Policy Research Working Paper 9195 The Relationship between Female Labor Force Participation and Violent Conflicts in South Asia Raymond Robertson Gladys López-Acevedo Matías Morales Poverty and Equity Global Practice March 2020 Policy Research Working Paper 9195 Abstract This paper explores the link between the prevalence of gender labor participation gap. The paper tests the add- violent conflicts and extremely low female labor force ed-worker effect theory—which posits that violence might participation rates in South Asia. The Labor Force Sur- increase female labor force participation as women try to veys from Bangladesh, Sri Lanka, India, and Pakistan are make up for lost household income—and finds mixed evi- merged with the Global Terrorism Database to estimate dence: greater prevalence of attacks may encourage married the relationship between terrorist attacks and female labor women to work more hours, but when the environment supply. Geographical data on exposure to violence are used gets more risky, all women work fewer hours. The paper also to compare administrative units exposed to attacks with finds that violence decreases female labor participation less those not exposed. The analysis finds that one additional where it was already higher and has a progressively greater attack reduces female labor force participation rates by impact on lowering female labor participation where the about 0.008 percentage point, on average. Violence has number of attacks is higher. less impact on male labor participation, thus widening the This paper is a product of the Poverty and Equity Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at gacevedo@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team The Relationship between Female Labor Force Participation and Violent Conflicts in South Asia Raymond Robertson (Texas A&M University and IZA), Gladys López-Acevedo (The World Bank and IZA) & Matías Morales (The World Bank) Keywords: Conict, Terrorism, Female Labor Force Participation, Added-worker eect, South Asia JEL Classication: J21, F51, O53 We thank Hans Timmer (Chief Economist for South Asia) and Maurizio Bussolo (Lead Economist in the Chief Economist Oce for South Asia) for valuable comments. The ndings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development (IBRD) and its aliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. 1 Introduction Countries in the South Asia region exhibit the lowest female labor force participation (FLFP) rates, and violent conicts in the region have persisted for 70 years. This paper studies the relationship between conict-related violence and women's decisions to work. Determinants of FLFP have been extensively studied in developed countries, where it has been either high or increasing over the last 50 years (see Killingsworth and Heckman [1987] for a comprehensive literature review). FLFP rates in South Asia remain lower than in other countries with similar GDP per capita, and these countries have experienced limited progress during the last 16 years. As shown in Figure 1, The Islamic Republic of Iran and Pakistan have the lowest FLFP rates in the region. In India and Sri Lanka FLFP rates keep falling. The literature cites several possible explanations, including: i. Social and cultural norms and attitudes, since women hold prime responsibility for housework and childcare whereas men are expected to become the "breadwinners" (Field and Vyborny [2015]) ii. Information frictions, as women face longer search times and smaller networks (Schaner and Das [2016]) iii. Lack of appropriate human capital1 . iv. Discrimination, as women earn between 50% and 70% of what men earn for the same work (Pande et al. [2016]) v. Safety concerns and mobility barriers, since the presence of women in public spaces is still rare in these countries (Sudarshan [2009]). This paper focuses on how violent conict aects FLFP. More specically, we focus on the role that terrorism plays in shaping women's decisions about work.2 Although there is no single universally accepted denition of terrorism, this paper we denes a "terrorist attack" as any threatened or actual use of illegal force and violence by a non-state actor to attain a political, economic, religious, or social goal through fear, coercion, or "intimidation".(START [2018]). As follows from this denition, terrorism aims to frighten the population and induce a change in behavior. The directional impact of violence on FLFP, positive or negative, is not obvious, but Figure 2 suggests a negative relation between terrorist attacks incidence and FLFP rates across countries on the surface. There are at least two possible reasons for this negative relationship. On the one hand, Hudson and Leidl [2015] argue that societies that build and support institutions that inhibit FLFP tend to be more violent because reducing the rights of women is conducive to conict.This argument suggests that long-run, systematic dierences across countries increase the probability of violence 1 Mitra and Singh [2006] describe the pathological situation that has taken place in India known as the "Kerala Puzzle" where womens' education has improved but its integration to the labor market has remained low. 2 Conicts that started decades ago in South Asia have not came to a denite end. Instead, they have reached an equilibrium in which terrorism is used as a tactic of war by radical groups but there are not two clear-cut sides in conict. 2 in regions where women have fewer, or worse, labor market opportunities. On the other hand, violence increases the costs of working through the perceived risk of violence, and these costs may be especially salient for workers on the margin of the decision to work. Such a relationship would emerge in time-series data where regions with dierent exposure levels to violence would experience dierent FLFP rates accordingly. For example, Khan et al. [2017] nd a short-lived negative eect of terrorism on labor market outcomes for Pakistani women that the authors attribute to fear.3 Using national-level panel data and novel instrumental variables, Berrebi and Ostwald [2016] nd a negative impact of terrorism over female labor force participation. When their cross-country results are updated for the period 1990-2016, however, they prove to be sensitive to the inclusion of explanatory variables in the most parsimonious specications. As Table 1 shows, the negative relation between violence (approximated by the three variables used throughout this paper) and FLFP loses signicance after the inclusion of control variables, which hints the existence of some other factors that drive both violence and FLFP. We argue that analyzing local labor markets as we do alleviates the confounding factors problem by exploiting within-country variations in FLFP and violence. Focusing on the within-country variation in violence, however, may reveal a positive relationship between violence and FLFP. For instance, Gallegos [2012] nds that exposure to conict between 1980 and 2000 increased the probability of women working in Peru as a mechanism to complement household's income. Menon and van der Meulen Rodgers [2015] show how civil war in Nepal aected women's decisions about engaging in work outside the home. Kreibaum and Klasen [2015] estimate the eect of Vietnam's war on FLFP but they use a cohorts approach to nd that the "added worker" eect is positive for people directly aected by the war and smaller for those individuals entering working age after the