Policy Research Working Paper 9126
Narrow Incumbent Victories
and Post-Election Conflict
Evidence from the Philippines
Benjamin Crost
Joseph H. Felter
Hani Mansour
Daniel I. Rees
Development Economics
Knowledge and Strategy Team
January 2020
Policy Research Working Paper 9126
Abstract
Post-election violence is a common form of conflict, but a pattern consistent with electoral manipulation. There is
its underlying mechanisms are not well understood. Using no evidence that the increase in post-election violence is
data from the 2007 Philippine mayoral elections, this paper related to the incumbent’s political platform or their perfor-
provides evidence that post-election violence is particularly mance in past elections. These results provide support for
intense after narrow victories by incumbents. Using a den- the notion that post-election violence is triggered by elec-
sity test, the study shows that incumbents were substantially tion fraud or by the failure of democratic ways of removing
more likely to win narrow victories than their challengers, unpopular incumbents from office.
This paper is a product of the Knowledge and Strategy Team, Development Economics. It is part of a larger effort by the
World Bank to provide open access to its research and make a contribution to development policy discussions around the
world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may
be contacted at bencrost@illinois.edu.
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Produced by the Research Support Team
Narrow Incumbent Victories and Post-Election Conflict:
Evidence from the Philippines
Benjamin Crost, Joseph H. Felter, Hani Mansour, and Daniel I. Rees
Benjamin Crost (corresponding author) is an assistant professor at the University of Illinois at
Urbana-Champaign; his email is bencrost@illinois.edu. Joseph H. Felter is senior research scholar
at the Center for International Security and Cooperation, Stanford University; his email address
is joseph.felter@stanford.edu. Hani Mansour is an associate professor at the University of
Colorado Denver and an IZA research fellow; his email is hani.mansour@ucdenver.edu. Daniel I.
Rees is a professor at the University of Colorado Denver and an IZA research fellow; his email is
daniel.rees@ucdenver.edu.
JEL classification codes: D72, D73, D74
Keywords: Civil conflict, election fraud, institutional weaknesses, Philippines
The anti-communist propaganda is calculated to pave the way for cheating the
progressive forces and their allies and cutting down their votes. The impen-
ding electoral fraud at their expense will only further discredit the ruling system
and will further justify the people’s determination to intensify the revolutionary
armed struggle.
– Prof. Jose Maria Sison, Chief Political Consultant, National Democratic Front
of the Philippines
1 Introduction
Election-related violence has long been associated with civil conflict and is believed to be
responsible for a substantial number of casualties worldwide in the wake of competitive nati-
onal and local elections. In the past decade alone, there are numerous cases in which election
outcomes allegedly incited post-election violence. Some examples include the January 2007
elections in Bangladesh where the discovery of 12 million voters illegally listed on the electo-
ral roll precipitated violence and civil unrest; the December 2007 elections in Kenya where
violence erupted after the Orange Democratic Movement, a popular political party, accused
the Election Commission of Kenya (ECK) of large scale vote rigging; and the 2009 electi-
ons in Iran where scores of people were killed and injured as they protested the suspected
fraudulent reelection of President Mahmoud Ahmadinejad.
A growing literature attempts to explain the relationship between elections and violence.
Callen and Long (2015) argue that fraudulent elections can exacerbate civil conflict by un-
dermining democratic processes and increasing popular support for non-democratic, and
potentially violent, political actors. A World Bank report notes that “leaders lacking trust
in ‘winner-take-all’ scenarios may manipulate [election] outcomes, which can trigger serious
violence” (World Bank, 2012). Several recent papers find that the likelihood of post-election
violence increases when international observers denounce elections as fraudulent (Daxecker,
1
2012; Hyde and Marinov, 2014; Borzyskowski, 2013).
More generally, a number of political scientists argue that bad governance, abuse of power,
and grievances stemming from political suppression are the principal causes of civil conflict
(Schock, 1966; Hegre et al., 2001; Henderson and Singer, 2000; Tucker, 2007).1 While there is
strong evidence of a positive association between elections and civil conflict (Weidmann and
Callen, 2013), this association could potentially be due to cultural, historical, or institutional
factors. Nevertheless, international donors spend substantial amounts on election monitoring
and related programs in an effort to dampen civil conflict and its attendant ills (Kelley, 2008).
Understanding whether and how elections can lead to violence is thus an important issue
relevant to development policy. Civil conflict is a major impediment to development and
poverty-reduction (World Bank, 2012). Its effects include reductions in economic growth
(Abadie and Gardeazabal, 2003), educational attainment (Leon, 2012), height-for-age Z-
scores (Akresh et al., 2012), and birth weight (Mansour and Rees, 2012). According to the
World Bank (2012), “[p]overty reduction in countries affected by major violence is on average
nearly a percentage point slower per year than in countries not affected by violence” (p. 59).
The enormous potential gains from ending or preventing civil conflict have led development
researchers and practitioners to renew their focus on better understanding its causes and
developing prescriptions for effectively addressing it.2
This study examines whether elections narrowly won by incumbents are associated with
post-election violence in an existing civil conflict between government security forces and
insurgents.3 The analysis is based on data from the 2007 Philippine mayoral elections, the
1
For instance, Tucker (2007) argued that election fraud galvanized the public, triggered protests, and
eventually led to the so-called “colored revolutions” in Eastern Europe.
2
See Blattman and Miguel (2010) for a review of the extensive literature on civil conflict.
3
The fact that this study examines organized violence in an already existing conflict sets it apart from
most previous papers in the literature on the relationship between electoral manipulation and violence, which
tend to focus on less organized forms of violence such as riots and protests.
2
last to be counted manually. The empirical analysis begins by showing that incumbent
mayors were more likely to win tightly contested elections than their challengers. This could
be an indication that the election outcomes were manipulated (McCrary, 2008; Grimmer
et al., 2011; Blakeslee, 2013), and is consistent with the observation that incumbents in
developing countries are typically better positioned to manipulate close elections (Pastor,
1999; Schedler, 2002; Kulkarni, 2012; Weidmann and Callen, 2013). It is also consistent with
anecdotal evidence from the Philippines that incumbent mayors can exert a fair amount
of influence on local election commissions (Arnaiz et al., 2013). Incumbents may also win
narrowly if they have better information about their standing in the race, and, using this
information calibrate their campaigning efforts to secure a narrow victory (Grimmer et al.,
2011).
Next, the study shows that narrow incumbent victories in the 2007 elections were associated
with a substantial increase in post-election conflict-related casualties and violent incidents
but were essentially unrelated to pre-election levels of violence. Narrow incumbent victories
could have led to post-election conflict through a number of channels. For instance, if voters
believed that incumbents won the elections fraudulently, this could have increased support
for violent insurgent groups and emboldened them to intensify their operations (Berman et
al., 2011a; World Bank, 2012). There is, in fact, anecdotal evidence that the New People’s
Army (NPA), one of the main insurgency groups operating in the Philippines, actively uses
allegations of fraud to recruit members and to gain popular support (see Section 2 for more
details). Alternatively, it is possible that incumbents who were involved in closely fought
elections (perhaps due to perceptions of corruption or the actual abuse of power) were able
to secure victory through technically legal means, but the opposition turned to violence
because the election outcome was not to their liking.
A number of analyses are conducted in an effort to better understand what drove the observed
relationship between narrow incumbent victories and post-election violence. First, the study
3
shows that narrow incumbent victories are associated with more conflict in the poorer half of
Philippine municipalities, where incumbents appears to have had the advantage. In wealthier
municipalities, where incumbents and their challengers were equally likely to win, narrow
incumbent victories are not associated with post-election violence. Second, the study shows
that decisive incumbent victories are not associated with an increase in post-election violence.
This latter result suggests that the estimates do not simply reflect the effect of incumbent
victories per se. Third, the study shows that the relationship between narrow incumbent
victories and post-election violence does not depend on party affiliation or family connections
to previous incumbents, suggesting that insurgents were not reacting to the candidate’s
political platform or out of a distaste for political dynasties.
In addition to providing evidence that narrow incumbent wins are associated with an increase
in post-election conflict, this paper adds to a growing literature on forensic measurements
of corruption and institutional weaknesses.4 Contributions to this literature recognize that
observer reports of corruption cannot always be trusted because, “the agents engaged in a
particular behavior prefer to keep it hidden” (Zitzewitz, 2012), and that measures of institu-
tional weaknesses may be systematically influenced by unobserved variables (Olken, 2009).
Researchers have therefore devised methods to detect corruption and institutional weaknes-
ses without relying on observer reports. Examples of the forensic approach applied to the
measurement of corruption include Olken (2007) and Bertrand et al. (2007). Olken (2007)
measured corruption by observing the amount of building materials used in the construction
of roads in Indonesia. Bertrand et al. (2007) indirectly observed corruption by experimen-
tally varying incentives for rapidly obtaining driver’s licenses in India.5 The paper adopts a
forensic approach because election irregularities are committed in secret and because other
institutional failures may be systematically present in areas prone to post-election violence.
4
See Zitzewitz (2012) for an overview of forensic methods in economics.
5
See also Fisman and Miguel (2007), who measure corruption by observing the number of unpaid parking
violations by diplomats in New York City.
4
This approach facilitates the identification of electoral irregularities and other possible go-
verning failures without relying on potentially biased reports from election observers.
2 Background: Elections and Armed Conflict in the
Philippines
The Philippines is a constitutional republic with a population of more than 90 million.
The presidential elections are scheduled every 6 years, while congressional, provincial and
municipal elections are held every 3 years. The focus of this paper is on the midterm elections
held on May 14, 2007. Although the presidency was not in contention, 1,598 mayors were
elected, 755 of whom were incumbents.6 The mayoral candidates belonged to over 40 different
parties, most of which were affiliated with one of the two major political camps: the center-
right governing coalition around then-president Gloria Macapagal-Arroyo’s KAMPI party,
and the opposition led by the center-left Liberal Party.
The 2007 Philippine elections were the last to be counted manually (USAID, 2008). Accor-
ding to a report by USAID (2008), voters had to indicate the candidates they supported
on ballots, which typically entailed writing between 20 and 30 names by hand. At the end
of the voting process, the ballots were read aloud at each polling station by the Board of
Election Inspectors and recorded on a tally sheet or black board. The results from each
polling station were sent to the local Board of Canvassers for tabulation and then on to the
Philippine Commission on Elections (COMELEC).
