Supplementary Appendix.

Transcription

Appendix. Data Considerations, Reporting Processes, and Additional Empirical and Qualitative Evidence for Land Reform as a
Counterinsurgency Policy: Evidence from Colombia
This appendix provides additional background information and robustness tests to probe the
quality of our data and results. The appendix is organized into five sections.
In Section 1 we document the sources, definitions, and coding of armed group actions for both
the Colombian government data as well as data collected independently by the Jesuit think tank
Centro de Investigación y Educación Popular, or CINEP. We discuss how these data are collected
and reported, and present details on the fidelity of the process as well as potential biases based on
documentation on reporting processes as well as interviews with government officials.
In Section 2 we discuss the threats to inference posed by reporting biases, as well as different
methods for estimating guerrilla activities. We introduce separate, aggregate data collected by human
rights monitors, and conduct an analysis of the most “visible” guerrilla activities that are arguably
less subject to potential reporting bias. We also address how the possibility of reciprocity between
guerrilla actions and government actions may affect the estimations.
In Section 3 we conduct a series of overlap analyses in an effort to gain a better understanding
of potential reporting biases. We also discuss the nature of the reporting bias that must exist for
it to be problematic for our results, and then implement a series of sensitivity analyses to explore
the robustness of our results to potential biases in the reporting of guerrilla attacks. These exercises
suggest the nature and degree of bias would have to be rather severe to alter our conclusions.
In Section 4 we discuss additional robustness tests that are relevant to the manuscript’s findings,
including several that are cited but not reported in the manuscript. The logic of each of these is
discussed in this section.
In Section 5 we provide additional qualitative evidence on the link between land reform and
guerrilla activity based on case studies from Araucan towns.
1
1.1
Definitions and Coding for Armed Group Actions
Data on Armed Group Actions from Colombian Government
Colombian government data on armed group actions is taken from the Colombian Vice-Presidencys
Human Rights Observatory and Sánchez (2007). Sánchez’s (2007) data comes from information from
both the Colombian Government and the non-profit organization Fundación Social. Data from these
two sources are now managed by the government’s Human Rights Observatory.
The definitions of armed group actions are consistently applied across the country (Human Rights
Observatory 2006), and therefore do not vary by locale. Furthermore, the categorization of guerrilla,
paramilitary, and government actions limits the possibility of their activities being conflated with
activities of other armed groups. Data for each of these groups are coded separately. Finally, the
definition of guerrilla activities specifically targeted at civilians such as homicides, political homicides,
displacement, massacres, and kidnappings.
The following definitions were translated from the Spanish from:
Observatorio del Programa Presidencial de Derechos Humanos y DIH. Nota Metodológica: Bitácora
Semana de Prensa. Bogotá: Vicepresidencia de la República de Colombia.
http://www.derechoshumanos.gov.co/observatorio de DDHH/nota metodologica.asp
1
1.1.1
Actions of Guerrilla and Paramilitary Groups
The following are considered actions of illegal armed groups: attacks on infrastructure and property, attacks on aircraft, roadblocks, demobilizations, ambushes, clashes, voluntary surrenders, use
of human shields, harassment, theft, other indiscriminate attacks, armed strikes, car theft, forced
recruitment, illegal roadblocks, attacks against population centers and other terrorist events. All
guerrilla structures as well as paramilitaries are considered illegal armed groups.
Actions against the security forces or other armed groups:
Attack on installations and public security forces: Attacks committed by illegal armed
groups directed against fixed posts of the public security forces such as military bases, posts
of the National Police and/or security agencies.
Ambush: Attack by an illegal armed group, against a moving patrol, moving unit or members
of the public security forces, by surprise and whose intensity is higher than the expected
response.
Harassment: The attack by an illegal armed group against fixed or moving units of members
of the public security forces, by surprise and whose intensity is less than the expected response.
Attacks on private property (and infrastructure): The following are considered attacks on
infrastructure and assets: those acts performed by members of illegal armed groups against cultural property, healthcare facilities and goods, state assets, objects indispensable to the survival
of the population, private goods, property for public use, commercial and financial infrastructure,
infrastructure for communications, education, energy and oil, roads, facilities containing dangerous
emissions and/or radiation and against commercial and public transport.
Roadblocks: Action through which an illegal armed group prevents or obstructs traffic on the
tracks, using illegal methods, which includes incineration of vehicles or the installation of explosive
devices.
Confrontation: Armed clashes that develop between members of illegal armed groups.
Incursions on Population Centers: Any action taken by illegal armed groups against a town
where there is presence of security forces, and whose effects are dualamong civilians and the militaryon goods or people. The effect on goods may be indicated by the partial or total destruction
of dwellings, places of worship, government facilities, and the total or partial destruction of police
stations. The effect on military personnel or civilians refers to the murder and /or personal injury,
among others, on said individuals.
Other terrorism attack (bombing, arson, other): Acts of violence directed against the civilian
population, whose main purpose is to terrorize, through the use of explosive devices, and that result
in death and/or injury of civilians and/or damage to the property of civilians.
Robbery (highway): Action taken in public through which one or more subjects appropriates a
vehicle and/or the property and valuables within it.
Roadblock: Actions taken by illegal armed groups carried out on roads with the purpose of blocking
free passage, and ultimately intended to commit crimes such as kidnappings, murders, and thefts,
among others, or get information on the movement of regular troops, other armed groups, and the
public.
1.1.2
Actions By Law Enforcement and Other Public Officials
The following are considered actions of the Security Forces and other public servants: bombings, arrests, fighting, explosive ordnance disposal, the dismantling of camps, the destruction of illicit crops,
laboratories and other illegal infrastructure, actions arising out of friendly fire, seizures, mine deactivation incidents, damage to infrastructure and property, forced releases and rescues of abducted
2
victims. The following are considered members of the Armed Forces: members of the Army, Navy,
Air Force and National Police; officials of national government entities such as DAS, CTI, the Fiscalı́a
(Attorney General), as well as other officials who are linked to the government.
Capture: The apprehension of a person to deprive him of liberty, whether he was caught in the act
or through an arrest warrant.
Explosive Ordinance Disposal: The controlled neutralization and/or deactivation of explosive
devices conducted by members of the security forces and other public servants.
Raids (dismantling of camps): The planned efforts by members of the security forces and other
public servants to destroy and dismantle a camp used to house members of illegal armed groups. Also
includes the planned effort by members of the security forces and other public servants, to destroy
and/or dismantle facilities, infrastructure and equipment used for illegal purposes (destruction of
other illegal infrastructure).
Destruction of laboratories: The planned efforts by members of the security forces and other
public servants to destroy facilities and inputs used to process illicit drugs.
Seizure: The planned efforts by members of the security forces and other public servants to legally
confiscate material used illegally, for illegal purposes, or acquired illegally. These include weapons,
fuel, communications equipment, explosives, war material, narcotics and precursors, propaganda,
among others.
Rescue and Liberation by Force: Action through which a kidnapper is forced to free the victim
by military pressure exerted by members of security forces and other public servants. Also offensive
action through which members of the security forces and other public servants secure the release of
a kidnapped/abducted victim.
1.1.3
Data Collection and Reporting Processes
The process of how data on conflict incidents are collected by the Colombian Armed Forces, from
the unit on the ground up to the General Command, is detailed in the following document:
“Procesos de Información Comando General de las Fuerzas Militares” (Information Processes of
the General Command of the Armed Forces), Colombian Ministry of Defense, Office of Strategy and
Planning, Sectoral studies group.
The document describes in detail the process of how data on conflict incidents are collected by
the Colombian Armed Forces, from the unit on the ground all the way up to the General Command.
There is substantial overlap in these processes among the various branches of the armed forces, although there are also some slight variations in reporting for groups such as mobile units (e.g., naval
vessels and aircraft). As an example of these processes, below we have outlined the process of the
Army, which consists of no fewer than three levels of review.
1. Tactical unit sends radio report of incidents to the Brigade.
2. Brigade reviews information for accuracy and sends information to the Division (either electronically or by radio), which verifies and refines information on a daily basis for completeness.
3. The division enters the information into the Information System of the Army Command
(SICOE) and a paper report is sent to the General Command of the Armed Forces, ensuring there
are no duplicates of reports.
4. Officers at the Central Command review information for accuracy and combine reports from
