Conflict, Road Insecurity and Self-Employed Shopkeepers in
Rural Peru
Julia Seiermann
∗
November 8, 2012
Abstract
This paper examines how violent conflict affects the decision to become a self-employed
shopkeeper or vendor, proposing road insecurity as a new channel of conflict transmission.
In a conflict situation, the destruction of roads and the risk of assaults increase the cost of
transporting goods. The impact of distance to the next market on the probability of opening a
shop is thus expected to differ between conflict and non-conflict districts. I test this prediction
in the context of the Peruvian armed internal conflict. Using a logit model, I find that the
probability of opening a shop decreases with distance to the next market in conflict districts. In
a 2SLS specification, distance to Ayacucho and temperature are used as instruments for conflict
and its interaction with distance to market. The coefficient of interest becomes insignificant,
but exogeneity cannot be rejected. Several robustness checks provide further support to the
logit result.
∗ Graduate
Institute for International and Development Studies, Geneva
1
“. . . y regresamos a nuestra casa de miedo; de cierto ya no salı́amos ya a la calle o no habı́a
negocio. Ya no habı́a ni qué hacer comer a nuestros hijos, ya no entraba ni a la tienda nadie, se
ha cerrado no más ya, adentro estábamos con miedo.“
“. . . frightened, we returned to our house; we did not go out in the street anymore, neither was
there any business. There was not even anything left to feed to our children, nobody entered the
shop anymore, it was simply closed, and we were inside, frightened.“
Julia Castillo Jopa
Shopkeeper from Puquio district, Ayacucho
Public Hearing of the Truth and Reconciliation Commission
Lima, June 21, 2002
1
Contents
1 Introduction
3
2 Literature Review
4
3 The Peruvian Armed Internal Conflict
5
4 Data
4.1 Data Sources . . . . . . . . . . . . . . . . . .
4.2 Descriptive Statistics . . . . . . . . . . . . . .
4.2.1 Shopkeepers and Street Vendors . . .
4.2.2 Conflict versus Non-Conflict Districts
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
6
6
6
7
5 Theory and Empirical Strategy
5.1 Conflict, Transport Cost and the Occupation Decision
5.2 The Baseline Equation . . . . . . . . . . . . . . . . . .
5.3 Control variables . . . . . . . . . . . . . . . . . . . . .
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
7
8
9
11
6 Instrumenting for Conflict
6.1 Is There a Risk of Omitted Variable Bias?
6.2 Identification Strategy . . . . . . . . . . .
6.2.1 Distance to Ayacucho . . . . . . .
6.2.2 Suitability for Coca Cultivation . .
6.3 Technique . . . . . . . . . . . . . . . . . .
6.4 Results: 2SLS Model . . . . . . . . . . . .
6.4.1 First Stage . . . . . . . . . . . . .
6.4.2 Second Stage . . . . . . . . . . . .
6.4.3 Tests . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
12
12
14
14
14
15
16
16
16
16
7 Robustness Checks
7.1 Different Conflict Periods . . . . . . . . .
7.2 Conflict Intensity . . . . . . . . . . . . . .
7.3 Other Non-Agricultural Self-Employment
7.4 Outliers and Market Villages . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
19
19
20
21
21
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8 Conclusion
22
References
23
A Appendix
26
2
1
Introduction
This paper assesses the impact of violent conflict on self-employment in rural Peru of the 1990s.
While the conflict literature has so far emphasized the creation or destruction of employment
opportunities (Justino [2010]) and migration (Bozzoli et al. [2011]), I examine a new channel through
which conflict can affect the self-employment decision, namely road destruction and insecurity.
In rural Peru, over 50% of the non-agricultural self-employed work as shopkeepers or street vendors. Many of the goods they sell are not produced in the same village, but need to be transported
there. In a peaceful context, distance to the next market might increase the profitability of a shop,
since distance makes it more efficient to bundle transportation. During the armed internal conflict, however, the insurgent group1 Sendero Luminoso (Shining Path) destroyed roads and other
infrastructure. Trying to impose a subsistence economy and “starve out the cities“, the insurgents
prohibited villagers to participate in markets, persecuted those who did not comply, and assaulted
vehicles transporting goods (Del Pino [1996]). Thinking of road insecurity as an increase in transportation cost, it should lead to a reduction in vendors’ profits. Ceteris paribus, road insecurity
should thus deter people from becoming a self-employed shopkeeper or vendor.
Yet, conflict also affects the employment decision through other channels. People may be pushed
into self-employment following the destruction of alternative employment opportunities. A reduction of the local labor force through migration or deaths may enable people to switch from selfemployment to wage employment. In order to test the road security channel, I therefore focus on
distance to the next market. I expect its impact on self-employment to differ between conflict and
non-conflict districts. Using a logit model, I regress self-employment as a shopkeeper or street vendor on a district-level conflict dummy, the distance between the village and the next market, and
an interaction term between conflict and distance to market. The results suggest that distance to
market has a negative impact impact on the probability of working as a self-employed shopkeeper
in conflict districts. This finding is in line with my hypothesis, according to which conflict affects
the occupation decision via the channel of road insecurity. It is robust to a variety of specifications,
in which I control for other factors influencing the occupation decision.
The concern that conflict and the self-employment decision might both be endogenous to local
economic conditions is addressed by using an instrumental variables (IV) procedure. Given that my
regression equation includes an interaction term, I need at least two instruments for conflict. First,
I use distance to Ayacucho, a department capital in the central Andes. It was at the local university
that philosophy professor Abimael Guzmán split from the Maoist wing of the Communist Party
of Peru (PCP) to found Partido Comunista del Perú - Sendero Luminoso (PCP-SL), which would
later initiate the armed internal conflict. Second, I use temperature, as a proxy for the suitability
of a region to cultivate coca. Given fragile state control, insurgents of various groups were able to
establish a stronghold in coca-producing areas. In the 2SLS model, the coefficient on the interaction
term between conflict and distance to market is not significant. Yet, the hypothesis that conflict
1 In
its decree to establish the Truth and Reconciliation Commission (CVR), the Peruvian government qualified
Sendero Luminoso as a terrorist organization. The CVR itself, however, distinguishes between terrorist and nonterrorist actions in its analysis. As this paper does address the legal dimension of the conflict, I use the term
insurgency when referring to Sendero Luminoso.
3
and its interaction with distance to market are exogenous cannot be rejected, which suggests that
the logit results are unbiased.
Several robustness checks provide support for the road insecurity hypothesis, and raise interesting further questions on the underlying mechanisms. The negative effect of distance to market
on the decision to work as a vendor is neither driven by unobserved common characteristics of
the districts around Ayacucho (the epicenter of the conflict) nor by a general effect of distance to
market on non-agricultural self-employment. Conflict intensity, measured as the number of fatal
victims in a district, increases the negative impact of distance to market. Significant results on past
conflict raise the question of the persistence of the effect, but cannot be disentangled from conflict
intensity in the present analysis.
2
Literature Review
The self-employment literature has traditionally focused on the importance of credit constraints,
such as the classic papers by Evans and Jovanovich [1989] and Blanchflower and Oswald [1998].
Other determinants of self-employment which have been highlighted include age, household characteristics, parents’ employment status, education, experience, and alternative labor market options
(Georgellis et al. [2005], Le [1999]).
Most of the literature is concerned with industrialized countries, but a number of recent studies
have analyzed the determinants of self-employment in developing country contexts. Paulson and
Townsend [2004] find that financial constraints restrict entrepreneurial activity in Thailand. The
importance of infrastructure for the household decision to establish a nonfarm enterprise is emphasized in Deininger et al. [2007] for Sri Lanka, and in Jin and Deininger [2008] for Tanzania. Both
studies report significant positive effects of different types of infrastructure, such as electrification,
distance to a bank, credit access, public transport, and road quality. Escobal [2001] studies the
determinants of income shares by activities in Peruvian households using the 1997 Living Standards
Measurement Survey (LSMS). He finds a positive impact of education, electricity and credit access
and a negative impact of distance to market on the share of income stemming from non-agricultural
self-employment activities.
As to my knowledge, only two studies have so far addressed the impact of conflict on selfemployment. Using data from Uganda, Deininger [2003] finds that civil strife reduces non-agricultural
enterprise start-ups at the household level. Bozzoli et al. [2011] focus on the impact of conflict on
self-employment in Colombia. They find a negative impact of homicide rates on men’s, but not
women’s self-employment. Considering the channel of displacement, they find that individuals are
less likely to be self-employed in sending communities and more likely to be self-employed in host
communities.
This paper ties in with the existing literature by suggesting a new channel through which
violent conflict affects the self-employment decision, namely road insecurity. Since this channel is
of particular importance for rural shopkeepers and vendors, who bring in goods in order to sell them
4
in their village, my analysis concentrates on these professions. While highlighting road insecurity, I
control for other determinants of self-employment, such as education, credit access and agricultural
wages.
3
The Peruvian Armed Internal Conflict
Peru was just undergoing a transition from a military dictatorship to democracy and had elected a
new government when the Maoist insurgent group Sendero Luminoso declared a “people’s war“ and
initiated violence in 1980 in the Central Andean province of Ayacucho. Initially, the conflict did
not figure among the government’s main preoccupations, but was perceived as a local delinquency
problem. But when Sendero Luminoso enlarged the scope of its activities, and the police was not
able to contain them, the government decided to deploy the military to the conflict zones in 1982.
The conflict spread from Ayacucho to other regions of the country, mainly neighboring Andean
provinces, Lima and the central jungle region. Parties to the conflict included Sendero Luminoso
and Movimiento Revolucionario Túpac Amaru (MRTA), another left-wing insurgent group, as well
as police, armed forces and rural self-defense committees (rondas campesinas). Violence escalated,
and human rights violations by both insurgents and state forces were numerous. The total number
of dead and disappeared during the conflict is estimated 70,000 by the Truth and Reconciliation
Commission (Comisión de la Verdad y Reconciliación, CVR).
The main violent phase of the conflict ended in 1992 with the capture of Abimael Guzmán,
the head of Sendero Luminoso, in Lima. However, violence continued on a lower scale during
the autocratic government of Alberto Fujimori, which lasted until 2000 (Comisión de la Verdad y
Reconciliación [2003]). Until today, isolated attacks take place in the Andean and jungle regions of
the country, where Sendero Luminoso is still cultivating links to drug trafficking (Huerto Amado
[2012]).
According to the CVR, the majority of victims were among the rural, historically excluded
population, which was symptomatic of the divisions within the Peruvian society. The extensive
research of the CVR focused on the numerous human rights violations committed in the conflict,
paying close attention to the juridical dimension of the conflict, but also the context which enabled
its emergence and duration. While these atrocities were certainly its most abominable aspect,
the conflict also caused profound disruptions of public life and social and economic interactions
(Theidon [2001]). The destruction of roads, the assaults, and the ideologically motivated effort of
Sendero Luminoso to suppress market activity have affected the local economy, especially in rural
areas. The aim of this paper is to contribute to a better understanding of their consequences, which
could have persistent implications on poverty.
5
4
4.1
Data
Data Sources
I use data from the 1994 Living Standards Measurements Survey (conducted by the Instituto Nacional de Estadı́stica e Informática [1994]) for employment status, distance to market and individual,
household and village-level controls. The LSMS comprises individual and household modules on a
variety of topics, as well as a community questionnaire. For my analysis, I restrict the sample to
adults (15 years and older) from rural areas who are active in the labor market and for whom all
necessary information is available.
96% of the individuals in my initial sample live in villages which are at most 26 km away from
the next market. The remaining 4% live in those 5 villages (out of 143) which lie between 48
and 84 km away from the next market. I exclude these observations, since they considerably alter
the magnitude and significance of the regression coefficients (see robustness check 4) and might be
caused by errors. My final sample comprises 2,122 observations.
For conflict intensity, I use district-level data from the Peruvian Truth and Reconciliation Commission on deaths and disappearances during the armed internal conflict. Distance to Ayacucho is
measured with respect to each district capital. I constructed the data using the online tool “Distance Calculator Peru“, which measures the air distance between locations based on their latitude
and longitude. Altitude is obtained for each district capital from Google Earth. Temperature,
humidity and precipitation are measured at department level, by constructing the average annual
value between 1997 and 2006. The climate variables are obtained from the Compendio Estadistico
2007, an extensive statistical publication of the Instituto Nacional de Estadı́stica e Informática
[2007]. District-level data on education is constructed from the 1993 national census conducted by
the Instituto Nacional de Estadı́stica e Informática [1993].