end of the conict. Although these last two articles depart from our specic focus on terrorism, they are still relevant because they (i) clearly develop the argument of the "additional worker eect" that induces women to participate in the labor force as men are displaced from the household, dead or disabled; (ii) focus on countries with similar work morale, culture, and income level; and (iii) focus at the sub-region level and have similar methodological concerns as we do. This paper focuses on four countries with particularly low FLFP rates. Terrorist attacks have been persistent in these countries after the end of the crudest phase of civil conicts experienced decades ago. Terrorist attacks in these countries are motivated mainly by political and religious reasons that, fortunately for our empirical approach, are not caused by FLFP per se. Indeed, our identication strategy rests on the fact that terrorist attacks are typically focused in specic places within countries and are not a function of an individual woman's decision to work. Due to data limitations, we are not able to provide a formal account of the motivations for each attack. However, we know that the main driver of the India-Bangladesh conict is rooted in India's 1947 geographic partition, which led to creation of an Islamic majority population in Bangladesh and Pakistan. This issue has remained unsolved, and borders have been disputed and subject 3 Khan et al. [2017] use individual level data to capture characteristics of the labor force but use district level data for violence measures. We argue this is awed, as it delivers articially signicant estimates of the eect, and deal with it explicitly in our estimation approach. 3 to armed conicts ever since. In the case of Sri Lanka, terrorist groups such as the Liberation Tigers of Tamil Eelam (LTTE) and Janatha Vimukthi Peramuna (JVP) have remained active after being defeated in the Sri Lankan Civil War (1983 to 2009), motivated by ethnic disputes between a Tamil minority and the Sinhalese majority. As a result, terrorism in SriLanka and India has concentrated in specic areas This paper aims to estimate the relationship between violence and women's decisions to work in four South Asian countries. In doing so, we aim to extend the literature in two directions. First, while other studies use national-level data, we look at within-country dierences in violence using micro-data on employment and (i) number of attacks by administrative division; (ii) number of deaths, and (iii) number of people wounded as proxy measures for violence. Second, we pay particular attention to the distinction between the eect over the extensive and intensive margins; that is, whether women work or do notwork or to not work vis-Ã -vis the number of hours supplied, respectively. As such, our paper extends Menon and van der Meulen Rodgers [2015] and Kreibaum and Klasen [2015] by allowing for the possibility that the added worker hypothesis is at work in the context of terrorism in South Asian countries. In Section 2 we present two simple models that illustrate the relationship between violence and the extensive and intensive margins of female labor supply. Section 3 describes the data use in this paper. Section 4 explains the methods used to estimate the eect of our three violence measures on FLFP. Section 5 describes the results and their implications. Section 6 draws conclusions. 2 Theoretical Considerations Labor supply decisions are distinguished between the extensive margin, the decision to work or not, and the intensive margin, the decision of how much to work. In this section, we describe the theoretic intuition for each concept in turn. 2.1 The extensive margin of labor supply Women decide to work if the market wage, w, is above the sum of their reservation home value, h, and the costs of working (paid child care, transportation, and other factors) that we summarize in c. Therefore, the decision to work is represented by a simple equation: Work in market if w − h − c ≥ 0 (1) Work at home if w − h − c < 0 While the presence of violent conicts may increase the cost c of working in the market, we argue it is dierent from other costs because it introduces risk. As long as women are averse to risk, we can decompose the cost of violent conict into a level component that is included in c and the variance (as a component of risk) that we denote by σ 2 . Thus, we can augment equation (1) by adding in the variance, or the uncertainty caused by conicts. The new equation 4 for the decision for a woman to work becomes: Work in market if w − h − c − σ 2 ≥ 0 (2) Work at home if w − h − c − σ 2 < 0 Importantly, the variance has a non-linear relationship with the probability of a terrorist attack. Specically, it is very low for low and high probabilities of a terrorist attack, as specied by the solid line in Figure 5. Of course, a full specication of risk would consider the probability of attack times the impact of the event. Thus, the risk would be higher if deaths were involved, as represented by the dashed line. Note, however, that risk rises and then falls due to "banality"; or the psychological adjustment workers make when terrorist events become common and therefore aects behavior similarly to very low-probability terrorist events. In other words, people get used to the violence. In the middle of the range, there is much more uncertainty, which itself is a cost. If we allow the variables in (1) to be randomized and then aggregated across individuals, simple comparative static predictions match those suggested in the literature. Specically, the randomized and aggregated version of (1) predicts that, holding all else equal, higher wages (w) will be associated with higher FLFP. In addition, having a working partner (generally a male) will increase h and make FLFP less likely. When a working mate is lost, the reservation wage h falls, making FLFP more likely. This is the added worker eect described in the literature, which could be a result of terrorism (if a mate were a victim of terrorism, for example). Increases in c due to terrorism, such as rising transportation costs, will drive down FLFP linearly. And as discussed above, the nonlinear variance eect may introduce nonlinearity in the results. We evaluate all of these predictions using the data described above using the estimation methodology described in Section 4. 2.2 The intensive margin of labor supply Instead of choosing whether to participate or to not participate in the labor market, women could choose the number of hours they work in a dierent way in the presence of violence. To formally allow for this possibility we follow Ashenfelter [1980] and Serneels [2002]. Consider a representative household that maximizes household utility subject to a shared budget constraint. After maximizing, an individual's labor supply (in number of hours) can be written as a function of her wage (wi ), a time constraint (H i ), labor income from the other household members (wj Hj ), and a xed cost of entering the labor force (fi ).4 Hi = f (wi , H i , wj Hj , fi ) i = 1, .., n; i = j (3) 4 This cost may include the opportunity cost (e.g. reservation wage or value of foregone household work) and safety (terrorism risk). 