different branches of the armed forces.
5. Data is then given to the civilian officials in the Sectoral Studies unit of the Ministry of
Defense.
3
Figure 1: Information Processes of the General Command of the Armed Forces
The diagram below (Figure 1) details the flow of information through the General Command of
the Armed Forces.
In order to help confirm reporting procedures and potential sources of bias in reporting, we
conducted interviews with top government officials in the Ministry of Defense as well as with retired
military field intelligence (and other) officers who were on active duty during the 1990s through the
early 2000s. The officers ranged in rank from Lieutenants up to the mid-level officer rank of Captain.
Two were intelligence officers and one was the liaison between his battalion and brigade. Among
these officers, tours of duty were served in 11 of Colombias 32 departments, including Arauca and
other highly conflictive zones that had been INCORA colonization zones.
Statements from these interviews closely confirm the process of reporting conflict events displayed
in Figure 1. As one officer noted, “The communication centers of the military units (Battalions,
Brigades, Divisions, and the General Command) function 24 hours a day, seven days a week. The
patrol reports the new incident and it is the commander of the Battalion who decides which information to provide to higher-level commanders.” As an example of this process, a battalion in the high
conflict zone of the southern part of the department of Santander recorded and reported a broad
variety of incidents in 1992, including a case when guerrillas attacked officers on guard duty at the
base. As one of the officers noted, it was important to report information to unit commanders “for
strategic use in operations against the enemy.” He also thought that the process of reporting incidents was “basically the same over time.” In the manuscript, of course, we include as a precaution
year fixed effects in the models to account for possible secular changes in reporting over time.
Most importantly for understanding possible data biases, these interviews suggest that while
there may be opportunities for manipulation of information, these tend to be rare and are often
4
detected and punished severely. For instance, while one officer thought it would likely be more
difficult to manipulate reports of some kinds of events than others, with “massive kidnappings or
attacks on populations” being relatively difficult to manipulate, he also noted that, “Underreporting
adverse events could lead to an officer’s retirement from active service.” A second officer corroborated
this, noting that the few cases of attempted manipulation he knew of triggered punishments and that
subsequent actions were taken to limit such actions in the future: “Misinformation was later detected
and punished with strong disciplinary sanctions. As a result of this, the use of technology increased to
process information at all levels of the Command to avoid false information at all costs and legitimize
the military to the Colombian people.” He noted that commanders at every level are cautious about
reporting information that is not true or verified since it could provide cause for criticism from the
media or the “enemy.”
In sum, while these interviews only represent a small sample of operating officers, they suggest
there are considerable efforts to control attempts at manipulation of information and do not give
reason to believe that any manipulation that may occur is systematically correlated with levels of
land reform.
1.2
Data on Armed Group Actions from Centro de Investigación y Educación
Popular (CINEP)
The definitions of armed group actions as coded by CINEP correspond with those used by the
Colombian government.
The CINEP conflict data is based on events published in 20 national and regional press sources.
The national newspapers include El Tiempo, El Espectador, La República, El Nuevo Siglo, Voz
(weekly). The regional news sources include El Colombiano and El Mundo (Medellı́n), El Paı́s (Cali),
El Heraldo (Barranquilla), Vanguardia Liberal (Bucaramanga), La Opinión (Cúcuta), La Nación
(Neiva), Nuevo Dı́a (Ibagué), La Patria (Manizales), El Liberal (Popayán), Llano 7 Dı́as (Meta),
Boyacá 7 Dı́as (Tunja), La Tarde (Pereira), El Universal (Cartagena), Hoy Diario del Magdalena
(Santa Marta).
CINEP also verifies their data on conflict events with local clergy and church officials. Further
details on CINEPs data collection and categorization methodology can be found in:
CINEP. 2008. Marco Conceptual Banco de Datos de Derechos Humanos y Violencia Polı́tica. Second
Edition, October 2008, Bogotá D.C.: Centro de Investigación y Educación Popular.
2
2.1
Estimating Guerrilla Activities
Multiple Systems Estimation
One technique for estimating the true values of a phenomenon that possibly suffers from hidden
cases and misreporting is multiple systems estimation (MSE; see e.g. Lum et al. 2010). For multiple
systems estimation to be implemented most rigorously, it ideally requires at least three independent
data sources of guerrilla activity disaggregated into individual events. Unfortunately, events-level
data for the OVP (Government) and CINEP variables are not available with complete coverage, and
there has been no third independent collection of guerrilla activity that spans this time period with
full country coverage. We cannot therefore conduct a viable MSE to approximate the true severity
of potential reporting biases. It is unlikely that events-level data from multiple independent sources
could be collected at the national level in this case due to data availability issues. An alternative
approach of aggregating many series of local-level data collection efforts would likely suffer substantial
5
missing coverage of certain areas. Short of using MSE, we take several additional steps to probe the
quality of our data, the nature and degree of bias that would be problematic for our inferences, and
the sensitivity of the results to different assumptions about possible reporting bias.
2.2
Robustness to Using More “Visible” Guerrilla Activities
Since some kinds of guerrilla activities are more likely to be susceptible to bias or manipulation
than others, we conduct several additional robustness tests that separate out these kinds of events.
Guerrilla activities fall into a series of different categories (see Section 1). Some of these categories
capture events that, unlike events such as human disappearances, are more publicly visible and are
therefore less likely subject to the same degrees of misreporting and bias than certain other guerrilla
actions. In other words, we are more certain to have close to the complete universe of events across
units and time and to be comparing “apples to apples.” We here separate out these activities from
other less visible activities and re-run the empirical analyses on these subsets of the data. We find
similar results.
The distribution of the component attacks that comprise our aggregated measure of guerrilla
activity is presented in Figure 2 to show how the 13,629 total actions break down by type. We use
these categories of actions to consider two different additional ways of constructing our key dependent
variable of guerrilla actions. The use of these two series aims to limit the types of actions included to
those where there should be reason to believe they are less susceptible to underreporting due to fear
or manipulation and private information. We construct both a “most visible” and a less restrictive
“likely visible” categorization of these more “publicly visible” events.
Figure 2: Armed Guerrilla Actions in Colombia by Event Category, 1988-2000
In comparison with the complete series of guerrilla actions, these new series omit more “private”
event categories, which could be subject to underreporting (Assault on private property, Robbery,
6
Incursions on population), event categories that are more ambiguous (Other terrorist attacks), and
events where there might be only (a few) military official witnesses and/or it might be embarrassing to
report (Raids and Ambushes). As a result, we construct the “most visible” categorization to include
only Explosive Attacks, Incendiary Attacks, Attacks Against Installations, and Illegal Roadblocks.
These are highly visible events, which comprise only about one-third (4,302) of the total actions
registered in the dataset for this time period. In comparison with the strict categorization, the
“likely visible” set includes several additional kinds of actions that could possibly be concealed or
misreported, but where there is reason to believe this is unlikely or rare given it would require the
collaboration of multiple troops involved in the incident. This likely visible set additionally includes
Political Attacks, Armed Contact with government troops, and Combat with government troops, for
a total of 9,770 actions. Since these events are larger and more public, the resulting variables may
also indicate the occurrence of more severe guerrilla activity in a given town. Even with the data
reconstructed in these different ways, the total events index and “most visible” and “likely visible”
indices are highly correlated (r=0.834). In additional robustness tests, displayed in Table 1, we find
that our results remain consistent and significant when both the most visible and likely visible counts
of guerrilla actions are used as the dependent variable in regressions. The coefficients of the land
reform variables are slightly higher when the “most visible” series is used.
In sum, an analysis of the more visible guerrilla actions that are less likely to suffer from misreporting biases yield results very similar to those in the manuscript.
7
Table 1: Land Reform and More Visible Guerrilla Actions in Colombia,
1988-2000
Most Visible
Likely Visible
Model 1
Model 2
Model 3
Model 4
Plots reformed
0.498*** 1.043*** 0.390***
0.867***
(0.154)
(0.179)
(0.126)
(0.150)
Prior plots
0.450*** 0.557*** 0.551***
0.645***
(0.123)
(0.110)
(0.102)
(0.090)
Plots*Prior plots
-0.350***
-0.315***
(0.053)
(0.039)
Paramilitary attacks
0.216
0.212
0.155
0.152
(0.146)
(0.144)
(0.102)
(0.103)
Government attacks 0.338*** 0.333*** 0.391***
0.387***
(0.057)
(0.058)
(0.052)
(0.053)
Poverty
-0.724*
-0.761*
0.475
0.463
(0.404)
(0.398)
(0.339)
(0.335)
Population density
0.423**
0.432**
0.371*
0.379*
(0.207)
(0.210)
(0.202)
(0.203)
Other tenancy
-2.225*** -2.152*** -1.530***
-1.498***
(0.525)
(0.510)
(0.453)
(0.449)
Coca
0.225
0.229
0.454***
0.464***