4.2
4.2.1
Descriptive Statistics
Shopkeepers and Street Vendors
19% of the people in my sample list non-agricultural self-employment as their primary or secondary
activity, with a significantly higher proportion in conflict (22%) than in non-conflict districts (17%).
More than half of them work as shopkeepers or street vendors. (See appendix table 6.)
A high percentage of shopkeepers and vendors are female (65%, compared to 42% of women in
the entire sample). Shopkeepers and vendors are more likely to have completed primary education
than people working in other occupations. Their households have higher than average consumption
expenditures, credit access and electricity. The shopkeepers in my sample tend to live in villages
which are closer to markets: 5.6 km on average, compared to a sample mean of 6.3 km. Their
villages are relatively poor (indicated by a higher likelihood of investment by the Peruvian Social
Fund, FONCODES), and labor market conditions are less likely to have improved over the previous
years. (See appendix table 7.)
6
4.2.2
Conflict versus Non-Conflict Districts
In conflict districts, 12% of individuals work as shopkeepers or vendors, compared to 9% in nonconflict districts. Individuals in conflict districts are more likely to have completed primary education, and the general education level obtained from the 1993 national census is higher in conflict
than in non-conflict districts. In my sample, villages in conflict districts are further away from
the next market (7.2 km, compared to 5.7 km for those in non-conflict districts) and have had
a relatively favorable labor market evolution. It has become easier to find a job in more conflict
than non-conflict villages, while it has become more difficult to find a job in more non-conflict than
conflict villages. (See appendix table 8.)
5
5.1
Theory and Empirical Strategy
Conflict, Transport Cost and the Occupation Decision
As discussed above, the literature on self-employment identifies a number of variables which influence the occupation decision. Conceptually, they can be divided into two categories: individual
and outside factors. Individual factors include personal characteristics such as age, education and
physical and intellectual capabilities, but also family characteristics such as household composition,
wealth, and access to credit. Outside factors comprise labor market conditions (i.e., alternatives to
self-employment) and the profitability of different professional activities given local conditions.
This paper emphasizes one particular outside factor, the profitability of working as a selfemployed vendor. I assume that profits from vending activities are altered by conflict via the
channel of road insecurity. I do not directly estimate profits, but I assume that, ceteris paribus, a
person is more likely to choose an occupation if expected profits are high. Under the assumption
that the other important factors are controlled for, the decision to work as a self-employed vendor
can serve as an indicator of profitability of this activity.
Distance to market and transport costs are assumed to play a significant role in the decision
to become a self-employed vendor. The importance of market access for agricultural producers in
developing countries has been highlighted by numerous studies, for example Jacoby [2000]. To my
knowledge, the impact of distance to market on the profitability of vending activities has not yet
been investigated. In peaceful times, distance to market may even raise the profitability of vending
activities, since it increases the efficiency of bundling the transportation of those goods which are
not produced and sold in the same village.
Conflict-induced road insecurity increases transport cost and thus the cost of supplying products. Note that the increase in transport costs, as an abstract concept, includes not only higher
prices for public transport or the risk of losing one’s goods, but also the risk of being injured or
killed in an assault. Anecdotal evidence shows that fear from the insurgents (or the armed forces)
caused people to substantially reduce economic and social interaction. (Comisión de la Verdad y
Reconciliación [2003] But conflict also affects profitability through other channels, such as clients’
income, and influences other factors of the self-employment decision, such as individual and labor
7
market characteristics. The aim of my estimation strategy is thus to isolate the effect of road
insecurity on the occupation decision.
5.2
The Baseline Equation
I use a logit model to regress a binary variable for working as a self-employed shopkeeper or vendor
on a conflict dummy, distance to market and the interaction between these two. Standard errors are
clustered at the village level, and different sets of control variables are included. As the coefficients
from binary choice models cannot be interpreted directly, I report OLS results to provide additional
intuition. Finally, I estimate a complementary log-log (cloglog) model, since the dependent variable
only takes the value 1 for 10 % of the observations.2
vendori = β1 · distmktv + β2 · conf lictd + β3 · conf lictd ∗ distmktv + β4 · controls
(1)
The subscript i indicates that the variable is measured at the individual level. v denotes village,
and d district level. The different sets of controls vary at individual, household, village, district and
department level (for a complete list, see appendix table 5). The dependent variable, vendor, takes
the value 1 if a person is working as a self-employed shopkeeper or street vendor as her primary or
secondary occupation. conflict is a binary variable which takes the value 1 when the number of dead
and disappeared at district level between 1990 and 1994 is positive, and 0 when no fatal victims
were registered in this time period. Although the conflict started in 1980, I use only the years
immediately preceding the LSMS survey, assuming that these are most relevant for the occupation
decision. The role of past conflict is investigated in detail in robustness check 1. Distance to market
(distmkt) is measured at village level.3 Control variables include several individual, household and
village-level variables. They are discussed in detail in section 5.3., and a list of all variables and
their definitions is included in the appendix.
As conflict is measured at district level, the conflict variable is subsumed by the district fixed
effects. However, the interaction term between conflict and distance to market can still be estimated, as there is variation in distance to market between the villages within a same district.
The interpretation of the interaction coefficient is the same as in the model without fixed effects.
In 19 districts, there are no observations of self-employed shopkeeper and vendors in the dataset.
The district fixed effect would predict the outcome perfectly in these cases, which is why they are
excluded from the logit and cloglog analyses. The sample size is reduced to 1,735 observations.
Self-employed shopkeepers and vendors are now overrepresented with respect to the original sample; and other systematic differences between included and excluded observations might cause the
results to differ substantially. A table with summary statistics of all explanatory variables in the
included and excluded sample is provided in in the appendix.
2 The complementary log-log model is asymmetric around zero, which is why it is used when the outcome of
interest is rare. Conditional probability in this model is given by the cdf of the extreme value distribution: C(x0 β) =
1 − exp(−exp(x0 β)) (Cameron and Trivedi [2005]).
3 To capture the time-dimension of transport costs, it would be ideal to construct a measure of “effective“ distance
to market by combining geographic distance with measures of road access and the availability of public transport.
Unfortunately, the data is not detailed enough for this endeavor.
8
I expect the impact of distance to market to differ between conflict and non-conflict districts. If
β3 , the coefficient on the interaction term between conflict and distance, is negative and significant,
this supports my hypothesis that conflict-induced road insecurity deters people from becoming a
self-employed vendor. The interpretation of interaction terms in binary choice models is, however,
far from straightforward and will be discussed in further detail in section 5.4.
5.3
Control variables
I control for basic characteristics at the individual and household level, namely age, gender, and
household composition (number of children (0-14) and adults in the household). Furthermore, I
introduce a number of control variables aimed to capture other important factors in the occupation decision: education, access to capital, infrastructure, migration, labor market conditions and
agricultural opportunities.
A person’s education is an important determinant of her range of employment options. In
particular, skills such as literacy and numeracy may give her a comparative advantage in vending
occupations. A higher education level may make these occupations less desirable in comparison
to alternative opportunities. To control for these effects, I introduce a dummy variable for complete primary education. Additional controls for complete secondary and tertiary education were
insignificant, which is why I do not include them in the final specification anymore.
Access to capital has been highlighted by the literature as one of the most decisive determinants of self-employment. Individual ability to raise capital is determined by a person’s (and her
household’s) wealth, and by access to credit. I approximate household wealth by including annual
consumption expenditures in the regression equation. This measure is not completely exogenous to
the occupation decision and the income generated in each individual member’s activity. However,
alternative measures, such as farm size, durable asset ownership or size of the household dwelling
are not completely exogenous either, and consumption expenditures present the advantage of being
a more contemporaneous measure of the household’s financial situation. Furthermore, the magnitude and significance level of the coefficients of interest remain the same when using the number
of rooms in the household dwelling instead of consumption expenditures. To control for credit
constraints, I include a self-reported binary measure of credit access. Only 15% of the individuals
in my sample live in households which reported access to some form of credit. This low percentage
may be explained by the fact that the Peruvian microcredit sector only took off in the early to
mid-1990s, and might not have reached many rural areas in 1994. However, credit access might still
be underreported, which is why this measure needs to be interpreted with caution. Since the development literature highlights the importance of infrastructure for establishing a nonfarm household
enterprise, I include a binary variable indicating whether the household dwelling has electricity or
not.
Individuals with entrepreneurial talent or motivation might migrate to villages in which vending
activities are particularly profitable, but also be more likely to migrate to escape violence. Selective
migration could bias the coefficients of interest. Thus, I controlled for individual migration using
a binary variables which takes the value 1 if a person was born in the current place of residence.
This variable was insignificant in all specifications tested, which is why I exclude it from the final
9
specification. Bozzoli et al. [2011] highlight that conflict may affect local labor markets through
displacement. As a proxy for migratory movements at a more aggregate level, I include dummy
variables indicating whether the number of dwellings in the village has increased or decreased since
1991.
The labor market situation in a village is probably one of the most important factors in the
individual occupation decision. One potential measure of alternative employment opportunities in
rural areas are agricultural wages. However, their impact on the occupation decision may go into
both directions. High agricultural wages may prevent people from becoming vendors, but they may
also make vending activities more profitable by raising the income of potential clients. Since my
analysis is not directly interested in the impact on wages, I use a more direct measure of alternative
opportunities. I combine an indicator of how the labor market in the village has evolved since 1991
with a proxy for the village’s economic situation in the beginning of the 1990s.
The LSMS community questionnaire contains a question whether it was easier or more difficult to
find a job in 1994 than in 1991. This variable the perceived evolution of the labor market situation
during those years, but does not provide a level-level comparison between villages. In order to
control for the relative economic situation, I use a binary variable indicating whether the Peruvian
Social Fund (FONCODES) has invested in a village until the time of the survey. FONCODES was
created in 1991 and started funding a variety of community-based projects in 1992. Resources were
allocated using a district-level poverty index and informal on-site assessments. Looking at school
infrastructure investments, Paxson and Schady [2002] find that FONCODES effectively targeted
poor districts, which is why its activities can be used as a proxy for low economic development.
Together, the FONCODES dummy and the recent labor market evolution should therefore capture
a village’s economic and labor market situation.
Education, an important factor in the occupation decision, played a crucial role in the advent
of the Peruvian conflict. The country experienced an unprecedented education expansion between
the 1940s and 1960s. With the increased access to education, rural and lower-class Peruvians were
hoping for progress and social mobility. The frustration of these expectations made the population
more receptive to violent, radical proposals such as those of Sendero Luminoso. Degregori describes
the Ayacuchan “Movimiento del 1969“, a protest against the abolition of the gratuity of education
by the military government, as a “clarion call, which announced the possibility of the apparition of
a phenomenon like Sendero Luminoso, and of its expansion to other places“ (Degregori [2007], p.
6). Furthermore, given their distinguished position in the villages, school teachers (many of them
educated at Universidad Nacional San Cristóbal de Huamanga, where the leader of Sendero Luminoso was a professor in the department of education) were instrumental in diffusing the movement’s
ideology and gaining support among the rural population (Comisión de la Verdad y Reconciliación
[2003]). Education could thus be an important source of omitted variables bias, which is why I
introduce a measure for the education level in each district, namely the percentage of persons 15
and older with at least one year of secondary education.
Finally, agriculture is the most important source of (self-)employment in rural areas, and might
alter the relative attractiveness of non-agricultural self-employment. I therefore include log(altitude)
of the district capital and average annual precipitation at the department level as proxies for
agricultural opportunities.
10
5.4
Results
The results from the different models are in line with my hypothesis. The coefficient on the interaction term between conflict and distance to market is negative and significant at the 5% level
in all cases. The inclusion of district fixed effects substantially increases the coefficient of interest. As discussed above, this effect could be caused by systematic differences between excluded
and included districts in the logit and cloglog cases. A comparison between the OLS, logit and
cloglog fixed effect results reveals that all but one coefficient (More difficult to find a job) are of
the same sign and statistical significance, which alleviates this concern about potential bias.The
main result is robust to a variety of specifications, in which I include different sets of the control
variables described above (see appendix table 6 for a detailed list, and tables 10-15 for reports of
all coefficients). It indicates that the likelihood of working as a self-employed shopkeeper decreases
with distance to market in conflict districts.
