by views of corruption or legitimacy, for example by voting against

politicians they viewed as closely allied with the sitting government. While the implication of this

argument in a presidential election would be anti-incumbent voting, in the context of the 2010

parliamentary election, this argument’s empirical implications in terms of voting behavior are not

as clear. Afghanistan lacked political parties, a large number of candidates ran for multiple seats

within the province, many sitting members of the Wolesi Jirga were seen as anti-

government/Karzai given other political allegiances, and many politically powerful candidates not

in government nonetheless held close ties to the government (Callen and Long, 2015; Rashid,

2008). For these reasons, it is impossible to gain empirical leverage on how to measure which

candidates would have been seen as “pro” or “anti” government at that time, particularly for the

individual voter (for which data do not exist). Moreover, our theory, hypotheses, data, and

empirical analysis focus on the decision to turn out, not the determinants of the vote for citizens

who cast ballots.

31

heightened awareness of violence during this time. “More of the people we interviewed and spoke

with in 2010 seemed to be taking the threat of violence seriously than they did in 2009,” and “it

was more the threat of violence in the days leading up to voting [in 2010] that reshaped individual

choices than the actual instances of violence that did eventually occur [on election day].”

To rule this out as a systematic explanation of behavior, we look at our surveys prior to, and

after, the election and make use of responses to a question about why respondents did not vote (or

plan to vote, in the baseline survey). Those respondents who reported that they would not vote

(baseline) or did not vote (endline) in the 2010 election were asked for a reason. We regress the

number of respondents indicating “insecurity; I fear/ed attacks” on a binary variable indicating the

polling center’s security classification (medium/high vs. low security). As in the other models, we

include violence over the previous five months, as well as the squared and cubed terms, on the

right hand side. Results are reported in SI Table 7.

Based on the survey data, there is no evidence that people living around medium or high

security-classified polling centers planned on not voting because they anticipated higher violence

there on election day (column 1). (Indeed, if anything, previous violence is actually negatively

correlated with identifying insecurity as a reason for not voting.) Similarly, when queried after the

election had passed, being located near a medium/high security polling center is not associated

with a higher likelihood of pointing to violence and insecurity as a reason for not voting (column

2).

This systematic evidence against this mechanism actually is corroborated by assessments of this

question from scholars who observed behavior first-hand and expected that anticipated violence

would drive people away from the polls. For example, while Coburn and Larson (2014) point to

the threat of violence as a factor in voting behavior, they also revealingly admit that in spite of this

32

threat, “relatively few of the people we talked to appeared to actually decide not to vote based on

this threat”, a point strongly supportive of our evidence and argument.

Our evidence in this section supports the interpretation of the negative effect that police had

on turnout in the 2010 election, and bear directly on the legitimacy of the Afghan state. The

evidence supports two key arguments. First, descriptively, many people view the ANP as corrupt,

dishonest, predatory, and even brutal. Second, views of the ANP deteriorated markedly around

polling centers that were to receive more police (SI Table 5), and exposure to the police also

negatively impacted Afghans’ view of the police and government performance to provide two key

public goods essential to state-building: justice and redistribution (SI Table 6). The deployment of

police might therefore reduce citizen’s expression of consent and desire to participate in practices

that support the government, like voting.

Conclusion

The U.S. and other donors spend millions of dollars a year on democracy assistance, but

conferring legitimacy via elections is particularly difficult in places like Afghanistan. As one scholar

writes, “in countries with protracted and deep-rooted conflicts, elections alone can rarely, if ever,

confer legitimacy on a particular government” (Gillies, 2011, xx). Echoing this sentiment, the UN

envoy to Afghanistan, speaking on the eve of the election in 2010, observed, “This is probably one

of the worst places and the worst times to have an election anywhere in the world” (quoted in

Farmer, 2010). Since the invasion of Afghanistan in 2001, the international community and the

Afghan government have focused enormous attention and money on holding elections and

strengthening the police. In this article, we propose a framework for understanding how the push

for democratization to increase state legitimacy through elections relates to the push for security

provision through policing. This theoretical link should inform how we think about strengthening

33

both of these institutions going forward in Afghanistan and other conflict prone countries,

particularly given our results and the public’s perceptions of security forces.

On one hand, our results are encouraging. To the extent that police were deployed to places

insurgents wanted to attack, a lack of correlation between deployments and violence suggests that

the police, in spite of endemic corruption and charges of incompetence, deterred election-related

violence. Our systematic assessment of police performance on this dimension is in line with

independent observers of the election, who reported that “security agents effectively protected

polling centers”, and ninety-two percent of polling stations opened in spite of threats of violence

(DI, 2011, 30-31). This fulfilled a main goal of the extra rollout, and may reflect the ANP basic

training curriculum’s heavy focus on counterinsurgency activities and much less emphasis on

human rights, as well as the fact that some police were trained by NATO military personnel, not

police officers (NDI, 2011, 21-22). We are unaware of any other empirical evidence showing that

Afghan security forces (army or police) have been successful in deterring insurgent violence.25 The

fact that the police had this effect is no small feat. Throughout the history of modern state building,

attaining this capability has proven to be a monumental challenge (Ferejohn and Rosenbluth, 2010;

Herbst, 2000; Scott, 2010; Thomson, 1994), and while the context of this performance is limited

and we caution against generalizing beyond it, it stands as a positive sign in the development of the

ANP.

On the other hand, our analysis is discouraging for the prospects of democratization and the

legitimization of the Afghan state. While the police may have deterred violence, they also deterred

25 While ISAF and organizations like the UN track insurgent violence over time across districts and

provinces, we should not interpret changes over time as the causal result of improved or

diminished security force capability.

34

voter turnout and hurt key indicators of democratic growth. This is problematic for the

development of participatory democracy, especially given that others have demonstrated the

painstakingly slow pace of political change in Afghanistan even with sustained and concentrated

programs designed to promote such attitudinal and behavioral transformation (Beath, Christia, and

Enikolopov, 2013). The quantitative analysis presented here agrees qualitatively with the concerns

of election observers and the Afghan election commission about turnout and the implied lack of

government support (Aikins and Hewad, 2010; Maroney, 2010; Rubin and Gall, 2010). Our

interpretation of this unfortunate result is that citizens avoided the polls where they knew extra

police were present, not because they feared higher violence, but because they see police as

corrupt and therefore do not want to interact with them or express consent by electing a

government they do not support.