(0.155)
(0.153)
(0.120)
(0.118)
New colonized
0.556**
0.518**
0.550***
0.513***
(0.236)
(0.225)
(0.197)
(0.189)
Altitude
-0.467*** -0.457*** -0.195**
-0.184**
(0.099)
(0.097)
(0.088)
(0.087)
Percent minorities
0.016
0.059
0.219
0.264
(0.673)
(0.675)
(0.533)
(0.536)
Year Effects
YES
YES
YES
YES
Dept. Fixed Effects
YES
YES
YES
YES
Observations
10694
10694
10694
10694
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
Standard errors clustered by municipality in parentheses. Constants
and time dummies are not shown. All independent variables
lagged by one period. Estimations calculated using Stata 9.2.
2.3
Conformity Between Colombian Government Data and Human Rights Data
To examine the quality of our data for the dependent variable of guerrilla activities in greater detail,
we have also been able to obtain a good approximation of the true “reference population” of conflict
events in Colombia.
Information from the Colombian government’s National Census of Personerı́as conducted by
the National Prosecutors (Procuradurı́a) office is available for 1993, around the mid-point of our
data. Personerı́as are government employees that are present in every town. They are charged with
monitoring the conflict conditions in their respective towns and carrying out the legal processes of
recording and reporting cases of victimization. As independent human rights monitors, they have no
formal interaction with the armed forces. According to Camilo Echandı́a, one of Colombia’s foremost
experts on the armed conflict, the data from the Presidency and Vicepresidency (OVP) “tends to
8
coincide with the results of the National Census of Personerı́as from the last semester of 1993, where
half of Colombian municipios registered guerrilla presence” (Echandı́a 1999). Similar to Echandı́a’s
finding, our data conforms closely with the independently collected Personerı́as reports when an
indicator of guerrilla presence is constructed (proxying for guerrilla presence if at least one armed
action is registered). According to our calculations, 51% of all municipios in Colombia had at least
one recorded guerrilla action during the five year period from 1989-1993 leading up to the Personerı́a
census. If one examines the three year period from 1992 to 1994, 46% of towns had guerrilla presence
(about an 8% undercount). Both of these estimates of aggregate guerrilla presence come quite close
to the 50% mark estimated by the Personerı́a census.
In sum, the Personerı́a data suggest that the degree of bias in the reporting of guerrilla actions
is likely low. But because there could possibly still be biases (e.g. if the Personerı́a data should
suffer some misreporting or if the government data deviate from the Personerı́a data later in the
period), we now focus on the nature and degree of potential bias that would be problematic for the
manuscript’s findings.
2.4
Reciprocity between Government and Guerrilla/Paramilitary Actions
We conducted further tests to examine a potential endogenous relationship between guerrilla actions
and government actions given that guerrilla actions may lead to state repression and that state
repression may spur behavioral challenges such as guerrilla action. Government actions may also
have a reciprocal relationship with paramilitary actions. Most importantly for the results we present
in the paper, land reform (plots reformed) is correlated with paramilitary and government actions
at less than .02. If the coefficients on paramilitary or government actions are therefore somewhat
suppressed or increased after accounting for reciprocity with guerrilla activity, this will not affect
the estimated coefficient of land reform. Of course, not accounting for the reciprocal relationship
could still bias the estimated coefficients on Paramilitary and Government Attacks. In a basic OLS
model with controls, Guerrilla actions predict paramilitary actions (.025, p<.001) and government
actions (.123, p<.001). When instrumenting for guerrilla actions as in Table 5 of the paper using
La Violencia and a count of ANUC land invasions, the predicted coefficients for paramilitary actions
(.032, p=.006) and government actions (.180, p=.002) are larger. The one-way causal effect (i.e.
stripped of reciprocal causation) of guerrilla actions on government and paramilitary actions is greater
than the unadjusted effect. This is most likely because guerrilla actions result in a response by the
government and paramilitaries in the form of more actions, which may suppress guerrilla actions
over the longer term.
Accounting for the reciprocal relationship would somewhat depress the estimated coefficients
of paramilitary and government attacks, although they should retain the same sign. Since that is
effectively similar to an omitted variable bias of “reciprocal conflict” that has a negative sign, and
because paramilitary actions are positively correlated with plots reformed and government actions
negatively correlated with plots reformed, accounting for reciprocal conflict with these variables would
lead to different directions of bias on the plots reformed variable. However, since the correlation
between plots reformed and the paramilitary and government actions variables is so close to 0, the
net effect of including the reciprocal relationship between guerrillas and paramilitaries/government
forces for the plots reformed coefficient should be negligible. Furthermore, because paramilitary and
government actions are slightly positively correlated with prior plots, our estimated coefficient on
prior plots should be biased downward (against our hypothesis).
9
3
Addressing Reporting Biases
3.1
Overlap Analysis
In the absence of a full multiple-recapture analysis, we conduct a series of overlap analyses in an
effort to gain a better understanding of potential reporting biases.
3.1.1
Correlating Colombian Government (OVP) Data and CINEP Data
To compare the Government (OVP) and CINEP data series, we first conduct a bivariate regression
that includes department and year fixed effects to account for possible noise and differences in
reporting and coverage between these sources across time and space (and as included in our full
specifications of the effect of land reform). The model (Table 2) indicates a reasonably large and
statistically significant correlation between the two datasets. The CINEP data generally report fewer
guerrilla actions than OVP.
Table 2: Correlation Between OVP and CINEP Data
(DV: Number of Guerrilla Actions)
CINEP Guerrilla Actions
0.531***
(0.181)
Department Fixed Effects
YES
Year Effects
YES
Observations
12950
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
Models estimated by OLS. Standard errors clustered by
municipality in parentheses.
To help provide a more detailed sense of where the OVP and CINEP datasets converge and where
they differ we conducted several additional sets of disaggregated correlations. We first ran correlations
between the OVP and CINEP municipality-year guerrilla actions data by department (Colombia has
30 departments, excluding Bogotá and the San Andrés islands). The correlations across departments
range from -.02 to .81, and are smallest in small, peripheral, sparsely populated departments - areas
where CINEP has poor press coverage. The distant, minimally populated departments of Amazonas
and Vaupés have no coverage at all in CINEP’s dataset. By contrast, the correlations are higher
in moderate to highly conflictive areas, moderate sized departments, and where there is newspaper
coverage, such as in the departments of Norte de Santander, Santander, Huila, and Cauca. Yet even
in a few large, central departments, such as Antioquia (with over one-tenth of the municipios in
the country), correlations are found to be more attenuated, perhaps because they have too many
events and too much terrain for the press to provide as good coverage. Comparing correlations by
whether the departments have newspaper reporting is telling: in departments where CINEP has at
least one newspaper source, the average correlation between datasets (weighted by municipality-year
observations) is approximately twice as large as the average correlation in departments with no press
source. This speaks to the importance of including department (and in some cases municipality)
fixed effects in our regression models.
We also analyzed how the correlation between the two datasets changes over time. Figure 3 shows
the coefficients from a series of regressions using CINEP guerrilla actions to predict OVP actions
(adjusting for department-level differences using department fixed effects given the discussion above).
The coefficients are larger in the early and later periods of the sample, with lower correlations in the
mid-1990s. The number of guerrilla actions by year (lower line) is also displayed in the graph (1000s
10
of attacks), which highlights that the time-trend of the relationship between CINEP and OVP data
tracks closely with the change in guerrilla actions. This is likely because in more conflictive years such as the years in the early and late 1990s - the press does a better job of covering events. To deal
with these possible differences in reporting across time, all of our regression models in the manuscript
contain year fixed effects.
Figure 3: Coefficients of CINEP Guerrilla Actions Predicting OVP Guerrilla Actions
3.1.2
Factor Analysis: Are OVP and CINEP Data Capturing the Same Concept?
This section tests more formally whether the OVP and CINEP data are capturing the same underlying dimension of guerrilla attacks. To do so, we conduct a factor analysis to investigate whether
variations in these two variables reflect the variations in a single unobserved latent variable.
The results are found in Tables 3.1 and 3.2. The negative eigenvalue for the second factor in Table
3.1 indicates that the dimensionality of the factor space is only one. Only one factor is retained.
The factor loadings in Table 3.2 indicate how the OVP and CINEP guerrilla action variables are
weighted for the retained factor as well as the correlation between these variables and the factor. The
uniqueness column indicates the proportion of the common variance of each variable not associated
with the factor.
Tables 3.1 and 3.2 indicate that the OVP and CINEP data are indeed measuring the same
underlying concept of guerrilla actions. This provides more formal justification for using CINEP as
an alternative to OVP to measure guerrilla actions.
Table 3.1: Factor Analysis of OVP and CINEP Data
(Principle factors method)
Eigenvalue Difference Proportion
Factor 1
0.41140
0.62653
2.0961
Factor 2