The coefficient on distance to market is insignificant for the baseline specifications (without
fixed effects). These results suggest that distance to market does not influence the decision to work
as a shopkeeper in non-conflict districts, but deters individuals from choosing this occupation in
conflict districts. With the inclusion of district fixed effects, however, the coefficient on distance to
market turns positive and strongly significant. This suggests that, within a given district, distance
to market makes individuals more prone to work as a shopkeeper in non-conflict settings. This
positive effect of distance is offset in the presence of conflict, in which case the net effect is negative.
Conflict per se appears to increase the likelihood of being a vendor, as the coefficient is positive and
significant. The reason for this effect cannot be easily determined. While it is, for example, possible
that the conflict curtailed other employment opportunities, note that there is no such positive effect
for other non-agricultural self-employment activities (see robustness check 3).
As mentioned above, the interpretation of coefficients on interaction terms in binary choice
models is difficult. Norton et al. [2004] highlight that the marginal effect depends on the other
covariates, and may even switch signs. Furthermore, its statistical significance also varies between
observations. The authors suggest reporting the interaction effects (i.e., the cross partial derivative
of the probability) for each observation and their significance plotted against the predicted probability of observing the outcome of interest for each observation. Kolasinski and Siegel [2010] criticize
the use of the cross partial derivative and show that the switch in signs can be a mechanical effect
driven by the restriction of the outcome variable between zero and one. Greene [2010] questions
the usefulness of testing the statistical significance of the interaction effect for each individual, and
suggests the use of graphics for a more meaningful analysis. In line with this suggestion, I construct
graphs to illustrate the relation between increasing distance to market and the likelihood of working
as a vendor in a non-conflict and conflict district respectively. The graphs depict the predictions
from the logit regression (with district fixed effects) for a “representative“ individual who is likely
to become a vendor: a 38-year old woman with completed primary education, living in a household
with 2 children and 2 other adults, with average consumption expenditures and neither credit access
nor electricity. All village, district and geographical controls are set at the sample mean. The solid
line represents the predicted likelihood of working as a vendor from the logit estimation, and the
dashed lines represent the 95% confidence interval.
11
Table 1: Results of different specifications, all control variables included
OLS
VARIABLES
Logit
Cloglog
(1)
(2)
(3)
(4)
(5)
(6)
0.087***
(0.032)
0.003
(0.003)
-0.008**
(0.004)
-
0.117***
(0.030)
-0.166**
(0.077)
0.918***
(0.283)
0.037
(0.030)
-0.090**
(0.040)
-
0.011***
(0.004)
-0.013***
(0.004)
1.011***
(0.313)
0.041
(0.033)
-0.100**
(0.043)
0.101***
(0.025)
-0.141**
(0.071)
Individual controls
Household controls
Village controls
District controls
X
X
X
X
X
X
X
-
X
X
X
X
X
X
X
-
X
X
X
X
X
X
X
-
District fixed effects
-
X
-
X
-
X
Conflict
Distance to market
Conflict*Distance to market
Observations
2,122
2,122
2,122
1,735
2,122
1,735
(Pseudo) R-squared
0.075
0.158
0.115
0.205
.
.
Log-pseudolikelihood
-411.4
-311.4
-627.3
-525.7
-628.3
-524.1
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05, * p<0.1.
12
6
6.1
Instrumenting for Conflict
Is There a Risk of Omitted Variable Bias?
The validity of the regressions presented in the previous section relies on the assumption that the
incidence of district-level conflict between 1990 and 1994 is exogenous to the occupation decision.
A potential concern is that difficult economic or labor market circumstances may at the same time
affect the occupation decision and induce people to support insurgent groups. A careful discussion
of the circumstances of the conflict is thus necessary to evaluate the concern of omitted variables
and devise a valid identification strategy.
Violence was initiated in Ayacucho in 1980, and spread to other parts of the central Andes, to
the jungle regions and to coastal cities over time. To understand the causal mechanisms behind the
geographic dispersion of the conflict, it is helpful to distinguish two questions: Why did the conflict
start in Ayacucho, and why did it spread to other parts of the country? The second question is less
controversial: As observed by León [2012], Sendero Luminoso extended the domain of its activities
according to a politico-military strategy, aiming to access political influence (Lima and other coastal
cities) and resources (the coca regions in the jungle). The first question - why did Sendero Luminoso
emerge in Ayacucho? - has motivated academic research from different disciplines, such as history,
sociology and anthropology. The ideological radicalism of Sendero Luminoso, and its apparent
disdain for human life, sets it apart from other insurgent groups in Latin America.4 Therefore,
Sendero Luminoso has regularly been depicted as alien to the Peruvian society. Degregori [2007], for
example, describes its evolution as an “ideological mutation“ (p. 182) towards “a kind of minuscule
star, those in which matter agglutinates almost without interatomic spaces, thereby attaining great
weight, disproportionate for its size“ (p. 180).5 Given these descriptions, one is tempted to regard
Sendero Luminoso as exogenous to social and economic conditions, a strange historical accident.
But while Sendero Luminoso remained a relatively small, autarkic movement, it is unlikely that
the conflict could have reached a similar magnitude without any support from the local population.
Degregori [2007] argues that the combination of the low profitability of latifundios (large agricultural
estates) and the expansion of education can explain the success of Guzmán’s ideology in Ayacucho.
As agricultural profits were meager and other professional alternatives lacking, education was seen
as the only path to progress, and the frustration of the high expectations was particularly strong.
Local economic conditions which influence the occupation decision may thus be directly related
to the conflict. I attempt to account for these factors using village and district control variables
for education and the economic and labor market situation, as described in section 4.3. However,
some of these measures are rather coarse, and might not be sufficient to prevent omitted variables
bias. To address this concern, I implement an instrumental variable strategy, using distance to
Ayacucho and average annual temperature as instruments for conflict and the interaction term
4 Sendero Luminoso was responsible for the majority (54%) of victims of the conflict. This contrasts with other
Latin American conflicts, in which state and paramilitary forces caused more victims than the insurgency. Unlike
other groups, including MRTA, the other main subversive movement in the Peruvian conflict, Sendero Luminoso also
attacked the defenseless population.(Comisión de la Verdad y Reconciliación [2003])
5 Translation by the author. The original wording in Spanish is: “una especie de estrella enana, esas donde
la materia se apelmaza casi sin espacios interatómicos, alcanzando ası́ un gran peso, desproporcionado para su
tamaño“.
13
between conflict and distance to market.
6.2
6.2.1
Identification Strategy
Distance to Ayacucho
The armed internal conflict was initiated by the Maoist insurgent group Sendero Luminoso in
1980 in the province of Ayacucho, where their leader Abimael Guzmán had been a philosophy
professor at Universidad Nacional San Cristóbal de Huamanga (UNSCH).6 While the combination
of local economic conditions and education explain the movement’s success, these conditions were
not unique to Ayacucho, as emphasized by Degregori [2007]. My identification strategy relies on
the assumption that Guzmán’s appointment as a professor in Ayacucho, and thus the fact that the
conflict started there, was a historical coincidence.
UNSCH was founded by the Spanish colonizers as the second university of the country in 1680,
when Huamanga was a prosperous mining city on the commercial route between Lima and Cuzco.
Given its relatively low number of students and professors and the lack of government resources
following the War of the Pacific, UNSCH was closed in 1876. It was only reopened in 1959, at
the very beginning of an expansionary wave of public tertiary education in Peru (Luja Peralta and
Zapata Tejerina [1988]). UNSCH distinguished itself from the existing “traditional“ universities
by emphasizing its mission to serve the development of the region and maintaining a certain distance from the local landowning elites. It attracted renowned intellectuals from the entire country
(Comisión de la Verdad y Reconciliación [2003]). Given the left-wing orientation of UNSCH, it
is probably not a coincidence that Guzmán, an active member of the Communist Party of Peru
(PCP), was appointed there in 1962. However, one of the reasons why UNSCH became such a
distinguished project is that it was the first university to be opened in the context of a general
movement towards the modernization and democratization of tertiary education. Its early opening,
in turn, was possible because the university had already existed, such that its buildings were still in
place. Since UNSCH was founded in the colonial era, its geographic location is arguably exogenous
to the social and economic circumstances of the region in the middle of the 20th century. Therefore, distance to Ayacucho, where Guzmán was able to gain ideological support among university
students and build the core of what would later become Sendero Luminoso, can be considered a
valid instrumental variable for conflict.
6.2.2
Suitability for Coca Cultivation
One element of Sendero Luminoso’s expansion strategy was the aim to extract resources from the
coca business. In the coca producing regions of Peru, the weak presence of the state, its failure to
effectively combat drug trade and the corruption of police and bureaucracy facilitated the success of
both insurgent organizations, Sendero Luminoso and MRTA. Their links with the drug traffickers
enabled them to obtain weapons from Colombia. In exchange, the insurgents provided “order“
and protection to coca cultivators and traffickers. The police and armed forces also cultivated
6 Until 1825, the city of Ayacucho was called San Juan de la Frontera de Huamanga, or, in its short form,
Huamanga.
14
links with the coca business, for example by providing runways for planes operating between Peru
and Colombia. However, they seem to have exercised less control over the entire supply chain
than the insurgents. This situation resulted in numerous violent confrontations between insurgents,
police and the military (which was deployed to the region to combat both, drugs and insurgency)
(Comisión de la Verdad y Reconciliación [2003]).
Therefore, the geographical and climatic suitability for coca cultivation can predict conflict in the
Peruvian context. According to the (relatively scarce) literature on coca cultivation, Erythroxylum
coca, the most common species in Peru, grows best at an altitude of 500-2000m. Furthermore, it
requires relatively high temperatures, precipitation and humidity (Morello and Matteucci [2001],
Ismatullayev [2011]). These conditions are met in the tropical parts of the Peruvian Andes.
Like distance to Ayacucho, altitude of the district capital and department-level aggregates of
temperature, precipitation and humidity over 1997-2006 predict conflict well in the Peruvian context. But the necessity of instrumenting the interaction term between conflict and distance to
market as well complicates the identification strategy. A standard approach would involve using an
instrument for conflict and the interaction of this instrument with distance to market. However,
this strategy leads to under- and weak identification in the present setup, which is why I require at
least two different instruments. When including a combination of three or more instruments, the
Kleibergen-Paap statistic for weak identification is very low. Therefore, I use only two instruments,
distance to Ayacucho and temperature, in my final specification.
In order to be a valid IV, temperature should not directly influence the dependent variable, the
choice to become a self-employed shopkeeper or vendor. Two concerns might arise with respect to
this condition: First, substantial differences in agricultural employment opportunities with respect
to other regions might alter the attractiveness of self-employment, and second, the presence of coca
cultivation and trade might offer new employment opportunities or make vending activities more
profitable. Angrist and Kugler [2008] analyze the consequences of the disruption of the so-called
“air bridge“ for transporting coca paste in 1994, following which the production of coca leaves for
cocaine production shifted from Peru and Bolivia to Colombia. Focusing on coca-producing rural
areas of Colombia, they find that the income from self-employment increased, but employment
status, wage and salary earnings and hours worked did not change significantly. These results
suggest that even important changes in the price of coca do not affect the employment decision,
which is my dependent variable. General growing conditions, which are presumably more or less
stable across time, might still affect the occupation choice. However, other geographical control
variables capturing agricultural opportunities (log(altitude and precipitation) are not significant in
the logit, cloglog and 2SLS specification, which alleviates this concern.
6.3
Technique
I employ a 2SLS procedure, with standard errors clustered at the village level. I use the two different
instruments described above, distance of the district capital to Ayacucho (distaya) and average
annual temperature in a department (temperature). I cannot estimate the 2SLS model including
district fixed effects, as the second instrument is at department level, whereas both endogenous
variables vary at district or village level.
15
I estimate a 2SLS model with two simultaneous first-stages for the endogenous regressors, conflict
and its interaction with distance to market:
conf lictd = α1 · distayad + α2 · temperaturedpt + β1 · distmktv + β2 · controlsihvd
(2)
conf lictd ∗ dmktv = α1 · distayad + α2 · temperaturedpt + β1 · distmktv + β2 · controlsihvd (3)
The second stage is equivalent to the baseline specification, in which conflict and the interaction
term are replaced by the predicted values from the first stages:
\
vendori = β1 · distmktv + β2 · conf
lictd + β3 · conf lict\
d ∗ distmktv + β4 · controlsivhd
6.4
6.4.1
(4)
Results: 2SLS Model
First Stage
The instruments predict conflict and its interaction with distance to market reasonably well. As
expected, the coefficient on distance to Ayacucho is negative, and the coefficient on temperature is
positive. The R2 lies between 0.18 and 0.27 in the different specifications, which is not very high.