We believe our results speak to core theoretical and policy issues that many countries

besides Afghanistan face when they hold elections in the shadow of violence and corruption.

Regardless of the perceived benefits or problems of having elections in these contexts, the

empirical record clearly demonstrates that the international community and domestic governments

will continue to use elections as a core benchmark to transition to democracy after conflict. While

scholars and policymakers debate the merits of electoral timing in post-conflict settings and the

importance of increasing the security sector, a resumption of violence is not the only important

consideration: how citizens view the police responsible for keeping elections safe will also impact

the degree to which such processes potentially strengthen or erode democratization.

35

REFERENCES

Adams, G., and R. Sokolsky. 2015. Governance and Security Sector Assistance: The Missing Link -- Part I. Lawfare Blog, July 12.

Adler, E. 2002. Constructivism and International Relations. In Handbook of International Relations, edited by W. Carlsnaes, Thomas Risse and Beth A. Simmons. London: Sage Publications.

Aikins, M., and G. Hewad. 2010. Losing Legitimacy after Afghanistan's Elections. The AfPak Channel, October 24.

Beath, A., F. Christia, and R. Enikolopov. 2013. "Empowering Women through Development Aid: Evidence from a Field Experiment in Afghanistan." American Political Science Review 107 (3):540-557.

Bellows, J., and E. Miguel. 2008. "War and Institutions: New Evidence from Sierra Leone." African Economic Development 96 (2):394-399.

Berman, E., M. Callen, J. H. Felter, and J. N. Shapiro. 2011. "Do Working Men Rebel? Insurgency and Unemployment in Afghanistan, Iraq, and the Philippines." Journal of Conflict Resolution 55 (4):496-528.

Berman, E., M. J. Callen, C. C. Gibson, and J. D. Long. 2014. "Election Fairness and Government Legitimacy in Afghanistan." NBER Working Paper No. 19949.

Berman, E., J. H. Felter, J. N. Shapiro, and E. Troland. 2013. "Modest, Secure and Informed: Successful Development in Conflict Zones." American Economic Review: Papers and Proceedings 103 (3):512-517.

Berrebi, C., and E. Klor. 2006. "On Terrorism and Electoral Outcomes: Theory and Evidence from the Israeli-Palestinian Conflict." Journal of Conflict Resolution 50 (6):899-925.

Biddle, S., J. A. Friedman, and J. N. Shapiro. 2012. "Testing the Surge: Why Did Violence Decline in Iraq in 2007?" International Security 37 (Summer):7-40.

Binkerhoff, D. W., ed. 2007. Governance in Post-Conflict Societies. London: Routledge. Blair, G., K. Imai, and J. Lyall. 2014. "Comparing and Combining List and Endorsement

Experiments: Evidence from Afghanistan." American Journal of Political Science 58 (4):1043-1063.

Blattman, C. 2009. "From Violence to Voting: War and Political Participation in Uganda." American Political Science Review 103 (2):231-247.

Brancati, D., and J. L. Snyder. 2011. "Rushing to the Polls: The Causes of Premature Postconflict Elections." Journal of Conflict Resolution 55 (3):469-492.

Brancati, D., and J. L. Snyder. 2013. "Time to Kill: The Impact of Election Timing on Postconflict Stability." Journal of Conflict Resolution 57 (5):822-853.

Bratton, M., R. Mattes, and E. Gyimah-Boadi. 2004. Public Opinion, Democracy, and Market Reform in Africa. New York: Cambridge University Press.

Bremer, L. P. 2006. My Year in Iraq. New York: Simon & Schuster. Bush, S. S. 2015. Taming of Democracy Assistance. New York: Cambridge University Press. Callen, M., and J. D. Long. 2015. "Institutional Corruption and Election Fraud: Evidence from a

Field Experiment in Afghanistan." American Economic Review 105 (1):354-381. Carothers, T. 2007. "The 'Sequencing' Fallacy." Journal of Democracy 18 (1):12-27. Chandra, K. 2004. Why Ethnic Parties Succeed: Patronage and Ethnic Head Counts in India.

Cambridge: Cambridge University Press. Chandrasekaran, R. 2007. Imperial Life in the Emerald City. New York: Vintage Books.

36

Coburn, N., and A. Larson. 2014. Derailing Democracy in Afghanistan. New York: Columbia University Press.

Collier, P. 2009. Wars, Guns, and Votes. New York: Harper. Collier, P., and P. C. Vicente. 2014. "Votes and Violence: Evidence from a Field Experiment in

Nigeria." Economic Journal 124 (574):F327-F355. Condra, L. N. 2014a. Interview with Afghanistan Ministry of Interior official. Kabul,

Afghanistan, June 4. Condra, L. N. 2014b. Interview with former Afghanistan Independent Elections Commission

official. Kabul, Afghanistan, June 7. Condra, L. N. 2014c. Interview with former Afghanistan Ministry of Interior official. Kabul,

Afghanistan, June 2. Condra, L. N., and J. N. Shapiro. 2012. "Who Takes the Blame? The Strategic Effects of

Collateral Damage." American Journal of Political Science 56 (1):167-187. Cox, G. W. 1997. Making Votes Count. Cambridge: Cambridge University Press. Cox, G. W., and M. D. McCubbins. 1986. "Electoral Politics as a Redistributive Game." Journal

of Politics 48 (2):370-389. DI. 2011. Democracy International Election Observation Mission Afghanistan Parliamentary

Elections 2010: Final Report. In Available from: http://www.democracyinternational.com/sites/default/files/DI%20Afghanistan%202010%20EOM%20Final%20Report_web.pdf. Bethesda, MD: Democracy International, Inc.

Di Tella, R., and E. Schargrodsky. 2004. "Do Police Reduce Crime? Estimates Using Allocation of Police Forces after a Terrorist Attack." American Economic Review 94 (1):115-133.