-0.21513
-1.0961
Observations
12950
Retained Factors
1
11
Cumulative
2.0961
1.0000
Table 3.2: Factor Loadings and Unique Variances
(Principle factors method)
Factor 1 Uniqueness
OVP Guerrilla Actions
0.4535
0.7943
CINEP Guerrilla Actions
0.4535
0.7943
3.1.3
Histograms of Government and CINEP Data Overlap
Two histograms below provide a visualization of the distribution of overlap in reported guerrilla
actions by the government and CINEP. The degree of overlap is calculated at the municipal-year
level as the difference between the number of events in the government and CINEP datasets (positive
values indicate more recorded government actions than recorded CINEP actions). Given the absence
of case-level events data, perfect overlap as defined here indicates equal counts of guerrilla actions
in both datasets. The available data cannot adjudicate as to whether the reported cases themselves
are the same.
In terms of raw numbers, overlap is high: a total of 74% of municipality-years have perfect
overlap, where both sources report the same number of guerrilla actions. A full 88% of municipalityyears differ by at most one guerrilla action, 93% by at most two guerrilla actions, and 96% differ by
three or fewer actions.
The first histogram in Figure 4 shows this high degree of overlap for all municipality-year observations. The large number of observations with zero difference in guerrilla actions between these
datasets indicates that the majority of observations completely overlap. This is in part because
guerrilla actions are relatively rare and only occur in certain parts of the country, and therefore the
majority of observations in both datasets record zero guerrilla actions. The relatively greater number
of observations with positive differences in reporting between datasets shows that event reporting in
the government data is greater than the reporting by CINEP. This is likely because CINEP is based
on press coverage.
The second histogram in Figure 4 displays the overlap only for those municipality-years where
some non-zero number of events were reported in each dataset. Even restricting to this subset of
municipality-years, the majority of observations differ by one guerrilla action or less between datasets.
3.1.4
Cases of High and Low Overlap Between Government and CINEP Data
While the above sections indicate that there is significant overlap between the OVP and CINEP
data, they do not provide a sense of the municipality-years where these data sources diverge most.
This section takes a closer look at the departments and municipalities where overlap is particularly
high and low to shed greater light on differences between reported guerrilla action between the two
datasets.
Figure 5 displays the variation in overlap at a broader level by Colombia’s 32 departments. In
this table, perfect overlap corresponds to zero difference between government and CINEP reporting.
Positive values indicate relatively greater government data coverage. Mean annual overlap rates are
generally high at the department level, although there are three departments where it is somewhat
lower: Guaviare, Bogotá, and Arauca.
Figures 6 and 7 provide a more detailed picture of the overlap between the datasets by listing
the municipalities with the least amount of overlap as indicated by the mean difference between
Government and CINEP datasets from 1990-2000. Figure 6 shows the top 20 municipalities where
CINEP relatively over-reports events and Figure 7 shows the top 20 municipalities where Government
data relatively over-reports. These outlier cases are consistent with the municipality-year histogram
of overlap between datasets which indicates that most cases do not suffer from poor overlap.
12
Figure 4: Histograms of Government and CINEP Guerrilla Actions Data Overlap
(a) Full Municipality-Year Government and CINEP Data Overlap
(b) Municipality-Year Government and CINEP Data Overlap where
Guerrilla Actions Non-zero
Note: Positive values where Government guerrilla actions are greater than those in CINEP. Extreme values
are cut. Figure b restricts data to municipality-years where both Government and CINEP guerrilla actions
are both greater than zero.
13
Figure 5: Department Mean Annual Overlap 1988-2000
A cursory examination of Figure 7 suggests first that the government may relatively over-report
events in areas that have traditionally been guerrilla strongholds such as in the municipalities of
San Vicente de Chucurı́ (ELN) and San José del Guaviare (FARC) – places where press sources
may have more restricted access. In the regression tests, this discrepancy would be accounted for by
municipality fixed-effects, the historical indicator of guerrilla actions from 1985, and the indicator for
the La Violencia conflict of the 1950s. Second, the government may relatively over-report events in
large metropolitan areas such as Medellı́n, Bogotá, Bucaramanga, and Cali. The reporting in these
urban areas should not have great influence on our results since they are unlikely to be areas where
land reform is prevalent and interacting with processes of insurgency. They would also be accounted
for in the regression through municipality fixed-effects and variables for population and distance to
department capitals. The municipalities in Figure 6 of CINEP over-reporting do not immediately
indicate a clear pattern although some cases such as Turbo and Necoclı́ are known for having active
Diocese.
Even within this set of areas with poor overlap between datasets we are able to find some
indication of support for our hypotheses from a closer examination of the cases. An additional
and noteworthy pattern from Figures 5 and 7 is that the department of Arauca and several of its
municipalities exhibit extreme levels of government over-reporting of guerrilla actions. We briefly
study some of these municipalities from Arauca as qualitative cases at the end of the article with
14
Figure 6: Top 20 Municipalities with Greatest Over-reporting in CINEP Data Relative to Government
the expectation that they would experience differing effects of land reform since they are similar
neighboring cases and received similar levels of significant land reform in the period prior to 1988
but then diverged in their levels of reform in the current period from 1988-2000. The qualitative
analysis helps provide a more in-depth assessment of our hypotheses and provides confirmation that
even in these cases with lower overlap between data sources, the relative differences in land reform are
correlated with differences in insurgent activity. In the regression analysis, poor overlap is implicitly
accounted for both with department fixed effects as well as in the difference-in-difference models at
the municipal level.
3.1.5
Overlap Analysis: Where the Government and CINEP Data Diverge
This section performs a more formal overlap analysis, broadly examining in a regression framework
the municipality-years where these data sources diverge most, and what effects this might have on
the results. We use the land reform variables and covariates to predict differences in the reported
number of guerrilla actions in the OVP and CINEP data at the municipality-year. The models
focus on municipality-years in which one of these sources reported any non-zero amount of guerrilla
activity. A substantial number of municipalities (20%) never experienced guerrilla activity during
the period, and others only experienced activity for a small part of the period. Because most of
these areas were beyond the reach of the conflict, and are coded as such by OVP, CINEP, and the
National Census of Personerı́as, they are unlikely to be subject to reporting bias and are dropped
15
Figure 7: Top 20 Municipalities with Greatest Over-reporting in Government Data Relative to
CINEP
from the analysis. Including them would only strengthen the number of observations with overlap
in the overlap analysis.
Table 4 displays the results of the overlap analysis. Model 1 is specified similarly to the main
model in Column 3 of manuscript Table 2, but excludes paramilitary and government attacks. Model
2 includes these variables.
16
Table 4: Overlap Analysis of OVP and CINEP Data
(Dependent Variable: OVP-CINEP Guerrilla Actions)
Model 1
-0.213
(0.313)
0.346
(0.225)
Model 2
Plots reformed
-0.139
(0.334)
Prior plots
0.267
(0.220)
0.610
(0.467)
Government attacks
0.643***
(0.151)
Poverty
-3.220**
-3.482**
(1.375)
(1.454)
Population density
0.345
0.290
(0.426)
(0.410)
Other tenancy
-1.491
-1.408
(1.038)
(0.975)
Coca
0.205
0.109
(0.346)
(0.347)
New colonized
1.665*
1.628*
(0.864)
(0.873)
Altitude
-0.648*
-0.716*
(0.351)
(0.370)
Percent minorities
-0.218
-0.241
(0.578)
(0.569)
Year Effects
YES
YES
Dept. Fixed Effects
YES
YES
Observations
3123
3123
Municipalities
703
703
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
Standard errors clustered by municipality in parentheses.
Constants and time dummies are not shown.
Estimations calculated using Stata 9.2.
The first point to note is that the difference in the OVP and CINEP data are not statistically
significantly correlated with the land reform variables. This implies that any biases in the differences
between the sources would not affect the coefficients of the land reform variables or the inferences
that we draw. Poverty and government attacks, however, are significant predictors of the differences
between data sources at the .05 level. Differences in reporting are greater in areas where poverty is
lower and government attacks are higher. The result on poverty may be related to CINEP’s more
complete network of local church officials in rural areas where poverty is higher. More government
attacks predicts greater differences in reporting. This may be because the government has better
access to information in these conflict zones than outside sources, or because they provide a relatively
more complete count of guerrilla activities in these areas.
Table 5 examines how land reform affects guerrilla actions in areas where OVP and CINEP data
are more likely and less likely to overlap. This table supplements the robustness tests conducted
for the first revision that split the data based on ruralness, state presence, and the distance of a
17
municipality from its department capital. The Table 5 models are specified similarly to Model 3 of
Table 2, and the analysis is split by level of poverty and guerrilla attacks. Models 1 and 2 indicate
that both current Plots reformed and Prior plots are positively and statistically significantly linked
to guerrilla actions in both high poverty and low poverty regions when split by the mean level of