5 We dene the labor status of the family members by introducing a discrete variable as follows : ui = g (H i , wi Hi ) =1 if Hi = 0 (4) =0 if Hi ≥ 0 Then, we can express the number of hours worked by household member i as a function of her own labor income, household income, labor status of the other members, and a xed cost of working. Hi = f (wi , u1 , ..., un , fi ) (5) Note that this equation only holds whenever the wage income is high enough to cover the xed cost of working. In other words, the number of hours is left-censored, so in the data we observe: Hi = Hi∗ if Hi > 0 (6) =0 if Hi∗ ≤ 0 In this model, the heterogeneity summarized in fi makes the less productive individuals move in and out of the labor force or supply more or fewer hours. In addition, when a shock such as a terrorist attack hits a household, women face two choices: (i) Supply more hours of work to compensate for income lost by the main earner (because he eventually joins one of the bands in conict or dies as a consequence of it, switching uj ), which is known as "the added worker eect" or (ii) Supply fewer hours of work, eventually leaving work for safety reasons (by changing fi ). 3 Data We combine several sources of data, including the Global Terrorism Database and national labor force surveys. Each are described in turn in this section. 3.1 Terrorism in South Asia The Global Terrorism Database (GTD, START [2018]) systematically records all attacks since 1972 around the world. For an incident to be included in the data set three conditions are necessary: i. It must be intentional; ii. It must entail some level of violence or immediate threat of violence and iii. Its perpetrators must be sub-national actors. 6 Additionally, at least two of the following three criteria must hold: a. The act must be aimed at attaining a political, economic, religious, or social goal; b. There must be evidence of an intention to coerce, intimidate, or convey some other message to a larger audience (or audiences) than the immediate victims; c. The action must be outside the context of legitimate warfare activities. The data set contains dierent variables that characterize each attack by type: assassination, bombing, kidnapping, and others; and by weapons used, target, perpetrator, and consequences. Critically for this paper, the availability of the number of killed and people wounded as a result of each attack allows measuring the level of violence and the exact coordinates of the incidents to match them to household location (Districts, States or Zilas, depending on availability for each country). Between 1972 and 2016, Bangladesh suered 1,605 terrorist attacks, India 10,978, and Sri Lanka 2,981 (GTD, 2016). Businesses have been the target in 10 percent of cases in India and 5 percent in both Sri Lanka and Bangladesh. Government institutions were targeted in 20 percent of cases in Bangladesh, 15 percent in India, and 12 percent in Sri Lanka. Citizens and their property - which includes public areas, markets, commercial streets, busy intersections and pedestrian malls - were targeted in 25 percent of cases in Bangladesh, 26 percent in India, and 19 percent in Sri Lanka. Transportation was targeted in approximately 8 percent of attacks in all three countries. The rest of the target types include tourists, utilities, and educational institutions, among others. Interestingly, religious sites are not particularly targeted, making up less than 4 percent of terrorist cases in all three countries. Lastly, military targets are important only in Sri Lanka (25 percent, versus 1 percent in Bangladesh and 7 percent in India). As noted earlier, our identication strategy exploits the geographic variation in terrorist attacks. Figure 3 illustrates the argument using as example Bangladesh and Sri Lanka. In the former, the cumulative number of attacks since 1972 has been scattered throughout the dierent zilas while in the later, terrorist groups have targeted the Sinhalese-populated cities of Colombo and Batticaloa. Therefore, when estimating the eect of the incidence of terrorism on FLFP we compare the administrative units that are not attacked to those attacked after controlling for other relevant explanatory variables. 3.2 Labor Force Surveys FLFP rates and distinct individuals' characteristics are measured using labor force surveys (LFS) that, due to our empirical strategy, we later collapse at dierent administrative divisions depending on the country. Importantly, labor force participation is not measured identically in the dierent country surveys and even within countries through years 7 but we do our best to capture the notion of the willingness of the individual to work during the previous 7 days to the interview. For Bangladesh we have data for 64 zilas for years 2005, 2010 and 2013; for India we have data of 31 states in years 1999, 2004, 2007, 2009 and 2011; for Sri Lanka we have 17 or 25 districts depending on the year between 1992 and 2015 (data for 2005, 2009 and 2010 are not available). We merge the GTD to the LFS for each country using year and administrative division. Thus, we have individuals' characteristics for any given year along with the total number of attacks committed, and people killed and wounded for each district, zila or state. While individual-level data vary substantially, violence-related data do not since these are measured at the administrative division level. Therefore, we collapse the original data set at the administrative unit level to build a panel-like data set with variables appropriately transformed to average age, share of males, and share of workers in the manufacturing sector of each administrative unit every year. Our resulting data set has two shortcomings: (i) there are many observations in earnings variables missing, which is an important weakness since market wages are a determinant of labor force participation5 and (ii) we lack data on the number of hours worked in India, so we drop this country from our estimations involving hours worked. 3.3 Summary Statistics Table 2 contains the summary statistics. We obtain the gures presented here for individuals' (older than 15) characteristics by averaging the multiple geographical areas over time and by country separately. Violence variables also represent the average of the total number of attacks and deaths and people wounded in a given year and area (note that when we do not observe attacks, we impute a "zero" to all three variables). We observe a wide gender gap in labor force participation in all four countries, but the gap between men and women is especially large in Pakistan, where the female LFP rate is only 14 percent and the male rate is 79 percent. As expected, we do not observe much variation in the share of males in each district. Bangladesh's economy is the most agriculturally oriented, with approximately 27 percent of employment in agriculture, as opposed to Pakistan where agriculture represents one-fth of employment. The manufacturing sector share of employment is similar throughout these countries, ranging between 3 and 8 percent of total employment. Regarding our violence measures, we observe that India is the country with the most attacks and wounded and people killed, and Bangladesh has the fewest. Figure 4 shows the that there is substantial variation in FLFP rates across administrative units in dierent countries over time which will be crucial for the empirical strategy described in the following section. Figure 6 shows male and female labor force participation rates in selected administrative units of Sri Lanka, Pakistan, and India that suered a given number of attacks in a year and had no attacks two periods before and two after the violence. In Hambantota, Sri Lanka, it appears that women left the workforce for safety reasons. Labor participation 5 However, we are able to calculate average wages at the administrative unit level needed for some specications. 