Diamond, L. 2006. "Promoting Democracy in Post-Conflict and Failed States." Taiwan Journal of Democracy (2):93-116.

Doyle, M. W., and N. Sambanis. 2006. Making War and Building Peace. Princeton, NJ: Princeton University Press.

Draca, M., S. Machin, and R. Witt. 2011. "Panic on the Streets of London: Police, Crime, and the July 2005 Terror Attacks." American Economic Review 101 (5):2157-2181.

Dunning, T. 2012. Natural Experiments in the Social Sciences. New York: Cambridge University Press.

Farmer, B. 2010. Afghan voters face blizzard of threats ahead of Saturday's parliamentary elections. The Telegraph, September 17.

FEFA. 2011. Afghanistan Parliamentary Election Observation Mission 2010: Free and Fair Election Foundation of Afghanistan.

Felbab-Brown, V. 2013. Aspiration and Ambivalence. Washington, DC: Brookings Institution Press.

Ferejohn, J. A., and F. M. Rosenbluth. 2010. War and State Building in Medieval Japan. In War and State Building in Medieval Japan, edited by J. A. Ferejohn and F. M. Rosenbluth. Stanford, CA: Stanford University Press.

Finnemore, M., and M. N. Barnett. 1999. "The Politics, Power, and Pathologies of International Organizations." International Organization 53 (4):699-732.

Fortna, V. P. 2008. Does Peacekeeping Work? Shaping Belligerents' Choices after Civil War: Princeton University Press.

Fukuyama, F. 2004. State-Building: Governance and World Order in the 21st Century. Ithaca, NY: Cornell University Press.

37

Gillies, D. 2011. Electoral Democracy and the Paradoxes of Peacebuilding. In Elections in Dangerous Places, edited by D. Gillies. Ithaca, NY: McGill-Queen's University Press.

Giustozzi, A. 2008. Koran, Kalashnikov, and Laptop. New York: Columbia University Press. Giustozzi, A., and M. Isaqzadeh. 2012. Policing Afghanistan. New York: Columbia University

Press. Gottfredson, M. R., and T. Hirschi. 1990. A General Theory of Crime. Stanford, CA: Stanford

University Press. Hafner-Burton, E. M., S. D. Hyde, and R. S. Jablonski. 2014. "When Do Governments Resort to

Election Violence?" British Journal of Political Science 44 (1):149-179. Herbst, J. 2000. States and Power in Africa: Comparative Lessons in Authority and Control.

Princeton, NJ: Princeton University Press. Höglund, K., A. K. Jarstad, and M. S. Kovacs. 2009. "The Predicament of Elections in War-Torn

Societies." Democratization 16 (3):530-577. Horowitz, D. L. 1985. Ethnic Groups in Conflict. Berkeley: University of California Press. Huntington, S. P. 1996. The Clash of Civilizations and the Remaking of World Order. New York:

Simon & Schuster. Hyde, S. D. 2007. "The Observer Effect in International Politics: Evidence from a Natural

Experiment." World Politics 60 (1):37-63. Hyde, S. D. 2011a. "Catch Us If You Can: Election Monitoring and International Norm

Diffusion." American Journal of Political Science 55 (2):356-369. Hyde, S. D. 2011b. The Pseudo-Democrat's Dilemma: Why Election Monitoring Became an

International Norm. Ithaca, NY: Cornell University Press. Hyde, S. D., and N. Marinov. 2012. "Which Elections Can Be Lost?" Political Analysis 20:191-

210. Iyengar, R. K. 2010. Meeting with Afghan Ministry of Interior officials. Kabul, Afghanistan,

April 22. Jung, D., and J. D. Long. 2015. Democracy, Insurgency, and Voter Turnout in Afghanistan.

University of Washington and Emory University. Kasara, K., and P. Suryanarayan. 2015. "When Do the Rich Vote Less Than the Poor and Why?

Explaining Turnout Inequality across the World." American Journal of Political Science 59 (3):613-627.

Kelley, J. 2012a. Monitoring Democracy: When International Election Observation Works, and Why It Often Fails. Princeton, NJ: Princeton University Press.

Kelley, J. G. 2012b. "International Influences on Elections in New Multiparty States." Annual Review of Political Science 15:203-220.

Klick, J., and A. Tabarrok. 2005. "Using Terror Alert Levels to Estimate the Effect of Police on Crime." Journal of Law and Economics 48 (1):267-280.

Lake, D. A. 2009. Hierarchy in International Relations. Ithaca, NY: Cornell University Press. Leach, M., and D. Kingsbury, eds. 2013. The Politics of Timor-Leste: Democratic Consolidation

After Intervention. Ithaca, NY: Southeast Asia Program Publications. Levi, M. 1988. Of Rule and Revenue. Berkeley, CA: University of California Press. Levitt, S. D. 1997. "Using Electoral Cycles in Police Hiring to Estimate the Effect of Police on

Crime." American Economic Review 87 (3):270-290. Lindberg, S. I. 2006. Democracy and Elections in Africa. Baltimore, MD: Johns Hopkins

University Press.

38

Lindberg, S. I. 2009. Democratization by Elections: A New Mode of Transition. Baltimore, MD: Johns Hopkins University Press.

Mansfield, E. D., and J. Snyder. 2005. Electing To Fight: Why Emerging Democracies Go To War. Cambridge, MA: MIT Press.

Maroney, S. 2010. Observers Debate Legitimacy of Afghanistan Election. Voice of America, September 19.

McFaul, M. 2010. Advancing Democracy Abroad. Lanham, MD: Rowman & Littlefield Publishers.

NDI. 2011. The 2010 Wolesi Jirga Elections in Afghanistan. Washington, DC: National Democratic Institute for International Affairs.