There is ample anecdotal evidence of widespread vote manipulation in the 2007 elections in
6
Although there are 1,630 municipalities in the Philippines, the study only uses election results from
the 1,598 races where the election outcomes were approved by the Philippine Commission on Elections
(COMELEC). The remaining 32 elections were declared as failed and were re-scheduled for a later date.
5
the form of ballot stuffing, miscounting of votes at the polling station, and mistabulation
of votes by the Board of Canvassers (USAID, 2008). In fact, election fraud through mis-
counting/mistabulation of votes is so common in the Philippines that a specific term has
developed for it: “dagdag-bawas,” which is literally translated as “subtract-add.”7 Inter-
national election observers of the 2007 election noted that ballots were tallied by hand and
characterized the counting process as “prone to fraud and misuse.” One observer reported
“a general feeling among voters that their votes would not be counted, a sentiment provoked
by acts of fraud and violence allegedly committed by politicians, election officials, and armed
groups” (Solidarity Philippines Australia Network, 2007). A report by USAID (2008) stated
that the antiquated method of vote counting in the Philippines opened the door for various
fraudulent activities such as substitution of ballots at the precinct level and tampering with
returns. These types of fraud are common in developing countries and were also documented
during the most recent Afghan elections (Callen and Long, 2015).
There is also evidence that many local election offices were understaffed, underfunded, and
relied on financial support from the municipal government (Arnaiz et al., 2013). Moreover,
the municipal treasurer and the school district supervisor, both of whom answered to the
mayor, typically held two out of the three seats on the Municipal Board of Canvassers.
Arnaiz et al. (2013) and others have argued that incumbent mayors were, as a consequence,
well positioned to influence the outcome of closely contested elections.
During the 2007 elections, there were two major organized armed groups active in the Phi-
lippines: the New People’s Army (NPA) and the Moro Islamic Liberation Front (MILF).
Formed in 1981, the MILF is a separatist movement fighting for an independent Muslim
state in the Bangsamoro area of the island of Mindanao. Because of its narrow geographic
7
Other mechanisms of fraud include vote buying and intimidation. However, these methods are unlikely
to be responsible for the results in this paper since they take place before the votes are counted. Incumbents
would therefore need extremely precise polling information to calibrate their vote-buying efforts in order to
ensure victories with very narrow margins.
6
focus, no MILF-related incidents are included in the data used for the study.8 The NPA
is the armed wing of the Philippine communist movement. Since taking up arms in 1969,
the NPA has relied on selective ambushes and harassment tactics rather than larger-scale
conventional battlefield confrontations against government forces. It operates in rural areas
across most of the country and extorts considerable sums of money from businesses and
private citizens (Rabasa et al., 2011). The NPA’s political wing is the Communist Party
of the Philippines (CPP), which is not recognized by the government and therefore cannot
participate in elections. While a number of parties affiliated with the CPP fielded candidates
for the House of Representatives in 2007, none fielded mayoral candidates.9 The only far-left
party that fielded mayoral candidates was Akbayan, which is a rival to the CPP. However,
only two of their mayoral candidates were incumbents and neither of them were involved in
tightly contested elections.
Many experts believe that vote manipulation decreases trust in democracy and leads to
increased support for non-democratic and potentially violent actors (World Bank, 2012;
Donno, 2013; Norris, 2014; Hall et al., 2015). In the case of the Philippines, the NPA
specifically uses allegations of fraud in their propaganda material and when recruiting. For
instance, a statement published on the NPA website (philippinerevolution.net) less than a
week before the 2007 elections read:
[Philippine] elections has [sic] always been rotten, bloody and dirty which is but
a result and extension of a degenerating rotten system beset by a chronic and
worsening social, political and economic crisis. But to stand up and fight for
truth, for justice and the democratic rights of the people will always be truly
liberating, patriotic and noble (Madlos, 2007).
A statement released ten days after the election read:
[T]he people must remain firmly united, vigilant and militant against the con-
8
The data also does not contain any incidents involving the Abu-Sayyaf Group, a smaller Islamist terrorist
group.
9
These parties are: Bayan Muna, Anakbayan, Gabriela, Migrante, Anakpawis and Sura (Holden, 2009).
7
tinuing manipulation of the real election results to favor the regime... The pro-
tests against the violent and fraudulent elections will surely continue as part of
the struggle to bring to account the corrupt, fascist and puppet regime (Salas,
2007).
Anecdotal evidence that propaganda based on allegations of election fraud was successful
comes from an interview conducted by the Armed Forces of the Philippines (AFP) with
an anonymous former high-ranking NPA commander following his voluntary surrender to
government forces. Describing the process through which he was recruited, he said:
[The NPA organizers] frequently visited me and briefed me on how rotten our
government was. I believed that they were telling the truth because during the
election in 1969, it was publicized that it’s prohibited to buy and sell votes.
However, the Barangay Captain himself, went to see me and gave me money [to
buy my vote]. That’s why I believed that the organizers were telling the truth
that the government is rotten.10
This anecdotal evidence is consistent with the argument that election fraud and underlying
perceptions of fraud are associated with an increase in conflict by decreasing trust in the
government and increasing support for the NPA; as increased support for insurgent mo-
vements facilitates recruitment and makes the population less likely to share information
with the government forces (Berman et al., 2011a). This evidence is also consistent with
the notion that other social and political factors such as bad governance and/or corruption
could have encouraged vulnerable incumbents to engage in vote manipulation and also led
to post-election violent incidents.
10
From an interview conducted by an AFP intelligence unit after this individual voluntarily surrendered
to government forces.
8
3 Empirical Analysis
3.1 Narrow Incumbent Victories
Several papers in the literature on electoral manipulation and civil conflict examine the
role of election observers and find that if international observers publicly declare that an
election was manipulated, the number of post-election violent incidents increased substanti-
ally (Daxecker, 2012; Hyde and Marinov, 2014; Borzyskowski, 2013). Although this result is
consistent with the hypothesis that election fraud leads to post-election unrest and violence,
international observers are not randomly assigned to elections. If international observers
were more likely to have been assigned to countries that were prone to post-election vio-
lence, the association between election fraud and post-election violence will overstate the
true, causal relationship.11
One method of addressing this issue would be to use so-called “forensic” measures of fraud.
The most often-used forensic test compares the distribution of digits found in vote counts
to the distribution predicted by Benford’s Law (Mebane, 2006).12 An alternative method
focuses on vote counts ending in 0 or 5, suggesting rounding (Weidmann and Callen, 2013;
Beber and Scacco, 2012). While these approaches have the advantage of being independent of
whether election observers were assigned to a particular polling station, they require detailed
polling-level vote counts which, in many countries, are hard to obtain. Instead of relying
on observer reports or digit-based forensic evidence, the analysis below uses the density test
developed by McCrary (2008) to detect evidence for vote manipulation. Specifically, the
test examines the probability density function of the incumbent’s vote margin – that is, the
11
It is, of course, also possible that international organizations avoid elections in violence-prone countries
in an effort to protect observers. In that case, the result would underestimate the effect of election fraud on
violence.
12
In addition, this approach requires detailed data on election returns at the polling station level, which
the study does not have access to in the present context.
9
difference between the incumbent’s vote share and that of the most successful challenger. If
this difference is greater than zero, the incumbent won the election; if it is smaller than zero,
the incumbent lost. In the absence of manipulation, the winner of a close election should
have been, in effect, randomly assigned and therefore the probability of an incumbent victory
should be equal to the probability of an incumbent loss. More precisely, the probability
density function of the incumbent’s vote margin should be smooth across the threshold of
zero, so that (in the limit) the incumbent was equally likely to win or lose close elections.
There are other possible reasons aside from fraud for why incumbents might have an ad-
vantage in close elections. An alternative explanation is that incumbents have a resource
advantage over their challengers and, given accurate polling information, can expend preci-
sely enough effort campaigning to secure a narrow victory. Consistent with this explanation,
Grimmer et al. (2011) found that resource-advantaged candidates in races for the U.S. House
of representatives were about 10 percent more likely to win elections decided by margins of
between 0 and 5 percentage points. However, aside from elections to the U.S. House of Re-
presentatives, there is little evidence that incumbents (or other candidates with a resource
advantage) can target campaigning effort in this fashion. Eggers et al. (2015) who analyzed
over 40,000 tightly contested mayoral and legislative races in the United States and seven
other countries, found no evidence that incumbents were more likely to win closely contested
elections. Consistent with this result, this paper finds no evidence that narrow incumbent
victories were associated with higher turnout, a plausible proxy for campaigning effort.13
The analysis follows the method developed by McCrary (2008) to explore whether incum-
bents were more likely to win closely fought elections. Specifically, observations are grouped
into bins of equal width and the following regression is estimated:
Yb = γ0 + γ1 Db + γ2 Xb + γ3 (Db × Xb ) + ηb . (1)
13
Results available from authors on request.
10
Here, Yb is the fraction of observations in bin b, Xb is the incumbent’s victory/loss margin
(in percentage points) at the bin’s midpoint, and Db is an indicator for whether the bin is
above the incumbent’s victory threshold (all bins have to be either entirely above or entirely
below zero). If the density of the incumbent’s vote margin is smooth across the threshold
of zero, the bins just above the threshold should contain as many observations as those just
below the threshold and the parameter γ1 will not be statistically distinguishable from zero.
The election data are publicly available and were obtained from the Republic of the Philip-
pines Commission on Elections (COMELEC). Mayoral election results from 2007 were used
to identify winners/losers in each municipality and to calculate margins of victories/losses;
returns from the 2004 election were used to identify incumbents.14 The sample is limited
to municipalities in which the incumbent mayor’s vote margin was within a 5 percentage-
point bandwidth of the zero threshold. This resulted in a sample of 153 municipalities. The
incumbent won reelection in 81 of these municipalities and lost in the remaining 72.
14
The majority of mayors elected in 2004 were matched with the 2007 election results using an automated
process. Unmatched observations were mostly due to the fact that not all mayors elected in 2004 ran for
re-election and to spelling errors. In the latter case, incumbent mayors were matched manually.
11
Figure 1. Did Incumbents Manipulate Close Elections?
.1
.08
.06
pdf
.04
.02
0
−.05 −.04 −.03 −.02 −.01 0 .01 .02 .03 .04 .05
Margin of victory
Mean Nonparametric fit
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The figure presents the probability density of incumbent mayors’ margin of victory in the 2007 election. Municipalities
were grouped into 20 bins of equal width according to the incumbent’s margin of victory. Each scatter dot represents one bin.