The problem of weak identification which may result from the low explanatory power of the first
stage is discussed more in detail in section 6.4.3 (Tests).
In the first stage predicting conflict, the coefficient on temperature however turns insignificant
when including all sets of control variables. In turn, the control variable of precipitation, while
insignificant in all other specifications of all models, turn positive and significant at the 10% level in
this particular case. See appendix table 16) Precipitation is positively correlated with temperature,
and may thus capture a common effect of the two, a problem which may be aggravated by the
high aggregation level of the climatic variables. While precipitation is also expected to predict
the suitability of a region for coca cultivation, it is not a suitable instrument for conflict and its
interaction to market due to under- and weak identification problems.
6.4.2
Second Stage
The main result from the logit regression is not robust to the IV specification. The coefficient on the
interaction term between conflict and distance to market is now insignificant in all specifications.
The same is true for the coefficient on conflict. (The magnitude of the coefficients cannot be
compared between logit and 2SLS). This result indicates that distance to market has no impact
on the likelihood of working as a shopkeeper or vendor. As I will discuss more in detail in the
next section, the insignificance of the coefficients is however likely to be caused by the relative
imprecision of the 2SLS estimator, as opposed to the simple logit.
16
Table 2: 2SLS: First stage on conflict (only main coefficients)
VARIABLES
Distance to Ayacucho in 100km
Temperature
Distance to market
Individual controls
Household controls
Village and district controls
Geographical controls
Dependent variable: Conflict
(3)
(4)
(1)
(2)
-0.079***
(0.017)
0.021**
(0.009)
0.011*
(0.006)
-0.082***
(0.017)
0.021**
(0.008)
0.011*
(0.006)
-0.068***
(0.018)
0.016*
(0.009)
0.009
(0.007)
X
-
X
X
-
X
X
-
(5)
(6)
-0.082***
(0.016)
0.020**
(0.010)
0.007
(0.007)
-0.070***
(0.018)
0.016*
(0.009)
0.009
(0.007)
-0.070***
(0.018)
0.016
(0.010)
0.005
(0.008)
X
X
X
X
X
X
-
X
X
X
X
R2
0.184
0.202
0.224
0.229
0.241
0.266
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05, * p<0.1.
Table 3: 2SLS: First stage on conflict*distance to market (only main coefficients)
VARIABLES
Distance to Ayacucho in 100km
Temperature
Distance to market
Individual controls
Household controls
Village and district controls
Geographical controls
(1)
Dependent variable: Conflict*Distance to market
(2)
(3)
(4)
(5)
(6)
-0.596***
(0.172)
0.343***
(0.101)
0.524***
(0.114)
-0.610***
(0.178)
0.347***
(0.101)
0.517***
(0.116)
-0.451***
(0.175)
0.274***
(0.092)
0.501***
(0.112)
-0.608***
(0.176)
0.359***
(0.103)
0.499***
(0.123)
-0.450**
(0.175)
0.273***
(0.090)
0.495***
(0.112)
-0.429**
(0.176)
0.304***
(0.093)
0.484***
(0.120)
X
-
X
X
-
X
X
-
X
X
X
X
X
X
-
X
X
X
X
R2
0.184
0.202
0.224
0.229
0.241
0.266
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05, * p<0.1.
17
Table 4: 2SLS: Second stage (only main coefficients)
VARIABLES
Conflict
Conflict*Distance to market
Distance to market
Individual controls
Household controls
Village and district controls
Geographical controls
Dependent variable: Vendor
(3)
(4)
(5)
(1)
(2)
0.104
(0.111)
-0.005
(0.013)
0.000
(0.007)
0.073
(0.110)
-0.005
(0.012)
0.001
(0.007)
0.160
(0.121)
-0.005
(0.014)
0.001
(0.007)
0.103
(0.121)
-0.009
(0.014)
0.003
(0.007)
0.119
(0.118)
-0.003
(0.013)
0.001
(0.007)
0.181
(0.119)
-0.011
(0.014)
0.004
(0.007)
X
-
X
X
-
X
X
-
X
X
X
X
X
X
-
X
X
X
X
0.059
0.056
0.053
0.0009
0.0025
0.0005
7.23
5.26
8.59
0.92
0.46
0.46
Adjusted R2
0.040
0.058
0.030
Underidentification
P-value
0.0002
0.0002
0.0026
Weak identification
Kleibergen-Paap rk Wald F-stat
7.81
8.49
5.05
Endogeneity test
P-value
0.56
0.85
0.28
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level.
Stock-Yogo critical values: 7.03 (10%), 4.58 (15%), 3.95 (20%).
6.4.3
(6)
*** p<0.01, ** p<0.05, * p<0.1.
Tests
The underidentification test is based on the Kleibergen-Paap rk LM statistic. The null hypothesis
that the regression is underidentified can be rejected in all specifications of the model, in which I
include different sets of control variables.
The weak identification measure reported is the Kleibergen-Paap rk Wald F statistic, since
the Cragg-Donald Wald F statistic is not valid with clustered standard errors. The Stock-Yogo
critical values for weak identification, however, apply to the Cragg-Donald case, and can thus not
be directly compared to the Kleibergen-Paap statistic. There are no corresponding critical values
for weak identification in the presence of clustered standard errors yet (Schaffer [2012]), which is
why I report the Stock-Yogo values for orientation.
In the present specification, I cannot test for overidentification, since the model is exactly
identified. When trying sets of three or more instruments, namely by including more geographical
variables as proxies for coca cultivation, problems of weak identification occurred, but the null
hypothesis that the instruments are valid could not be rejected. This may be interpreted as cautious
support for the assumption that distance to Ayacucho and temperature are valid instruments in
the present specification as well.
18
As suggested by Cameron and Trivedi [2005], I report the cluster-robust variant of the DurbinWu-Hausman test provided by the endog option of the ivreg2 command in Stata to test the endogeneity of conflict and its interaction with distance to market. The null hypothesis that these
regressors can be treated as exogenous cannot be rejected. The control variables for education
and the economic situation may be sufficient to avoid omitted variable bias in the simple logit. In
this case, it is neither necessary nor optimal to use an IV strategy in the present analysis. The
non-significance of the 2SLS results is likely to be caused by the loss of precision inherent to the instrumental variables procedure, which is aggravated in the case of weak instruments.(Cameron and
Trivedi [2005]) Again, these results need to be interpreted with caution. The Durbin-Wu-Hausman
test compares the coefficients resulting from OLS and 2SLS. Thus, the null hypothesis which is
not rejected is that the coefficients are equal. However, in the presence of weak instruments, IV
estimates are biased in the same direction as OLS in finite samples, as highlighted by Bound et al.
[1995]. The result of the Durbin-Wu-Hausman test could also be driven by weak instruments, and
allow no definite conclusion about the exogeneity of the explanatory variables.
7
Robustness Checks
The present section discusses the results of a variety of robustness checks, which I conduct to
corroborate my results and to gain more detailed insights in the causal mechanisms underlying the
relation between distance to market, conflict, and the occupation decision. As the necessity of an
IV procedure is not evident, I use the baseline logit model including all sets of control variables for
the robustness checks. Tables reporting the respective results are included in the appendix.
7.1
Different Conflict Periods
In order to verify whether contemporaneous conflict is indeed more relevant for the occupation
decision than past conflict, I replace the dummy for conflict between 1990 and 1994 with dummies
for earlier stages of the conflict: 1980-84 and 1985-89. The results are similar for all periods, but
the coefficient on the interaction term between conflict and distance to market is larger for the
earlier periods. The negative impact of distance to market seems to be larger in those districts
which were affected by the conflict in its early stages. It is important to note that most of these
districts still experienced violence in the 1990-94 period, since the conflict expanded gradually and
attained its largest geographical coverage in the early 1990s. Of the 80 districts in my dataset,
49 never experienced conflict. Only 9 districts experienced conflict in 1980-84, compared to 23 in
1985-89 and 29 in 1990-94. Only 2 districts which were affected in the 1980s did not have any fatal
victim in 1990-94. The significance of the results on the earlier stages of the conflict can thus partly
be explained by the considerable overlap between districts affected in the different periods, which
makes it difficult to conclude on the relative importance of past and contemporaneous conflict in
the occupation decision.
Nonetheless, the unexpected larger magnitude of the coefficients on earlier periods raises the
question whether the negative effect of distance to market observed in conflict districts could be
19
caused by other factors than the violence. The districts around the epicenter of the conflict, which
were affected in its early stages, are culturally and geographically similar. With the expansion of
the conflict, the characteristics of the conflict districts became more diverse. To verify whether
the coefficient of interest is not merely driven by regional specificities of the core conflict districts,
which could explain why its magnitude declines with the geographic expansion, I implement a
further robustness check. I run the baseline logit regression with all control variables, however
replacing the conflict dummy with distance to Ayacucho, and the interaction term with distance to
Ayacucho multiplied by distance to market. In a first specification, I use the subsample of districts
never affected by the conflict, in a second one the subsample of districts not affected in 1990-94,
and in a third one the full sample.
If the coefficient on the interaction term between distance to Ayacucho and distance to market
were positive and significant, this would entail that the negative effect of distance to market on the
likelihood of being a vendor is a general characteristic of the region around Ayacucho, independently
from the conflict. However, the coefficient is negative and significant in the first two specifications
(subsamples of non-conflict districts). This means that the effect of distance to market, which
on its own is positive in the non-conflict samples, decreases with distance to Ayacucho. When
including both conflict and non-conflict districts, the coefficient on the interaction between distance
to Ayacucho and distance to market is insignificant. Therefore, the negative effect of distance to
market in conflict districts does not seem to be driven by unobserved regional specificities of the
districts at the epicenter of the conflict. The larger magnitude of the coefficient on the interaction
of distance to market with dummies of conflict in 1980-84 and 1985-89 raises the question whether
the effects of the disruption of social and economic life due to the conflict are more persistent than
initially expected. This question cannot be finally answered in the scope of this paper, but points
towards an interesting direction for future research.
7.2
Conflict Intensity
A competing explanation for the larger coefficient on distance to market in districts with conflict in
the 1980s is that conflict intensity in 1990-94 was considerably higher in districts affected over the
entire conflict period than in those affected only in the early 1990s. In the districts affected in both
the early and late stages, the mean number of victims over the 1990-94 period is 27, compared to
10 in the districts only affected in the 1990s.
To explore the effect of conflict intensity, I replace the conflict dummy in the baseline logit
regression by the number of victims per district between 1990 and 1994. In further specifications,
I use dummies for different conflict intensities, namely for more than 10, 20 and 50 victims per
district between 1990 and 1994. As pointed out by León [2012], the conflict data collected by the
CVR is likely to suffer from measurement error, given that it is self-reported and collected in public
hearings. The share of victims caused by different perpetrators (insurgent groups, state agents
and others) differs considerably between the CVR database and those of other institutions, such as
human rights organizations and ministries (Comisión de la Verdad y Reconciliación [2003]). This
might be caused by selective reporting of cases according to the mandate of each organization or
its perception by the public. Therefore, the results on conflict intensity need to be interpreted with
caution.
20
The coefficient on the interaction term between conflict intensity and distance to market is
negative and significant. The same is true for the interactions between different conflict intensity
dummies and distance to market. For these, the magnitude of the coefficient increases with the
lower bound of the number of victims. Thus, the negative impact of distance to market in conflict
districts seems to increase with conflict intensity. As the districts which experienced higher levels
of violence are also those which were affected in the early stages of the conflict, it is difficult to
disentangle the aspects of persistence and intensity.
7.3
Other Non-Agricultural Self-Employment
Non-agricultural self-employment is generally more prevalent in the conflict districts, as shown
in the summary statistics. If the differential impact of distance to market applied not only to
shopkeepers and vendors, but also to other self-employment activities, this would cast doubt on the
hypothesis that the transmission channel is road insecurity and the increase in transport cost. To
verify this, I run the baseline logit regression on a dependent variable which takes the value 1 if a
person is self-employed in a non-agricultural activity, but not a shopkeeper or vendor. In a second
specification, I use a binary dependent variable for all types of non-agricultural self-employment
(i.e., including shopkeepers and vendors). In both cases, the coefficient on the interaction term
between conflict and distance to market is insignificant. This result indicates that the negative
impact of distance to market is indeed proper to vendors and shopkeepers. It can be interpreted as
support for the road insecurity hypothesis, since transport costs are likely to affect these professions
in particular.