poverty. Models 3-6 split the data using two different thresholds for government attacks. Models
3-4 split the data by whether a given municipality-year experienced any government attacks or not.
Models 5-6 split the data by whether a given municipality experienced any government attacks during
the entire period 1988-2000 or not. Prior plots is positive and statistically significant across Models
3-6. Plots reformed is positive and significant in Model 4 among municipality-years that experienced
no government attacks. It is insignificant in Model 3, although the size of the sample is substantially
reduced. Plots reformed regains its positive sign and significance in Model 5 among municipalities
that experienced some goverment attacks in the period. It is positive but insignificant in Model 6.
Table 5: Land Reform and Guerrilla Actions in Colombia, Split by Factors that Predict Differences
in OVP and CINEP
High
Low
Any Govt No Govt
Any Govt
No Govt
Poverty
Poverty
Attacks
Attacks Attacks (Per) Attacks (Per)
Model 1
Model 2
Model 3
Model 4
Model 5
Model 6
Plots reformed
0.302**
0.756***
-0.155
0.440***
0.215*
0.242
(0.134)
(0.213)
(0.099)
(0.142)
(0.121)
(0.290)
Prior plots
0.278*** 0.811***
0.122**
0.512***
0.362***
0.147*
(0.099)
(0.124)
(0.058)
(0.103)
(0.076)
(0.082)
0.204
0.146*
0.181***
0.414**
0.221***
0.714
(0.125)
(0.089)
(0.060)
(0.171)
(0.085)
(0.673)
Government attacks 0.215*** 0.276***
(0.037)
(0.034)
Poverty
-1.235***
0.113
-0.417
0.956*
(0.422)
(0.326)
(0.358)
(0.516)
Population density
-3.048*
0.418**
0.316***
0.365*
0.394***
0.313
(1.629)
(0.196)
(0.102)
(0.206)
(0.114)
(0.322)
Other tenancy
-1.850*** -2.262*** -1.130** -1.821***
-1.806***
-1.291**
(0.549)
(0.592)
(0.559)
(0.430)
(0.470)
(0.609)
Coca
0.355***
0.247
-0.034
0.441***
0.194*
0.479
(0.135)
(0.184)
(0.122)
(0.127)
(0.113)
(0.295)
New colonized
0.540***
0.670
0.883***
0.585***
0.584***
0.376
(0.150)
(0.454)
(0.191)
(0.196)
(0.177)
(0.240)
Altitude
-0.091
-0.280*** -0.328*** -0.274***
-0.378***
-0.100
(0.091)
(0.097)
(0.095)
(0.084)
(0.085)
(0.117)
Percent minorities
-0.130
-0.542
0.054
0.078
0.153
-0.507
(0.488)
(0.602)
(0.303)
(0.632)
(0.390)
(0.724)
Year Effects
YES
YES
YES
YES
YES
YES
Dept. Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
6643
5328
948
10692
5472
6168
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
Standard errors clustered by municipality in parentheses. Constants and time dummies not shown.
Models 3 and 4 split sample by whether a municipality-year had any government attacks.
Models 5 and 6 split sample by whether a municipality had any government attacks in the period.
All independent variables lagged by one period. Estimations calculated using Stata 9.2.
18
While these analyses highlight some similarities across the two guerrilla action datasets and the
Personerı́a data from Section 2.3 provides some confirmation that the OVP data may well approximate the “true” number of attacks, there could still be a threat to inference if both the OVP and
CINEP data are biased in the same direction. As a result, we conduct a series of analyses that
address the nature and degree of bias that would be problematic for the inferences we draw in the
paper.
3.2
Nature of Bias: At What Level Must Bias Exist to Change the Results?
What is the nature of bias that must exist in the data on guerrilla attacks for it to be problematic
for our results? The main models in the paper include department fixed effects and year dummies,
and taking into account similar results using CINEP data, any problematic remaining biases in the
data must be non-uniform municipality-specific biases that are shared by both government and nongovernment data and that are in some way correlated with land reform and guerrilla activity. That
the biases must be correlated with land reform is important given that “classical” measurement error
(i.e. random error that is uncorrelated with the error term) leads to attenuation bias (see e.g. Imai
and Yamamoto 2010), leading to an underestimate of the true effect of land reform on guerrilla
activity.
Here we subject our results to an even stricter specification than using department and year fixed
effects to see if our results remain and to gain a more precise sense of the kind of remaining bias
that would have to exist to destabilize our results. More specifically, we re-estimate the main results
from Model 3 of manuscript Table 2 and Model 4 of Table 5 using department*year fixed effects,
and using municipality-year difference-in-difference models using department*year fixed effects. The
results are displayed in Tables 6 and 7. They are substantively and statistically similar to the results
in the manuscript.
3.2.1
Models Using Department*Year Fixed Effects
As a result of including an interaction between geographic- and time-specific fixed effects, Table
6 indicates that any bias in the data on guerrilla activities must be municipal-level time-varying
biases that deviate from department-level yearly trends and are also correlated with municipal-level
land reform for it to be problematic for the results we draw in our empirical analyses. Furthermore,
similar empirical results using municipal fixed effects models (as noted in the manuscript) mitigate
concerns of bias across municipalities. The type of bias that could be problematic for the finding
of a positive effect of current and prior land titling on guerrilla actions is therefore quite restrictive:
municipality-year biases across both government and CINEP sources that result in consistent overreporting where land reform is high (or under-reporting where land reform is low) on a yearly-varying
basis within a given municipality. The only way to truly address this possibility would be to run a
model with municipality*year fixed effects, which would themselves perfectly predict the data and
make it impossible to measure the effect of any covariates, including land reform, on guerrilla activity.
19
Table 6: Land Reform and Guerrilla Actions in Colombia,
1988-2000, Using Dept.*Year Fixed Effects
Plots reformed
Prior plots
Model 1
0.449***
(0.130)
0.443***
(0.097)
Plots*Prior plots
Model 2
0.926***
(0.159)
0.523***
(0.084)
-0.296***
(0.038)
0.226***
(0.082)
0.299***
(0.025)
-0.065
(0.317)
0.383**
(0.194)
-1.835***
(0.415)
0.359***
(0.111)
0.620***
(0.180)
-0.268***
(0.078)
0.052
(0.508)
YES
11640
0.229***
(0.084)
Government attacks
0.305***
(0.025)
Poverty
-0.047
(0.323)
Population density
0.374*
(0.192)
Other tenancy
-1.882***
(0.425)
Coca
0.358***
(0.113)
New colonized
0.653***
(0.188)
Altitude
-0.274***
(0.079)
Percent minorities
0.022
(0.508)
Dept.*Year Fixed Effects
YES
Observations
11640
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
3.2.2
Municipal Difference-in-Difference Models Using Department*Year Fixed Effects
Estimating difference-in-difference models at the municipal level provides an even stricter bound
on potentially problematic bias than that discussed above in Section 3.2.1. Municipal differencein-difference models address the possibility of unobservable heterogeneity in reporting patterns (or
other unobservable factors that may introduce bias) at the municipal level. We therefore conducted
a series of these models on municipality-years, including department-year fixed effects as in Table
6. These are quite stringent empirical tests given that the length of the time-series (13 years) is
relatively short and much smaller than the cross-section. This approach drops information contained
in cross-municipality level differences, and makes it more difficult to estimate the effects of slowly
changing variables such as poverty, population density, tenancy, and to some extent the prior plots
variable. Because the variables for coca, colonization zones, altitude, and minority presence are not
20
time-varying, they are dropped from the estimation.
Table 7 displays the results of these models, which capture how yearly changes in municipallevel land titling (and other independent variables) affect yearly changes in municipal-level guerrilla
actions that deviate from department-year specific trends. The main results remain robust to this
strict specification. Current and prior land titling has a positive effect on guerrilla actions, although
the latter loses significance (p<0.30) in Model 2, which includes the land reform interaction terms.
Model 2 indicates that there is a negative marginal effect of current reform when a certain threshold
in prior reform has been reached, consistent with the main findings presented in the paper.
Table 7 provides an even more restrictive condition than Table 6 on the type of bias that would
be problematic for the finding of a positive effect of current and prior land titling on guerrilla action:
municipal-level yearly-varying biases that deviate from department-level yearly trends and are also
correlated with yearly changes in municipal-level land reform. Furthermore, we know the direction of
bias that would have to exist to be problematic: consistent increases in over-reporting on a yearlyvarying basis within a given municipality where yearly changes in land reform are positive and large
(or more under-reporting where yearly changes in land reform are negative).
Table 7: Land Reform and Guerrilla Actions in Colombia, 1988-2000
Municipal Difference-in-Difference Models Using Dept.*Year Fixed Effects
Plots reformed
Prior plots
Model 1
.394*
(0.236)
1.563**
(0.740)
Plots*Prior plots
Model 2
0.531**
(0.159)
0.970
(0.923)
-0.100*
(0.054)
0.486
(0.596)
0.161***
(0.059)
-0.203
(0.670)
-0.630
(0.868)
-1.013
(0.764)
YES
8999
0.486
(0.594)
Government attacks
0.162***
(0.059)
Poverty
0.076
(0.556)
Population density
0.431
(0.844)
Other tenancy
-0.767
(0.702)
Dept.*Year Fixed Effects
YES
Observations
9835
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
in first differences. Coefficients on land reform variables calculated
using the first difference and lagged first difference.
Estimations calculated using Stata 9.2.
21
3.3
3.3.1
Sensitivity Analyses: How Much Bias is Necessary to Undermine the Results?
Sensitivity Analyses Based on Overlap Between OVP and CINEP Data
The first sensitivity analysis we conduct is based on an overlap analysis of the OVP and CINEP
datasets. Rather than predict the determinants of the differences between these data as in Table 4
above, we examine how sensitive the results are to removing observations where these data sources
report the most divergence. We estimate a series of regression models where the samples of each
successive model are increasingly restricted to only those observations with high levels of overlap