Its horizontal coordinate represents the midpoint of the bin, and its vertical coordinate represents the fraction of municipalities
for which the incumbent’s margin of victory was within the bin. Dashed lines are linear fits, separately estimated on both sides
of the eligibility threshold. Solid lines are nonparametric fits from a local linear regression that uses triangular kernels with a
bandwidth of 5 percentage points, separately estimated on both sides of the eligibility threshold.
12
Table 1. McCrary Test for Election Manipulation by Incumbent Mayors in the 2007 Elections
OLS Estimates
Dependent variable: Fraction of municipalities within bin
Full Sample Poor Municipalities Rich Municipalities
(1) (2) (3)
Incumbent victory 0.016** 0.025* -0.0061
(0.0079) (0.013) (0.013)
Margin -0.17 -0.22 0.46
(0.27) (0.46) (0.48)
Margin× incumbent victory -0.24 -0.56 -0.62
(0.39) (0.65) (0.61)
Constant 0.028*** 0.027*** 0.044***
(0.0056) (0.0094) (0.010)
Municipalities 153 75 78
Observations (bins) 30 30 30
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: Results from a probability density test for manipulation of the running variable (McCrary, 2008) are reported. The
running variable is the incumbent margin of victory in the 2007 mayoral elections. All regressions are weighted by a triangular
kernel with a bandwidth of 0.05. Observations are 30 bins of equal width. The dependent variable is the fraction of municipalities
with an incumbent margin of victory that falls within the bin. Standard errors are in parentheses. ∗ , ∗∗ and ∗∗∗ denote statistical
significance at the 10%, 5% and 1% levels, respectively.
13
Figure 1 plots the probability that a municipality falls into a particular bin against the
margin of victory at the midpoint of the bin, and provides evidence that incumbent mayors
were more likely to win elections decided by a margin of between 0 and 1.5 percentage points.
Table 1 reports McCrary test results – i.e., estimates of equation (1), based on 30 bins, 15
on each side of the threshold. The constant term represents the height at which the pdf
intercepts the threshold of zero from below, which is an estimate of how likely incumbents
were to lose tightly contested elections. The coefficient of the incumbent victory indicator
reflects the discontinuous change in the pdf at the zero threshold. For the entire sample,
the analysis estimates a constant of 0.028 and a discontinuous increase of 0.016, which is
statistically significant at the 5 percent level. Thus, incumbents were approximately 57
percent (0.016/0.028) more likely to win tightly contested elections than to lose them. This
constitutes evidence that incumbents were able to manipulate tightly contested elections in
order to secure victory. Below, the paper shows that this pattern of close incumbent electoral
wins closely corresponds to the pattern of violence observed after the election.
The results presented in Table 1 are based on a bandwidth of 5 percentage points. In
general, estimates with smaller bandwidths should be closer to the true value since they are
more strongly informed by observations closer to the threshold (McCrary, 2008). However,
reducing the bandwidth also decreases the effective sample size and therefore leads to less
precision. The results in Appendix Table A.1 provide evidence that the estimates are robust
to limiting the bandwidth to 3 or 4 percentage points around the zero threshold.
Previous evidence suggests that poorer regions tend to be more corrupt and are more likely
to have weak institutions (Olken and Pande, 2012; Ferraz and Finan, 2008; Besley and
Burgess, 2002). The paper explores whether incumbents are more likely to win in poorer
versus richer municipalities. The classification of municipalities as rich or poor is based on
the 2006 municipal-level poverty estimates provided by the Philippine Statistics Authority.15
15
These data are available at https://data.humdata.org/dataset/
14
If a municipality’s poverty estimate was below the median, it is designated as poor; if the
poverty estimate was above the median, it is designated as rich. The results in columns
(2) and (3) of Table 1 suggest that incumbents were more likely to win close elections
primarily in poor municipalities. Specifically, when the sample is restricted to the poorer
half of municipalities, a constant of 0.027 is obtained and there is a discontinuous increase of
0.025 in the pdf at the zero threshold, indicating that incumbents were about twice as likely
to win tightly contested elections than to lose them. In the richer half of municipalities,
the discontinuous decrease of 0.0061 is small relative to the constant of 0.044 and is not
statistically significant at conventional levels. Below, the paper shows that the pattern of
violence observed after the election is also concentrated in poor municipalities.
3.2 Narrow Incumbent Victories and Civil Conflict
The data on conflict violence come from incident reports collected by the Armed Forces of the
Philippines (AFP); they include information on insurgents, government forces and civilian
casualties. These incident-level data have been used to analyze the impact of aid and the
impact of economic conditions on conflict intensity (Crost et al., 2014; Berman et al., 2011b)
and are similar to SIGACTS – the U.S. incident-level military data, which have been used
to study the insurgency activities in Afghanistan and Iraq (Berman et al., 2011a; Iyengar
et al., 2011; Beath et al., 2011). The data also contain information on the date and the
location of the incident and identify the insurgent group involved. During the period under
study (November of 2006 through November 2007), 2,745 conflict incidents leading to 1,045
casualties were reported by units belonging to the Armed Forces of the Philippines deployed
throughout the country. The focus of this paper is on the number of casualties and the
number of violent incidents that occurred in the 153 municipalities in which the incumbent’s
philippines-small-area-poverty-estimates-2012-2009-2006
15
vote margin was within a 5 percentage-point bandwidth of the zero threshold.16 In these
municipalities, a total of 521 incidents occurred, leading to 315 casualties.17 Information
on the other municipality characteristics included in the estimation were obtained from the
2000 Philippines Census, available from the National Statistics Office of the Philippines.
16
A violent incident is defined as an incident that leads to at least one casualty.
17
As Crost et al. (2014) noted, it is possible that the AFP troops misreported the number of casualties
related to an incident. This type of misreporting, however, is likely to be limited since the field data were
originally collected to be used internally by the AFP for intelligence purposes and for the planning of future
operations. Moreover, casualties are typically easy to verify which makes misreporting more difficult, even
if individual AFP units had an incentive to do so.
16
Table 2. Summary Statistics
Incumbent victory Incumbent loss RD balance test
(1) (2) (3)
Casualties in 12 months before election 0.70 0.74 0.29
[1.92] [1.79] (0.69)
Violent incidents in 12 months before election 0.33 0.40 0.12
[0.68] [0.84] (0.27)
At least 1 casualty in 12 months before election 0.12 0.24 0.12
[0.32] [0.43] (0.20)
At least 3 casualties in 12 months before election 0.072 0.098 -0.093
[0.26] [0.30] (0.14)
At least 5 casualties in 12 months before election 0.026 0.039 0.050
[0.16] [0.19] (0.042)
Casualties in 12 months after election 0.61 0.35 1.61
[2.28] [1.34] (0.98)
Violent incidents in 12 months after election 0.20 0.18 0.44*
[0.62] [0.62] (0.25)
At least 1 casualty in 12 months after election 0.13 0.1 0.20**
[0.34] [0.30] (0.10)
At least 3 casualties in 12 months after election 0.084 0.057 0.21**
[0.28] [0.23] (0.09)
At least 5 casualties in 12 months after election 0.024 0.029 0.08
[0.15] [0.17] (0.06)
Population 40594 32265 -1742
[62379] [20504] (8165)
Area (km2 ) 210 220 77
[267] [260] (74)
Poverty rate 32.8 35.0 3.0
[16.0] [16.7] (5.1)
Incumbent vote share in 2004 election 0.536 0.534 -0.025
[0.146] [0.133] (0.049)
Affiliated with national governing coalition 0.90 0.77 0.18
[0.30] [0.42] (0.14)
Insurgent presence 0.265 0.271 0.056
[0.44] [0.45] (0.18)
Share of households with electricity 0.57 0.55 -0.18
[0.26] [0.25] (0.081)**
Share of households with piped water 0.44 0.36 0.012
[0.27] [0.27] (0.085)
Share of households with indoor plumbing 0.66 0.61 -0.095
[0.21] [0.24] (0.072)
Share of buildings with walls made of “strong” materials 0.69 0.67 -0.069
[0.19] [0.17] (0.062)
Share of buildings with roofs made of “strong” materials 0.64 0.59 -0.18
[0.24] [0.24] (0.072)**
Located in Luzon 0.52 0.64 -0.045
[0.50] [0.48] (0.18)
Located in Visayas 0.25 0.15 0.095
[0.44] [0.37] (0.14)
Located in Mindanao 0.23 0.20 -0.049
[0.42] [0.40] (0.15)
Located in ARMM 0.036 0.029 0.028
[0.19] [0.17] (0.022)
Municipalities 83 70 153
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: Summary statistics for the sample of 153 municipalities in which incumbents won or lost the 2007 election by a margin
of 5 percentage points or less. Following the classification of the 2000 Census of the Philippines, “strong” building materials for
walls are defined as concrete, brick, stone, wood, galvanized iron or aluminum, asbestos and glass. “Strong” building materials
for roofs are defined as concrete, galvanized iron or aluminum, clay tiles and asbestos. Standard deviations are reported in
brackets. Column 3 reports results of RD regressions using the municipality characteristics as outcomes. Standard errors of
these regression estimates are reported in parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1%
levels.
17
Descriptive statistics for the variables used in the analysis by whether incumbents won a
tightly contested election are provided in Table 2. The descriptive statistics show that
the number of casualties and violent incidents in the pre-election period was similar in
municipalities in which incumbents went on to win and municipalities in which the incumbent
went on to lose. To further test the validity of the empirical approach, a test for differences
in a number of municipality characteristics by whether incumbents won a tightly contested
election is provided in column (3). As shown in Table 2 there are no statistically significant
differences between the two types of municipalities in population, area size, poverty rate,
the incumbent vote share in the 2004 election, the presence of insurgent groups, whether the
incumbent is associated with the national governing coalition, or the region in which they
are located. There is some evidence that municipalities in which incumbents won had better
access to infrastructure, although the only statistically significant difference is for share of
households with electricity and share of buildings with roofs made of strong materials.
The second step in the analysis is to test whether conflict increased when incumbents won
close elections. Specifically, the following equation is estimated, based on a regression dis-
continuity (RD) design, for a subsample of municipalities within a small bandwidth around
the incumbent’s victory threshold:
Yi = β0 + β1 Di + β2 Xi + β3 (Di × Xi ) + Zi + εi . (2)
Here, Yi is the number of conflict casualties or the number of violent incidents in municipality
i experienced in the 12 months after the 2007 elections, Xi is the incumbent’s vote margin,
Di is an indicator for whether the incumbent won the election, and Zi is a vector of mu-
nicipality characteristics. The parameter β1 represents the difference in casualties between
municipalities in which the incumbent won a tightly contested election and municipalities in
which the incumbent narrowly lost.
