Policy Research Working Paper 9691 Public Work Programs and Gender-Based Violence Evidence from Lao PDR Elizaveta Perova Erik Johnson Aneesh Mannava Sarah Reynolds Alana Teman East Asia and the Pacific Region Gender Innovation Lab June 2021 Policy Research Working Paper 9691 Abstract Public workfare programs targeted at women have the the government randomized implementation of a public potential to empower them economically by providing workfare program targeted at rural women who received an jobs. However, the impact of public workfare programs average payment of US$550 over 18 months. The findings on gender-based violence is theoretically ambiguous. They show that the program was successful in increasing female may contribute to its reduction through lowering financial income, but it did not change women’s experience of gen- stress or improving a woman’s bargaining position due to der-based violence: comparing program participants and independent income. Yet, a woman’s higher income may control group women, there is no differences in self-reports also create incentives to use violence for extractive pur- of intimate partner violence (controlling behavior, emo- poses; putting women in a position of provider at home tional violence, or physical violence), violence from other and in male dominated sectors outside the home may create members of the household, or violence from perpetrators a backlash because these positions violate gender norms. outside the household. Some design aspects of this partic- Working outside the home could reduce exposure to an ular program may have resulted in the lack of impacts on abusive spouse, but it may increase harassment or assault gender-based violence. Changes in the design and imple- outside the household. This paper analyzes the impacts of mentation of public workfare programs are needed for them a public workfare program in the Lao People’s Democratic to work as a mechanism to reduce gender-based violence. Republic, a lower-middle-income Asian country, where This paper is a product of the Gender Innovation Lab, East Asia and the Pacific Region. 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 eperova@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 Public Work Programs and Gender-Based Violence: Evidence from Lao PDR Elizaveta Perova1*, Erik Johnson1, Aneesh Mannava, Sarah Reynolds2, and Alana Teman JEL codes: D1, D04, I15, J12, J16 Key words: public work programs, gender-based violence, Southeast Asia, randomized intervention 1 World Bank 2 University of California, Berkeley, School of Public Health *Corresponding Author eperova@theworldbank.org We gratefully acknowledge funding from the Umbrella Facility for Gender Equality (UFGE) and Australia’s DFAT to carry out this work. The data collection protocols were approved by Health Media Labs IRB. We thank Aline Coudouel, Patrick Premand, Stephanie Kuttner and participants of the East Asia and Pacific Chief Economist Office seminar for insightful comments. We thank Amber Peterman, Sofia Amaral and S Anukriti in advance for reviewing the work. All errors are ours. 1. Introduction Women’s economic empowerment has been theorized to reduce intimate partner violence (IPV), a large component of the broader category of gender-based violence (GBV). Theories based on Nash-bargaining suggest that improving women’s options outside a relationship also increases their bargaining power within it (Manser and Brown 1980), and is likely to improve their outcomes within marriage, including exposure to IPV. However, instrumental violence theory, which states that IPV can be used for extractive purposes (Bloch and Rao 2002; Bobonis, González-Brenes, and Castro 2013), suggests a possibility of increasing IPV. Similarly, psychological theories of male backlash, which postulate that violence may be used for maintaining a socially constructed identity of masculine superiority (Macmillan and Gartner 1999; Buller et al. 2018), also predict an increase in IPV. This potential backlash is not limited to the marital sphere. GBV perpetrated by men against women who violate traditional gender norms is also possible beyond domestic life and between strangers. Incidents range from harassment (Akerlof and Kranton 2000) to, in the extreme, feminicide, as in the case of garment factory workers in Mexico (Corradi et al. 2016). The mechanisms are quite similar to those behind IPV. Low-skilled women may put up with harassment in the work force due to lack of alternative opportunities in the job market. Similarly, men may feel their labor market security is threatened by women’s participation and act in a violent manner to discourage or punish women who work (Pillinger 2017). Although employment may reduce IPV, it could be replaced with GBV experienced in the workplace or while commuting. In this paper we investigate the effects on both domestic violence and violence outside the home, of temporary employment in a Laotian Public Works Program (PWP) that employed poor women in routine maintenance work on ancillary rural roads. Exploiting a randomized control trial, we are able to provide evidence on the causal relationship between women’s temporary employment opportunities and their exposure to IPV, violence from other