between the two datasets. We compare the main results from Model 3 of manuscript Table 2 with
successive models that only include observations where the difference in reported guerrilla actions
between OVP and CINEP is less than 10 (including a number of observations with low overlap),
less than 9, etc., down to only less than 2 (including only observations with high overlap) in a
given municipality-year.∗ The results of these tests are summarized in Figure 8. This figure shows
the estimated coefficients and 95% confidence intervals for the effects of our land reform variables
(Current Plots and Prior Plots) in the successive regression models.
As the observations are increasingly resticted to those with greatest overlap, the estimated coefficients of the land reform variables shown in Figure 8 decrease. Only the lower 95% confidence
interval for Current Plots crosses zero when using the most restrictive overlap criteria, but even then
the coefficient remains statistically significant at the 90% level. In none of the remaining models–not
even in the tests using the samples with the most restrictive overlap criteria–does the lower confidence
interval cross zero.
In sum, our principal results are robust to excluding the observations that could most be subject
to misreporting based on poor overlap between the OVP and CINEP datasets.
3.3.2
Regression Calibration Estimating Reporting Bias
The second set of sensitivity analyses more directly addresses the possibility of an omitted variable
(“Reporting Bias”) that predicts bias in the reporting of guerrilla activities. This variable could
capture incentives to misreport or sources of incompleteness that may result in biases in the dependent
variable of guerrilla attacks. We introduce proxies for these incentives or tendencies to misreport
that reasonably suffer from measurement error and then use regression calibration (see e.g. Carroll
et al. 1995, Hardin et al. 2003) and simulation extrapolation (see e.g. Cook and Stefanski 1990)
to estimate the resulting measurement error models. The first stage of these methods results in a
calibration function for estimating the unobserved covariate of “Reporting Bias.” This unobserved
covariate is then replaced by its predicted values from the calibration model. Finally, the standard
errors are adjusted to account for the estimation of the unobserved covariate.
We use two different approaches to estimate the unobserved covariate of Reporting Bias. First,
we use a variable called Delinquency from the OVP dataset, which is a count of actions committed
by common criminals and self-defense groups. These are unaffiliated groups and gangs, some of
which later formed paramilitary groups. To the extent that local military officers had incentives
to underreport guerrilla actions, it is conceivable that they could have reported guerrilla actions
or actions that were ambiguously attributable to delinquent group actions. Conversely, they could
have reported delinquent group actions as guerrilla actions in places where there were incentives or
tendencies to overreport. As a result, Delinquency should lead to lower predicted guerrilla actions.
Second, we use both Delinquency and Homicides to predict reporting bias. Homicide data are taken
∗
The models did not reach convergence when restricting to perfectly matching observations; however, the majority of these were the municipality-years that registered no guerrilla activity.
22
Figure 8: Sensitivity of Main Results to Overlap in OVP and CINEP Data
(a) Coefficient on Current Plots as Data are Restricted to
Municipality-Years of Greatest Overlap
(b) Coefficient on Prior Plots as Data are Restricted to MunicipalityYears of Greatest Overlap
Note: The figure plots the coefficients and 95% confidence intervals for Current Plots and Prior Plots from a
series of regression models specified as in Model 3 of Table 2 but with varying data samples. The data sample
for each model is restricted to municipality-years where the difference in OVP and CINEP reported guerrilla
activities is less than the values listed on the x-axis.
23
from the National Police and include all homicides - by criminals, domestic killings, and armed
groups.
The results of a series of regression calibration models are found in Table 8. The models are
specified as in Models 2 and 3 of Table 2, and include regional fixed effects and year dummies.†
Table 8: Land Reform and Guerrilla Actions, Using Regression
Calibration to Estimate Reporting Bias
Model 1
0.439***
(0.088)
0.647***
(0.040)
-0.005**
(0.002)
0.501***
(0.159)
0.339***
(0.028)
0.854***
(0.133)
0.326***
(0.066)
-2.366***
(0.216)
Model 2
Model 3
Model 4
Plots reformed
0.394*** 0.383*** 0.334***
(0.093)
(0.117)
(0.095)
Prior Plots
0.441*** 0.691*** 0.473***
(0.042)
(0.049)
(0.040)
Reporting Bias
-0.004**
-0.004*
-0.005***
(0.002)
(0.002)
(0.002)
0.378*** 0.981*** 0.974***
(0.122)
(0.254)
(0.261)
Government attacks
0.318*** 0.414*** 0.395***
(0.032)
(0.042)
(0.046)
Poverty
0.299**
0.634***
0.081
(0.130)
(0.152)
(0.156)
Population density
0.285*** 0.448*** 0.425***
(0.067)
(0.134)
(0.114)
Other tenancy
-1.837*** -2.435*** -1.847***
(0.210)
(0.249)
(0.191)
Coca
0.620***
0.616***
(0.058)
(0.067)
New colonized
0.616***
0.618***
(0.075)
(0.084)
Altitude
-0.309***
-0.323***
(0.034)
(0.041)
Percent minorities
0.053
0.114
(0.203)
(0.182)
Assumed M.E. Variance
2
2
Year Effects
YES
YES
YES
YES
Regional Fixed Effects
YES
YES
YES
YES
Observations
11709
11640
11709
11640
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
Bootstrapped standard errors in parentheses. Constants and time
dummies are not shown. All independent variables lagged by one
period. Estimations calculated using Stata 9.2.
Because Models 1 and 2 only use one replicate measurement (Delinquency) to measure Reporting
Bias, we must assume a measurement error variance. We choose a measurement error variance of 2,
which is sizeable relative to the variance of 7.5 in guerrilla actions. Given that guerrilla actions spans
a large range of activity with a sizeable number of municipality-years experiencing no activity and
†
Results were similar using department fixed effects, but were much more computationally intensive.
24
a significant portion with high guerrilla activity, and given similar results using the most “visible”
events (see Section 2.2) and CINEP data, we should expect the measurement error variance to be
considerably less than the variance in guerrilla activity.‡ When the estimated “Reporting Bias”
variable is included in the model, it is negative and statistically significant, but does not strongly
impact the estimated effect of the other covariates. The land reform variables Plots reformed and
Prior plots remain positive and statistically significantly linked to greater guerrilla actions after
introducing Reporting Bias, and their substantive effect is similar to that in Table 2 of the manuscript.
The coefficients of the other covariates are generally similar in size and statistical significance to those
reported in manuscript Table 2. Models 3 and 4 introduce two replicate measurements for Reporting
Bias in the form of Delinquency and Homicides. As a result, the assumption of known measurement
error variance on Reporting Bias can be relaxed and is now estimated from the replicates. The
results are similar to those in Models 1 and 2.
We next estimate Model 1 of Table 8 using simulation extrapolation (SIMEX) to obtain a graph
that shows how the amount of measurement error affects the estimated coefficients. Figure 9 shows
the estimated coefficients for the Plots Reformed and Prior Plots variables based on the Model 1
specification in Table 8. As detailed above, Model 1 includes a measure of reporting bias that is
measured with error. The graph illustrates the extrapolated point estimates for the land reform
variables in the fitted model, as well as how these estimates vary according to the variance in the
measurement error of Reporting Bias. A Lamba of -1 corresponds to the case of no measurement
error. The key point is that land reform variables remain positive and significant across the range
of measurement error in Reporting Bias. Furthermore, the sizes of these coefficients are relatively
stable. The estimated coefficients of the land reform variables from the simulation extrapolation are
only slightly lower than the naive estimates that do not account for potential reporting bias.
‡
Results were nonetheless similar assuming a larger (e.g. 3) or smaller (e.g. 1) measurement error
variance as well.
25
Figure 9: Sensitivity of Land Reform Variables to Measurement Error in Estimated Reporting Bias
Note: The figure plots the extrapolated point estimates for the Plots Reformed and Prior Plots variables from
Model 1 of Table 8, as well as how these estimates vary according to the variance in the measurement error of
Reporting Bias. A quadratic extrapolant function is used. Measurement error in Reporting Bias is increasing
in Lamda.
26
4
Supplementary Robustness Tests
This section provides a discussion and presents the results of a series of additional robustness tests
relevant to the manuscript, most of which are cited but not reported in the manuscript due to space
considerations. A brief description of each robustness table is given here along with the findings.
Table 9: This table uses municipal-level data on homicide rates from the National Police for the
period 1990-2000 to empirically test whether the effect of land reform was different for civilian
violence than for guerrilla actions. Importantly, we find that the estimated effect of land reform on
homicide rates differs from its effect on guerrilla activity, suggesting that the indicator for guerrilla
activity used in the manuscript is not simply a proxy for violence more generally.
Table 9: Land Reform and Civilian Homicide Rates
(Dependent Variable: Homicide rate per 100,000 residents)
Plots reformed
Prior Plots
Model 1
-7.170
(4.514)
1.776
(2.774)
Model 2
-9.526**
(4.620)
-0.397
(2.943)
Plots*Prior plots
12.076*
(6.505)
10.633***
(3.304)
-16.988
(13.522)
12.050***
(3.275)
26.421
(22.278)
Model 3
-11.306**
(5.071)
1.136
(2.964)
1.811
(2.132)
12.171*
(6.518)
10.703***
(3.299)
-16.522
(13.504)
12.020***
(3.272)
25.899
(22.272)
Model 4
-13.993***
(5.235)
-1.124
(3.120)
1.949
(2.149)
12.507**
(6.279)
10.534***
(3.176)
-26.537*
(14.494)
11.148***
(3.143)
43.757**
(22.055)
20.032***
(6.999)
-1.739
(7.361)
-8.891***
(2.953)
-12.017
(12.716)
YES
10818
12.420**
(6.271)
Government attacks
10.464***
(3.180)
Poverty
-26.959*
(14.522)
Population density
11.185***
(3.145)
Other tenancy
44.316**