18
The discontinuous jump in the probability of an incumbent victory at the zero threshold
documented in the previous section is a clear violation of the standard RD assumption that
unobserved variables are continuous across the threshold and would bias the estimate of β1
upwards if incumbents were more likely to win narrow elections in municipalities especially
prone to violence. Two tests are conducted in an effort to explore this possibility. First,
equation (2) is estimated using the number of casualties that occurred in the 12 months
prior to the 2007 election as the outcome. If municipalities above the zero threshold were
more violent for reasons unrelated to the incumbent’s victory, it is expected they would
have experienced more casualties before the election took place, not just afterwards. Se-
cond, the analysis estimates a panel regression model that explicitly controls for unobserved
municipality characteristics.
Figure 2 plots casualties during the 12 months before and the 12 months after the 2007
election against the incumbent’s margin of victory. Scatter dots represent the mean of
casualties in bins of equal width and are sized to reflect the number of municipalities in each
bin. The top panel shows that municipalities experienced substantially more casualties after
the incumbent won an election by a margin of between 0 and 1.5 percentage points than
after losing by a similar margin. The bottom panel shows that casualties in the 12 months
before the election were essentially unrelated to incumbent victories.
19
Figure 2. Did Narrow Incumbent Victories Cause Conflict?
1 1.5
Casualties
.5
0
−.05 −.04 −.03 −.02 −.01 0 .01 .02 .03 .04 .05
Margin of victory
Mean Nonparametric fit
1.5
1
Casualties
.5
0
−.05 −.04 −.03 −.02 −.01 0 .01 .02 .03 .04 .05
Margin of victory
Mean Nonparametric fit
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The figure presents the relationship between the incumbent’s margin of victory and casualties experienced in the 12
months before and after the 2007 election. Scatter dots represent the mean of casualties per month and are sized to reflect the
number of municipalities in each bin. Dashed lines are linear fits, separately estimated on both sides of the eligibility threshold.
Solid lines are nonparametric fits from a local linear regression that uses triangular kernels with a bandwidth of 5 percentage
points, separately estimated on both sides of the eligibility threshold.
20
Table 3. Incumbent Victories and Conflict Casualties: RD Design
Post-Election Period Pre-Election Period
Casualties Violent Incidents Casualties Violent Incidents
(1) (2) (3) (4) (5) (6) (7) (8)
Incumbent victory 1.22*** 0.89*** 0.40** 0.31*** 0.29 0.059 0.023 –0.057
(0.44) (0.29) (0.16) (0.12) (0.62) (0.44) (0.24) (0.19)
Margin –17.3*** –10.1 –7.95*** –5.67** 1.24 4.73 –0.77 1.96
(6.52) (6.56) (2.79) (2.45) (16.8) (11.0) (7.53) (6.09)
Margin × incumbent victory –17.5 –13.2 –0.43 –0.43 –22.4 –27.8 –5.66 –9.85
(26.4) (16.9) (7.95) (6.30) (26.8) (20.1) (10.6) (8.59)
Mean of dep. var. 0.50 0.50 0.20 0.20 0.72 0.72 0.36 0.36
Control variables No Yes No Yes No Yes No Yes
No. of municipalities 153 153 153 153 153 153 153 153
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The running variable of the RD design is the incumbent margin of victory. All regressions are weighted by a triangular
kernel with a bandwidth of 0.05. Reported values are marginal effects from Poisson regressions, calculated at the sample
means. Standard errors are in parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1% levels. Control
variables are: population, area, as well as share of households with electricity, piped water, indoor plumbing, roof made of strong
building material, and wall made of strong building material. Following the classification of the 2000 Census of the Philippines,
“strong” building materials for walls are defined as concrete, brick, stone, wood, galvanized iron or aluminum, asbestos and
glass. “Strong” building materials for roofs are defined as concrete, galvanized iron or aluminum, clay tiles and asbestos.
21
The results in Table 3 report Poisson estimates of the regression discontinuity (RD) des-
cribed by equation (2). The coefficient associated with incumbent victories represents the
discontinuous increase in casualties across the incumbent’s victory threshold. The results in
columns (1) and (2) suggest that municipalities experienced between 0.89 and 1.22 additio-
nal casualties in the 12 months after a narrow incumbent victory as compared to a narrow
loss. These estimates are quite large as compared to the average of 0.50 casualties in the 12
months after the election and are statistically significant at the 1 percent level. Similarly,
the results in columns (3) and (4) indicate that in the 12 months after a narrow incumbent
victory, municipalities experienced between 0.31 and 0.40 additional violent incidents. The
results in columns (5) - (8) provide no evidence of a statistically significant relationship bet-
ween narrow incumbent victories and pre-election casualties or violent incidents. Appendix
Table A.2 shows that the results are robust to limiting the bandwidth to 3 or 4 percentage
points around the threshold.18
One potential concern with the results reported in Table 3 is that the estimated effect of in-
cumbent victories decreases when controls for observable municipal characteristics are added
in columns (2) and (4). This raises the possibility that the results are driven by unobserved
differences between municipalities in which incumbents won narrowly and municipalities in
which they did not. In an effort to explore this hypothesis, the paper estimates a panel
regression model that explicitly controls for systematic differences in unobserved characte-
ristics. Specifically, the following equation is estimated based on observations from the 12
months before and the 12 months after the 2007 election:
Yit = β0 + β1 Di + β2 Xi + β3 (Di × Xi ) + β4 (P ostt )
(3)
+ γ1 (Di × P ostt ) + γ2 (Xi × P ostt ) + γ3 (Di × Xi × P ostt ) + εit .
18
The paper also estimates a Probit model based on equation (2) where the dependent variable is an
indicator for having experienced 1 or more casualties, 3 or more casualties, and 5 or more casualties. The
results, reported in Appendix Table A.9, show that a narrow incumbent victory increases the likelihood of
having 1 or more casualties and 3 or more casualties.
22
Here, Yit is the number of conflict casualties or violent incidents that municipality i expe-
rienced in month t, Xi is the incumbent’s vote margin, and Di is an indicator for whether
the incumbent won the election. P ostt is an indicator for whether the observation was made
in the post-election period. The parameter β1 represents the difference in pre-election casu-
alties between municipalities in which incumbents narrowly won and municipalities in which
incumbents narrowly lost. The parameter γ1 represents the increase in casualties following
a narrow incumbent victory.
23
Table 4. Incumbent Victories and Conflict Casualties: Panel Regressions
Poisson Estimates
Dependent variable: Casualties
(1) (2) (3) (4) (5)
Incumbent victory × post-election 0.12** 0.10** 0.10** 0.13** 0.14**
(0.055) (0.047) (0.052) (0.055) (0.065)
Incumbent victory 0.017 0.0046 0.0039 0.0077 –0.052
(0.037) (0.028) (0.027) (0.025) (0.043)
Margin × post-election –2.07* –1.81* –1.81* –1.76* –1.58*
(1.26) (1.03) (1.02) (1.01) (0.88)
Margin 0.074 0.29 0.29 0.28 0.30
(1.01) (0.76) (0.75) (0.73) (0.64)
Margin × incumbent victory × post-election –0.68 –0.25 –0.25 –0.25 –0.22
(3.16) (2.35) (2.35) (2.28) (2.04)
Margin × incumbent victory –1.35 –1.47 –1.47 –1.43 –1.26
(1.57) (1.27) (1.26) (1.22) (1.08)
Post-election period (12 months) –0.091** –0.078** –0.071** –0.095*** –0.11***
(0.043) (0.037) (0.028) (0.032) (0.040)
Time trend (months) –0.00047 0.0013
(0.0022) (0.0016)
Time trend squared (months) 0.00026**
(0.00012)
Time trend × incumbent victory –0.000094 –0.0018
(0.0038) (0.0036)
Time trend squared × incumbent victory –0.00026
(0.00019)
Gov. coalition × inc. victory × post-election –0.061
(0.068)
Governing coalition × incumbent victory 0.057
(0.039)
Governing coalition × post-election 0.057
(0.045)
Governing coalition 0.0034
(0.019)
Mean of dep. var. 0.051 0.051 0.051 0.051 0.051
Control variables No Yes Yes Yes Yes
No. of municipalities
No. of observations 3672 3672 3672 3672 3672
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: All regressions are weighted by a triangular kernel with a bandwidth of 0.05. The unit of observation is the municipality-
month. The sample is restricted to observations within 12 months of the 2007 election (the month of the election is dropped).
Marginal effects calculated at the sample means are reported. Standard errors, clustered at the municipality level, are in
parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1% levels. Control variables are the same as in
Table 3.
24
Results from estimating equation (3) are presented in Table 4. Those reported in the first
column show that narrow incumbent victories are associated with 0.123 additional post-
election casualties per month (or 1.48 per year). Including the control variables reduces
this estimate to 0.103 casualties per month (or 1.24 per year). Both of these estimates are
statistically significant at the 5 percent level. Similarly, the first two columns of Table 5
show that narrow incumbent victories lead to a 0.046-0.052 increase in the number of violent
incidents (or 0.55-0.62 per year).
25
Table 5. Incumbent Victories and Violent Incidents: Panel Regressions
Poisson Estimates
Dependent variable: Violent incidents
(1) (2) (3) (4) (5)
Incumbent victory × post-election 0.052** 0.046** 0.031 0.045** 0.060*
(0.022) (0.019) (0.020) (0.021) (0.035)
Incumbent victory 0.0012 –0.0034 0.0034 0.0038 –0.013
(0.012) (0.010) (0.011) (0.011) (0.021)
Margin × post-election –1.01* –0.93** –0.92** –0.89* –0.86**
(0.54) (0.46) (0.46) (0.45) (0.43)
Margin –0.040 0.10 0.10 0.097 0.12
(0.40) (0.32) (0.31) (0.30) (0.29)
Margin × incumbent victory × post-election 0.24 0.35 0.35 0.33 0.31
(1.08) (0.85) (0.84) (0.81) (0.74)
Margin × incumbent victory –0.30 –0.46 –0.46 –0.44 –0.41
(0.56) (0.45) (0.45) (0.43) (0.41)
Post-election period (12 months) –0.049*** –0.044*** –0.043*** –0.057*** –0.063***
(0.017) (0.015) (0.014) (0.016) (0.022)
Time trend (months) –0.000054 0.00097
(0.00091) (0.00090)
Time trend squared (months) 0.00013**
(0.000062)
Time trend × incumbent victory 0.0011 0.000039
(0.0013) (0.0012)
Time trend squared × incumbent victory –0.00013
(0.000078)
Gov. coalition × inc. victory × post-election –0.021
(0.030)
Governing coalition × incumbent victory 0.017
(0.018)
Governing coalition × post-election 0.012
(0.019)
Governing coalition 0.0021
(0.0083)
Mean of dep. var. 0.023 0.023 0.023 0.023 0.023
Control variables No Yes Yes Yes Yes
No. of municipalities
No. of observations 3672 3672 3672 3672 3672
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: All regressions are weighted by a triangular kernel with a bandwidth of 0.05. The unit of observation is the municipality-
month. The sample is restricted to observations within 12 months of the 2007 election (the month of the election is dropped).