members of the household, and violence outside the home. The public workfare program targeted women from the poorest households in villages in rural Lao PDR, where few wage opportunities exist. Additionally, jobs in infrastructure are traditionally male occupations, so this program challenged typical gender norms. Although there was almost complete participation among the treated and large increases in household income, we find no change in self-reported exposure to IPV, in GBV from other household members (including controlling behavior, emotional, sexual or physical violence) nor in GBV outside the home. We are only aware of one paper that has studied women’s labor market participation and GBV outside the home in lower- and middle-income countries. Amaral et al. find a government-sponsored employment program primarily benefitting poor women increased kidnappings, sexual harassment, and domestic violence, though dowry deaths decreased (Amaral, Bandyopadhyay, and Sensarma 2015). Other causal studies examine the broad category of GBV but do not distinguish between domestic violence and 2 violence outside the home (e.g. Aizer 2010; Amaral 2019, Perova, Reynolds and Schmutte, 2021). 1 A larger body of literature from developing country settings on women’s economic empowerment and IPV yields mixed findings. At an aggregate scale in Mexico, an increase in women’s employment increased homicides of married women (Davila 2019), while in India improved labor force participation (Chin 2012) for women reduced IPV. Examination of the impacts of economic empowerment at an individual level by Bulte and Lensink (2019) found that women who received empowerment and business training were more likely to be victims of IPV; the mechanism identified was more conflict over the increased income. Kotsadam and Villanger (2020) randomly assigned manufacturing jobs to equally qualified female applicants in Ethiopia and found that being offered a job had no aggregate effect on experiencing IPV. However, three months after employment, job offers reduced emotional abuse and there were indications of heterogeneous effects whereby women with low bargaining power at baseline experienced increased risks of abuse if offered a job. Heath (2014) provides one example of mixed results in the large correlational literature on women’s employment and IPV. She finds evidence in Bangladesh that a woman’s initial level of bargaining power can explain heterogeneity in impacts: better labor market options for women decrease the likelihood of experiencing domestic violence for women who have higher baseline bargaining power, but increase the likelihood of experiencing domestic violence among women with lower baseline bargaining power. Lessons from this literature suggest that workfare programs should be cautious. Workfare programs often target a relatively vulnerable population that typically is poorer, less educated and has lower access to the traditional job market. For women, these characteristics are often correlated with lower empowerment and bargaining power. 2 Instrumental violence theory and empirical work imply that this group may be particularly vulnerable to an increased exposure of IPV. In contrast to the population in these employment studies however, participants in our rural sample have fewer outside options than the industrial park settings studied by Kotsadam and Villanger (2020), and the villages studied by Heath (2014), a 30-minute journey from Bangladesh’s capital: jobs are scarce in the studied PWP program area. In general, the target population of workfare programs like PWPs traditionally lack access to the labor market for reasons such as limited supply of wage jobs and low level of skills. In this sense, the target population of the PWPs is similar to the target population of cash transfers (CTs). Buller et al. (2018) conducted a review of 14 cash transfer programs and IPV in lower-middle-income countries, and found that in most cases IPV decreased, with a few exceptions. However, unlike conditional cash transfers, the conditions of which focus on children and typically re-enforce women’s duties in the domestic sphere, PWPs require participants to work, 1 Aizer and Perova, Reynolds and Schmutte take the same strategy of analyzing medical injuries for assault as proxies for IPV, though they are unable to differentiate between IPV specifically, GBV outside the home, and incidental crime against women. 2 Doss (2017) reviews the literature on proxies for bargaining power. 