7.4
Outliers and Market Villages
As mentioned in section 4, 5 out of the 143 villages in my sample lie very far away from the next
market (between 48 and 84 km, compared to a mean of 6.3 km in the rest of the sample). I suspect
that these values might be caused by errors, for example misunderstanding concerning the definition
of a “market“ or “fair“. Unfortunately, I cannot verify their accuracy, since the LSMS data does
not provide village names. While I consider it justifiable to exclude the outliers from my analysis,
I report the results of the baseline logit on the full sample for academic transparency. The main
result is not robust to the inclusion of the outliers. The coefficient on the interaction term between
conflict and distance to market decreases in magnitude and becomes insignificant. The outliers
exert considerable leverage on the results.
Selective migration is an important concern when studying the impact of conflict on the occupation choice, as mentioned in section 5.3. If persons desiring to work as vendors migrate selectively
to villages close to markets, and the same “type“ of person is also more likely to migrate to escape
conflict, this can bias the result of interest. I conduct robustness checks using a sample excluding
market villages (i.e., for which distance to the next market equals zero), and a sample excluding
markets which are at most 2 km from the next market. The interaction term remains negative,
significant, and of similar magnitude as with the usual sample, which corroborates my main result.
21
8
Conclusion
In sum, this paper examined road insecurity as a transmission channel of conflict, and illustrated
how it alters individual occupation decisions at the example of self-employed shopkeepers and
vendors in rural areas of Peru. The initial hypothesis was that conflict-induced road insecurity
raises the cost of transporting goods to a village, and thus prevents individuals from working as
shopkeepers or vendors if they live relatively far from the next market.
In line with this hypothesis, the results of the logit regressions display a negative relation between
distance to market and the decision to become a self-employed shopkeeper or vendor, but only in
districts affected by conflict. This result is however not robust to the 2SLS specification, which
was proposed to alleviate concerns about omitted variable bias. The coefficient on the interaction
term between conflict and distance to market becomes insignificant, potentially due to the relative
inefficiency of the estimator. The endogeneity test results, however, suggest that conflict and its
interaction with distance to market can be treated as exogenous in the present specification, in
which case the results of the logit model are unbiased. Together with the robustness checks, which
show, among others, that the result is neither driven by the geographic and cultural similarity of the
districts close to the conflict’s epicenter nor by general effects on self-employment, these results can
be viewed as evidence for the existence of a road insecurity channel through which conflict affects
rural livelihoods. There seems to be a “transport cost“ imposed on shopkeepers by conflict-induced
road insecurity, although, in the absence of data on assaults, it is not possible to say whether the
fear of loosing one’s life or the risk of loosing one’s goods in an assault was more important in the
Peruvian context.
Given that non-agricultural income often represents more than 50% of rural households’ budgets (Davis et al. [2010]), the impact of conflict on the occupation decision can have important
implications for poverty. Understanding its underlying mechanisms is thus relevant for development policy. As revealed by the robustness checks, further research is necessary to investigate the
potential persistence of the effects of conflict-induced road insecurity. If there is such a thing as
“learning by market participation“, and if market participation is inhibited due to road insecurity,
conflict regions may suffer a permanent fall-back in terms of productivity. Policies to specifically
tackle this problem could constitute a useful component of post-conflict reconstruction efforts. In
order to provide insights for policy-makers, further research could focus on the relevance of the road
insecurity channel for other activities (such as agriculture or manufacturing in rural areas) and the
persistence of its effects.
From a broader perspective, this paper highlights one particular way in which conflict can
disrupt local economic interaction and affect the daily life of rural populations. By focusing on one
aspect of the micro-economic cost of conflict, it shows how conflict restrains individuals’ economic
opportunities, and may thereby cause persistent adverse effects on their well-being.
22
References
Joshua D. Angrist and Adriana D. Kugler. Rural windfall or a new resource curse? Coca, income
and civil conflict in Colombia. The Review of Economics and Statistics, 90(2):191–215, 2008.
David G. Blanchflower and Andrew J. Oswald. What Makes an Entrepreneur. Journal of Labor
Economics, 16(1):26–60, 1998.
John Bound, David A. Jaeger, and Regina M. Baker. Problems With Instrumental Variable Estimation When the Correlation Between the Instruments and the Endogenous Explanatory Variable
is Weak. Journal of the American Statistical Association, 90(430):443–450, 1995.
Carlos Bozzoli, Tilman Brueck, and Nina Wald. Self-Employment and Conflict in Colombia. DIW
Discussion Papers, 1098, 2011.
A. Colin Cameron and Pravin K. Trivedi. Microeconometrics: Methods and Applications. Cambridge
University Press, New York, 2005.
Comisión de la Verdad y Reconciliación. Informe Final. Comisión de la Verdad y Reconciliación,
Lima, 2003.
Benjamin Davis, Paul Winters, and Gero Carletto. A Cross-Country Comparison of Rural Income
Generating Activities. World Development, 38(1):48–63, 2010.
Carlos Iván Degregori. ¿Por qué apareció Sendero Luminoso en Ayacucho? El desarrollo de la educación y la generación del 69 en Ayacucho y Huanta. In Anne Pérotin-Dumon, editor, Historizar
el pasado vivo en América Latina. 2007. URL http://www.historizarelpasadovivo.cl/.
Klaus Deininger. Causes and Consequences of Civil Strife. Micro-Level Evidence from Uganda.
World Bank Policy Research Working Paper, 3045, 2003.
Klaus Deininger, Songqing Jin, and Mona Sur. Sri Lanka’s Rural Non-Farm Economy: Removing
Constraints to Pro-Poor Growth. World Development, 35(12):2056–2078, 2007.
Ponciano Del Pino. Tiempos de guerra y de dioses. Ronderos, evangélicos y senderistas en el valle
del rı́o Apurı́mac. In Carlos Iván Degregori, José Coronel, Ponciano Del Pino, and Orin Starn,
editors, Las rondas campesinas y la derrota de Sendero Luminoso, pages 114–185. Instituto de
Estudios Peruanos, Lima, 1996.
Distance Calculator Peru. URL http://distancecalculator.globefeed.com/Peru_Distance_
Calculator.asp.
Javier Escobal. The Determinants of Nonfarm Income Diversification in Rural Peru. World Development, 29(3):497–508, 2001.
David S. Evans and Boyan Jovanovich. An Estimated Model of Entrepreneurial Choice under
Liquidity Constraints. Journal of Political Economy, 97(4):808–827, 1989.
Yannis Georgellis, John G. Sessions, and Nikolaos Tsitsianis. Self-Employment Longitudinal Dynamics: A Review of the Literature. Economic Issues, 10:291–296, 2005.
23
William Greene. Testing hypotheses about interaction terms in nonlinear models. Economics
Letters, 107, 2010.
Hans Huerto Amado. Combatir narcotráfico es clave para erradicar a Sendero del VRAE, según
especialistas. El Comercio, February 12, 2012.
Instituto Nacional de Estadı́stica e Informática.
Censos Nacionales IX de Población y
IV de Vivienda, 1993. URL http://iinei.inei.gob.pe/iinei/RedatamCpv1993.asp?id=
ResultadosCensales?ori=C.
Instituto Nacional de Estadı́stica e Informática. Encuesta Nacional de Hogares sobre Medición
de Niveles de Vida (Living Standards Measurement Survey), 1994. URL http://microdata.
worldbank.org/index.php/catalog/617.
Instituto Nacional de Estadı́stica e Informática. Perú. Compendio Estadı́stico 2007. INEI, Lima,
2007.
Otabek Ismatullayev. Erythroxylum coca, 2011. URL http://bioweb.uwlax.edu/bio203/2011/
ismatull_otab/index.htm.
Hanan G. Jacoby. Markets and the Benefits of Rural Roads. The Economic Journal, 110(465):
713–737, 2000.
Songqing Jin and Klaus Deininger. Key Constraints for Rural Non-Farm Activity in Tanzania:
Combining Investment Climate and Household Surveys. Journal of African Economies, 18(2):
319–361, 2008.
Patricia Justino. War and Poverty. Microcon Research Working Paper, 32, 2010.
Adam C Kolasinski and Andrew F. Siegel. On the economic meaning of interaction term coefficients in non-linear binary response regression models, 2010. URL http://ssrn.com/abstract=
1668750orhttp://dx.doi.org/10.2139/ssrn.1668750.
Anh T. Le. Empirical Studies of Self-Employment. Journal of Economic Surveys, 13(4):381–416,
1999.
Gianmarco León. Civil Conflict and Human Capital Accumulation: Long Term Consequences of
Political Violence in Peru. Journal of Human Resources (forthcoming), 2012.
Héctor Luja Peralta and Mario Zapata Tejerina. La educación superior en Perú. CRESALC UNESCO, 1988.
Jorge Morello and Silvia Diana Matteucci. Aspectos ecológicos del cultivo de coca. Encrucijadas
(UBA), 1(8):82–91, 2001.
Edward C. Norton, Hua Wang, and Chunrong Ai. Computing interaction effects and standard
errors in logit and probit models. The Stata Journal, 4(2):154–167, 2004.
Anna L. Paulson and Robert Townsend. Entrepreneurship and financial constraints in Thailand.
Journal of Corporate Finance, 10:229–262, 2004.
24
Christina Paxson and Norbert R. Schady. The Allocation and Impact of Social Funds: Spending
on School Infratstructure in Peru. The World Bank Economic Review, 16(2):297–319, 2002.
Mark E. Schaffer. Statalist entry: st: RE: Interpreting Kleibergen Paap weak instrument statistic,
2012. URL http://www.stata.com/statalist/archive/2012-06/msg01166.html.
Kimberly Theidon. Terror’s Talk: Fieldwork and War. Dialectical Anthropology, 26:19–35, 2001.
25
A
Appendix
26
Table 5: List of Variables
VARIABLE
DEFINITION
LEVEL
Dependent variable
Vendor
Dummy: Self-employed shopkeeper or vendor
Individual
Independent variables
Conflict
Distance to market
Conflict*Distance to market
Dummy: No. of victims between 1990-94 > 1
Distance in km
Interaction term
District
Village
Individual controls
Female
Age
Primary education
Dummy
Age in years
Dummy: Completed primary school
Individual
Individual
Individual
Household controls
Children
Adults
Consumption expenditures
Credit access
Electricity
Numbe of children (0-14) in household
Other adults in household
Yearly, in 100 Peruvian Soles
Dummy: Self reported access to credit
Dummy: Dwelling has electricity
Household
Household
Household
Household
Household
Village and district controls
More households in village
Less households in village
FONCODES
Easier to find job
More difficult to find job
Education level
Dummy: No. of households increased since 1991
Dummy No. of households decreased since 1991
Dummy: FONCODES invested in village
Dummy: Easier to find a job than in 1991
Dummy: More difficult to find a job than in 1991
% of adults (≥15) with secondary education
Village
Village
Village
Village
Village
District
Geographical controls
log(altitude)
Precipitation
log(altitude of district capital in m)
Average yearly precipitation in mm, 1997-2006
District
Department
Instruments for conflict
Distance to Ayacucho
Distance of district capital to Ayacucho, in km
District
Temperature
Average temperature in C◦ , 1997-2006
Department
Sources:
Conflict: Comisión de la Verdad y Reconciliación (CVR)
Individual, household and village-level variables: Living Standards Measurement Survey 1994 (LSMS)
Education level: National Census 1993, INEI
Altitude: Google Earth
Temperature, precipitation: Instituto Nacional de Estadı́stica e Informatı́ca 2007 (INEI)
Distance to Ayacucho: Peru Distance Calculator (PDC)
27
Table 6: Summary statistics - Occupations
VARIABLES
Full sample
Non-agricultural self-employment
Observations
Only self-employed
Shopkeeper or vendor
Tailor, sewer, or other textile professional
Services
Artisan or craftsman
Food production
Technical professional
(1)
Full sample
mean
(sd)
(2)
No conflict
mean
(sd)
(3)
Conflict
mean
(sd)
(4)
Difference
(2)-(3)
0.19
(0.39)
0.17
(0.37)
0.22
(0.42)
-0.06***
2122
1297
825
-
0.54
(0.50)
0.14
(0.35)
0.11
(0.31)
0.11
(0.31)
0.06
(0.22)
0.05
(0.22)
0.54
(0.50)
0.10
(0.31)
0.10
(0.33)
0.11
(0.32)
0.08
(0.27)
0.05
(0.22)
0.54
(0.50)
0.19
(0.39)
0.12
(0.31)
0.09
(0.29)
0.02
(0.13)
0.05
(0.22)
0.01
-0.08**
-0.02
0.03
0.06***
0.00
Observations
412
229
183
Asterisks denote the significance level of the t-test on the difference between conflict and
non-conflict districts: *** p<0.01, ** p<0.05, * p<0.1.