Marginal effects calculated at the sample means are reported. Standard errors, clustered at the municipality level, are in
parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1% levels. Control variables are the same as in
Table 3.
26
While the panel models in columns (1) and (2) of Tables 4 and 5 control for unobserved
differences between municipalities in which incumbents won/lost a narrow election, a remai-
ning concern is that the two groups of municipalities were on non-parallel time trends. To
explore this possibility, controls for group-specific linear and quadratic time-trends are ad-
ded to the model. The estimates in columns (3) and (4) of Table 4 are robust to controlling
for group-specific trends. While the estimate in column (3) of Table 5 is only marginally
significant, the results controlling for quadratic time trends in column (4) are similar to
those in columns (2) and are statistically significant at the 5 percent level. Specifically, the
results indicate that an incumbent victory is associated with 0.126 additional casualties per
month (or 1.51 additional casualties per year) and 0.045 additional violent incidents (or 0.54
additional casualties per year).
3.3 Heterogeneous Effects and Sources of Violence
This section examines different margins of heterogeneity to better understand the relati-
onship between narrow incumbent victories and post-election violence. First, the paper
estimates the relationship between incumbent victories and post-election conflict-related ca-
sualties in poor and rich municipalities (measured using the 2006 municipal-level poverty
estimates provided by the Philippine Statistics Authority as described in Section 3.1). The
results in columns 2 and 3 of Table 1 provided evidence that narrow incumbent victories
were concentrated in poor municipalities. If narrow incumbent victories are associated with
increased civil conflict, it would be expected that post-election violence to be highest in
these municipalities. The RD estimates in column (1) and (2) of Table 6 suggest that an
incumbent victory in a poor municipality was, on average, associated with 1.65 additional
casualties and 0.56 additional violent incidents during the 12 months after the election (Panel
C). By contrast, the estimates in columns (3) and (4) provide little evidence that incumbent
27
victories in poor municipalities were related to pre-election violence. In rich municipalities
there is no statistically significant relationship between incumbent victories and violence in
either the pre- or the post-election period (Panel B).
28
Table 6. Incumbent Victories and Conflict: Rich vs. Poor Municipalities
Poisson RD Estimates
Post-Election Period Pre-Election Period
Dependent variable: Casualties Violent Incidents Casualties Violent Incidents
(1) (2) (3) (4)
Panel A: All Municipalities (153 municipalities)
Effect of Incumbent Victory 0.89** 0.31*** 0.059 -0.057
(0.29) (0.12) (0.44) (0.19)
Mean of dep. var. 0.50 0.20 0.72 0.36
Panel B: Rich Municipalities (78 municipalities)
Effect of Incumbent Victory 0.40 0.12 -0.77 -0.26
(0.36) (0.15) (0.73) (0.37)
Mean of dep. var. 0.17 0.09 0.46 0.28
Panel C: Poor Municipalities (75 municipalities)
Effect of Incumbent Victory 1.65** 0.56** -0.38 -0.15
(0.79) (0.27) (0.25) (0.18)
Mean of dep. var. 0.84 0.31 0.99 0.44
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: Regressions are based on the 2007 Philippine mayoral elections. All models flexibly control for incumbent margin of
victory on both sides of the threshold. Specifications are identical to those reported in columns 2, 4, 6 and 8 of Table 3. All
regressions are weighted by a triangular kernel with a bandwidth of 0.05. Marginal effects calculated at the sample means
are reported. Standard errors are in parentheses. ∗ , ∗∗ and ∗∗∗ denote statistical significance at the 10%, 5% and 1% levels.
Standard errors in column (3) are clustered at the municipality level. All specifications include the same control variables that
were used in Tables 3 and 4.
29
Second, the results in Table 10 provide evidence that narrow incumbent victories are asso-
ciated with a larger increase in the number of post-election casualties in areas with active
insurgency presence.19 Specifically, the results in column (2) suggest that narrow incumbent
victories in areas with no active insurgency presence are associated with a statistically in-
significant 0.21 increase in the number of post-election casualties. By contrast, the effect in
areas where insurgents are active suggests an additional increase of 0.53, significant at the
5 percent level. Reassuringly, there is no relationship between narrow incumbent victories
and post-election violence in areas with or without insurgency presence. The data also al-
lows the analysis to distinguish between violent incidents involving the NPA or other lawless
elements. The results in Table 7 indicate that the increase in violence after an incumbent’s
narrow victory involves both insurgents belonging to the NPA as well as other lawless ele-
ments. There is no such relationship in the pre-election period. The results of estimating
the relationship between narrow incumbent victories and violence based on the group who
initiated the violent incidents are reported in Table 8. The analysis distinguishes between
incidents initiated by AFP or rebel groups. The results in columns (1)-(4) indicate that both
the AFP and rebel groups initiate incidents of violence after a narrow incumbent victory.
However, as shown in Table 9, the AFP is the group that suffers most of the casualties after
a narrow incumbent win.
19
Based on an intelligence assessment conducted in 2001, the AFP assigned villages into three categories
of rebel activity: threatened, influenced and infiltrated. Insurgent presence is defined as having at least one
village in the municipality in any of the three categories.
30
Table 7. Conflict by Insurgent Group
Post-Election Period Pre-Election Period
Casualties Violent Incidents Casualties Violent Incidents
NPA LE NPA LE NPA LE NPA LE
(1) (2) (3) (4) (5) (6) (7) (8)
Incumbent victory 0.30** 0.24* 0.18** 0.095** 0.017 0.010 –0.11 0.055
(0.14) (0.12) (0.072) (0.038) (0.19) (0.34) (0.11) (0.12)
Margin –0.60 –4.12 –2.42 –1.90** –3.84 11.8 1.03 0.83
(2.32) (2.60) (1.58) (0.93) (3.68) (12.9) (3.59) (4.46)
Margin × incumbent victory –3.62 –11.6 0.47 –2.26 –5.37 –26.6* –5.13 –4.85
(3.81) (7.21) (2.92) (3.20) (10.4) (15.7) (5.59) (5.33)
Mean of dep. var. 0.25 0.25 0.11 0.08 0.30 0.42 0.19 0.17
No. of municipalities 153 153 153 153 153 153 153 153
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The running variable of the RD design is the incumbent margin of victory. All regressions are weighted by a triangular
kernel with a bandwidth of 0.05 and include the same control variables as in Table 3. Reported values are marginal effects,
calculated at the sample means. The dependent variables in columns marked “NPA” are based on conflict involving the New
People’s Army. The dependent variables in columns marked “LE” are based on conflict with other lawless elements. Our sample
does not contain any casualties in incidents with the Moro-Islamic Liberation Front or Abu Sayyaf Group. Standard errors are
in parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1% levels.
31
Table 8. Who Initiates the Incidents?
Post-Election Period Pre-Election Period
Casualties Violent Incidents Casualties Violent Incidents
AFP Rebel AFP Rebel AFP Rebel AFP Rebel
(1) (2) (3) (4) (5) (6) (7) (8)
Incumbent victory 0.24* 0.30** 0.16*** 0.15** 0.010 0.017 –0.025 –0.036
(0.12) (0.14) (0.053) (0.075) (0.34) (0.19) (0.16) (0.12)
Margin –4.12 –0.60 –2.92** –1.79 11.8 –3.84 3.60 –0.94
(2.60) (2.32) (1.37) (1.45) (12.9) (3.68) (6.51) (3.20)
Margin × incumbent victory –11.6 –3.62 –1.45 0.19 –26.6* –5.37 –7.95 –2.91
(7.21) (3.81) (3.38) (2.90) (15.7) (10.4) (7.21) (5.23)
Mean of dep. var. 0.25 0.25 0.11 0.12 0.42 0.30 0.18 0.18
No. of municipalities 153 153 153 153 153 153 153 153
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The running variable of the RD design is the incumbent margin of victory. All regressions are weighted by a triangular
kernel with a bandwidth of 0.05 and include the same control variables as in Table 3. Reported values are marginal effects,
calculated at the sample means. The dependent variables in columns marked “AFP” are based on incidents initiated by the
Armed Forces of the Philippines. The dependent variables in columns marked “Rebel” are based on incidents initiated by
insurgent groups. Standard errors are in parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1%
levels.
32
Table 9. Who Suffers the Casualties?
Post-Election Period Pre-Election Period
Casualties suffered by: Casualties suffered by:
AFP Rebel AFP Rebel
(1) (2) (3) (4)
Incumbent victory 0.61** 0.10 –0.012 0.14
(0.30) (0.071) (0.33) (0.11)
Margin –9.37* –1.89 5.53 –2.25
(4.98) (1.33) (9.46) (3.24)
Margin × incumbent victory –2.98 0.33 –24.5 –1.92
(12.9) (1.36) (15.6) (4.39)
Mean of dep. var. 0.38 0.08 0.50 0.10
No. of municipalities 153 153 153 153
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The running variable of the RD design is the incumbent margin of victory. All regressions are weighted by a triangular
kernel with a bandwidth of 0.05 and include the same control variables as in Table 3. Reported values are marginal effects,
calculated at the sample means. The dependent variable in columns marked “AFP” is the number of casualties suffered by
the Armed Forces of the Philippines. The dependent variable in columns marked “Rebel” is the number of casualties suffered
by insurgent groups. Standard errors are in parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1%
levels.