3 often in low-skilled jobs, to receive transfers. Transfers from PWPs may induce different effects compared with those of unearned income from cash-transfers, since women take on a new role outside the household, which could challenge social and gender norms. When these programs place women in employment traditionally reserved for men, the participants may be more at risk of assault or harassment. This evaluation of a PWP contributes to the existing literature on GBV and women’s economic opportunities in two ways. First, to our knowledge, this is the only study exploring the causal relationship between a workfare program and gender-based violence through a Randomized Controlled Trial. Second, the inclusion of domestic violence (physical or sexual) by a non-partner and violence (physical or sexual) outside the home along with the study of IPV (emotional, physical, sexual violence and controlling behavior) provides a broader analysis of GBV than is typically explored in economic empowerment studies. This consideration is important for women’s safety, particularly when employment programs may violate traditional gender norms in the public sphere. Our results of neither a positive nor negative impact on gender-based violence inside and outside the home have at least two policy implications. First, they suggest that PWPs can be used as a mechanism for increasing household income without imposing safety risks on participating women. Second, while some social protection instruments, such as cash transfers, can also serve as a policy tool for lowering IPV, we cannot make a similar case for PWPs. As complementary programming, such as behavioral change communication, increases the potential of cash transfers to reduce GBV (Roy et al. 2019), complimentary interventions may be needed for the PWPs to work as GBV- reducing interventions. 2. Context: Gender-Based Violence in Lao PDR Lao PDR ranks at 113 of 162 countries on gender equality, according to the Gender Inequality Index (GII), produced by UNDP. 3 This rank is far behind some countries in Southeast Asia: Singapore (12), Malaysia (59), Vietnam (65) and Thailand (80), though other Southeast Asian countries have similar rank: the Philippines (104), Cambodia (117), Myanmar (118) and Indonesia (121). Despite this low ranking on gender equality, Lao PDR has relatively low rates of IPV, compared to a set of countries for which comparable data are available. The most recent source of nationally representative data on GBV in Lao PDR is the Lao National Survey on Women’s Health and Life Experiences, which uses the standardized World Health Organization methodology. The survey results suggest that 15.3% of ever-partnered women experience physical and/or sexual violence from an intimate partner in their lifetime. To contextualize this 3 The GII measures gender inequalities in three aspects of human development—reproductive health, empowerment and economic status. To do so, it aggregates 5 indicators: maternal mortality ratio, adolescent birth rates, proportion of parliamentary seats occupied by females, proportion of females and males aged 25 years and older with at least some secondary education and labor force participation rate of female and male populations aged 15 years and older. 4 prevalence rate, we compared this to a similar indicator from 47 countries for which a Demographic and Health Survey was administered in the last decade, and included a GBV module based on the WHO methodology. Lifetime exposure to physical and/or sexual violence from intimate partners in these 47 countries ranges from 6.4% in the Comoros to 57.1% in Papua New Guinea. Lao PDR ranks in the 15% of countries with the lowest prevalence of IPV, with a prevalence rate similar to neighboring Philippines (12.2%), Myanmar (16.3%) and Cambodia (18.2%). The same national survey estimates that 5.1% and 5.3% of Laotian women experience physical and sexual violence, respectively, from a non-partner. For approximately 60% of respondents who reported this violence, the violence was committed by family members. Notably, close to two-thirds of these family members were women, especially mothers and stepmothers. This suggests women in Lao PDR experience relatively low levels of violence by non-partnered men. Qualitative work (from the Laotian National Commission for Advancement of Women, 2015) outlines four key triggers of IPV: (i) failure of a woman to meet the expectations of gender roles, which assert superiority of a husband; (ii) jealousy, mistrust or infidelity by the husband; (iii) alcohol or drug use; and (iv) financial stress. Notably, women’s participation in a PWP program could affect at least two of these triggers: it may violate gender roles and is likely to affect the financial situation of the household. 3. Road Management Group (RMG) Program and Experiment Design The RMG program emerged from an evaluation of the government’s Poverty Reduction Fund (PRF), a community-driven development program aimed at reducing poverty through infrastructure improvements in education, drinking water, irrigation, health, and transport. An evaluation carried out in 2015 showed that the PRF had been successful in improving infrastructure in rural areas but that post-completion, road quality deteriorated quickly. Under pressure from harsh rainy seasons and vehicular traffic, the access roads typically fell into extreme disrepair within a couple years and often became unusable. Aiming to provide a part-time supplementary income-earning opportunity to at-need households and to extend the lives of the access roads, a pilot Road Maintenance Group (RMG) program was established, in which female villagers from poor households would be trained and formed into work units to provide basic road maintenance services. Two eligibility criteria were used to recruit interested RMG members from each village: I. Candidates must belong to a poor household, based on PRF poverty ranking, assigned in 2016 and updated by village heads on the day of the selection. 