(22.058)
Coca
20.026***
(6.997)
New colonized
-2.041
(7.376)
Altitude
-8.885***
(2.953)
Percent minorities
-11.622
(12.710)
Year Effects
YES
YES
YES
Observations
10884
10818
10884
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
All models estimated by OLS with department fixed effects.
27
Table 10: Given potential concerns of misreporting bias or manipulation of armed group action
data by the Colombian military, we use municipal-level data on guerrilla activities, paramilitary
activities, and government actions that was gathered independently by the Jesuit think tank Centro
de Investigación y Educación Popular (CINEP). This organization codes their data in the same way
as the government, making their data comparable with the government data. Furthermore, given
that CINEP has different incentives from the government, and an extensive network of local sources
even in rural areas, biases in the military and CINEP data sources are likely to differ. Substituting
the CINEP data for the government data in the Table 10 models yields similar results to those found
in the manuscript using the Colombian military’s data.
Table 10: Land Reform and Guerrilla Actions Using CINEP Data
(Dependent Variable: Number of guerrilla attacks, CINEP)
Model 1
0.231**
(0.117)
0.334***
(0.085)
Model 2
0.180+
(0.115)
0.271***
(0.089)
Model 4
Plots reformed
0.547***
(0.162)
Prior plots
0.328***
(0.090)
Plots*Prior plots
-0.130***
(0.043)
0.233**
0.164
0.165
(0.102)
(0.110)
(0.109)
Government attacks 0.303*** 0.290***
0.284***
(0.033)
(0.035)
(0.035)
Poverty
1.117***
0.814**
0.785**
(0.310)
(0.350)
(0.346)
Population density
0.267**
0.261**
0.264**
(0.130)
(0.128)
(0.129)
Other tenancy
-1.408*** -1.340**
-1.312**
(0.515)
(0.528)
(0.526)
Coca
0.519***
0.523***
(0.126)
(0.125)
New colonized
0.314*
0.274
(0.174)
(0.172)
Altitude
-0.042
-0.037
(0.097)
(0.097)
Percent minorities
-0.024
0.020
(0.416)
(0.410)
Year Effects
YES
YES
YES
YES
Observations
10884
10818
10884
10818
+ p < 0.15; * p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
All models use negative binomial estimator with department fixed effects.
28
Model 3
0.633***
(0.170)
0.390***
(0.085)
-0.144***
(0.043)
0.232**
(0.101)
0.296***
(0.033)
1.068***
(0.308)
0.269**
(0.130)
-1.379***
(0.512)
Table 11: This table presents a series of random effects negative binomial and tobit models to estimate the effect of land reform on guerrilla actions. The results are similar to the negative binomial
and tobit models with department fixed effects that are presented in the manuscript.
Table 12: This table presents a series of negative binomial and tobit models specified similarly to
Table 2 in the manuscript, but with an alternative definition of offensive government actions. We
limit government actions to those that are plausibly more offensive in nature (capture of armaments,
rescue or freeing of kidnap victims, anti-narcotic operations, and raids). The results are similar
to the negative binomial and tobit models with department fixed effects that are presented in the
manuscript.
Table 13: This table examines whether minority presence in an area influences how land titling
impacts guerrilla activity. The first two models indicate that the coefficients on the land reform
variables are robust to dropping those municipalities that have 50% or more minorities. Models 3
and 4 include an interaction between land reform and ethnic minority variables. This interaction is
not significant and the results for the effects of land reform on insurgent activity are stable. This
suggests that the effects of land reform on insurgency (according to our data which consists of titles to
individuals and not collective territories such as resguardos) are consistent regardless of variations in
levels of demand for reform that depend on the composition of the population in certain municipios.
In other words, they are not conditional on the size of the ethnic minority population in a municipio.
However, in some tobit models the interaction term coefficient is significant and negative (and the
main results on the effects of land reform hold). This would be expected according to our theory of
the importance of the ratio of beneficiaries to non-beneficiaries since municipios with greater minority populations would exhibit relatively less demand for traditional forms of individual land titling.
Given a constant level of supply of land reform across these areas and non-minority areas, the overall
number of (non-minority) titling beneficiaries should be proportionally higher and should thus more
likely demonstrate a negative effect of tilting on insurgency in these areas. This result would be
consistent with the threshold effect we find in our main set of interaction tests between the current
plots reformed and prior plots reformed variables.
Table 14: This table estimates a series of Heckman two-stage selection models in order to first
predict land reform and then its impact on guerrilla actions. The first stage model includes variables for the land area of a municipality as well as the percentage of overused land according to the
Colombian geography agency (IGAC) as instruments to predict selection into land reform treatment.
These regressors are excluded in the second stage outcome equations. Model 1 shows the determinants of receiving any land reform at all, whereas Models 2-4 show a series of second-stage models
that predict the number of guerrilla actions. The results are similar to those in the manuscript.
Table 15: This table includes two additional measures of state presence in the basic negative
binomial models presented in the manuscript. The first is a measure of bureaucratic strength in
the form of government officials per capita. The second is a measure of police stations (in the town
center, measured in 1995). Whereas the measure of government officials is statistically insignificant,
police stations are positive and statistically significant, likely because police stations are either placed
in areas that have a greater propensity for guerrilla activity, or because these stations become targets
for guerrillas. The results for the land reform variables are similar to those in the manuscript.
29
30
0.460***
(0.076)
0.054
(0.041)
-0.002
(0.012)
0.753***
(0.197)
0.157**
(0.074)
-1.930***
(0.317)
Model 2
0.094*
(0.050)
0.432***
(0.061)
0.041
(0.042)
-0.004
(0.012)
0.735***
(0.195)
0.170**
(0.074)
-1.861***
(0.312)
Model 3
0.057
(0.050)
0.281***
(0.064)
0.030
(0.041)
-0.009
(0.011)
0.287
(0.211)
0.204**
(0.083)
-1.307***
(0.319)
0.616***
(0.119)
0.407***
(0.130)
-0.162***
(0.054)
-0.241
(0.246)
Model 4
0.067
(0.050)
0.370***
(0.066)
0.080*
(0.041)
0.000
(0.011)
0.313
(0.205)
0.119*
(0.071)
-1.199***
(0.311)
0.655***
(0.116)
0.398***
(0.127)
-0.228***
(0.052)
-0.071
(0.239)
0.889***
(0.082)
YES
11640
955
0.498***
(0.028)
0.110**
(0.054)
0.061***
(0.015)
0.578***
(0.103)
0.138***
(0.040)
-1.049***
(0.175)
Model 5
Model 6
0.135***
(0.052)
0.423***
(0.024)
0.102*
(0.054)
0.060***
(0.015)
0.540***
(0.102)
0.141***
(0.040)
-1.022***
(0.173)
Model 7
0.116**
(0.052)
0.308***
(0.026)
0.093*
(0.054)
0.058***
(0.016)
0.224**
(0.111)
0.142***
(0.034)
-0.688***
(0.183)
0.427***
(0.064)
0.439***
(0.072)
-0.099***
(0.025)
-0.015
(0.134)
Tobit
Model 8
0.130**
(0.053)
0.341***
(0.032)
0.135**
(0.054)
0.057***
(0.016)
0.324***
(0.112)
0.094**
(0.040)
-0.818***
(0.181)
0.466***
(0.064)
0.397***
(0.073)
-0.145***
(0.025)
0.132
(0.132)
0.833***
(0.059)
YES
11640
955
Year dummies
YES
YES
YES
YES
YES
YES
Observations
11709
11709
11640
11709
11709
11640
Municipalities
962
962
955
962
962
955
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
Tobit models use log attacks. Standard errors in parentheses. Constants and time dummies are not shown.
All independent variables lagged one period. Estimations calculated using Stata 9.2.
Neighbor violence
Percent minorities
Altitude
New colonized
Coca
Other tenancy
Population density
Poverty
Government attacks
Prior plots
Plots reformed
Model 1
Negative Binomial
Table 11: Land Reform and Guerrilla Actions in Colombia 1988-2000, Random Effects Models
31
0.553***
(0.113)
0.148
(0.151)
0.173***
(0.029)
1.134***
(0.263)
-0.053
(0.076)
-1.879***
(0.452)
Model 2
0.317***
(0.120)
0.479***
(0.102)
0.126
(0.158)
0.165***
(0.029)
1.025***
(0.265)
-0.044
(0.076)
-1.811***
(0.447)
Model 3
0.298**
(0.118)
0.326***
(0.103)
0.045
(0.154)
0.152***
(0.026)
0.513*
(0.295)
-0.041
(0.073)
-1.452***
(0.426)
0.280**
(0.124)
0.574***
(0.167)
-0.208***
(0.078)
0.177
(0.590)
Model 4
0.328***
(0.115)
0.329***
(0.099)
0.084
(0.161)
0.152***
(0.026)
0.654**
(0.297)
-0.045
(0.077)
-1.460***
(0.428)
0.323***
(0.124)
0.572***
(0.167)
-0.241***
(0.078)
0.199
(0.545)
0.930***
(0.151)
YES
11640
955
0.348***
(0.026)
-0.009
(0.071)
0.105***
(0.005)
0.880***
(0.105)
-0.029
(0.033)
-1.220***
(0.173)
Model 5
Model 6
0.186***
(0.064)
0.287***
(0.024)
-0.012
(0.071)
0.103***
(0.005)
0.818***
(0.105)
-0.022
(0.033)
-1.167***
(0.172)
Model 7
0.167***
(0.064)
0.203***
(0.025)
-0.042
(0.071)
0.100***
(0.005)
0.475***
(0.111)
-0.024
(0.032)
-1.001***
(0.174)
0.304***
(0.053)
0.473***
(0.060)
-0.065**
(0.027)
0.001
(0.152)
Tobit
Year dummies
YES
YES
YES
YES
YES
YES
Observations
11709
11709
11640
11709
11709
11640
Municipalities
962
962
955
962
962
955
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
All models include department fixed effects. Tobit models use log attacks. Standard errors in parentheses.
Constants and time dummies are not shown. All independent variables lagged one period.
Neighbor Violence
Percent minorities
Altitude
New colonized
Coca
Other tenancy
Population density
Poverty
Government Attacks
Paramilitary Attacks
Prior Plots
Plots reformed
Model 1
Negative Binomial
Table 12: Land Reform and Guerrilla Actions, Alternative Measure of Offensive Government Actions
Model 8
0.193***
(0.063)
0.216***
(0.025)
-0.002
(0.070)
0.100***
(0.005)
0.600***
(0.111)
-0.036
(0.032)
-1.030***
(0.173)
0.330***
(0.053)
0.453***
(0.060)
-0.087***
(0.027)
0.032
(0.151)
0.740***
(0.073)
YES
11640
955
Table 13: Land Reform and Guerrilla Actions in Minority Areas
(Dependent Variable: Number of guerrilla attacks)
Plots reformed
Prior Plots
Dropping Munis with More
Than 50% Minorities
Model 1
Model 2
0.351***
0.374***
(0.125)
(0.122)
0.474***
0.481***
(0.101)
(0.100)
Plots Reformed*Minority
Paramilitary Attacks
Government Attacks
Poverty
Population density
Other tenancy
Coca
New colonized
Altitude
0.234**
(0.095)
0.271***
(0.030)
0.061
(0.356)
0.397**
(0.195)
-1.800***
(0.441)
0.344***
(0.123)
0.641***
(0.206)
-0.265***
(0.086)
0.248***
(0.089)
0.264***
(0.028)
0.191