33
Table 10. Heterogeneous Effects by Insurgent Presence
Post-Election Pre-Election
(1) (2) (3) (4)
Incumbent victory 0.37 0.21 0.11 0.0024
(0.42) (0.33) (0.89) (0.63)
Insurgent presence × incumbent victory 0.84** 0.53** 0.22 0.19
(0.36) (0.23) (0.82) (0.61)
Margin 0.70 7.38 14.8 11.5
(17.0) (15.2) (22.0) (13.7)
Insurgent presence × Margin –22.4 –22.7* –28.3 –17.9
(16.8) (12.8) (18.4) (14.9)
Margin × incumbent victory –4.30 –12.5 –32.7 –37.8
(19.6) (15.9) (39.6) (30.1)
Insurgent presence × Margin × incumbent victory –19.7 11.6 27.3 29.5
(50.3) (21.6) (46.4) (33.9)
Mean of dep. var. 0.50 0.50 0.72 0.72
Control variables No Yes No Yes
No. of municipalities 153 153 153 153
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The running variable of the RD design is the incumbent margin of victory. All regressions are weighted by a triangular
kernel with a bandwidth of 0.05. Reported values are marginal effects, calculated at the sample means. Standard errors are in
parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1% levels. Control variables are the same as in
Table 3.
34
3.4 Mechanisms and Robustness Tests
There appears to be a strong, positive association between narrow incumbent victories and
post-election violent incidents and casualties from civil conflict. One explanation for this
association is that aggrieved voters, believing that incumbents won fraudulently, increase
their support for insurgent groups (Berman et al., 2011a). Alternatively, narrow incumbent
victories may have led to post-election violence for reasons unrelated to fraud. For example,
underlying social triggers (e.g., underlying perceptions of corruption or the abuse of power)
may become more salient when incumbents win narrowly, leading to an increase in violence.
It is also possible that insurgents might have typically preferred the challenger’s political
platform to the incumbent’s and reacted with violence when their candidate lost. Below, the
paper describes several additional analyses aimed at gaining a better understanding of the
mechanisms behind the results reported thus far.
First, the paper tests whether the relationship between narrow incumbent victories and
violence was driven by the political platform of the incumbent by interacting an indicator
for whether the incumbent was affiliated with the governing coalition of President Arroyo
with the other variables in the panel regression. The results are presented in column (5)
of Tables 4 and 5.20 Out of the 153 incumbents in the sample, 116 were affiliated with the
governing coalition. The results suggest that the insurgent reaction did not depend upon the
incumbent’s political affiliation. Specifically, none of the interactions between the governing
coalition indicator and the variables in the panel regressions are statistically significant at
conventional levels.
A number of additional robustness tests are conducted in the Appendix. First, equations (1)
and (2) are re-estimated after redefining Xb to equal the victory/loss margin between the
20
As discussed in Section 2, the governing coalition consisted of center-right parties, while the opposition
mostly consisted of center-left parties. Only two incumbents in the 2007 mayoral elections were members of
a far left party and neither of them was involved in a tightly contested election.
35
most successful candidate from the governing coalition and the most successful candidate
from the opposition party. The results of this analysis, in Appendix Figures A.1–A.2 and
Tables A.3–A.4, show no evidence that coalition candidates are more likely to win close
elections or that narrow victories by coalition candidates lead to increased conflict.
Second, the paper examines the effect of decisive incumbent victories on civil conflict. As
noted above, the evidence for conflict comes from municipalities in which the incumbent won
by a margin of victory between 0 and 1.5 percentage points. When these municipalities are
excluded, the results of the McCrary test show no evidence that incumbents are more likely
to win narrow elections (Appendix Table A.5). If violence is due to incumbents who won
narrowly and not to incumbent victories in general, then incumbent victories should have
no effect on post-election violence when these municipalities are excluded. Appendix Table
A.6 shows that this the case: the estimates of γ1 become small, negative and statistically
insignificant when victories decided by a margin of between 0 and 1.5 percentage points are
excluded.21 The results of this robustness test suggest incumbents winning close elections,
not incumbent victories per se, drove the increase in post-election violence documented in
Tables 3 and 4.
Third, the paper examines the role of political dynasties. Mayors in the Philippines are
term-limited after three elections, and there is evidence that incumbency has a causal effect
on the probability that a direct relative runs in their place once the term-limits have been
reached (Querubin, 2016). To examine the role of family dynasties, equation (1) is estimated
on a sample of closely fought elections in which one of the top-two candidates is a relative
of the incumbent mayor.22
21
Similar results are obtained when using the sample of all municipalities in which the margin of victory
was outside of the range of 0 to 1.5 percentage points (including those municipalities in which incumbents
won with margins of victory above 5 and below -5 percentage points). When the sample is restricted to
municipalities within a bandwidth of 0 to 1.5 percentage points, the effect of incumbent victories is large
and statistically significant.
22
After limiting the sample to elections won with a bandwidth of 5 percentage points, relatives are identified
by manually comparing the last names of candidates to that of the incumbent mayor. This procedure detects
36
The results, reported in column (1) of Appendix Table A.7, provide no evidence that the
relatives of incumbent mayors were more likely to win closely contested elections. Restricting
the sample to elections decided by 3 percentage points or fewer does not change this basic
conclusion. In columns (3) and (4), closely fought elections involving a relative are added
to the main sample. Their inclusion reduces the magnitude of the discontinuity from 0.016
(reported in column (1) of Table 1) to 0.008 and the estimate becomes statistically insig-
nificant. However, the estimate based on a bandwidth of 3 percentage points, reported in
column (4), is very similar in both magnitude and significance to the estimate reported in
column (3) of Appendix Table A.1. The analysis then relates the narrow victory of relatives
to the number of post-election casualties using a sample of narrow incumbent victories which
also includes relatives. The results in columns (1) and (2) of Appendix Table A.8 are similar
to the results reported in Table 3 and Appendix A.2. This analysis suggests that narrow
victories of incumbent relatives are not associated with post-election violence.
4 Conclusion
There is a widespread belief among academics and practitioners that elections perceived as
fraudulent can increase support for non-democratic actors and serve as a catalyst for civil
conflict (World Bank, 2012; Donno, 2013; Norris, 2014; Hall et al., 2015). Over the past few
decades, international donors and the World Bank have intensified their efforts to ensure
that elections in fragile democracies are conducted fairly (Daxecker, 2012) and have spent
substantial amounts on election monitoring and other related programs (Kelley, 2008).
This study examines the association between narrow incumbent victories and civil conflict
using data from the 2007 mayoral elections held in the Philippines. With a forensic ap-
49 municipalities in which one of the candidates is a relative of the incumbent mayor. Limiting the sample
to a bandwidth of 3 percentage points reduces the number of such municipalities to 38.
37
proach, the paper finds that incumbent mayors were more likely to win tightly contested
elections compared to their challengers, an indication of election fraud (McCrary, 2008) or
the presence of other institutional failures that lead to violence following a contested race. In
addition, the results indicate that municipalities in which incumbent mayors were narrowly
elected experienced between 0.89 and 1.22 additional post-election casualties from civil con-
flict during the 12 month period after the election as compared to municipalities in which the
challenger won. However, there is no evidence of an association between narrow incumbent
victories and pre-election violence.
The paper conducted several tests in an effort to better understand the mechanisms under-
pinning the relationship between narrow incumbent victories and post-election violence. For
instance, the analysis showed that narrow incumbent victories were associated with post-
election violence only in poor municipalities. In richer municipalities, where there is little
evidence that incumbents disproportionately win narrow elections, there was essentially no
relationship between post-election violence and narrow incumbent victories. The paper also
showed that decisive incumbent victories do not lead to an increase in post-election violence.
Finally, the paper showed that the increase in post-election violence was unrelated to the
party affiliation of the incumbent, suggesting that the increase in violence does not simply
reflect dissatisfaction with the incumbent’s political platform.
Although it is difficult to determine the exact mechanism at work, anecdotal evidence sug-
gests that electoral fraud (or the perception of it) and other institutional weaknesses such as
corruption or the abuse of power may increase support for insurgents and make it easier for
them to recruit, which is consistent with a “hearts-and-minds” model of insurgency (Berman
et al., 2011a). The results support the argument that elections held during (and potentially
immediately after) civil conflicts can have unanticipated negative consequences Flores and
Nooruddin (2012); Brancati and Snyder (2011); Matanock (2016) and that election monito-
ring and automated vote counting could help to dampen post-election unrest. In addition,
38
they suggest that future election monitoring should not focus on presidential and congressi-
onal races to the exclusion of local races if the goal is to reduce post-election violence.
39
References
Abadie, Alberto and Javier Gardeazabal, “The economic costs of conflict: A case study
of the Basque Country,” American Economic Review, 2003, 93 (1), 113–132.
Akresh, Richard, Leonardo Lucchetti, and Harsha Thirumurthy, “Wars and Child
Health: Evidence from the Eritrean-Ethiopian Conflict,” Journal of Development Econo-
mics, 2012, 99 (2), 330–340.
Arnaiz, Jani C., Glamorfe L. Calicdan, Aizeline M. David, Che de los
Reyes, Sarita Kare-Telado, Julius D. Mariveles, Jay D. Mendoza, To-
nette T. Orejas, Eladio D. Perfecto, and Aireen Perol-Jaymalin, “The Co-
melec: A reality check - Far too much work, far too few personnel with poor provisi-
ons,” Philippine Center for Investigative Journalism, 2013. http://pcij.org/stories/
far-too-much-work-far-too-few-personnel-with-poor-provisions/.
Beath, Andrew, Fotini Christia, and Ruben Enikolopov, “Winning Hearts and Minds
through Development Aid: Evidence from a Field Experiment in Afghanistan,” Centre for
Economic and Financial Research at New Economic School Working Paper No 166, Oc-
tober 2011. https://www.princeton.edu/~pcglobal/conferences/methods/papers/
beath.pdf.
Beber, Bernd and Alexandra Scacco, “What the Numbers Say: A Digit-Based Test for
Election Fraud,” Political Analysis, 2012, 20 (2), 211–234.
Berman, Eli, Jacob N. Shapiro, and Joseph H. Felter, “Can Hearts and Minds Be
Bought? The Economics of Counterinsurgency in Iraq,” Journal of Political Economy,
August 2011a, 119 (4), 766–819.
40
, Michael Callen, Joseph H. Felter, and Jacob N. Shapiro, “Do Working Men
Rebel? Insurgency and Unemployment in Afghanistan, Iraq, and the Philippines,” Journal
of Conflict Resolution, 2011b, 55 (4), 496–528.
Bertrand, Marianne, Simeon Djankov, Rema Hanna, and Sendhil Mullainathan,
“Obtaining a Driver’s License in India: An Experimental Approach to Studying Corrup-
tion,” The Quarterly Journal of Economics, 2007, 122 (4), 1639–1676.