28
Table 7: Summary statistics - All regression variables, by vendor
VARIABLES
Conflict
Distance to market
Female
Age
Primary
Children
Adults
Consumption expenditures
Credit access
Electricity
More households in village
Less households in village
FONCODES
Easier to find job
More difficult to find job
Education level
Altitude
Precipitation
Distance to Ayacucho
Temperature
(1)
Full sample
mean
(sd)
(2)
Non-vendors
mean
(sd)
(3)
Vendors
mean
(sd)
(4)
Difference
(2)-(3)
0.39
(0.49)
6.28
(6.26)
0.42
(0.49)
36.58
(15.85)
0.47
(0.50)
2.34
(1.81)
2.56
(1.76)
47.38
(43.22)
0.15
(0.36)
0.23
(0.42)
0.66
(0.47)
0.06
(0.24)
0.47
(0.50)
0.17
(0.37)
0.59
(0.49)
45.54
(15.15)
1718
(1436)
758.65
(681.38)
591
(293)
16.8
(5.8)
0.38
(0.49)
6.35
(6.30)
0.39
(0.49)
36.47
(16.00)
0.46
(0.50)
2.35
(1.84)
2.60
(1.78)
46.06
(42.94)
0.14
(0.35)
0.22
(0.41)
0.66
(0.47)
0.07
(0.25)
0.46
(0.50)
0.18
(0.38)
0.59
(0.49)
45.44
(15.14)
1738
(1444)
762.57
(677.05)
597
(292)
16.8
(5.7)
0.45
(0.50)
5.60
(5.93)
0.65
(0.48)
37.57
(14.50)
0.63
(0.48)
2.28
(1.61)
2.17
(1.50)
58.71
(44.04)
0.23
(0.42)
0.33
(0.47)
0.67
(0.47)
0.03
(0.16)
0.56
(0.50)
0.11
(0.32)
0.60
(0.49)
46.38
(15.21)
1543
(1350)
724.90
(718.29)
538
(300)
16.4
(6.11)
-0.7**
.75*
-.26***
-1.10
-.18***
.07
.43***
-12.65***
-.09***
-.11***
-.01
.04**
-.10***
.06**
-.01
-.94
195*
37.67
59***
0.4
Observations
2,122
1,901
221
Asterisks denote the significance level of the t-test on the difference between conflict and
non-conflict districts: *** p<0.01, ** p<0.05, * p<0.1.
29
Table 8: Summary statistics - All regression variables, by conflict
VARIABLES
Vendor
Distance to market
Female
Age
Primary education
Children
Adults
Consumption expenditures
Credit access
Electricity
More households in village
Less households in village
FONCODES
Easier to find job
More difficult to find job
Education level
Altitude
Precipitation
Distance to Ayacucho
Temperature
Observations
(1)
Full sample
mean
(sd)
(2)
No conflict
mean
(sd)
(3)
Conflict
mean
(sd)
(4)
Difference
(2)-(3)
0.10
(0.31)
6.28
(6.26)
0.42
(0.49)
36.6
(15.9)
0.47
(0.50)
2.34
(1.81)
2.56
(1.76)
47.38
(43.22)
0.15
(0.36)
0.23
(0.42)
0.66
(0.47)
0.06
(0.24)
0.47
(0.50)
0.17
(0.37)
0.59
(0.49)
45.54
(15.15)
1718
(1436)
758.65
(681.38)
591
(293)
16.8
(5.8)
0.09
(0.29)
5.67
(5.74)
0.40
(0.49)
37.5
(16.2)
0.45
(0.50)
2.16
(1.79)
2.60
(1.79)
47.37
(44.75)
0.15
(0.36)
0.23
(0.42)
0.65
(0.48)
0.06
(0.25)
0.48
(0.50)
0.08
(0.27)
0.67
(0.47)
43.39
(14.41)
1745
(1447)
647.42
(669.27)
667
(254)
16.6
(5.5)
0.12
(0.33)
7.22
(6.90)
0.44
(0.50)
35.2
(15.2)
0.51
(0.50)
2.63
(1.82)
2.49
(1.71)
47.39
(40.73)
0.16
(0.36)
0.22
(0.41)
0.68
(0.46)
0.06
(0.23)
0.46
(0.50)
0.31
(0.46)
0.48
(0.50)
48.91
(15.66)
1677
(1417)
933.52
(663.70)
472
(310)
17
(6.1)
-.03**
2,122
1,297
825
-1.55***
-.04*
2.3***
-.05**
-.47***
.11
-.02
-.01
.01
-.03
.01
.02
-.23***
.19***
-5.52***
68
-286.01***
195***
-0.4
-
Asterisks denote the significance level of the t-test on the difference between conflict and
non-conflict districts: *** p<0.01, ** p<0.05, * p<0.1.
30
Table 9: Summary statistics - All regression variables, by inclusion in fixed effects sample
VARIABLES
Vendor
Conflict
Distance to market
Female
Age
Primary education
Children
Adults
Consumption expenditures
Credit access
Electricity
More households in village
Less households in village
FONCODES
Easier to find job
More difficult to find job
Education level
Altitude
Precipitation
Distance to Ayacucho
Temperature
Observations
(1)
Full sample
mean
(sd)
(2)
Included
mean
(sd)
(3)
Excluded
mean
(sd)
(4)
Difference
(2)-(3)
0.10
(0.31)
0.39
(0.49)
6.28
(6.26)
0.42
(0.49)
36.6
(15.9)
0.47
(0.50)
2.34
(1.81)
2.56
(1.76)
47.38
(43.22)
0.15
(0.36)
0.23
(0.42)
0.66
(0.47)
0.06
(0.24)
0.47
(0.50)
0.17
(0.37)
0.59
(0.49)
45.54
(15.15)
1718
(1436)
758.65
(681.38)
591
(293)
16.8
(5.8)
0.13
(0.33)
0.42
(0.49)
5.99
(5.79)
0.41
(0.49)
36.1
(15.5)
0.48
(0.50)
2.38
(1.74)
2.58
(1.77)
49.64
(46.37)
0.16
(0.37)
0.26
(0.44)
0.69
(0.46)
0.07
(0.26)
0.51
(0.50)
0.16
(0.36)
0.61
(0.49)
45.86
(15.52)
1520
(1391)
760.36
(737.07)
590
(299)
17.18
(5.97)
0
(0)
0.24
(0.43)
7.58
(7.92)
0.44
(0.50)
38.9
(17.1)
0.43
(0.50)
2.18
(2.10)
2.48
(1.73)
37.24
(21.84)
0.10
(0.30)
0.08
(0.27)
0.57
(0.50)
0.02
(0.14)
0.28
(0.45)
0.22
(0.41)
0.51
(0.50)
45.54
(13.25)
2608
(1288)
750.98
(717.75)
594
(265)
15.00
(4.23)
0.13***
2,122
1735
387
0.18***
-1.59***
-0.03
-2.8***
0.06**
0.20*
0.10
12.40***
0.07***
0.18***
0.12***
0.05***
0.23***
-0.06***
0.10***
1.75**
-1089***
9.38
-3.38
2.18***
-
Asterisks denote the significance level of the t-test on the difference between conflict and
non-conflict districts: *** p<0.01, ** p<0.05, * p<0.1.
31
Table 10: OLS with interaction terms - all coefficients reported
VARIABLES
Conflict
(1)
(2)
0.078**
(0.034)
0.002
(0.003)
-0.008**
(0.004)
0.097***
(0.015)
0.004**
(0.002)
-0.000*
(0.000)
0.061***
(0.016)
-
Dependent variable: Vendor
(3)
(4)
(5)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
0.077**
(0.034)
0.003
(0.003)
-0.008**
(0.004)
0.095***
(0.015)
0.004**
(0.002)
-0.000**
(0.000)
0.043**
(0.017)
-0.003
(0.004)
-0.011**
(0.005)
0.001*
(0.000)
0.036
(0.029)
0.044*
(0.025)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Education level
-
-
Log(altitude)
-
-
-0.009
(0.026)
-0.067*
(0.034)
0.032
(0.022)
-0.056*
(0.034)
-0.027
(0.025)
0.000
(0.001)
-
Precipitation
-
-
-
-0.077**
(0.035)
-0.073*
(0.037)
-0.074*
(0.044)
Distance to market
Conflict*Distance to market
Female
Age
Age2
Primary education
Children
Constant
0.083**
(0.033)
0.003
(0.003)
-0.008**
(0.004)
0.098***
(0.015)
0.004**
(0.002)
-0.000*
(0.000)
0.056***
(0.015)
-
0.082**
(0.033)
0.003
(0.003)
-0.009**
(0.004)
0.096***
(0.015)
0.004**
(0.002)
-0.000**
(0.000)
0.042**
(0.017)
-0.003
(0.004)
-0.011**
(0.005)
0.001
(0.000)
0.035
(0.030)
0.044*
(0.025)
-0.005
(0.006)
-0.000
(0.000)
-0.038
(0.056)
0.082**
(0.032)
0.003
(0.003)
-0.008**
(0.004)
0.096***
(0.015)
0.004**
(0.002)
-0.000**
(0.000)
0.038**
(0.017)
-0.005
(0.004)
-0.012**
(0.005)
0.001*
(0.000)
0.035
(0.029)
0.054**
(0.026)
-0.018
(0.025)
-0.079**
(0.037)
0.027
(0.022)
-0.060*
(0.031)
-0.037
(0.025)
-0.000
(0.001)
-0.029
(0.052)
(6)
0.087***
(0.032)
0.003
(0.003)
-0.008**
(0.004)
0.097***
(0.015)
0.004**
(0.002)
-0.000**
(0.000)
0.038**
(0.016)
-0.005
(0.004)
-0.012**
(0.005)
0.001
(0.000)
0.033
(0.029)
0.059**
(0.025)
-0.021
(0.025)
-0.080**
(0.035)
0.027
(0.022)
-0.056*
(0.031)
-0.038
(0.025)
-0.001
(0.001)
-0.006
(0.006)
-0.000
(0.000)
0.030
(0.070)
R-squared
0.049
0.065
0.057
0.066
0.074
0.075
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05, * p<0.1.
32
Table 11: OLS with district fixed effects - all coefficients reported
VARIABLES
Distance to market
Dependent variable: Vendor
(2)
(3)
(1)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
0.011**
(0.004)
-0.014***
(0.005)
0.103***
(0.016)
0.005***
(0.002)
-0.000**
(0.000)
0.050***
(0.017)
-0.008
(0.005)
-0.008*
(0.005)
0.001**
(0.000)
0.041
(0.029)
0.029
(0.032)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Yes
-0.018
(0.073)
Conflict*Distance to market
Female
Age
Age2
Primary education
Children
District fixed effects
Constant
0.010**
(0.004)
-0.013**
(0.005)
0.106***
(0.016)
0.004***
(0.002)
-0.000**
(0.000)
0.062***
(0.016)
-
0.011***
(0.004)
-0.012***
(0.005)
0.106***
(0.016)
0.004***
(0.002)
-0.000**
(0.000)
0.060***
(0.015)
-
0.015
(0.027)
-0.095
(0.067)
0.036
(0.023)
-0.003
(0.026)
-0.014
(0.025)
0.011***
(0.004)
-0.013***
(0.004)
0.103***
(0.016)
0.005***
(0.002)
-0.000**
(0.000)
0.048***
(0.017)
-0.008*
(0.005)
-0.008*
(0.005)
0.001**
(0.000)
0.045
(0.030)
0.027
(0.030)
0.013
(0.026)
-0.098
(0.065)
0.038
(0.025)
-0.018
(0.027)
-0.023
(0.026)
Yes
Yes
Yes
-0.045
(0.093)
-0.012
(0.072)
-0.023
(0.084)
-
R-squared
0.141
0.153
0.146
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level.