Norris, P. 2014. Why Electoral Integrity Matters. New York: Cambridge University Press. Posner, D. N. 2005. Institutions and Ethnic Politics in Africa. Cambridge, UK: Cambridge

University Press. Przeworski, A., S. C. Stokes, and B. Manin, eds. 1999. Democracy, Accountability, and

Representation. New York: Cambridge University Press. Rashid, A. 2008. Descent into Chaos. New York: Viking. Rashid, A. 2012. Pakistan on the Brink. New York: Viking. Roessler, P. G. 2005. "Donor-Induced Democratization and the Privatization of State Violence in

Kenya and Rwanda." Comparative Politics 37 (2):207-227. Rubin, A. J., and C. Gall. 2010. Widespread Fraud Seen in Latest Afghan Elections. New York

Times, September 24. Rubin, B. 2007. "Saving Afghanistan." Foreign Affairs 86 (1):57-78. Scott, J. C. 2010. The Art of Not Being Governed. New Haven, CT: Yale University Press. Serchuk, V. 2006. Cop Out: Why Afghanistan Has No Police. The Weekly Standard, July 17. Shaver, A., and A. Wright. 2016. "Are Modern Insurgencies Predictable? New Evidence from

Afghanistan and Iraq": Working paper, Princeton University. Sherman, L., and D. Weisburd. 1995. "General Deterrent Effects of Police Patrol in Crime 'Hot

Spots': A Randomized Study." Justice Quarterly 12 (4):625-648. Snyder, J. 2000. From Voting to Violence. New York: W.W. Norton. Stedman, S. J. 1997. "Spoiler Problems in Peace Processes." International Security 22 (2):5-53. Thomson, J. E. 1994. Mercenaries, Pirates, and Sovereigns. Princeton, NJ: Princeton University

Press. Torabi, Y., and L. Delesgues. 2007. Afghan Perceptions of Corruption: A Survey across Thirteen

Provinces. http://www.iwaweb.org/_docs/reports/psm/afghan_perception_of_corruption_2006.pdf.

UNODC. 2010. Corruption in Afghanistan: Bribery as Reported by the Victims. https://www.unodc.org/documents/data-and-analysis/Afghanistan/Afghanistan-corruption-survey2010-Eng.pdf.

Vendrell, F. 2011. Elections and the Future of Afghanistan. In Elections in Dangerous Places, edited by D. Gillies. Baltimore, MD: Johns Hopkins University Press.

Weidmann, N. B., and M. Callen. 2013. "Violence, Control and Election Fraud: Evidence from Afghanistan." British Journal of Political Science 43 (1):53-75.

Wilkinson, S. I. 2004. Votes and Violence. New York: Cambridge University Press. Zakaria, F. 1997. "The Rise of Illiberal Democracy." Foreign Affairs 76 (6):22-43.

39

Table 1A: Police's Effect on Turnout

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

Dependent Variables Turnout in 2010

Turnout in 2010

Turnout in 2010

Turnout Change (2010-2009)

Turnout Change (2010-2009)

Turnout Change (2010-2009)

Standarized Turnout Change

(2010-2009)

Standarized Turnout Change (2010-2009)

Standarized Turnout Change

(2010-2009)

Medium or High Security Deployment -33.448 -33.592 -33.854 -23.660 -23.614 -24.118 -0.115 -0.114 -0.117

(11.960) (12.030) (12.145) (12.518) (12.685) (12.804) (0.070) (0.070) (0.071)

SIGACTs Prediction -2.701 -0.603 -0.002 (13.641) (19.121) (0.105) SIGACTs Prediction Residual 2.642 3.056 0.016 (8.662) (11.832) (0.066) SIGACTs (1-week lag) -9.182 8.781 0.059

(17.744) (28.668) (0.161)

SIGACTs (2-week lag) 4.095 -31.356 -0.188

(16.666) (20.955) (0.115)

SIGACTs (3-week lag) -9.500 4.538 0.034

(19.497) (18.185) (0.096)

SIGACTs (4-week lag) -12.989 -6.805 -0.030

(22.963) (32.963) (0.182)

Total violence previous 5 months 28.271 -28.827 -0.191

(83.181) (81.357) (0.447)

Total violence squared 9.808 49.157 0.283

(51.966) (49.673) (0.268)

Total violence cubed -2.518 -6.460 -0.036

(6.083) (5.387) (0.029)

Constant 335.401 335.854 336.207 92.945 93.030 93.938 0.002 0.002 0.007

(2.460) (3.479) (4.005) (2.575) (4.294) (4.615) (0.014) (0.024) (0.025)

Observations 1823 1823 1823 1823 1823 1823 1823 1823 1823 R-squared 0.429 0.429 0.431 0.395 0.395 0.397 0.391 0.391 0.393 Notes: All regressions include district fixed effects. Robust standard errors are clustered at the district level.

40

Table 1B: Police's Effect on Turnout

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

Dependent Variables Turnout in 2010

Turnout in 2010

Turnout in 2010

Turnout Change (2010-2009)

Turnout Change (2010-2009)

Turnout Change (2010-2009)

Standardized Turnout Change

(2010-2009)

Standardized Turnout Change

(2010-2009)

Standardized Turnout Change

(2010-2009)

Medium Security Deployment -45.483 -45.643 -46.256 -30.486 -30.362 -32.417 -0.146 -0.145 -0.157

(16.921) (16.842) (16.987) (21.414) (21.223) (21.358) (0.120) (0.118) (0.119)

High Security Deployment -23.186 -23.365 -23.345 -17.840 -17.887 -17.085 -0.088 -0.088 -0.083 (14.244) (14.384) (14.483) (16.109) (16.442) (16.629) (0.091) (0.093) (0.093) SIGACTs Prediction -3.146 -0.852 -0.003 (13.683) (18.982) (0.105) SIGACTs Prediction Residual 2.977 3.246 0.017 (8.681) (11.758) (0.065) SIGACTs (1-week lag) -9.002 8.901 0.059

(17.759) (28.623) (0.160)

SIGACTs (2-week lag) 2.644 -32.327 -0.193*

(16.580) (20.772) (0.114)

SIGACTs (3-week lag) -9.938 4.245 0.032

(19.599) (18.211) (0.096)

SIGACTs (4-week lag) -13.492 -7.142 -0.032

(23.247) (32.971) (0.182)

Total violence previous 5 months 31.262 -26.826 -0.181

(85.447) (82.281) (0.451)

Total violence squared 9.032 48.638 0.280

(52.690) (50.065) (0.270)

Total violence cubed -2.396 -6.378 -0.036

(6.145) (5.405) (0.029)

Constant 335.504 336.032 336.354 93.003 93.129 94.036 0.002 0.002 0.007

(2.460) (3.466) (4.012) (2.613) (4.288) (4.632) (0.015) (0.024) (0.025)

Observations 1823 1823 1823 1823 1823 1823 1823 1823 1823 R-squared 0.429 0.429 0.432 0.395 0.395 0.397 0.391 0.391 0.393 Notes: All regressions include district fixed effects. Robust standard errors are clustered at the district level.