8 rates for both males and females increased in Bahawalpur, Pakistan and Galle, Sri Lanka, but these rates fell in Haryana, India. 4 Estimation Approaches We face several challenges to meaningfully estimate the eect of violence over FLFP. First, there is a mismatch in the level of variation between our two key variables of interest: incidence of terrorist attacks and individual FLFP. Second, we want to test whether there is a non-linearity in the eect of violence and also if this eect is greater(smaller) in administrative units with higher (lower) FLFP rates for which we need special econometric specications. Finally, we want to understand the eect on the intensive margin of labor supply (that is, hours worked). In this section we justify the use of each estimation strategy to answer dierent parts of the broader question on how violence aects FLFP. 4.1 Pseudo Panel Estimation Our goal is to estimate the eect of violence at the district level over the extensive margin of labor supply at the individual level; that is, does more violence increase or decrease female labor participation. However, the dierence in the level of variation in our two variables of interest would deliver articially high standard errors of the regression coecients, which is known as the "Moulton Problem". Hence, we solve it by collapsing the data-set at the district level so we get a panel-like dataset with administrative units in each country observed through time. For our regression analysis, we use three alternative measures of violence as independent variables: number of attacks, number of people killed and number of people wounded in a given year and area. For the dependent variable we use the FLFP rates of each area for each year, expressed as a number between "zero" and "one". We add or remove year and areas xed eects throughout the dierent specications to shed light on how much each dimension contributes to explaining changes in FLFP rates beyond their violence prevalence. Our interest lies in coecient β of equation 7 which we interpret as the marginal change in FLFP rate in administrative unit a at year t caused by a marginal increase in number of attacks, people wounded, or people killed. Matrix Xat includes the share of workers in the manufacturing sector, the share of the male population and the average age of workers in each administrative unit over time. FLFPat = α + β Violenceat + δ Xat + εat (7) In principle, our violence measures are exogenous to FLFPR as it is unlikely that labor force status of individual women inuences the terrorist incidence. Additionally, observing areas over time allows us to rule out the presence of non-observable variables that might drive our results. We use least squares weighted by the number of women in each area to deal with the "Moulton Problem" once the data are collapsed. 9 This empirical approach allows us to compare districts that are similar in observable characteristics with the only dierence being the exposure to violence while dealing with non-observable variables with the use of panel data. We run several models where the dependent variable is always FLFP rates expressed as a number between "zero" and "one". The variable of interest at the right hand side switches between dierent proxies of violence measured as ows at the district level for each year. In the Robustness section we provide alternative specications with the aim of shedding light on the non-linearity of the relation by using (i) a single dummy variable equal to "one" if there was an attack in a year in a given area; (ii) a set of dummy variables for dierent ranges of attacks number; (iii) number of attacks lagged one time period. 4.2 Quantile Regressions Violence may have dierent eects depending on the initial level of FLFP in each district. For instance, the occurrence of a terrorist attack could reduce participation relatively more in areas with higher initial participation or, alternatively, have more impact in areas with lower rates. Or there might be no dierence in the eect along the distribution of rates. To formally test this possibility, we estimate a set of quantile regressions which basically look for estimates of parameters by minimizing the dierence between the explanatory variables and a given quantile of the dependent variable (as opposed to OLS that minimizes with respect to the mean). We run the estimations using ve points of the FLFP rates distributions: quantiles 10, 25, 50, 75 and 90; and use our three measures of violence (attacks, people wounded, or people killed). 4.3 Tobit Estimation An empirical approximation to equation 6 consists of estimating a Tobit model augmented by the exposure to violence in each district. Due to the widely missing data on wages for each worker, we use "education" as a proxy. However, we include for control the average district wage, calculated using available data in each country (relative to the mean) to proxy for the market value of working (or outside option of remaining in domestic work). Finally, we control for the number of children under age seven in the household, as well as country, and time xed eects. it + β5 Violencedt · uit + εit Hit = β0 + β1 · wdt + β2 · N Kids<7it + β3 · Violencedt + β4 u1 1 (8) Under this setting, the interaction term β5 identies the average dierence in hours supplied between a woman whose spouse is employed compared to a woman whose spouse is unemployed when an attack occurs, measured as when one extra person dies or is wounded. We estimate equation (8) using only females who are not head of households, and we disregard presence or absence of other working household members. We remove India from this estimation, as mentioned, because we lack information on number of working hours, and we use the individual-level data (that is, we 10 do not "collapse" the data set.)6 5 Results 5.1 The extensive margin Tables 3, 4 and 5 show the results of our benchmark specications that use the level of FLFP rates as the dependent variable and our three measures of violence as explanatory variables: number of attacks, number of people killed and number of people wounded, respectively. Column (1) of Table 3 computes the simple correlation between our variables of interest: for each additional attack the FLFP rate falls 0.1 percentage point on average. When including dummies by area, however, this eect drops by three-fourths and when including year and country xed eects the point estimate is a negative 0.08 percentage point. In the last specication, we add control variables, but this does not signicantly aect the results. Using alternative measures of violence leads to similar results. For each additional wounded or killed individual FLFPR falls 0.02 and 0.07 percentage point, respectively. However, statistical signicance disappears and point estimates fall, respectively, when xed eects or controls are included in Tables 4 and 5. All in all, point estimates fall by 80 percent between columns (1) and (2) and by 30 to 40 percent between columns (1) and (3) or (4). We argue that what remains after including control variables is the share of the negative relation observed in the data (i.e. columns (1)) that can be reasonably attributed to violence. Therefore, on average, women tend to participate less in labor markets in the presence of violent conicts. However, when we include variables that capture specic characteristics of the dierent regions, part of the eect fades away as that variation explains FLFP. But since the eect of violence still persists, we claim that violence indeed has a causal eect on FLFP rate because, as shown in column (4) of Table 3, one extra attack reduces the FLFP rate by 0.008 percentage point on average. 