*** p<0.01, ** p<0.05, * p<0.1.
33
(4)
0.158
Table 12: Logit with interaction terms - all coefficients reported
VARIABLES
Conflict
(1)
(2)
0.854**
(0.342)
0.024
(0.033)
-0.100**
(0.049)
1.081***
(0.153)
0.058**
(0.026)
-0.001*
(0.000)
0.706***
(0.175)
-
Dependent variable: Vendor
(3)
(4)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
0.898***
(0.339)
0.037
(0.033)
-0.102**
(0.048)
1.071***
(0.157)
0.061**
(0.026)
-0.001**
(0.000)
0.501***
(0.194)
-0.035
(0.055)
-0.152**
(0.067)
0.004*
(0.002)
0.357
(0.252)
0.474*
(0.245)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Education level
-
-
Log(altitude)
-
-
-0.105
(0.278)
-1.005**
(0.486)
0.352
(0.244)
-0.590
(0.448)
-0.268
(0.263)
0.002
(0.007)
-
Precipitation
-
-
-
-4.574***
(0.497)
-4.486***
(0.518)
-4.508***
(0.560)
Distance to market
Conflict*Distance to market
Female
Age
Age2
Primary education
Children
Constant
0.877***
(0.329)
0.031
(0.031)
-0.094**
(0.046)
1.091***
(0.154)
0.057**
(0.026)
-0.001*
(0.000)
0.655***
(0.172)
-
0.958***
(0.336)
0.040
(0.034)
-0.108**
(0.048)
1.083***
(0.158)
0.059**
(0.025)
-0.001*
(0.000)
0.489**
(0.190)
-0.032
(0.056)
-0.156**
(0.065)
0.004
(0.002)
0.328
(0.261)
0.503**
(0.246)
-0.062
(0.061)
-0.000
(0.000)
-4.020***
(0.662)
(5)
(6)
0.948***
(0.316)
0.041
(0.031)
-0.093**
(0.043)
1.087***
(0.158)
0.062**
(0.026)
-0.001**
(0.000)
0.465**
(0.192)
-0.051
(0.057)
-0.156**
(0.070)
0.004**
(0.002)
0.346
(0.243)
0.646***
(0.250)
-0.208
(0.270)
-1.267**
(0.534)
0.331
(0.248)
-0.660*
(0.394)
-0.404
(0.262)
-0.004
(0.007)
-
1.011***
(0.313)
0.041
(0.033)
-0.100**
(0.043)
1.101***
(0.158)
0.061**
(0.025)
-0.001**
(0.000)
0.450**
(0.187)
-0.053
(0.056)
-0.162**
(0.068)
0.004*
(0.002)
0.310
(0.248)
0.743***
(0.252)
-0.240
(0.271)
-1.271**
(0.520)
0.320
(0.257)
-0.622
(0.382)
-0.436
(0.266)
-0.008
(0.007)
-0.084
(0.068)
0.000
(0.000)
-3.238***
(0.814)
-4.025***
(0.630)
Pseudo-R2
0.076
0.097
0.088
0.099
0.113
0.115
Log-pseudolikelihood
-655.0
-640.2
-646.4
-639.2
-628.7
-627.3
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05, * p<0.1.
34
Table 13: Logit with district fixed effects - all coefficients reported
VARIABLES
Distance to market
Dependent variable: Vendor
(2)
(3)
(1)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
0.114***
(0.028)
-0.211***
(0.077)
1.340***
(0.179)
0.071***
(0.023)
-0.001***
(0.000)
0.645***
(0.213)
-0.118*
(0.063)
-0.131*
(0.067)
0.007***
(0.003)
0.454
(0.284)
0.243
(0.345)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Yes
-4.094***
(0.747)
Conflict*Distance to market
Female
Age
Age2
Primary education
Children
District fixed effects
Constant
0.104***
(0.031)
-0.185**
(0.076)
1.331***
(0.176)
0.064***
(0.022)
-0.001**
(0.000)
0.811***
(0.193)
-
0.107***
(0.033)
-0.154*
(0.082)
1.331***
(0.176)
0.064***
(0.022)
-0.001**
(0.000)
0.798***
(0.195)
-
(4)
0.255
(0.288)
-1.335
(0.828)
0.430
(0.281)
0.039
(0.474)
-0.408
(0.260)
0.117***
(0.030)
-0.166**
(0.077)
1.342***
(0.179)
0.071***
(0.024)
-0.001**
(0.000)
0.611***
(0.215)
-0.121*
(0.066)
-0.139**
(0.068)
0.008***
(0.003)
0.501*
(0.295)
0.288
(0.335)
0.249
(0.284)
-1.539*
(0.886)
0.475
(0.320)
-0.283
(0.486)
-0.548*
(0.281)
Yes
Yes
Yes
-4.377***
(1.068)
-4.221***
(0.843)
-4.121***
(0.989)
-
Pseudo-R2
0.174
0.195
0.182
0.205
Log-pseudolikelihood
-546.8
-533.0
-541.0
-525.7
Number of observations: 1,735. 19 districts were omitted, as there were no vendors
among the observations in these, such that the district dummy predicted failure perfectly.
Robust standard errors in parentheses, clustered at village-level.
*** p<0.01, ** p<0.05, * p<0.1.
35
Table 14: Cloglog with interaction terms - all coefficients reported
VARIABLES
Conflict
(1)
(2)
0.778**
(0.315)
0.022
(0.031)
-0.092**
(0.047)
1.006***
(0.146)
0.052**
(0.026)
-0.001
(0.000)
0.658***
(0.164)
-
Dependent variable: Vendor
(3)
(4)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
0.827***
(0.312)
0.034
(0.030)
-0.094**
(0.046)
0.984***
(0.147)
0.055**
(0.025)
-0.001*
(0.000)
0.465**
(0.183)
-0.028
(0.049)
-0.137**
(0.063)
0.003**
(0.001)
0.339
(0.229)
0.431*
(0.225)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Education level
-
-
Log(altitude)
-
-
-0.090
(0.257)
-0.945**
(0.457)
0.331
(0.229)
-0.532
(0.428)
-0.234
(0.240)
0.002
(0.006)
-
Precipitation
-
-
-
-4.436***
(0.477)
-4.312***
(0.491)
-4.393***
(0.529)
Distance to market
Conflict*Distance to market
Female
Age
Age2
Primary education
Children
Constant
0.787***
(0.301)
0.029
(0.029)
-0.086**
(0.044)
1.010***
(0.147)
0.051**
(0.025)
-0.001
(0.000)
0.608***
(0.161)
-
0.870***
(0.306)
0.036
(0.031)
-0.099**
(0.045)
0.994***
(0.148)
0.053**
(0.025)
-0.001*
(0.000)
0.451**
(0.179)
-0.026
(0.051)
-0.142**
(0.062)
0.002*
(0.001)
0.308
(0.238)
0.463**
(0.227)
-0.054
(0.055)
-0.000
(0.000)
-3.906***
(0.613)
(5)
(6)
0.870***
(0.289)
0.037
(0.029)
-0.084**
(0.041)
0.992***
(0.148)
0.055**
(0.025)
-0.001*
(0.000)
0.430**
(0.178)
-0.045
(0.052)
-0.143**
(0.066)
0.003**
(0.001)
0.325
(0.218)
0.598***
(0.230)
-0.197
(0.253)
-1.198**
(0.496)
0.325
(0.229)
-0.581
(0.382)
-0.367
(0.240)
-0.004
(0.006)
-
0.918***
(0.283)
0.037
(0.030)
-0.090**
(0.040)
1.003***
(0.148)
0.054**
(0.024)
-0.001*
(0.000)
0.413**
(0.174)
-0.046
(0.051)
-0.149**
(0.065)
0.003**
(0.001)
0.292
(0.224)
0.686***
(0.234)
-0.221
(0.254)
-1.198**
(0.488)
0.311
(0.237)
-0.550
(0.370)
-0.399
(0.245)
-0.007
(0.007)
-0.071
(0.061)
0.000
(0.000)
-3.217***
(0.762)
-3.892***
(0.589)
Log-pseudolikelihood
-655.4
-641.3
-646.8
-640.4
-629.6
-628.3
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05, * p<0.1.
36
Table 15: Cloglog with district fixed effects - all coefficients reported
VARIABLES
Distance to market
Dependent variable: Vendor
(2)
(3)
(1)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
0.105***
(0.021)
-0.195***
(0.072)
1.184***
(0.164)
0.056***
(0.021)
-0.001**
(0.000)
0.576***
(0.194)
-0.106*
(0.056)
-0.124**
(0.060)
0.006***
(0.002)
0.434*
(0.238)
0.251
(0.311)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Yes
-3.838***
(0.676)
Conflict*Distance to market
Female
Age
Age2
Primary education
Children
District fixed effects
Constant
0.095***
(0.025)
-0.168**
(0.069)
1.165***
(0.163)
0.051**
(0.021)
-0.000*
(0.000)
0.721***
(0.178)
-
0.091***
(0.029)
-0.134*
(0.075)
1.162***
(0.162)
0.051**
(0.021)
-0.000*
(0.000)
0.719***
(0.181)
-
(4)
0.214
(0.263)
-1.288*
(0.760)
0.428*
(0.247)
0.093
(0.421)
-0.339
(0.231)
0.101***
(0.025)
-0.141**
(0.071)
1.185***
(0.164)
0.056***
(0.021)
-0.001**
(0.000)
0.549***
(0.195)
-0.110*
(0.057)
-0.136**
(0.061)
0.007***
(0.002)
0.478*
(0.251)
0.289
(0.300)
0.228
(0.256)
-1.552*
(0.839)
0.489*
(0.287)
-0.276
(0.455)
-0.472*
(0.251)
Yes
Yes
Yes
-4.232***
(0.982)
-4.012***
(0.748)
-3.875***
(0.891)
-
Log-Pseudolikelihood
-548.1
-532.7
-541.7
-524.1
Number of observations: 1,735. 19 districts were omitted, as there were no vendors
among the observations in these, such that the district dummy predicted failure perfectly.
Robust standard errors in parentheses, clustered at village-level.
*** p<0.01, ** p<0.05, * p<0.1.
37
Table 16: 2SLS: First stage on conflict (all coefficients)
VARIABLES
Distance to Ayacucho in 100km
(1)
(2)
-0.079***
(0.017)
0.021**
(0.009)
0.011*
(0.006)
0.019
(0.015)
-0.002
(0.003)
-0.000
(0.000)
0.014
(0.037)
-
Dependent variable: Conflict
(3)
(4)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
-0.082***
(0.017)
0.021**
(0.008)
0.011*
(0.006)
0.017
(0.015)
-0.004
(0.003)
0.000
(0.000)
0.026
(0.036)
0.029***
(0.010)
0.006
(0.014)
-0.000
(0.000)
0.045
(0.067)
-0.074
(0.098)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Education level
-
-
Log(altitude)
-
-
0.071
(0.100)
-0.069
(0.166)
0.017
(0.078)
0.214*
(0.123)
-0.048
(0.096)
0.001
(0.003)
-
Precipitation in 100mm
-
-
-
0.477***
(0.147)
0.456***
(0.147)
0.398**
(0.190)
Temperature
Distance to market
Female
Age
Age2
Primary education
Children
Constant
-0.068***
(0.018)
0.016*
(0.009)
0.009
(0.007)
0.017
(0.013)
-0.002
(0.003)
0.000
(0.000)
0.026
(0.034)
-
-0.082***
(0.016)
0.020**
(0.010)
0.007
(0.007)
0.005
(0.014)
-0.004
(0.003)
0.000
(0.000)
0.016
(0.037)
0.021**
(0.010)
0.004
(0.014)
-0.000
(0.000)
0.045
(0.064)
-0.063
(0.094)
0.010
(0.030)
0.013*
(0.008)
0.347
(0.362)
(5)
(6)
-0.070***
(0.018)
0.016*
(0.009)
0.009
(0.007)
0.013
(0.013)
-0.004*
(0.003)
0.000
(0.000)
0.032
(0.035)
0.030***
(0.010)
0.002
(0.013)
-0.000
(0.000)
0.045
(0.063)
-0.065
(0.096)
0.087
(0.096)
-0.045
(0.165)
0.016
(0.078)
0.210*
(0.123)
-0.034
(0.096)
0.002
(0.003)
-
-0.070***
(0.018)
0.016
(0.010)
0.005
(0.008)
0.002
(0.012)
-0.004
(0.002)
0.000
(0.000)
0.022
(0.035)
0.022**
(0.010)
-0.001
(0.013)
-0.000
(0.000)
0.049
(0.061)
-0.052
(0.091)
0.081
(0.093)
-0.071
(0.167)
-0.001
(0.077)
0.168
(0.126)
-0.062
(0.094)
0.002
(0.003)
0.015
(0.033)
0.012*
(0.007)
0.199
(0.436)
0.335*
(0.188)
R2
0.184
0.202
0.224
0.229
0.241
0.266
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05, * p<0.1.