Besley, Timothy and Robin Burgess, “The political economy of government responsi-
veness: Theory and evidence from India,” Quarterly Journal of Economics, 2002, 117 (4),
1415–1451.
Blakeslee, David S., “Politics and Public Goods in Developing Countries: Evidence from
India,” Working Paper, 2013. http://www.columbia.edu/~dsb2108/davidsblakeslee_
job_market_paper.pdf.
Blattman, Christopher and Edward Miguel, “Civil War,” Journal of Economic Lite-
rature, 2010, 48 (1), 3–57.
Borzyskowski, Inken von, “Sore Losers? International Condemnation and Domestic In-
centives for Post-Election Violence,” Working Paper, 2013. http://www.polisci.wisc.
edu/Uploads/Documents/IRC/Borzyskowski_Election%20Violence.pdf.
Brancati, Dawn and Jack L. Snyder, “Rushing to the Polls: The Causes of Premature
Postconflict Elections,” Journal of Conflict Resolution, June 2011, 55 (3), 469 –492.
Callen, Michael and James D. Long, “Institutional Corruption and Election Fraud:
Evidence from a Field Experiment in Afghanistan,” American Economic Review, 2015,
105 (1), 354–381.
41
Crost, Benjamin, Joseph H. Felter, and Patrick B. Johnston, “Aid Under Fire:
Development Projects and Civil Conflict,” American Economic Review, 2014, 104 (6),
1833–1856.
Daxecker, Ursula E., “The Cost of Exposing Cheating: International Election Monitoring,
Fraud, and Post-Election Violence in Africa,” Journal of Peace Research, 2012, 49 (4),
503–516.
Donno, Daniela, Defending Democratic Norms: International Actors and the Politics of
Electoral Misconduct, Oxford University Press, 2013.
Eggers, Andrew C., Olle Folke, Anthony Fowler, Jens Hainmueller, Andrew B.
Hall, and James M. Snyder, “On The Validity Of The Regression Discontinuity Design
For Estimating Electoral Effects: New Evidence From Over 40,000 Close Races,” American
Journal of Political Science, 2015, 59 (1), 259–274.
Ferraz, Claudio and Frederico Finan, “Exposing corrupt politicians: The effects of
Brazil’s publicly released audits on electoral outcomes,” Quarterly Journal of Economics,
2008, 123 (2), 703–745.
Fisman, Raymond and Edward Miguel, “Corruption, Norms and Legal Enforcement:
Evidence from Diplomatic Parking Tickets,” Journal of Political Economy, 2007, 115 (6),
1020–1028.
Flores, Thomas E. and Irfan Nooruddin, “The Effect of Elections on Postconflict Peace
and Reconstruction,” Journal of Politics, 2012, 74 (2), 558–570.
Grimmer, Justin, Eitan Hersh, Brian Feinstein, and Daniel Carpenter, “Are Close
Elections Random?,” Working Paper, 2011. http://stanford.edu/~jgrimmer/cef2.
pdf.
42
Hall, Thad E., Susan Hyde, and Beth Wellman, “Election Quality and Public Confi-
dence in Political Institutions: Revisiting the Orange Revolution,” Working Paper, 2015.
http://susan.hyde.co/HallHydeWellman.pdf.
Hegre, Havard, Tanja Ellingsen, Scott Gates, and Nils Petter Gleditsch, “Toward
a Democratic Civil Peace? Democracy, Political Change, and Civil War, 1816-1992,”
American Political Science Review, 2001, 95 (1), 33–48.
Henderson, Errol A. and J. David Singer, “Civil War in the Post-Colonial World,
1946-92,” Journal of Peace Research, 2000, 37 (3), 275–299.
Holden, William N., “Ashes from the Phoenix: State Terrorism and the Party-List Groups
in the Philippines,” Contemporary Politics, 2009, 15 (4), 377–393.
Hyde, Susan D. and Nikolay Marinov, “Information and Self-Enforcing Democracy:
The Role of International Election Observation,” International Organization, 2014, 68
(2), 329–359.
Iyengar, Radha, Jonathan Monten, and Matthew Hanson, “Building Peace: The
Impact of Aid on the Labor Market for Insurgents,” NBER Working Paper 17297, August
2011. http://www.nber.org/papers/w17297.pdf.
Kelley, Judith, “Assessing the Complex Evolution of Norms: The Rise of International
Election Monitoring,” International Organization, 2008, 62 (2), 221–255.
Kulkarni, Parashar, “Electoral Fraud and Strategic Electoral Reform Politics,” Working
Paper, 2012. https://www.exeter.ac.uk/media/universityofexeter/elecdem/pdfs/
florence/Kulkarni_Electoral_Fraud_and_Strategic_Electoral_Reform.pdf.
Leon, Gianmarco, “Civil Conflict and Human Capital Accumulation: The Long Term
Effects of Political Violence in Peru,” Journal of Human Resources, 2012, 47 (4), 991–
1022.
43
Madlos, Jorge, “Fight systematic fraud being perpetrated by the
Arroyo ruling clique in the May 2007 elections,” Philippine Re-
volution Web Central, 2007. http://www.philippinerevolution.net/statements/20070508_
fight-systematic-fraud-being-perpetrated-by-the-arroyo-ruling-clique-in-the-may-2007-elections-br-a-call-for-sincere-officials-of-comelec .
Mansour, Hani and Daniel I. Rees, “Armed conflict and Birth Weight: Evidence from
the al-Aqsa Intifada,” Journal of Development Economics, 2012, 1 (99), 190–199.
Matanock, Aila, “Bullets for Ballots: Electoral Participation Provisions in Peace
Agreements and Conflict Recurrence,” Working Paper, 2016. http://matanock.
berkeley.edu/wp-content/uploads/Matanock_Bullets-for-Ballots_June2016.pdf.
McCrary, Justin, “Manipulation of the Running Variable in the Regression Discontinuity
Design: A Density Test,” Journal of Econometrics, 2008, 142 (2), 697–714.
Mebane, Walter, “Election Forensics: The Second-digit Benford’s Law Test and Recent
American Presidential Elections,” Election Fraud Conferece, Salt Lake City, Utah, Sep-
tember 29-30. http://www-personal.umich.edu/ wmebane/fraud06.pdf, 2006.
Norris, Pippa, Why Electoral Integrity Matters, Cambridge University Press, 2014.
Olken, Ben and Rohini Pande, “Corruption in Developing Countries,” Annual Review
of Economics, 2012, 117 (3), 479–505.
Olken, Benjamin, “Monitoring Corruption: Evidence from a Field Experiment in Indone-
sia,” Journal of Political Economy, 2007, 115 (2), 200–249.
, “Corruption Perceptions vs. Corruption Reality,” Journal of Public Economics, 2009, 93
(7-8), 950–964.
Pastor, Robert A., “The Role of Electoral Administration in Democratic Transitions:
Implications for Policy and Research,” Democratization, 1999, 6 (4), 1–27.
44
Querubin, Pablo, “Family and Politics: Dynastic Persistence in the Philippines,” Quarterly
Journal of Political Science, 2016, 11 (2), 151–181.
Rabasa, Angel, John IV Gordon, Peter Chalk, Audra K. Grant, Scott K.
McMahon, Stephanie Pezard, Caroline Reilly, David Ucko, and Rebecca S.
Zimmerman, “From Insurgency to Stability, Volume II: Insights from Selected Case
Studies.,” Santa Monica, California: Rand Corporation, 2011.
Salas, Santiago, “NDF-EV condemns virtual martial rule in troop deploy-
ment to Tacloban City and military machinations in the election period,”
Philippine Revolution Web Central, 2007. http://www.philippinerevolution.net/statements/20070523_
ndf-ev-condemns-virtual-martial-rule-in-troop-deployment-to-tacloban-city-and-military-machinations-in-the-election-period.
Schedler, Andreas, “Elections without Democracy: The Menu of Manipulation,” Journal
of Democracy, 2002, 13 (2), 36–50.
Schock, Kurt, “A Conjunctural Model of Political Conflict: The Impact of Political Oppor-
tunities on the Relationship between Economic Inequality and Violent Political Conflict,”
Journal of Conflict Resolution, 1966, 40 (1), 98–133.
Solidarity Philippines Australia Network, “Philippines Elections 2007: A Climate of
Resignation, Fraud, and Violence,” KASAMA, 2007, 21 (2). http://cpcabrisbane.org/
Kasama/2007/V21n2/IOM-InitialFindings.htm.
Tucker, Joshua A., “Enough! Electoral Fraud, Collective Action Problems, and Post-
Communist Colored Revolutions,” Perspectives on Politics, 2007, 5 (3), 535–551.
USAID, Electoral Security Assessment - Philippines, United States Agency for International
Development, Washington, D.C., 2008.
Weidmann, B. Nils and Michael Callen, “Violence, Control and Election Fraud: Evi-
dence from Afghanistan,” British Journal of Political Science, 2013, 43, 53–75.
45
World Bank, World Development Report 2011: Conflict, Security, and Development 2012.
Zitzewitz, Eric, “Forensic Economics,” Journal of Economic Literature, 2012, 50 (3), 731–
769.
46
Figure A.1. Did the Governing Coalition Manipulate Elections?
.1
.08
.06
pdf
.04
.02
0
−.05 −.04 −.03 −.02 −.01 0 .01 .02 .03 .04 .05
Margin of victory
Mean Nonparametric fit
Linear fit
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The figure presents the probability density of coalition candidates’ margin of victory in the 2007 election. Municipalities
were grouped into 20 bins of equal width according to the incumbent’s margin of victory. Each scatter dot represents one bin.
Its horizontal coordinate represents the midpoint of the bin, and its vertical coordinate represents the fraction of municipalities
for which the incumbent’s margin of victory was within the bin. Dashed lines are linear fits, separately estimated on both sides
of the eligibility threshold. Solid lines are nonparametric fits from a local linear regression that uses triangular kernels with a
bandwidth of 5 percentage points, separately estimated on both sides of the eligibility threshold.
47
Figure A.2. Did Narrow Victories by Coalition Candidates Cause Conflict?