4 Only one woman from each such household could participate in the program. 4 Dervisevic et. al. (2020) describe this process and confirm targeting is effective. 5 II. Candidates must be between 18 and 50 years old, though the upper and lower age limits were not strictly enforced, as long as the woman was deemed capable of performing the work. In every village, the number of eligible and interested women exceeded the number of available RMG jobs, with the excess demand varying across the villages. Therefore, a lottery was carried out to select RMG group members, with preference given to poorer women, according to their PRF rank. Specifically, PRF ranking divides all households into four groups: “poorest”, “poor”, “middle income” and “better off”. If there were enough interested women from the “poorest” households to fill the village’s RMG and waitlist spots, the lucky draw was restricted to women from these households only. 5 If there were not enough women from the “poorest” households, the lucky draw was opened up to households in the next poverty rank – “poor” – households, and so on. The pool then included women from both “poor” and “poorest” households, for example, and the resulting lottery meant that some waitlist women were “poorest” and some participants were “poor”. Thus, 339 women were selected to become RMG members and 843 women were waitlisted across 85 villages in 7 provinces. 6 The women were organized into RMGs of around three to five members, with group size depending on the length of road they were assigned to maintain; generally, one member per kilometer. The groups were tasked with carrying out routine road maintenance, such as clearing roads of vegetation, clearing the drainage system and making small repairs to the road surface. They were provided with simple training and basic hand tools. Payments were made monthly or quarterly, depending on the payment preference of each RMG, corresponding to the number of days worked. Typically, the RMGs worked a few days each month, with monthly variation based on maintenance needs as determined by the weather: the amount of work is generally higher during the wet season. All group members received the same wage, based on a fixed-daily rate set at slightly below the prevailing market wage in each village so as not to discourage private sector employment. The male rate was used because there were few employment opportunities for women; only 5% of the control group were regular earners. The PRF organized monthly road quality audits to assess the quality of the roads and make deduction to the wage payments in case of poor performance. Our respondents reported at endline that they did not experience any deductions. The road maintenance activities started in June 2018. The maintenance contracts ran for 18 months, from October 2018 up to the January 2020, covering two full wet seasons and providing an average of 75 5 The women did not necessarily need to be at the draw; they could be registered by another household member. But women who were selected for the RMG program had to be available for a training session over the two days following the draw. 6 Although most RMGs were intended to be drawn from only one village, some were drawn from two. This is why we ended up with a higher number of villages than road-segments. Further, we had targeted 87 villages when planning data collection but later learned that four villages were merged. Therefore, there were only 85 villages in our sample and not 87. 6 days of employment for each RMG member. This equates to just over four days of work per month. The RMG members were paid wages of around 60,000 kip/day (US$7.3/day), resulting in a total average income of around $550 per RMG member over the implementation period. Since this was a part-time activity, the expectation was that it provided supplementary income for participants. Indeed, the RMG payment corresponds to 40% of household income of waitlisted women. This is a meaningful increase in income: for comparison, conditional cash transfers received by the poorest decile in eight Latin American countries are in the range of 7.8% to 33.6% of household income (Amarante and Brun 2018). 4. Data Baseline data was collected between September and October 2018, before any salary payments were made. Women knew whether they were selected to be a part of a road maintenance group at the time of baseline data collection and had started work. We were able to interview 333 treatment households and women, and 813 control households and women. Not all women who were selected participated in the survey (6 treated households, 34 control households). We discuss attrition from selection to baseline data collection below. The baseline data allowed us to verify that treated and control samples are balanced across many characteristics. However, there are relatively fewer Hmong-Iumien women in the treated sample at baseline, and control group households are smaller and have somewhat higher monthly income 7 (Table A1 8 ). We control for these baseline characteristics in our estimates to account for potential bias from imbalance. Endline data collection took place in December 2019 and January 2020, 18 months after the program started and the end of the contract period. At endline, we were able to interview 323 treated households and 776 control households. There was attrition of 47 households (4 percent): 10 treated, 37 control. At both baseline and endline, we administered a household questionnaire to the person most knowledgeable about household matters, and an individual questionnaire to women enrolled as RMG and waitlist (control group) members. GBV experience was measured only at endline because we did not have enough time to pass a supplemental national ethics review for more sensitive data collection before the baseline survey. 