1

Damaging Democracy: Security Provision and Turnout in Afghan Elections

Supporting Information Draft: February 17, 2016

This section provides some robustness checks on our main results, as well as tests of alternate

explanations for them. While we have shown that the assignment of polling centers to categories of

police was not a function of previous levels of violence, and have argued for plausibly exogenous

variation in treatment assignment, there might be other omitted confounds. We now consider a set

of potential omitted variables, a set of four principal factors that might be expected to correlate

with police deployment levels and with turnout, such that estimated effects of police on turnout

and violence around election day that we report would reflect those factors and not the impact of

police deployments. Our estimates remain remarkably robust to the inclusion of variables

capturing these potential confounds, consistent with a causal interpretation of our estimates.

The first is the competitiveness of the election in 2009. It is not immediately clear how to sign

the bias in this case with respect to the theoretical effect that competitiveness in the previous

election should have on both turnout and police levels in the next election. Its effect on turnout

seems likely to be positive, based on the reasoning that one’s vote is more likely to be

consequential in a competitive area than in one where a candidate won handily last time. The

effect on police deployment is ambiguous.

Second, international election monitors deployed to some polling centers on election day and

we might expect their deployment plan to have affected where police were deployed, as well as

turnout levels. It seems reasonable to expect that, if there was an effect, the presence of election

monitors would correlate positively with both police deployment and turnout. The Afghan

government would want to ensure the safety of the monitors and so be more likely to designate

those polling centers to receive more police, and civilians might expect that the presence of

election monitors at a polling center is likely to decrease the level of fraud perpetrated there and so

be more likely to turn out and vote.

Third, we might worry about strategic deployment of police designed to help the political

fortunes of President Karzai and politicians close to him. In their study of fraud in the 2010

Afghan election, Callen and Long (2015) use measures of political connectedness of candidates

running in the 2010 election. They explain how candidates’ connections to Provincial or District

Election Officials might affect how much fraud is associated with the electoral returns reported for

a polling center, particularly fraud that is perpetrated at the Provincial Aggregation Center and

levels above the polling center itself. If this connectedness affected police deployment and turnout,

we should expect the bias to be in the positive direction. Given that these are factors operating at

the district- or province-level, we control for them through the inclusion of district- and province-

fixed effects in our model specifications.

Fourth, ethnicity might play a role in levels of turnout and police deployment. We might

expect more police to be allocated to areas where more violence is expected (heavily Pashtun areas

where the Taliban was more active) and higher turnout in areas where voters are more supportive

of non-Taliban rule of the state (non-Pashtun areas). We do not have data that would allow for us

to control for this at the polling center level given the lack of a recent publicly available census.

However, to the extent that there is a low level of heterogeneity in ethnic mixes of the population

across districts, the inclusion of district fixed effects controls for this factor.

Note that in the cases above where the direction of the bias is not ambiguous, the sign is

hypothesized to be positive and thus, presents less of a problem for our results. If the hypothesized

effect of the factor on police deployment is positive, then this constitutes a ‘hard’ test when we

examine the effect of police deployment on violence and find a weak negative effect. SI Table 8

summarizes these factors, their expected correlations with the treatment and outcome variables,

and the resulting expected bias.

To test the plausibility of these alternate explanations, we use data on the 2010 election to

replicate our main results from regressing police deployment levels on turnout (Table 1) and

include these other factors individually as controls on the right hand side of the regression

specifications. These results are shown in SI Table 9. In Panel A, the dependent variable is the

turnout in the 2010 election; in Panel B, the dependent variable is the difference in turnout

between the 2010 election and the 2009 election. As we can see, in both panels there is little

evidence for any of those concerns: the core negative effect of police on turnout is consistently

strong. In column 7, we include a measure of the competitiveness of each polling center in the

2009 election, the log of the difference between President Karzai’s vote share and challenger

Abdullah Abdullah’s vote share. Including this measure slightly reduces the size of the estimated

effect of police on turnout.

To test whether the presence of election monitors might be responsible for the core result, we

include a dummy variable for whether Democracy International, a democracy-promoting

organization, assigned international election monitors to the polling center in the 2010 elections.1

Column 5 includes a dummy indicating that the polling center was assigned treatment; column 6

includes a dummy indicating that monitors actually were present. The inclusion of these variables

does not change the core effect of police on turnout.

1 They had the largest deployment in terms of number of stations and provinces visited of all

international observers.

Including district or province fixed effects (columns 1 and 4) – to control for connections of

election officials to President Karzai – does not noticeably change the estimated coefficient on the

security classification variable.

We also include dummy variables coding the majority ethnicity of the district in which the

polling center is located. In column 2, we include a dummy for whether the district is Pashtun-

majority or not. In column 3, we include dummies for the other main ethnic groups, as well.

Province fixed effects are included in each model. Controlling for ethnicity in this way does not

alter our estimates of the police’s effect on turnout.

Finally, it is plausible that citizens interpreted the presence of police as a signal that violence

was likely, and so avoided voting in polling centers with police. In August 2010, we asked a sample

of voters the question: “Please tell me whether you think the following are likely or unlikely to

occur on the upcoming election day: Violence in your neighborhood.” We code those responding

“very likely” or “somewhat likely” to this question as ‘1’ on a binary variable (those responding

“somewhat unlikely” or “very unlikely” are coded ‘0’). In column 8 of SI Table 9 (Panel A only),

we include the average response to this question (by polling center) as an additional covariate in

our core regression of 2010 turnout on the security classification the polling center. As one would

expect, the measure negatively predicts turnout, suggesting that voters’ beliefs about the likelihood

of violence influenced their decision of whether and where to vote. However, adding this measure

to the regression does not affect the estimated coefficient on the security classification variable,

indicating that expectations impacted voters’ decisions separately from police deployments.