5.2 The intensive margin In Figure 8 we show the predicted hours of work obtained after estimating equation 8. We nd that that when no attacks occur - that is, when no one is killed nor wounded - predicted number of working hours is signicantly higher for women whose spouses are unemployed, consistent with the predictions of Equation 6. However, for high levels of violence, the dierence in hours is not signicantly dierent. According to panel (a), hours of women with employed spouses increase with the number of attacks, while hours of women with unemployed spouses only slightly increase beyond 40 hours. This result suggests that women do complement spouse incomes in violent environments. According to 6 Dealing with the Moulton Problem under this setting by collapsing the data would be more troublesome as we would eliminate all variation in the number of hours worked by women. Therefore, we opt to work with the individual-level data. 11 panel (b), however, when using the number of people wounded, hours exerted by both women with unemployed spouses and women whose spouses are employed fall with the level of violence. This result indicates that violent environments signicantly discourage women from working, to the extent that not even women whose spouses are unemployed work more. Finally, when using the number of people killed in panel (c), the same outcome holds, although hours exerted by women with employed spouses are more sensitive to the number of deaths compared to number of people wounded. 5.3 Non-Linearities Table 6 switches the violence measure to a binary variable summarizing whether an area was hit or was not hit by a terrorist attack in a given year regardless of the number of attacks, killed or people wounded. The justication for performing this estimation is that risk perception is not caused by any particular number of attacks. Dierently put, there is "non-linearity" related to whether an area is attacked only once or more times. We obtain high and signicant point estimates in three specications indicating that suering an attack reduces FLFP rates between 5.1 and 5.6 percentage points. Signicance disappears, however, and the point estimate drops to one-fth of the benchmark, when using area xed eects but recovered with country and year xed eects plus control variables. This dierence in results in columns (2) and (4) is likely due to the fact that the "Dummy Attack" variable does not have enough variation beyond that of the area xed eects. Tables 7 through 9 show the results of the Quantile Regressions Estimations which we include to explore the possibility that violence may have dierent eects depending on the initial level of FLFP in each district. In other words, we test whether violence impacts dierently at distinct points of the FLFP distribution. All specications include country and year xed eects. We nd point estimates of the marginal eect at the dierent points of the FLFP distribution that are similar to those on the average. Moreover, the marginal eect of an attack is not statistically signicant at the bottom of the distribution and reaches a peak at the median (-0.01 percentage point). In other words, the impact of an attack is higher in districts where women's work is more prevalent. When measuring violence as the number of people killed or wounded, a similar pattern emerges. Point estimates are not signicant for quantiles 10 and 25 of the FLFP distribution. However, the maximum eect is found at the right end of the distribution indicating that the higher number of people killed and - in particular - people wounded, the higher the drop in FLFP rates. In the Appendix, we show the results of this same set of quantile regressions but adding explanatory variables where point estimates slightly reduce and signicance is gained also in the 25th percentile. To test whether a non-linear relationship exists between the number of attacks and FLFP rates we use a set of dummy variables splitting the range of attacks distribution (0, (0, 5], (5, 15] and (15, +)). The results depicted in Table 10 show that higher levels of violence have greater eect on FLFP rates. In particular, according to columns (1) through (4) the FLFP rate drops between 4 and 12 percentage points when 15 or more attacks take place a given year in a region relative to another where no attacks occur, on average. Note that signicance remains robust when including 12 explanatory variables and area, country and year xed eects. A complementary approach to test the non-linearity hypothesis is to use two dierent splines of the violence mea- sures.7 This approach delivers marginal eects that depend on the point at which we evaluate the violence (non-linearly in the second one), as shown in Figure 9. When using the square of violence measures, we nd that for all of them the eect on FLFP rates decreases along with the level of violence. When using the cube of violence, we nd a pattern not consistent with the hypothesis depicted in Figure 5 according to which the impact should be smaller for low and high levels of violence where the variance of attacks is smaller (see Figure 9 in the Appendix for the results of the same exercise but including control variables). Dierently put, the marginal eect of one extra terrorist attack deters FLFP more when there are relatively few or relatively many attacks. We speculate that on one hand, an isolated event might be stranger and hence have a greater impact; on the other hand, when there is an extremely high level of violence women get particularly discouraged from participating in the labor market. 5.4 Robustness Violence episodes also threaten the security of men, not only of women. Hence, it is plausible to expect violence to aect male labor participation even though we know that men are the main households earners in South Asia and we would expect their labor supply to be less elastic. Tables 11, 12 and 13 show the results of our benchmark estimations using the male labor force participation rate as dependent variable instead of FLFP. We nd signicant negative impact on men's labor participation, but it is between two and seven times lower than the negative impact for females. Therefore, a violent attack tends to widen the gender labor gap. One last robustness check consists in running our main results using individual level data. The main reason to use the aggregated version was to deal with the Moulton problem since our violence measure varies only at the administrative unit level. Hence, we run slightly modied versions of our benchmark estimations clustering standard errors at the violence level variable to show the extent of the problem. While our collapsed version of the estimations controls for the share of males, the share of workers in manufacturing and the average age of each district, the individual level data versions control only for the individual's age. Most of our results are statistically signicant (see Tables 14, 15 and 16). Moreover, in terms of the orders of magnitude, we get very similar point estimates to our benchmark estimates. When clustering standard errors and including years xed eects, for each attack FLFP falls 0.147 percentage point; for each wounded person FLFP falls 0.0148 percentage point and for each person killed FLFP falls 0.047 percentage point. 