9 A special audio-computer self-interviewing (ACASI) module was developed for administration of GBV questions in order to protect participants’ privacy. 10 Our team specifically designed 7 The baseline data collection took place after the work had started but before the first monthly payment had been made. It is plausible that we observe lower incomes in the treatment group because RMG women had to give up some other work in order to engage in RMGs. 8 Tables starting with A are in the Online Appendix. 9 We passed external IRB with HML IRB Research & Ethics and Lao PDR ethics review for both baseline and endline data collection. 10 It is common for neighbors, relatives, other family members to stop by when interviews are conducted. Given construction of the dwellings (very thin walls), even being in a closed room did not guarantee complete privacy, as one could hear the questions from outside or adjacent rooms. 7 the ACASI module to suit this study population’s needs, considering their illiteracy and minority languages. Questions were recorded on electronic tablets in multiple ethnic languages and participants were taught during survey administration how to select answers by pressing the appropriate color-coded symbols (e.g. green star = yes). A test module on non-sensitive questions included at the beginning of the GBV module confirmed that the participants had learned it correctly. Participants listened to the questions with headphones and selected their own answers, ensuring confidentiality. Studies that have compared the likelihood of revealing victimization experience suggest that the likelihood of reporting GBV is either higher or at the same level with ACASI, as with traditional face-to-face direct questioning, commonly used in demographic and health surveys (Cullen 2020; Dervisevic et al. 2020). 5. Empirical Approach The analytical approach was registered as a pre-analysis plan with 3ie RIDIE (study ID: RIDIE- STUDY-ID-5e61e44365ed9). This paper focuses only on the GBV outcomes, and we note deviations from the pre-analysis plan when they occur. We measure the impact of the RMG program on women and their households through intent-to-treat (ITT) and treatment on the treated (ToT) models. ITT allows us to estimate the population-level effectiveness of the intervention when participants may not comply perfectly with the intervention activities, while ToT identifies the results on those who engaged in the intervention and shows the efficacy that can be achieved if participants comply. We estimate the ITT effect on GBV through the following estimating equation: 1 = + 0 + 0 + (1) where 1 is the value of the outcome of interest at endline, 0 is a dummy that takes the value of 1 for the treatment group (RMG) as assigned at baseline, and 0 contains a set of baseline characteristics (age, marital status (1/0) ever went to school, ethnicity dummies, household size, winsorized household monthly income and baseline level outcome 0 , when available (as for labor outcomes confirming participation). In the case of perfect compliance with treatment assignment, we can interpret the ITT as a Treatment-on-Treated (ToT) effect and this would be equal to the Average Treatment Effect (ATE). However, in our case, 14 women dropped out of the program and were replaced by 12 women from the waitlist. We have information on the women who dropped out, which allows us to track actual treatment. This allows us to instrument for take-up using the random assignment in a two-stage regression to estimate a ToT. Since this is only relevant for women who take up the intervention, it should be interpreted as a 8 Local Average Treatment Effect (LATE). In the first stage, we estimate the likelihood of taking up the treatment based on (exogenous) assignment to treatment through the randomization process: = + + (2) where is an indicator that takes the value of 1 for women assigned to the RMG and 0 for women assigned to the waitlist, and is equal to 1 if the woman participated in RMG, and 0 otherwise. In the second stage, � we regress the outcome of interest on the predicted take-up 1 . We use robust standard errors in all specifications. Since randomization was at the individual level, standard errors are not clustered. However, we also run the estimations with errors clustered by village as a robustness check to account for potential variation in treatment intensity due to the varying number of RMG members per village, and for the possibility that RMG women may be subject to public violence as a group. Recall our sample suffered from attrition at two points: first, we lost 40 households from selection to baseline