SI Table 1. Descriptive Statistics

Variables Observations Mean Std. Dev.

SIGACTs (1-week lag) 72,920 0.0470 0.344 SIGACTs (2-week lag) 72,920 0.0468 0.342 SIGACTs (3-week lag) 72,920 0.0465 0.341 SIGACTs (4-week lag) 72,920 0.0461 0.341 SIGACTs (5-week lag) 72,920 0.0455 0.338 SIGACTs (6-week lag) 72,920 0.0447 0.336 SIGACTs (7-week lag) 72,920 0.0440 0.334 SIGACTs (8-week lag) 72,920 0.0435 0.332 Standardized Turnout 2010 1,823 -0.176 0.849 Standardized Turnout 2009 1,823 -0.154 0.852 Does the Government do a Good Job with Resources? (Yes/No) (pre) 130 0.543 0.308 Does the Government do a Good Job with Resources? (Yes/No) (post) 130 0.451 0.288 Satisfaction with Afghan Democracy? (5-pt scale) (pre) 121 0.863 0.188 Satisfaction with Afghan Democracy? (5-pt scale) (post) 128 0.779 0.232 Is Afghanistan a Democracy? (pre) 130 0.657 0.287 Is Afghanistan a Democracy? (post) 130 0.694 0.244 Will you use Courts or Police to Solve a Dispute? (Yes/No) (pre) 130 0.278 0.234 Will you use Courts or Police to Solve a Dispute? (Yes/No) (post) 130 0.212 0.210 How Important is it to Pay Your Taxes? (Very/Not) (pre) 130 0.491 0.305 How Important is it to Pay Your Taxes? (Very/Not) (post) 130 0.454 0.265

SI Table 2. Treatment Assignment and Violence (1) (2) (3) (4)

Variables

Recent violence and cubic

polynomial in past violence

nearby

Adding Province FE

Adding District FE

Dropping last 4-weeks of violence

District FE

SIGACTs (1-week lag) 0.031 0.058 0.067

(0.072) (0.074) (0.089)

SIGACTs (2-week lag) 0.039 -0.001 -0.004

(0.065) (0.067) (0.086)

SIGACTs (3-week lag) -0.063 -0.080 -0.007

(0.058) (0.062) (0.071)

SIGACTs (4-week lag) 0.137 0.111 -0.169

(0.059) (0.061) (0.086)

Total violence previous 5 months 0.128 -0.019 -0.558 -0.626

(0.275) (0.323) (0.335) (0.239)

Total violence squared -0.203 -0.128 0.267 0.215

(0.155) (0.181) (0.183) (0.146)

Total violence cubed 0.018 0.013 -0.019 -0.021

(0.019) (0.022) (0.023) (0.019)

Constant 1.308 1.316 1.333 1.331

(0.029) (0.027) (0.007) (0.007)

Observations 1823 1823 1823 1823 R-squared 0.004 0.068 0.316 0.314 Notes: All regressions include district fixed effects. Robust standard errors are clustered at the district level.

SI Table 3. Police's Effect on Violence (Timing Change at Election) (1) (2) (3) (4) (5) (6) (7) (8)

Variables

Election week vs.

week before

Election week vs.

week before

Election week vs. 4-week pre-election average

Election week vs. 4-week pre-election average

4-week average

post vs. 4-week

average before

4-week average

post vs. 4-week

average before

2-month average

post vs. 2-month average before

2-month average

post vs. 2-month average before

Medium Security Deployment -0.009 0.019 0.006 0.014

(0.035) (0.037) (0.012) (0.013)

High Security Deployment 0.006 0.050 0.029 0.021

(0.056) (0.057) (0.025) (0.019)

Medium or High Security -0.021 -0.008 -0.013 0.009

(0.040) (0.041) (0.013) (0.012)

SIGACTs (2-week lag) 0.045 0.046

(0.184) (0.185)

SIGACTs (3-week lag) -0.047 -0.046

(0.130) (0.131)

SIGACTs (4-week lag) -0.063 -0.062

(0.130) (0.131)

SIGACTs (5-week lag) 0.343 0.343 0.251 0.250 -0.045 -0.046

(0.209) (0.210) (0.195) (0.195) (0.074) (0.074)

SIGACTs (6-week lag) -0.035 -0.034 -0.060 -0.059

(0.208) (0.208) (0.052) (0.053)

SIGACTs (7-week lag) 0.131 0.132 -0.034 -0.033

(0.158) (0.158) (0.069) (0.069)

SIGACTs (8-week lag) -0.156 -0.156 -0.079 -0.079

(0.176) (0.177) (0.062) (0.062)

Total violence previous 5 months 0.538 0.534 0.646 0.642 -0.365 -0.367 -0.639 -0.640

(0.678) (0.680) (0.601) (0.605) (0.248) (0.251) (0.197) (0.197)

Total violence squared 0.252 0.254 0.028 0.029 0.229* 0.230* -0.065 -0.065

(0.492) (0.492) (0.392) (0.392) (0.133) (0.132) (0.118) (0.118)

Total violence cubed -0.070 -0.070 -0.054 -0.054 -0.041 -0.041 0.008 0.008

(0.065) (0.065) (0.052) (0.052) (0.017) (0.017) (0.016) (0.016)

Constant 0.062 0.062 0.050 0.050 -0.005 -0.005 0.006 0.006

(0.017) (0.017) (0.018) (0.018) (0.006) (0.007) (0.008) (0.008)

Observations 1823 1823 1823 1823 1823 1823 1823 1823

R-squared 0.503 0.503 0.510 0.510 0.645 0.646 0.790 0.790 Notes: All regressions include district fixed effects. Robust standard errors are clustered at the district level.

SI Table 4. Summary Statistics of Afghans’ Perceptions of Police

Variables Observations

Mean

(% agree) Std. Dev.