7 at + δ Xat + εat and (2) FLFPRat = In concrete, we estimate two dierent regressions: (1)FLFPRat = α + β1 Violenceat + β2 Violence2 α + β1 Violenceat + β2 Violence2 at + β 3 Violence 3 at + δ Xat + ε at 13 6 Conclusions This paper aims to estimate to what extent violence explains the extremely low female labor force participation (FLFP) rates in South Asia. This is relevant for both academics and policy makers as FLFP is well-known to contribute to development. In addition to providing extra labor input and therefore increasing gross domestic product, women having earning power has been shown to increase investment in children, including education, which promotes future GDP growth (Duo [2012]). The direction of causality between violence and FLFP and its implications for identication represents a crucial challenge. One the one hand violence could increase FLFP through the added worker eect whereby women increasingly work outside the home to help support households during a crisis. On the other hand violence may increase the cost of working for women, either directly due to the risk of violence or indirectly by reducing economic opportunities. Therefore, economic theory does not oer a clear prediction about the relationship. At rst glance, the evidence supports the latter rather than the former eect, because in a cross-section of countries or regions a negative correlation between violence and FLFP emerges; that is, FLFP is lower in areas where violence is higher. Thus, looking at aggregate data shows little evidence for the added worker eect. We evaluate the validity of this argument using disaggregated data that track FLFP rates over repeated cross- sections. On the "extensive margin" of labor supply - that is, on women's decision whether to work or not work - we nd that one extra-attack reduces FLFP by 0.008 percentage point. This results in a widening of the gender labor participation gap as the eect of violence on men's labor participation is up to seven times smaller in our study. On the intensive margin - that is, on women's decision regarding how many total hours they work - the presence of violent attacks encourages married women to exert more working hours, but when the environment gets more risky, as the number of dead and people wounded increases, women reduce their working hours. We also nd that violence has a greater impact on discouraging women from working in areas where female work was more prevalent before the advent of violence. Moreover, the eect of violence is non-linear as a greater number of attacks generates a greater impact on FLFPR. Although we nd a robust negative eect of terrorist attacks on FLFP using dierent estimation approaches, we recognize that the sizes of the eects found are not high enough to attribute the extremely low FLFP rates in the region solely to violence. Although we provide evidence that violence is indeed one of the determinants of low FLFP, further research and richer data are needed to establish how the presence of violence inuences FLFP compared to drivers such as values, cultural norms and human capital dierences, among other factors. 14 References Mark R. Killingsworth and James J. Heckman. Female labor supply: A survey. In O. Ashenfelter and R. Layard, editors, Handbook of Labor Economics, volume 1 of Handbook of Labor Economics, chapter 2, pages 103204. Elsevier, 1987. URL https://ideas.repec.org/h/eee/labchp/1-02.html. Erica Field and Kate Vyborny. Female labor force participation in asia: Pakistan country study. Unpublished, 2015. Simone Schaner and Smita Das. Female Labor Force Participation in Asia: Indonesia Country Study. ADB Economics Working Paper Series 474, Asian Development Bank, February 2016. URL https://ideas.repec.org/p/ris/ adbewp/0474.html. Aparna Mitra and Pooja Singh. Human capital attainment and female labor force participation: The kerala puzzle. Journal of Economic Issues, 40(3):779798, 2006. ISSN 00213624. URL http://www.jstor.org/stable/4228291. R. Pande, E. Fletcher, and D. Ford. Female labor force participation in asia: Constraints and challenges. Technical report, Asian Development Bank, 2016. URL https://www.adb.org/sites/default/files/project-document/ 201716/47363-001-tacr-01.pdf. Ratna M Sudarshan. Through the magnifying glass: Women's work and labour force participation in urban delhi. Economic and Political Weekly, Vol. 44(Issue No. 48), 28 2009. START. Global terrorism database, 2018. Retrieved from https://www.start.umd.edu/gtd. Valerie M. Hudson and Patricia Leidl. The Hillary Doctrine: Sex and American Foreign Policy. Columbia University Press, 2015. URL http://www.jstor.org/stable/10.7312/huds16492. Sarah Khan, Jana Kuhnt, and Atika Pasha. Employment in the times of terror: A gendered perspective on pakistan. Mimeo, 2017. Claude Berrebi and Jordan Ostwald. Terrorism and the labor force: Evidence of an eect on female labor force participation and the labor gender gap. Journal of Conict Resolution, 60(1):3260, 2016. doi: 10.1177/0022002714535251 . URL https://doi.org/10.1177/0022002714535251. Jose Gallegos. The eects of civil conicts on women's labor force participation: A causal mechanism approach. Mimeo, 2012. Nidhiya Menon and Yana van der Meulen Rodgers. War and women's work: Evidence from the conict in nepal. Journal of Conict Resolution, 59(1):5173, 2015. doi: 10.1177/0022002713498699 . URL https://doi.org/10.1177/ 0022002713498699. 15 Merle Kreibaum and Stephan Klasen. Missing Men: Dierential Eects of War and Socialism on Female Labour Force Participation in Vietnam. Courant Research Centre: Poverty, Equity and Growth - Discussion Papers 181, Courant Research Centre PEG, July 2015. URL https://ideas.repec.org/p/got/gotcrc/181.html. Orley Ashenfelter. Unemployment as disequilibrium in a model of aggregate labor supply. Econometrica, 48(3):547564, 1980. ISSN 00129682, 14680262. URL http://www.jstor.org/stable/1913122. Pieter Serneels. The added worked eect and intra household aspects of unemployment. Technical report, 2002. Esther Duo. Women empowerment and economic development. Journal of Economic Literature, 50(4):105179, December 2012. doi: 10.1257/jel.50.4.1051. URL http://www.aeaweb.org/articles?id=10.1257/jel.50.4. 1051. 16 7 Figures & Tables Figure 1: Female Labor Force Participation Rates, Selected SA Countries Labor force participation rate (% of female ages 15+) 20 40 0 60 80 2000 2016 2000 2016 2000 2016 2000 2016 2000 2016 2000 2016 2000 2016 2000 2016 2000 2016 2000 2016 Iran Pakistan India Turkey Bangladesh Sri Lanka Afghanistan Indonesia Bhutan Nepal Source: The World Bank Figure 2: Number of Terrorists Attacks and FLFPR, worldwide 80 Female Labor Force Participation Rate (%) Kenya Thailand 60 Israel Colombia United Kingdom Nigeria Philippines Ukraine Afghanistan 40 Bangladesh Libya Sudan Lebanon India Pakistan 20 Somalia Iraq Yemen 0 0 1000 2000 3000 4000 Number of Attacks in 2014 Source: GTD and The World Bank, 2014. Countries with more than 100 attacks labeled. 