data collection, and then another 47 households from baseline to endline. We check whether attrition is affected by observable characteristics, and whether attrition varied by observable characteristics between treatment status by estimating: = 0 + 1 + ∑ =1 2 + ∑=1 3 + (3) where is a binary variable equal to 1 if individual i was lost to attrition, captures treatment or control status, and is a vector of observable baseline characteristics, such as age and marital status. Coefficients 2 capture whether attrition differs depending on the baseline characteristics. Coefficients 3 capture whether the likelihood of attrition, associated with a specific observable characteristic, is different depending on treatment status. We estimate equation 3 on two samples: the sample of all who participated in the lottery, and baseline sample. For the first sample, we have a very limited number of variables, which were recorded on the day of the lottery: PRF poverty ranking and women’s age. For the second sample, we include in the regression such characteristics as age, marital status, education, household size, household monthly income and ethnicity variables. Table 1 presents the results. We find no evidence of differential attrition from selection to baseline (Column 1). The RMG participants are not more or less likely to attrit between baseline and endline (Column 2). We also do not find evidence of differential attrition between treatment and control women for specific groups (p-value of F-test for joint significance of all interactions is 0.631). However, we reject the null that attritors and non-attritors are similar across the full set of observable characteristics (Column 2). Attritors are younger, less likely to be married, more likely to belong to ethnic minority groups and come from poorer households, as manifested by observable dwelling characteristics. We address potential bias 9 due to attrition, along with bias for non-response to be discussed later, by calculating Lee (2009) bounds and Kling-Liebman bounds (Kling & Liebman, 2004). 6. Results The 85 villages participating in the RMG program are among Lao PDR’s poorest and offer few job opportunities. According to the baseline data, only one-fifth of working-age individuals in RMG-eligible households had paid work outside the household. Over 70% of RMG-eligible women worked on the household farm or business in unpaid roles. When compared against rural households from a nationally representative Laos Expenditure and Consumption Survey (LECS) V, used for calculation of poverty rates in Lao PDR, the RMG-eligible households were poorer than the average Laotian rural households across a range of wealth indicators, including housing composition, ownership of durables, and nutrition consumption; wait-listed and RMG women are also less educated on average and less likely to be in the labor force (Baseline Report, East Asia & Pacific Gender Innovation Lab 2020). A supplementary survey confirmed that the program women are among the poorest in these already poor villages (Dervisevic et al. 2020). 6.1. Impact on Women’s Economic Situation The RMG program was meant to directly change labor market activities by giving participants a regularly paying job for 18 months. We first confirm that the intervention has worked as intended before testing how GBV was impacted. To do so, we estimate the impact of treatment on whether the women were listed among the household members that regularly contributed wage income, and on household average monthly wage income contributed over the 12 months preceding the endline survey. We find increases in regular work for RMG members. RMG members were 77.4% more likely than the waitlist group to be regular income earners (Table 2). Monthly income (winsorized at the 99th percentile to minimize the effects of outliers) increases by 173,558.24 kip ($23.56) for RMG households, over 40% greater than waitlist household income. We note that the value of the increase is only about 20% less than the amount of the payment to the RMG groups. This is notable within the PWP literature because other income generation is often displaced (Bertrand et al., 2017). Lee’s bounds estimates adjusting for attrition confirm positive outcomes (results not shown). 6.2. GBV We examine four types of IPV (controlling behavior, emotional violence, physical violence, and sexual violence), two types of non-IPV domestic violence (physical and sexual violence by household members other than intimate partner) and two types of GBV outside the home (physical and sexual). The single deviation from our pre-analysis plan regarding these outcomes is that we had initially grouped together the 10 non-IPV domestic violence into a single category. However, we decided to change this structure to improve parallelism with the other outcomes. As described in our pre-analysis plan, we constructed indicators of controlling behavior, emotional violence, physical violence and sexual violence as dummy variables, equal to 1 if any question about exposure to the specific type of violence in that category during the previous 12 months received an affirmative response. This structure of outcome measure is used worldwide (UN Department of Economic and Social Affairs