(0.362)
0.399**
(0.198)
-1.824***
(0.443)
0.383***
(0.125)
0.631***
(0.209)
-0.297***
(0.087)
0.803***
(0.146)
Neighbor Violence
Pc. minorities
Interacting Land Reform
with Minority Presence
Model 3
Model 4
0.412***
0.437***
(0.130)
(0.126)
0.455***
0.464***
(0.099)
(0.098)
-0.398
-0.426
(0.582)
(0.556)
0.230**
0.246***
(0.094)
(0.087)
0.274***
0.267***
(0.030)
(0.028)
0.064
0.194
(0.343)
(0.347)
0.392**
0.394**
(0.194)
(0.197)
-1.934***
-1.951***
(0.434)
(0.437)
0.356***
0.397***
(0.123)
(0.125)
0.628***
0.625***
(0.200)
(0.203)
-0.272***
-0.304***
(0.087)
(0.087)
0.879***
(0.151)
0.077
0.088
(0.582)
(0.542)
YES
YES
11640
11640
Year Effects
YES
YES
Observations
11295
11295
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
32
Table 14: Heckman Models for Selection into Land Reform
(Stage 1 DV: Land reform treatment)
(Stage 2 DV: Number of guerrilla attacks)
Selection Equation
Model 1
Plots reformed
Prior plots
Plots*Prior plots
Outcome Equations
Model 2 Model 3
Model 4
0.439*** 0.449*** 0.939***
(0.123)
(0.127)
(0.155)
0.512*** 0.460*** 0.560***
(0.107)
(0.105)
(0.093)
-0.326***
(0.042)
0.229
0.228*
0.230*
(0.143)
(0.122)
(0.121)
0.300*** 0.319*** 0.317***
(0.049)
(0.048)
(0.049)
0.239
0.006
-0.004
(0.340)
(0.349)
(0.343)
0.517***
0.435**
0.434**
(0.198)
(0.200)
(0.202)
-1.141** -1.573*** -1.622***
(0.522)
(0.464)
(0.458)
0.284**
0.306**
(0.144)
(0.140)
0.494**
0.497**
(0.211)
(0.202)
-0.236*** -0.240***
(0.091)
(0.091)
0.344
0.292
(0.626)
(0.626)
0.037
(0.066)
Government attacks
0.016
(0.039)
Poverty
-0.024
(0.186)
Population density
-0.126**
(0.050)
Other tenancy
-1.192***
(0.270)
Coca
0.310***
(0.086)
New colonized
0.582***
(0.097)
Altitude
-0.144***
(0.050)
Percent minorities
-1.216***
(0.308)
Area
0.090***
(0.024)
Percent land overuse
0.035
(0.114)
Year Effects
YES
YES
YES
YES
Observations
10658
10658
10658
10658
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
Model 1 uses a probit estimator with department fixed effects.
Models 2-4 use negative binomial estimators with department fixed effects,
and include the inverse Mills’ ratio from Stage 1 as a regressor.
33
Table 15: Land Reform and Guerrilla Actions, Alternative
Measures of State Presence
(Dependent Variable: Number of guerrilla attacks)
Plots reformed
Prior plots
Model 1
0.369***
(0.119)
0.646***
(0.113)
Model 2
0.341***
(0.120)
0.436***
(0.097)
Plots*Prior plots
0.232
(0.192)
0.261***
(0.024)
0.626**
(0.309)
-0.201
(0.239)
-2.145***
(0.462)
-0.003
(0.002)
0.081***
(0.027)
Model 3
0.899***
(0.157)
0.732***
(0.107)
-0.354***
(0.053)
0.221
(0.188)
0.255***
(0.024)
0.603**
(0.304)
-0.169
(0.229)
-2.092***
(0.450)
-0.003
(0.002)
0.078***
(0.025)
Model 4
0.803***
(0.145)
0.522***
(0.085)
-0.297***
(0.040)
0.122
(0.146)
0.249***
(0.026)
0.050
(0.323)
-0.135
(0.211)
-1.619***
(0.406)
-0.003
(0.002)
0.073***
(0.021)
0.349***
(0.115)
0.609***
(0.184)
-0.266***
(0.081)
-0.018
(0.544)
YES
11458
0.126
(0.149)
Government Attacks
0.255***
(0.025)
Poverty
0.058
(0.330)
Population density
-0.162
(0.218)
Other tenancy
-1.654***
(0.417)
Govt Officials
-0.003
(0.002)
Police Stations
0.075***
(0.022)
Coca
0.345***
(0.118)
New colonized
0.639***
(0.192)
Altitude
-0.274***
(0.082)
Percent minorities
-0.051
(0.543)
Year Effects
YES
YES
YES
Observations
11507
11458
11507
* p < 0.10; ** p < 0.05; *** p < 0.01 (two-tailed)
34
5
Additional Case Study Evidence from Araucan Towns
This section expands upon the discussion of the case municipalities discussed in the manuscript,
providing further background on land reform and guerrilla activity from the municipalities in Arauca.
From the 1960s through the 1980s there was a strong relationship between Araucan peasants and
INCORA (Carroll 2011). Although there was uniformly large (“prior”) land reform during this era
across the case municipios, their histories from the subsequent period exhibit greater divergence. In
1964, INCORA began the Arauca Project 1, which would come to title a total of 200,000 hectares.
Colonists promptly arrived and were given land titles. INCORA’s efforts reflected integral and
intensive development: the institution assisted economic development and helped provide additional
public goods, such as paved roads to help transport agricultural products and inputs to help make the
land productive and the communities prosper. As one resident of Saravena recalled, “The Institute
was very concerned about us. They would visit even the most far-flung villages... This is a big
region, but whichever part you venture to, anyone will tell you, ’This was made by INCORA.’
Bridges, schools, hospitals, airports. Saravena was practically built by INCORA.’ It helped us
organize as a community and in cooperative labor to keep up the plots” (INCORA 2002). Carroll
confirms the intensiveness of INCORA activity, noting that “the piedmont settlers’... [of Saravena,
Tame, and Arauquita] interests were largely ignored by the dominant plains politicians... Instead,
they derived benefits directly from INCORA” (Carroll 2011, 183). In later years, from the 1970s
and continuing into the early 1990s, ANUC-supported protests successfully demanded additional
responses from INCORA in the form of titling and credit. As Carroll (2011, 183) writes, “The first
major mobilization, a civil strike in 1972, concentrated five thousand protesters in Saravena and
succeed in not only bringing national authorities to the negotiating table but also in reactivating
INCORA’s loan and expenditure programs.”
The beginnings of land reform in Arauca were auspicious, but evidence from the latter period of
the 1980s and 1990s shows that titling was better sustained in some parts of the department, such
as Tame, than others. Sometimes, for random reasons, there were opportunities to easily title land,
such as the death of a large landowner. This happened in a case in Tame when Julio Delgado died
in 1997 (El Tiempo 1999). His latifundio of 103,000 hectares of highly productive land underwent
an extinction of domain in 1999 and was repartitioned to some 350 peasant families, some of whom
had previously invaded the area as early as 25 years prior.§
In other parts of the Department of Arauca – beyond Tame where the latifundio was successfully divided – INCORA was apparently less efficient and effective at titling land for bureaucratic
reasons (Rojas Sánchez 1994). Policy flexibility for additional reforms such as increasing funding for
local INCORA offices was impeded by centralized control by the national government and political
incentives to not concede to insurgent demands (Carroll 2011, 222). In one instance, when national
politicians balked at meeting ANUC protesters’ demands to bolster INCORA’s Arauca office with
additional resources, the departmental governor was able to overcome the bureaucratic impediments
by contributing revenues from oil proceeds (Carroll 2011, 221-22). However, as a 1992 report shows,
these kinds of contributions to supplement land titling were apparently not sustained. The Araucan
departmental manger of INCORA could not get sufficient attention and response from the national
government office to complete titling for the Sarare subregion of Saravena, one of the case municip§
As we control for in our models with variables for paramilitary attacks, the hacendado effect was
present in areas of Arauca where large landholders acted to defend their properties, “[In the plains]...
holdings tend to be very large...and their owners are wealthy and powerful, but frequently without
formal and universally recognized land titles. Instead, the large cattle rancher would have a right of
possession. The legal precariousness of the right to possession probably contributed to the vehemence
with which the wealthiest ranchers later reacted to the emergence of the peasant settler-based Left”
(Carroll, 2011, 180-181).
35
ios which suffered high levels of guerrilla actions (El Tiempo 1992). Even though this manager was
viewed as an “honest” official and had completed some public works projects in Saravena, ANUC
and other peasant organizations were left unsatisfied and placed the blame on him for the failure to
get the national government to address their needs.
In sum, while Tame benefitted from more sustained episodes of land titling during the 1990s, a
review of the archives of the national newspaper El Tiempo reveals fewer reports of similar INCORA
activity in the 1990s in other Araucan towns, such as Saravena.
36
6
References
El Tiempo. 1992. “INCORA del Arauca: El Peligro Está Acechando” (INCORA of Arauca: Being
Stalked by Danger). El Tiempo, April 4, 1992.
Carroll, Leah Anne. 2011. Violent Democratization: Social Movements, Elites, and Politics in
Colombia’s Rural War Zones, 1984-2008. South Bend: University of Notre Dame Press.
Carroll, R. J., D. Ruppert, and L. A. Stefanski. 1995. Measurement Error in Nonlinear Models.
London: Chapman & Hall.
Cook, J. and L. A. Stefanski. 1995. “A simulation extrapolation method for parametric measurement error models.” Journal of the American Statistical Association 89: 1314-28.
El Tiempo. 1999. “La Finca de Don Julio.” August 11, 1999.
Guzmán, Daniel, Tamy Guberek, Amelia Hoover, and Patrick Ball. 2007. “Missing People in
Casanare.” Unpublished manuscript.
Hardin, J. W., H. Schmiediche, and R. J. Carroll. 2003. “The regression-calibration method for
fitting generalized linear models with additive measurement error.” Stata Journal 3(4): 372-84.
Imai, Kosuke, and Teppei Yamamoto. 2010. “Causal Inference with Differential Measurement
Error: Nonparametric Identification and Sensitivity Analysis.” American Journal of Political Science
54(2): 543-60.
INCORA. 2002. Colombia, Tierra Y Paz: Experiencias Y Caminos Para La Reforma Agraria
Alternativas Para El Siglo XXI, 1961-2001. Bogotá: INCORA.
Lum, Kristian, Megan Price, Tamy Guberek, and Patrick Ball. 2010. “Measuring Elusive Populations with Bayesian Model Averaging for Multiple Systems Estimation: A Case Study on Lethal
Violations in Casanare, 1998-2007.” Statistics, Politics, and Policy 1(1): 1-26.
Rojas Sanchez, Martha. 1994. “Investigan al INCORA por Demorar Tı́tulos” (INCORA Being
Investigated for Delaying Titles). El Tiempo, April 13, 1994.
37

Supplementary Appendix.

Transcription

Similar documents

Guerrilla marketing – structuring the

Shark Attacks - International Wildlife Museum

REVISTA SEMANA

On the Privacy of Real-World Friend-Finder Services

November 17th, 18th and 19th, 2011

Guerilla Warfare - Che Guevara

ISACA Columbus April 2006

Shaolin Monks Fact Sheet

Automating the hacking process

Intrusion Detection and Prevention

The Ethical Hacking Guide to Corporate Security