Panel A: Kandahar survey (August-October 2010)

ANP officers in my area are illiterate 369 0.724 0.448 ANP officers treat members of the local community with respect 369 0.274 0.446 ANP officers are well respected by local people 369 0.263 0.441 ANP officers in my area sometimes beat people up 369 0.623 0.485 Most ANP officers are corrupt 369 0.637 0.482 ANP officers put the interests of their community before their own interests 369 0.298 0.458 Panel B: ANQAR survey (March and May/June 2010) Seen/experienced the police engage in corrupt acts (Pashtun respondents) 10,507 0.322 0.467 Seen/experienced the police engage in corrupt acts (Tajik respondents) 5,039 0.172 0.378 Seen/experienced the police engage in corrupt acts (Uzbek respondents) 1,204 0.0880 0.283

SI Table 5. Police's Effect on Views of Afghan National Police (ANP) and Afghan National Army (ANA) (1) (2) (3) (4) (5) (6) DV: Presence of [ANP/ANA] makes polling center (-1=Less Safe; 0=No Difference; 1=Safer)

ANP ANP ANP ANA ANA ANA No FE Province FE District FE No FE Province FE District FE

Medium or High Security Deployment -0.407 -0.235 -0.169 0.095 0.112 -0.099

(0.109) (0.147) (0.174) (0.113) (0.052) (0.253)

Total violence previous 5 months -0.059 4.688 4.584 2.118 6.407 6.444

(4.632) (4.392) (5.079) (3.884) (2.408) (2.529)

Total violence squared -3.709 -31.461 -29.216 -17.000 -39.809 -39.367

(23.016) (20.698) (24.369) (33.385) (28.573) (28.670)

Total violence cubed 12.230 54.260 49.319 29.687 64.820 63.512

(32.344) (29.376) (33.182) (63.495) (60.794) (60.936)

Constant 0.504 0.441 0.437 0.368 0.303 0.304

(0.110) (0.067) (0.079) (0.140) (0.052) (0.053)

Observations 130 130 130 130 130 130 R-Squared 0.016 0.249 0.276 0.005 0.269 0.294 Notes: Robust standard errors are clustered at the district level.

SI Table 6. Attitudes vis-à-vis the Police

Panel A: Pre-Election Responses (1) (2) (3) (4) (5)

Variables Is Afghanistan a

democracy? Satisfaction with

Afghan democracy

Will you use Courts or Police

to Solve Dispute?

How Important Is It to Pay Your

Taxes?

Is the central government doing

a good job?

Medium or High Security Deployment 0.184 -0.186 0.134 0.019 0.129

(0.049) (0.129) (0.082) (0.039) (0.156)

Total violence previous 5 months -0.822 0.089 -1.233 -1.030 -2.080

(2.337) (0.994) (1.913) (1.216) (1.693)

Total violence squared -3.646 -4.269 13.316 4.515 11.745

(15.833) (7.059) (14.321) (11.766) (16.417)

Total violence cubed 18.254 10.968 -24.238 -10.116 -17.586

(26.216) (11.847) (25.941) (23.061) (32.188)

Constant 0.679 0.872 0.273 0.515 0.569

(0.024) (0.010) (0.024) (0.012) (0.023)

Observations 130 121 130 130 130 R-squared 0.217 0.223 0.189 0.372 0.318 Panel B: Post-Election Responses

Medium or High Security Deployment 0.104 0.082 -0.027 -0.126 -0.260

(0.123) (0.174) (0.047) (0.062) (0.024)

Total violence previous 5 months 1.641 -0.476 2.524 1.171 1.843

(0.629) (0.999) (0.928) (1.101) (0.620)

Total violence squared -10.231 0.423 -19.786 -8.173 -18.565

(5.558) (6.848) (6.113) (9.443) (7.386)

Total violence cubed 10.825 -2.953 32.101 16.546 37.047

(10.856) (11.408) (9.831) (17.455) (15.746)

Constant 0.682 0.798 0.198 0.440 0.449

(0.008) (0.011) (0.010) (0.009) (0.013)

Observations 130 128 130 130 130 R-squared 0.431 0.243 0.343 0.188 0.320

Panel C: Change in Responses

Medium or High Security Deployment -0.079 0.274 -0.162 -0.145 -0.389

(0.158) (0.310) (0.052) (0.100) (0.176)

Total violence previous 5 months 2.463 -0.491 3.757 2.201 3.923

(2.516) (1.837) (2.136) (1.748) (2.140)

Total violence squared -6.586 1.421 -33.102 -12.688 -30.310

(18.402) (12.717) (13.516) (11.631) (15.929)

Total violence cubed -7.429 -6.253 56.338 26.662 54.633

(31.589) (21.035) (23.072) (19.251) (27.551)

Constant 0.003 -0.083 -0.074 -0.076 -0.119

(0.022) (0.018) (0.030) (0.019) (0.021)

Observations 130 120 130 130 130 R-squared 0.214 0.122 0.372 0.251 0.344 Notes: All regressions include district fixed effects. Robust standard errors are clustered at the district level.

SI Table 7. Anticipation of Violence as Reason for Not Voting (1) (2)

Variables Insecurity/Fear of Attacks (Baseline)

Insecurity/Fear of Attacks (Endline)

Medium or High Security Deployment 0.022 -0.200

(0.097) (0.129)

Total violence previous 5 months -1.426 -0.251

(3.043) (1.206)

Total violence squared -3.273 -2.748

(18.071) (7.495)

Total violence cubed 25.271 11.572

(28.089) (11.724)

Constant 0.120 0.220

(0.039) (0.014)

Observations 130 130 R-Squared 0.208 0.278 Notes: All regressions include district fixed effects. Robust standard errors are clustered at the district level.

SI Table 8. Expected Effect of Other Factors on Police and Turnout Factor Expected Effect on Police Expected Effect on Turnout Competitiveness of the 2009 election

? (+) Voters turn out if their vote will be consequential.

International election monitors in 2010

(+) Government has an interest in protecting monitors.

(+) Voters turn out if they believe vote will be free of fraud.