38
Table 17: 2SLS: First stage on conflict*distance to market (all coefficients)
VARIABLES
Distance to Ayacucho in 100km
(1)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
-0.610***
(0.178)
0.347***
(0.101)
0.517***
(0.116)
0.226
(0.146)
-0.018
(0.025)
0.000
(0.000)
0.012
(0.355)
0.151
(0.097)
-0.012
(0.100)
-0.004
(0.003)
0.016
(0.444)
-0.292
(0.707)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Education level
-
-
Log(altitude)
-
-
1.042
(0.848)
-1.008
(1.508)
0.091
(0.634)
2.540*
(1.325)
-0.270
(0.883)
0.025
(0.017)
-
Precipitation in 100mm
-
-
-
-2.512*
(1.423)
-2.351
(1.459)
-4.246**
(1.783)
Temperature
Distance to market
Female
Age
Age2
Primary education
Children
Constant
-0.596***
(0.172)
0.343***
(0.101)
0.524***
(0.114)
0.246
(0.153)
-0.003
(0.026)
-0.000
(0.000)
-0.028
(0.353)
-
Dependent variable: Conflict*Distance to market
(2)
(3)
(4)
(5)
-0.451***
(0.175)
0.274***
(0.092)
0.501***
(0.112)
0.217
(0.143)
-0.002
(0.024)
-0.000
(0.000)
0.082
(0.328)
-
-0.608***
(0.176)
0.359***
(0.103)
0.499***
(0.123)
0.162
(0.131)
-0.014
(0.025)
0.000
(0.000)
-0.022
(0.371)
0.115
(0.099)
-0.022
(0.101)
-0.002
(0.003)
0.057
(0.441)
-0.234
(0.688)
0.142
(0.193)
0.056
(0.085)
-3.828
(2.472)
-0.450**
(0.175)
0.273***
(0.090)
0.495***
(0.112)
0.182
(0.140)
-0.021
(0.023)
0.000
(0.000)
0.069
(0.335)
0.173*
(0.099)
-0.067
(0.086)
-0.005*
(0.003)
-0.006
(0.433)
-0.265
(0.702)
1.141
(0.810)
-0.958
(1.507)
0.059
(0.640)
2.560**
(1.293)
-0.189
(0.911)
0.030*
(0.017)
-4.160**
(1.847)
(6)
-0.429**
(0.176)
0.304***
(0.093)
0.484***
(0.120)
0.113
(0.129)
-0.016
(0.023)
0.000
(0.000)
0.036
(0.343)
0.148
(0.094)
-0.083
(0.085)
-0.003
(0.003)
0.113
(0.436)
-0.295
(0.678)
1.166
(0.803)
-1.045
(1.498)
-0.017
(0.617)
2.352*
(1.259)
-0.229
(0.862)
0.040**
(0.018)
0.283
(0.218)
0.043
(0.079)
-7.365**
(3.137)
R2
0.184
0.202
0.224
0.229
0.241
0.266
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05, * p<0.1.
39
Table 18: 2SLS: Second stage (all coefficients)
VARIABLES
Conflict
(1)
(2)
0.10
(0.111)
-0.01
(0.013)
0.00
(0.007)
0.10***
(0.015)
0.00**
(0.002)
-0.00*
(0.000)
0.06***
(0.016)
-
Dependent variable: Vendor
(3)
(4)
(5)
Adults
-
Consumption expenditures
-
Credit access
-
Electricity
-
More households in village
-
0.07
(0.110)
-0.01
(0.012)
0.00
(0.007)
0.09***
(0.015)
0.00**
(0.002)
-0.00**
(0.000)
0.04**
(0.017)
-0.00
(0.005)
-0.01**
(0.005)
0.00*
(0.000)
0.04
(0.030)
0.04*
(0.026)
-
Less households in village
-
-
FONCODES
-
-
Easier to find job
-
-
More difficult to find job
-
-
Education level
-
-
Log(altitude)
-
-
-0.02
(0.030)
-0.07
(0.047)
0.03
(0.024)
-0.09**
(0.042)
-0.02
(0.025)
-0.00
(0.001)
-
Precipitation
-
-
-
-0.08
(0.063)
-0.07
(0.059)
-0.06
(0.083)
Conflict*Distance to market
Distance to market
Female
Age
Age2
Primary education
Children
Constant
0.16
(0.121)
-0.00
(0.014)
0.00
(0.007)
0.09***
(0.015)
0.00**
(0.002)
-0.00*
(0.000)
0.05***
(0.016)
-
0.10
(0.121)
-0.01
(0.014)
0.00
(0.007)
0.10***
(0.015)
0.00**
(0.002)
-0.00**
(0.000)
0.04**
(0.016)
-0.00
(0.005)
-0.01**
(0.005)
0.00
(0.000)
0.03
(0.031)
0.04*
(0.026)
-0.01
(0.007)
-0.00
(0.000)
-0.04
(0.058)
0.12
(0.118)
-0.00
(0.013)
0.00
(0.007)
0.09***
(0.015)
0.00**
(0.002)
-0.00**
(0.000)
0.04**
(0.017)
-0.01
(0.005)
-0.01**
(0.005)
0.00*
(0.000)
0.03
(0.029)
0.06**
(0.026)
-0.03
(0.028)
-0.08*
(0.046)
0.03
(0.023)
-0.08**
(0.038)
-0.03
(0.024)
-0.00
(0.001)
-0.01
(0.077)
Adjusted R2
0.040
0.058
0.030
0.059
0.056
Underidentification
P-value
0.0002
0.0002
0.0026
0.0009
0.0025
Weak identification
Kleibergen-Paap rk Wald F-stat
7.81
8.49
5.05
7.23
5.26
Endogeneity test
P-value
0.56
0.85
0.28
0.92
0.46
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level. *** p<0.01, ** p<0.05,
Stock-Yogo critical values: 7.03 (10%), 4.58 (15%), 3.95 (20%).
(6)
0.18
(0.119)
-0.01
(0.014)
0.00
(0.007)
0.10***
(0.015)
0.00***
(0.002)
-0.00**
(0.000)
0.04**
(0.016)
-0.01
(0.005)
-0.01**
(0.005)
0.00*
(0.000)
0.03
(0.030)
0.06**
(0.026)
-0.03
(0.029)
-0.08*
(0.043)
0.03
(0.024)
-0.07*
(0.038)
-0.03
(0.025)
-0.00
(0.001)
-0.01
(0.008)
-0.00
(0.000)
0.05
(0.083)
0.053
0.0005
8.59
0.46
* p<0.1.
Table 19: Robustness check 1a: Different Conflict periods
Dependent variable: Vendor
VARIABLES
Distance to market
Conflict 1980-84
Conflict 80-84*Dist. market
Conflict 1985-89
(1)
(2)
(3)
(4)
(5)
0.018
(0.027)
1.800***
(0.338)
-0.173***
(0.062)
-
0.034
(0.027)
-
0.041
(0.033)
-
0.044
(0.033)
-
-
-
0.046
(0.033)
1.159***
(0.428)
-0.094
(0.065)
1.012**
(0.422)
-0.097*
(0.052)
-0.092
(0.396)
-0.022
(0.042)
-
-
Conflict 85-89*Dist. market
-
Conflict 1990-94
-
1.420***
(0.332)
-0.147***
(0.052)
-
Conflict 90-94 *Distance to market
-
-
Conflict 1980-94
-
-
1.011***
(0.313)
-0.100**
(0.043)
-
Conflict 80-94 *Distance to market
-
-
-
-
Individual controls
Household controls
Village and district controls
Geographical controls
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
0.134
-614.3
0.118
-625.2
-
Pseud-R2
0.124
0.125
0.115
Log pseudo-likelihood
-620.9
-620.5
-627.3
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level.
*** p<0.01, ** p<0.05, * p<0.1.
41
1.119***
(0.302)
-0.108**
(0.042)
Table 20: Robustness check 1b: Distance to Ayacucho
Sample:
VARIABLES
Distance to market
Distance to Ayacucho
Distance to Ayacucho*Distance to market
Individual controls
Household controls
Village and district controls
Geographical controls
No conflict
1980-94
(1)
No conflict
1990-94
(2)
Usual
sample
(3)
0.275***
(0.062)
0.002*
(0.001)
-0.000***
(0.000)
0.247***
(0.056)
0.001*
(0.001)
-0.000***
(0.000)
0.007
(0.063)
-0.001
(0.001)
-0.000
(0.000)
X
X
X
X
X
X
X
X
X
X
X
X
Observations
1,244
1,297
Pseudo-R2
0.161
0.135
Lof pseudo-likelihood
-311.8
-348.0
Robust standard errors in parentheses, clustered at village-level.
*** p<0.01, ** p<0.05, * p<0.1
42
2,122
0.106
-633.6
Table 21: Robustness check 2: Different Conflict Intensities
VARIABLES
Distance to market
Conflict
Conflict*Distance to market
Conflict intensity
Dependent variable: Vendor
(2)
(3)
(4)
(1)
0.041
(0.033)
1.011***
(0.313)
-0.100**
(0.043)
-
(5)
0.019
(0.028)
-
0.015
(0.027)
-
0.015
(0.027)
-
0.014
(0.026)
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Conflict intensity*Distance to market
-
Conflict>10
-
0.033***
(0.006)
-0.003**
(0.001)
-
Conflict>10*Dist. market
-
-
Conflict>20
-
-
1.171***
(0.369)
-0.095*
(0.049)
-
Conflict>20*Dist. market
-
-
-
Conflict>50
-
-
-
2.114***
(0.441)
-0.141**
(0.064)
-
Conflict>50*Dist. market
-
-
-
-
Individual controls
Household controls
Village and district controls
Geographical controls
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
0.127
-619.3
0.130
-616.6
Pseudo-R2
0.115
0.131
0.115
Log pseudo-likelihood
-627.3
-616.0
-627.7
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level.
*** p<0.01, ** p<0.05, * p<0.1.
43
2.437***
(0.391)
-0.193**
(0.083)
Table 22: Robustness check 3: Other Non-Agricultural Self-Employment
Dependent variable:
VARIABLES
Conflict
Distance to market
Conflict*Distance to market
Individual controls
Household controls
Village and district controls
Geographical controls
Pseudo-R2
Log Pseudo-likelihood
Vendor
(1)
Other non-agricultural
self-employment
(2)
All non-agricultural
self-employment
(3)
1.011***
(0.313)
0.041
(0.033)
-0.100**
(0.043)
-0.193
(0.377)
0.010
(0.030)
0.033
(0.035)
0.513*
(0.289)
0.030
(0.024)
-0.036
(0.030)
X
X
X
X
X
X
X
X
X
X
X
X
0.115
-627.3
0.073
-571.2
0.090
-936.3
Number of observations: 2,122.
Robust standard errors in parentheses, clustered at village-level.
*** p<0.01, ** p<0.05, * p<0.1.
44
Table 23: Robustness check 4: Outliers and market villages
Sample:
VARIABLES
Conflict
Distance to market
Conflict*Distance to market
Individual controls
Household controls
Village and district controls
Geographical
Usual
sample
(1)
Including
outliers
(2)
Excluding
market villages
(3)
Excluding
villages less than
2 km from market
(4)
1.011***
(0.313)
0.041
(0.033)
-0.100**
(0.043)
0.767***
(0.283)
-0.008
(0.010)
-0.055
(0.035)
1.146***
(0.354)
0.039
(0.035)
-0.113**
(0.047)
1.005**
(0.471)
0.050
(0.038)
-0.102**
(0.050)
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
Observations
2,122
2,212
1,951
Pseudo-R2
0.115
0.111
0.110
Log pseudo-likelihood
-627.3
-645.0
-570.2
Robust standard errors in parentheses, clustered at village-level.
*** p<0.01, ** p<0.05, * p<0.1
45
1,551
0.109
-428.9