17 Figure 3: Cumulative number of attacks by administrative division (a) Bangladesh (b) Sri Lanka 18 Figure 4: Female Labor Force Participation Rates by Administrative Division Bangladesh India .8 .6 .6 .4 mean mean .4 .2 .2 0 0 2004 2006 2008 2010 2012 2014 2000 2005 2010 Survey Year Survey Year Notes: (a) Each dot represents a different zila; (b) Country unweighted average in red triangle; (c) Magura Notes: (a) Each dot represents a different state; (b) Country unweighted average in red triangle; (c) Bihar has has the lowest rate in 2013 and Bandarban the highest the lowest FLFPR in 2011 and Sikkim the highest. (a) Bangladesh (b) India Pakistan Sri Lanka .8 .6 .5 .6 .4 mean mean .4 .3 .2 .2 .1 0 2006 2008 2010 2012 2014 1990 1995 2000 2005 2010 2015 Survey Year Survey Year Notes: (a) Each dot represents a different district; (b) Country unweighted average in red triangle (c) Hangu Notes: (a) Each dot represents a different district; (b) Country unweighted average in red triangle; (c) Mannar has the lowest rate in 2013 and Jhang the second highest has the lowest rate in 2015 and Anuradhapura the highest (c) Pakistan (d) Sri Lanka Figure 5: Probability and Risk (Variance) Risk(More deaths) Risk(Variance) Probability of Violence 19 Figure 6: Attacks timing and Labor Force Participation Hambantota, Sri Lanka Bahawalpur, Pakistan 1999 2009 2012 .38 .4 .42 .44 .46 .2 .25 .3 .35 .4 .45 2013 .8 .81 .82 .83 .84 .78 .79 .8 .81 .82 1998 Female LFPR Female LFPR Male LFPR Male LFPR 1995 1996 1994 2007 2008 -2 -1 0 1 2 -2 -1 0 1 2 Periods since attack Periods since attack 1996: 1 Attack, 0 Killed, 0 Wounded 2009: 3 Attacks, 0 Killed, 10 Wounded Haryana, India Galle, Sri Lanka 2004 .75 .755 .76 .765 .77 .78 .8 .82 .36 .37 .38 .39 .4 2008 .15 .2 .25 .3 .35 Female LFPR Female LFPR 2003 Male LFPR Male LFPR 1999 2007 2011 2009 2007 2011 2004 .76 -2 -1 0 1 2 -2 -1 0 1 2 Periods since attack Periods since attack 2007: 1 Attack, 66 Killed, 0 Wounded 2007: 1 Attack, 12 Killed, 47 Wounded Female LFPR Male LFPR 20 Figure 7: Conditional marginal eects of Violence -.0005 .002 -.001 Effects on Linear Prediction Effects on Linear Prediction 0 -.0015 -.002 -.002 -.004 -.0025 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 0 100 200 300 Number of attacks Number of attacks (a) Number of Attacks2 (b) Number of Attacks3 0 .001 Effects on Linear Prediction Effects on Linear Prediction -.0005 -.001 0 -.001 -.002 -.0015 -.003 0 50 100 150 200 0 100 200 300 Number of killed people Number of killed people (c) Number of Attacks2 (d) Number of Attacks3 .0002 -.0002 0 Effects on Linear Prediction Effects on Linear Prediction -.0003 -.0004 -.0002 -.0004 -.0006 -.0005 -.0008 0 100 200 300 0 100 200 300 400 500 Number of killed people Number of killed people (e) Number of Attacks2 (f) Number of Attacks3 Marginal eects from a model where FLFPR is at LHS and either the squared or the cube of Violence is at the RHS. All regressions at area level weighted by the number of women in each area. Year and country xed eects included. Robust standard errors. 21 Figure 8: Linear Prediction from Tobit estimations of Hours Worked 44 40 40 39 42 39 Linear Prediction of Hours Linear Prediction of Hours Linear Prediction of Hours 38 40 38 37 38 37 36 36 36 35 0 10 20 30 40 50 0 20 40 60 80 100 120 140 160 180 200 220 240 260 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 Spouse Unemployed Spouse Employed Spouse Unemployed Spouse Employed Spouse Unemployed Spouse Employed (a) Number of Attacks (b) Number of Wounded (c) Number of Killed 22 Table 1: Eect of Violence on FLFP Rates, Cross-country estimates (1) (2) (3) (4) (5) (6) FLFPR FLFPR FLFPR FLFPR FLFPR FLFPR N Attacks -0.0180∗∗∗ 0.00138 (-3.36) (1.23) N Attacks(t-1) -0.00261 -0.0000693 (-0.34) (-0.05) N Attacks(t-2) -0.00736 -0.00172 (-1.20) (-1.42) N Wounded -0.00236∗∗∗ 0.000161 (-3.76) (1.20) N Wounded(t-1) -0.00157∗ 0.000105 (-2.23) (0.75) N Wounded(t-2) -0.00140∗ 0.000134 (-2.08) (1.00) N Killed -0.00422∗∗∗ 0.000159 (-3.44) (0.63) N Killed(t-1) -0.00134 -0.0000186 (-0.81) (-0.06) N Killed(t-2) -0.00386∗∗ -0.0000438 (-2.67) (-0.15) Population Size/1M -0.0374∗∗∗ -0.0381∗∗∗ -0.0376∗∗∗ (-9.35) (-9.87) (-9.72) Urban Population(% of Total) 0.0563∗ 0.0605∗∗ 0.0571∗∗ (2.53) (2.73) (2.58) Fertility (births per woman) -0.0779 -0.0193 -0.0617 (-0.35) (-0.09) (-0.28) GDP per capita/1K 0.267∗∗∗ 0.266∗∗∗ 0.267∗∗∗ (10.58) (10.53) (10.56) Gov Spending(% of GDP) -0.0123 -0.0124 -0.0122 (-0.89) (-0.90) (-0.88) FDI(% of GDP) 0.000274 0.000266 0.000257 (0.07) (0.07) (0.07) Observations 3463 3463 3463 3463 3463 3463 Years 1990-2016; 142 countries. Country and Year xed eects included when using controls. A categorical 7-categories variable for Civil Liberties index included in the estimations not shown. FLFPR ∈ [0, 100]. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 23 Table 2: Summary Statistics (1) (2) (3) (4) Bangladesh India Pakistan Sri Lanka mean/sd mean/sd mean/sd mean/sd Female LFPR 0.33 0.32 0.14 0.38 0.12 0.13 0.13 0.10 Male LFPR 0.84 0.81 0.79 0.80 0.05 0.04 0.08 0.03 Share of Males 0.50 0.51 0.51 0.48 0.02 0.02 0.04 0.02 Share Agriculture 0.27 0.23 0.19 0.20 0.09 0.11 0.13 0.11 Share Manufactures 0.06 0.06 0.03 0.08 0.04 0.03 0.03 0.03 N of Attacks 0.86 10.70 6.57 1.02 5.75 24.54 25.36 2.58 N of Wounded 1.91 26.61 23.55 14.09 9.64 80.02 96.01 96.39 N of Killed 0.37 16.82 11.12 5.21 1.56 38.02 42.75 18.82 Observations 192 155 575 360 Unit of observation: district/year. When no attacks ocurr in a given district/year a zero is inputed to N of Attacks, Wounded and people killed. Table 3: Eect of Attacks on FLFP Rates (1) (2) (3) (4) N of Attacks -0.00119∗∗∗ -0.000265∗ -0.000811∗∗∗ -0.000775∗∗∗ (-5.72) (-2.23) (-3.59) (-4.05) Area/Country/Year/Controls No/No/No/No Yes/No/No/No No/Yes/Yes/No No/Yes/Yes/Yes Observations 1282 1280 1282 1282 Least Squares Weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1]. Robust standard errors Area, Country or Year xed eects included as indicated. Controls include, by area, (1) The share of workers in Manufacturing; (2) Share of males and (3) Average age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 24 Table 4: Eect of N Wounded on FLFP Rates (1) (2) (3) (4) N of Wounded -0.000205∗∗ -0.0000224 -0.000160∗∗∗ -0.000148∗∗∗ (-2.65) (-1.69) (-4.31) (-4.47) Area/Country/Year/Controls No/No/No/No Yes/No/No/No No/Yes/Yes/No No/Yes/Yes/Yes Observations 1282 1280 1282 1282 Least Squares Weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1]. Robust standard errors Area, Country or Year xed eects included as indicated. Controls include, by area, (1) The share of workers in Manufacturing; (2) Share of males and (3) Average age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 5: Eect of N Killed on FLFP Rates (1) (2) (3) (4) N of Killed -0.000753∗∗∗ -0.000167∗∗ -0.000513∗∗∗ -0.000477∗∗∗ (-7.69) (-2.92) (-5.26) (-5.45) Area/Country/Year/Controls No/No/No/No Yes/No/No/No No/Yes/Yes/No No/Yes/Yes/Yes Observations 1282 1280 1282 1282 Least Squares Weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1]. Robust standard errors Area, Country or Year xed eects included as indicated. Controls include, by area, (1) The share of workers in Manufacturing; (2) Share of males and (3) Average age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 6: Eect of the presence of an attack on FLFP Rates (1) (2) (3) (4) Dummy Attack -0.0560∗∗∗ -0.0115 -0.0516∗∗∗ -0.0512∗∗∗ (-5.24) (-1.76) (-5.45) (-5.77) Area/Country/Year/Controls No/No/No/No Yes/No/No/No No/Yes/Yes/No No/Yes/Yes/Yes Observations 1282 1280 1282 1282 Least Squares Weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1]. Robust standard errors Area, Country or Year xed eects included as indicated. Controls include, by area, (1) The share of workers in Manufacturing; (2) Share of males and (3) Average age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 7: Quantile Regression. Eect of Attacks on FLFP Rates Q(0.10) Q(0.25) Q(0.50) Q(0.75) Q(0.90) N of Attacks -0.000673 -0.000603 -0.00114∗∗∗ -0.000709∗ -0.000894∗∗∗ (-0.79) (-1.86) (-9.20) (-2.01) (-4.04) Observations 1282 1282 1282 1282 1282 Quantile Regression weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1] Country and Year Dummies included. Robust standard errors. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 25 Table 8: Quantile Regression. Eect of N Killed on FLFP Rates Q(0.10) Q(0.25) Q(0.50) Q(0.75) Q(0.90) N of Killed -0.000184 -0.000337 -0.000661∗ -0.000470∗∗∗ -0.000688∗ (-0.91) (-1.51) (-2.54) (-9.29) (-2.56) Observations 1282 1282 1282 1282 1282 Quantile Regression weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1] Country and Year Dummies included. Robust standard errors. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 9: Quantile Regression. Eect of N Wounded on FLFP Rates Q(0.10) Q(0.25) Q(0.50) Q(0.75) Q(0.90) N of Wounded -0.0000154 -0.000159 -0.000198∗∗∗ -0.000165∗∗∗ -0.000258∗∗∗ (-0.40) (-0.92) (-4.74) (-3.90) (-5.03) Observations 1282 1282 1282 1282 1282 Quantile Regression weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1] Country and Year Dummies included. Robust standard errors. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 10: Eect of N Attacks on FLFP Rates (1) (2) (3) (4) N Attacks ∈ (0, 5] -0.0269∗ -0.00837 -0.0372∗∗∗ -0.0383∗∗∗ (-2.17) (-1.22) (-3.33) (-3.66) N Attacks ∈ (5, 15] -0.102∗∗∗ -0.0321∗∗ -0.0621∗∗∗ -0.0606∗∗∗ (-4.52) (-3.26) (-3.57) (-3.55) N Attacks ∈ (15, +) -0.128∗∗∗ -0.0436∗∗∗ -0.108∗∗∗ -0.100∗∗∗ (-7.56) (-3.68) (-6.72) (-6.51) Area/Country/Year/Controls No/No/No/No Yes/No/No/No No/Yes/Yes/No No/Yes/Yes/Yes Observations 1282 1280 1282 1282 Least Squares Weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1]. t-statistics in parenthesis. Controls, Country, Area or Year Dummies included as indicated. Zero attacks is the base(excluded) category. Robust standard errors ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 11: Eect of Attacks on Male LFP Rates (1) (2) (3) (4) N of Attacks -0.000213∗∗∗ -0.000150∗∗ -0.000157∗∗ -0.000267∗∗∗ (-3.46) (-2.69) (-2.79) (-5.39) Area/Country/Year/Controls No/No/No/No Yes/No/No/No No/Yes/Yes/No No/Yes/Yes/Yes Observations 1282 1280 1282 1282 Least Squares Weighted by the number of men in each Area/Year. MLFPR ∈ [0, 1]. Robust standard errors Area, Country or Year xed eects included as indicated. Controls include, by area, (1) The share of workers in Manufacturing; (2) Share of males and (3) Average age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 26 Table 12: Eect of N Wounded on Male LFP Rates (1) (2) (3) (4) N of Wounded -0.0000596∗∗∗ -0.00000993 -0.0000493∗∗∗ -0.0000681∗∗∗ (-4.20) (-1.46) (-4.43) (-5.16) Area/Country/Year/Controls No/No/No/No Yes/No/No/No No/Yes/Yes/No No/Yes/Yes/Yes Observations 1282 1280 1282 1282 Least Squares Weighted by the number of men in each Area/Year. MLFPR ∈ [0, 1]. Robust standard errors Area, Country or Year xed eects included as indicated. Controls include, by area, (1) The share of workers in Manufacturing; (2) Share of males and (3) Average age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 13: Eect of N Killed on Male LFP Rates (1) (2) (3) (4) N of Killed -0.000180∗∗∗ -0.0000658∗ -0.000148∗∗∗ -0.000212∗∗∗ (-5.71) (-2.08) (-5.15) (-9.07) Area/Country/Year/Controls No/No/No/No Yes/No/No/No No/Yes/Yes/No No/Yes/Yes/Yes Observations 1282 1280 1282 1282 Least Squares Weighted by the number of men in each Area/Year. MLFPR ∈ [0, 1]. Robust standard errors Area, Country or Year xed eects included as indicated. Controls include, by area, (1) The share of workers in Manufacturing; (2) Share of males and (3) Average age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 14: Eect of Attacks on FLFP Rates using individual-level data (1) (2) (3) N of Attacks -0.00186∗∗∗ -0.00186∗∗∗ -0.00147∗∗∗ (-93.18) (-5.46) (-3.45) Cluster/Year FE No/No Yes/No Yes/Yes Observations 1437269 1437269 1437269 Ordinary Least Squares. FLFPR ∈ [0, 1]. Clustered S.E. at the district level and Year xed eects as indicated. All regressions control for individuals age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 15: Eect of N Wounded on FLFP Rates using individual-level data (1) (2) (3) N of Wounded -0.000164∗∗∗ -0.000164 -0.000192∗ (-43.29) (-1.52) (-2.59) Cluster/Year FE No/No Yes/No Yes/Yes Observations 1437269 1437269 1437269 Ordinary Least Squares. FLFPR ∈ [0, 1]. Clustered S.E. at the district level and Year xed eects as indicated. All regressions control for individuals age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 27 Table 16: Eect of N Killed on FLFP Rates using individual-level data (1) (2) (3) N of Killed -0.000956∗∗∗ -0.000956∗∗∗ -0.000829∗∗ (-74.47) (-3.45) (-2.98) Cluster/Year FE No/No Yes/No Yes/Yes Observations 1437269 1437269 1437269 Ordinary Least Squares. FLFPR ∈ [0, 1]. Clustered S.E. at the district level and Year xed eects as indicated. All regressions control for individuals age. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 28 8 Appendix Table 17: Quantile Regression. Eect of Attacks on FLFP Rates Q(0.10) Q(0.25) Q(0.50) Q(0.75) Q(0.90) N of Attacks -0.000900 -0.000659∗∗∗ -0.00109∗∗∗ -0.000558∗∗ -0.000794∗∗∗ (-1.44) (-10.21) (-4.61) (-3.26) (-5.22) Observations 1282 1282 1282 1282 1282 Quantile Regression weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1] Controls, Country and Year Dummies included. Robust standard errors. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 18: Quantile Regression. Eect of N Killed on FLFP Rates Q(0.10) Q(0.25) Q(0.50) Q(0.75) Q(0.90) N of Killed -0.000252 -0.000306∗∗ -0.000579∗∗ -0.000445∗∗∗ -0.000502∗∗ (-1.28) (-2.94) (-3.04) (-6.44) (-2.85) Observations 1282 1282 1282 1282 1282 Quantile Regression weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1] Controls, Country and Year Dummies included. Robust standard errors. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 Table 19: Quantile Regression. Eect of N Wounded on FLFP Rates Q(0.10) Q(0.25) Q(0.50) Q(0.75) Q(0.90) N of Wounded -0.0000365 -0.0000951∗∗∗ -0.000189∗∗ -0.000154∗ -0.000201∗∗∗ (-0.28) (-9.05) (-2.74) (-2.43) (-3.62) Observations 1282 1282 1282 1282 1282 Quantile Regression weighted by the number of women in each Area/Year. FLFPR ∈ [0, 1] Controls, Country and Year Dummies included. Robust standard errors. t-statistics in parenthesis ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 29 Figure 9: Conditional marginal eects of Violence -.0005 .002 -.001 Effects on Linear Prediction Effects on Linear Prediction 0 -.0015 -.002 -.002 -.004 -.0025 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 0 100 200 300 Number of attacks Number of attacks (a) Number of Attacks2 (b) Number of Attacks3 0 .001 Effects on Linear Prediction Effects on Linear Prediction -.0005 -.001 0 -.001 -.002 -.0015 -.003 0 50 100 150 200 0 100 200 300 Number of killed people Number of killed people (c) Number of Attacks2 (d) Number of Attacks3 -.0001 .0002 0 -.0002 Effects on Linear Prediction Effects on Linear Prediction -.0002 -.0003 -.0004 -.0004 -.0006 -.0005 -.0008 0 100 200 300 0 100 200 300 400 500 Number of killed people Number of killed people (e) Number of Attacks2 (f) Number of Attacks3 Marginal eects from a model where FLFPR is at LHS and either the squared or the cube of Violence is at the RHS. All regressions at area level weighted by the number of women in each area. Controls, Year and country xed eects included. Robust standard errors. 30