President Karzai’s interests ? ? Ethnic politics (+)

Pashtun areas expected to be more violence and could receive more police.

(+) Non-Pashtun areas supportive of non-Taliban rule and more likely to turn out.

SI Table 9. Addressing Potential Confounders

Panel A: Levels (2010 Turnout) (1) (2) (3) (4) (5) (6) (7) (8)

Variables Province FE

Pashtun Dummy (Province

FE)

Ethnicity Dummies (Province

FE)

District FE

DI Treatment Assignment

(District FE)

DI Monitors (District

FE)

Competitiveness (District FE)

Anticipated Violence (District

FE)

Medium or High Security Deployment -37.7 -37.6 -38.9 -33.9 -53.2 -54.5 -25.6 -58.3

(13.0) (13.0) (13.9) (12.1) (23.0) (24.0) (12.0) (20.0)

Ismaili -33.7

(25.0)

Mixed -65.1

(43.1)

Nuristani 21.9

(80.2)

Pashai -69.3

(58.5)

Pashtun -74.8

(54.3)

Tajik -236.6

(62.8)

Waziri -97.3

(60.5)

SIGACTs (1-week lag) -20.7 -20.5 -11.2 -9.2 -13.5 -10.8 -0.5 -7.1

(19.8) (19.7) (18.6) (17.7) (25.8) (22.8) (18.8) (23.6)

SIGACTs (2-week lag) 18.0 18.2 11.7 4.1 125.0 128.0 -4.0 124.3

(20.1) (20.1) (18.4) (16.7) (30.5) (32.5) (17.1) (32.9)

SIGACTs (3-week lag) -12.6 -12.6 -5.0 -9.5 -76.8 -79.3 -11.4 -79.2

(19.3) (19.3) (18.6) (19.5) (20.5) (15.1) (18.3) (18.8)

SIGACTs (4-week lag) 14.5 15.1 7.6 -13.0 -- -- -16.2 --

(21.2) (20.8) (19.8) (23.0) (24.2)

Total violence previous 5 months -162.3 -163.5 -127.7 28.3 894.0 922.9 37.0 899.7

(92.5) (93.2) (89.3) (83.2) (478.1) (529.3) (82.2) (529.6)

Total violence squared 103.1 103.2 84.8 9.8 -5295.7 -5154.2 3.8 -5158.0

(59.8) (59.8) (57.3) (52.0) (1809.8) (2280.7) (50.7) (2065.4)

Total violence cubed -14.4 -14.4 -12.0 -2.5 9289.2 8767.6 -1.3 8955.3

(7.3) (7.4) (7.0) (6.1) (2672.8) (3254.3) (5.8) (2964.9)

Pashtun Majority District -7.7

(24.1)

Assigned Election Monitors -15.8

(13.0)

Received Election Monitors -9.6

(28.6)

Log(|Karzai VS - Dr. Abdullah VS|) -23.2

(4.4)

Survey: Local Violence Likely -25.2 (36.6) Constant 340.7 343.6 415.2 336.2 330.1 326.4 203.3 339.4

(7.1) (12.4) (44.4) (4.0) (8.0) (9.2) (25.3) (20.3)

Observations 1823 1823 1823 1823 132 132 1817 130 R-squared 0.210 0.210 0.240 0.431 0.477 0.475 0.458 0.456 Panel B: Trends (2010-2009 Turnout)

Medium or High Security Deployment -39.2 -39.0 -37.8 -24.1 159.3 157.7 -30.4

(13.1) (13.2) (13.4) (12.8) (20.7) (22.0) (12.2)

Ismaili -76.1

(24.6)

Mixed -45.4

(43.4)

Nuristani -74.0

(78.2)

Pashai -102.5

(59.9)

Pashtun -81.6

(56.7)

Tajik -70.4

(66.5)

Waziri -118.1

(63.5)

SIGACTs (1-week lag) 12.6 12.9 14.8 8.8 -3.7 8.6 -1.3

(23.3) (23.1) (22.8) (28.7) (21.6) (37.5) (25.1)

SIGACTs (2-week lag) -15.9 -15.7 -17.5 -31.4 130.7 137.8 -27.5

(19.5) (19.4) (18.8) (21.0) (37.1) (42.8) (17.9)

SIGACTs (3-week lag) 17.7 17.8 16.6 4.5 -51.5 -62.5 8.3

(13.7) (13.6) (13.7) (18.2) (17.1) (34.9) (19.4)

SIGACTs (4-week lag) 15.6 16.7 16.6 -6.8 -- -- -9.1

(23.7) (23.9) (24.2) (33.0) (31.6)

Total violence previous 5 months -166.2 -168.3 -152.2 -28.8 1045.1 1023.0 -37.1

(86.9) (88.1) (84.1) (81.4) (663.0) (630.5) (74.0)

Total violence squared 125.9 126.1 117.9 49.2 -8659.5 -7869.4 58.0

(48.8) (49.1) (47.1) (49.7) (3268.9) (2334.3) (47.4)

Total violence cubed -17.5 -17.6 -16.7 -6.5 14821.3 13098.1 -7.7

(6.2) (6.2) (5.9) (5.4) (4946.0) (3242.9) (5.2)

Pashtun Majority District -13.9

(18.5)

Assigned Election Monitors -15.7

(9.0)

Received Election Monitors -46.4

(69.1)

Log(|Karzai VS - Dr. Abdullah VS|) 26.1

(5.8)

Constant 97.0 102.4 158.5 93.9 88.6 89.0 241.9

(6.6) (10.2) (47.2) (4.6) (7.7) (4.8) (31.7)

Observations 1823 1823 1823 1823 132 132 1817 R-squared 0.185 0.186 0.191 0.397 0.659 0.661 0.430 Notes: In Column 3 (Ethnicity dummies), Hazara is the reference category. Robust standard errors are clustered at the district level.

SI Figure 1. Violent Incidents from January 2010 – June 2011, by Police Deployment Level

REFERENCES Callen, M., and J. D. Long. 2015. "Institutional Corruption and Election Fraud: Evidence from a

Field Experiment in Afghanistan." American Economic Review 105 (1):354-381.