50
40 30
Casualties
20 10
0
−.05 −.04 −.03 −.02 −.01 0 .01 .02 .03 .04 .05
Margin of victory
Mean Nonparametric fit
Linear fit
20
15
Casualties
10
5
0
−.05 −.04 −.03 −.02 −.01 0 .01 .02 .03 .04 .05
Margin of victory
Mean Nonparametric fit
Linear fit
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The figure presents the relationship between the coalition candidate’s margin of victory and casualties experienced in
the 12 months before and after the 2007 election. Scatter dots represent the mean of casualties per month and are sized to
reflect the number of municipalities in each bin. Dashed lines are linear fits, separately estimated on both sides of the eligibility
threshold. Solid lines are nonparametric fits from a local linear regression that uses triangular kernels with a bandwidth of 5
percentage points, separately estimated on both sides of the eligibility threshold.
48
Table A.1. McCrary-Test for Election Manipulation by Incumbent Mayors: Robustness to
Varying Bandwidths
Dependent variable:
Fraction of municipalities within bin
(1) (2) (3)
Incumbent Victory 0.016** 0.021* 0.023**
(0.0079) (0.011) (0.0094)
Constant 0.028*** 0.025*** 0.022***
(0.0056) (0.0076) (0.0066)
Bandwidth 0.05 0.04 0.03
Bins 30 24 18
Municipalities 153 123 92
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: Results of a probability density test for manipulation of the running variable (McCrary, 2008) are reported. The running
variable is the incumbent margin of victory. Each regression is weighted by a triangular kernel with the bandwidth shown at
the top of the column. Observations are 30 bins of equal width. The dependent variable is the fraction of municipalities with
an incumbent margin of victory that falls within the bin. Standard errors are in parentheses. ∗ , ∗∗ and ∗∗∗ denote statistical
significance at the 10%, 5% and 1% levels, respectively.
49
Table A.2. Estimates from RD Regressions: Testing the Robustness to Varying Bandwidths
Poisson Estimates
Casualties Violent incidents
(1) (2) (3) (4) (5) (6)
Incumbent Victory 0.89*** 0.99*** 1.27** 0.31*** 0.34*** 0.45***
(0.29) (0.34) (0.52) (0.12) (0.12) (0.15)
Margin -10.1 -12.8 -38.0 -5.67** -7.1* -17.1
(6.6) (12.3) (24.7) (2.45) (3.8) (7.2)
Margin × incumbent victory -17.5 -13.6 17.1 -0.43 0.66 12.0
(26.4) (19.2) (23.22) (6.3) (9.4) (11.4)
Bandwidth 0.05 0.04 0.03 0.05 0.04 0.03
Mean of dep. var. 0.50 0.49 0.64 0.2 0.17 0.22
Municipalities 153 123 92 153 123 92
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: Each regression is weighted by a triangular kernel with the bandwidth shown at the top of the column. The sample
is restricted to observations within 12 months after the 2007 election. Marginal effects calculated at the sample means are
reported. Standard errors are in parentheses. ∗ , ∗∗ and ∗∗∗ denote statistical significance at the 10%, 5% and 1% levels. All
regressions include the same control variables as Table 3.
50
Table A.3. McCrary Test for Election Manipulation by Candidates Affiliated with the Go-
verning Coalition
OLS Estimates
Dependent variable: Fraction of municipalities within bin
(1)
Coalition victory -0.002
(0.009)
Margin 0.33
(0.30)
Margin× Coalition victory -0.045
(0.43)
Constant 0.034***
(0.006)
Municipalities 198
Observations (bins) 30
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: Results of a probability density test for manipulation of the running variable (McCrary, 2008) are reported. The running
variable of the RD design is the coalition candidate’s margin of victory. All regressions are weighted by a triangular kernel with
a bandwidth of 0.05. Observations are 30 bins of equal width. The dependent variable is the fraction of municipalities with
an incumbent margin of victory that falls within the bin. Standard errors are in parentheses. ∗ , ∗∗ and ∗∗∗ denote statistical
significance at the 10%, 5% and 1% levels, respectively.
51
Table A.4. Coalition Victories and Conflict: RD Design
Poisson Estimates: Casualties in
the 12 months after the 2007 election
Post-Election Pre-Election
(1) (2) (3) (4)
Coalition victory 0.085 0.040 0.083 -0.072
(0.451) (0.34) (0.51) (0.46)
Margin -14.9** -12.3* -14.3 -8.2
(6.91) (6.45) (12.5) (12.4)
Margin × coalition victory 30.6** 27.2** 31.8* 23.8
(13.3) (8.71) (18.7) (20.2)
Controls No Yes No Yes
Mean of dep. var. 0.68 0.68 0.88 0.88
Municipalities 198 198 198 198
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The running variable of the RD design is the coalition candidate’s margin of victory. All regressions are weighted by a
triangular kernel with a bandwidth of 0.05. Marginal effects calculated at the sample means are reported. Standard errors are
in parentheses. ∗ , ∗∗ and ∗∗∗ denote statistical significance at the 10%, 5% and 1% levels. Control variables are the same as
those used in Table 3.
52
Table A.5. McCrary Test for Election Manipulation: Excluding Very Narrow Incumbent
Victories
OLS Estimates
Dependent variable: Fraction of municipalities within bin
Whole Sample Poor Municipalities Rich Municipalities
(1) (2) (3)
Incumbent victory -0.002 -0.005 0.004
(0.005) (0.006) (0.008)
Constant 0.031*** 0.032*** 0.036***
(0.003) (0.0031) (0.005)
Municipalities 122 58 64
Observations (bins) 24 24 24
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The table compares the probability of incumbent victories with margins between 2 and 5 percentage points and incumbent
losses with margins between 0 and 5 percentage points. Observations are 24 bins of equal width. The dependent variable is the
fraction of municipalities with an incumbent margin of victory that falls within the bin. Standard errors are in parentheses. ∗,
∗∗ and ∗∗∗ denote statistical significance at the 10%, 5% and 1% levels, respectively.
53
Table A.6. RD Estimates Excluding Very Narrow Incumbent Victories
Post-Election Period Pre-Election Period
Casualties Violent Incidents Casualties Violent Incidents
(1) (2) (3) (4)
Incumbent victory 0.47 0.21 –0.16 0.13
(0.46) (0.15) (0.63) (0.40)
Margin –5.42* –3.46** 5.78 2.10
(3.06) (1.59) (9.52) (5.50)
Margin × incumbent victory –9.03 –1.39 –18.7 –17.0
(16.8) (5.47) (24.3) (15.2)
Mean of dep. var. 0.30 0.16 0.66 0.35
No. of municipalities 122 122 122 122
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The sample consists of municipalities with incumbent margins of victory between -5 and 5 percentage points, except
victories with a margin of victory smaller than 1.5 percentage points. The running variable of the RD design is the incumbent
margin of victory. All regressions are weighted by a triangular kernel with a bandwidth of 0.05 and include the same control
variables as in Table 3. Reported values are marginal effects, calculated at the sample means. The dependent variable in columns
marked is the number of casualties suffered by the Armed Forces of the Philippines. The dependent variable in columns marked
is the number of casualties suffered by insurgent groups. Our sample does not contain any casualties in incidents with the
Moro-Islamic Liberation Front or Abu Sayyaf Group. Standard errors are in parentheses. *, ** and *** denote statistical
significance at the 10%, 5% and 1% levels.
54
Table A.7. McCrary Test for Election Manipulation by Incumbent Mayors and their Relatives
OLS Estimates
Dep. var.: Fraction of municipalities within bin
Relatives Only Incumbents and Relatives
(1) (2) (3) (4)
Victory 0.0026 0.011 0.0080 0.021**
(0.016) (0.017) (0.0069) (0.0080)
Margin –0.14 –1.15 0.056 –0.99*
(0.69) (1.00) (0.24) (0.46)
Margin × victory –0.54 0.50 –0.36 0.36
(0.82) (1.40) (0.34) (0.65)
Bandwidth 0.05 0.03 0.05 0.03
No. of bins 30 18 30 18
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: Results from a probability density test for manipulation of the running variable (McCrary, 2008) are reported. The
running variable of the RD design is the incumbent (or his/her relative’s) margin of victory. Regressions are weighted by a
triangular kernel bandwidths of 0.05 (columns 1 and 3) and 0.03 (columns 2 and 4). Observations are 30 bins of equal width.
The dependent variable is the fraction of municipalities with an incumbent margin of victory that falls within the bin. Standard
errors are in parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1% levels, respectively.
55
Table A.8. Dynastic Victories and Conflict: RD Design
Poisson Estimates: Casualties in the
12 months before/after the 2007 election
Incumbents and Relatives
(1) (2)
Incumbent victory 1.04** 1.36***
(0.44) (0.52)
Margin –18.1*** –47.4**
(6.13) (18.6)
Margin × incumbent victory 4.69 24.5
(18.2) (31.0)
Bandwidth 0.05 0.03
No. of municipalities 202 130
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The running variable of the RD design is the incumbent (or his/her relative’s) margin of victory. Regressions are weighted
by a triangular kernel with bandwidths of 0.05 (column 1) or 0.03 (column 2). Reported values are marginal effects, calculated
at the sample means. Standard errors are in parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1%
levels. Following the classification of the 2000 Census of the Philippines, building materials for walls are defined as concrete,
brick, stone, wood, galvanized iron or aluminum, asbestos and glass. building materials for roofs are defined as concrete,
galvanized iron or aluminum, clay tiles and asbestos
56
Table A.9. Incumbent Victories and Conflict Casualties: Probability of at least 1, 3, or 5
casualties
Probit Estimates:
1 or more cas. 3 or more cas. 5 or more cas.
(1) (2) (3) (4) (5) (6)
Incumbent victory 0.22** 0.20** 0.20*** 0.17*** 0.059 0.032
(0.095) (0.091) (0.071) (0.061) (0.041) (0.027)
Margin –4.61** –3.92** –3.74*** –3.00** –0.84 –0.41
(1.95) (1.85) (1.40) (1.18) (0.68) (0.45)
Margin × incumbent victory 1.07 0.78 0.41 0.17 –2.01 –1.04
(3.41) (3.36) (2.54) (1.97) (2.08) (0.95)
Mean of dep. var. 0.12 0.12 0.072 0.072 0.026 0.026
Control variables No Yes No Yes No Yes
No. of municipalities 153 153 153 153 153 153
Source: Data for the study comes from field reports of the Armed Forces of the Philippines (AFP) and election returns from
the Philippines Commission on Elections (COMELEC).
Note: The running variable of the RD design is the incumbent margin of victory. All regressions are weighted by a triangular
kernel with a bandwidth of 0.05. Reported values are marginal effects, calculated at the sample means. Standard errors are in
parentheses. *, ** and *** denote statistical significance at the 10%, 5% and 1% levels. Control variables are the same as in
Table 3.
57