Statistical Office 2014). The indicators of physical or sexual violence by a household member were each generated by a single question. The indicators of physical or sexual violence by a non- household member included responses to questions regarding violence from any of the following individuals: friend or family, authority, random stranger, and other. If any question was answered within the set of questions comprising the index, we included that woman in the analysis. However, if she did not answer any of the questions within the category, we considered this a non-response and did not include her in the analysis of that particular index. Summary statistics from the “non-empowered” control group indicate that just under one half of women reported experiencing controlling behavior by their spouse, just under one third reported emotional and physical violence, and just under one fifth reported sexual violence (Table 3). These levels of IPV are significantly higher than those reported in the Lao National Survey on Women’s Health and Life Experiences, where 10%, 4% and 3.1% of respondents reported exposure to emotional violence, physical and sexual violence respectively in the last 12 months prior to the interview. Our data exceeds these rates by a factor of seven for physical violence, a factor of six for sexual violence, and by a factor of 3 for emotional violence. There are several possibilities that explain this discrepancy. Firstly, these poor, rural women may have a higher experience of violence than the national average. Secondly, and, perhaps, most importantly, more women may have felt comfortable revealing their experience of violence using our ACASI interviewing technique rather than the in-person approach used in the national survey. Summary statistics on non-IPV GBV also indicate higher levels than reported on the national survey. Between 7% and 12% of respondents reported having experienced physical or sexual violence, either inside the home with a non-partner or from someone outside the home (Table 3). These numbers are at least twice as high as the national prevalence rates reported in the Lao National Survey on Women’s Health and Life Experiences, for physical and sexual violence from a non-partner: at 5.1% and 5.3%, respectively.11 We note that due to the sensitive nature of these questions, non-response is high, as is typical for surveys on IPV (UN Department of Economic and Social Affairs Statistical Office 2014). For our IPV variables, non-response reached over 40% for some of the questions regarding controlling behavior, while the non- 11 The Lao National Survey on Women’s Health and Life Experiences does not collect data on exposure to non- partner physical and sexual violence over the last 12 months. 11 response rate was as low as 16% for some less conceptual questions on IPV. For the non-partner violence, non-response ranged from 13-21%. For each question, we tested whether non-response was balanced across RMG participants and waitlist women (Table 3) and did not find any significant differences. When analyzing the data, we drop observations only if all responses in a category of violence (e.g. emotional IPV) are missing. The main results show that none of the impacts of the program on violence — either IPV or non-IPV GBV — significantly differed from zero (Tables 4 and 5). We employed a variety of robustness checks that were not included in the pre-analysis plan. We first consider a different index construction: the percentage of affirmative responses in each category. For each woman, we calculate the percentage of affirmative responses, excluding questions in the category for which she did not respond. Results are very similar to our main specification (Tables A2 and A3). We also examined each question individually, reweighting the sample with inverse probability weights (Tables A4-A9). We find one significant result for controlling behaviors: the likelihood that a husband needs to know where the woman is increases due to RMG participation, and for sexual violence: the likelihood of being threatened into sexual acts increases due to participation in the program (Tables A4-A9). However, these are two indicators out of 30, which may happen by chance. We also considered a variation in combining the response to individual questions within categories using the index proposed by Kling et. al (2007) to account for correlation between the responses (Tables A10 and A11). Results from this specification do not reject the hypothesis that the PWP program had no impact on women’s experience of violence, although the confidence intervals are wider. We grouped the outcomes of each type (IPV & non-IPV GBV) into two indices indicating “ever experienced” for each type, but this specification also had null results (Table A12). Also yielding null results, for the IPV outcomes we considered the smaller subsample of partnered women and controlled for their partner’s characteristics (partner participates in the labor force, has ever attended school, and age. 12 In addition to the attrition concerns mentioned earlier, we note that a number of women declined to answer (see summary statistics table 3). In our main results we dropped observations when all responses in a category of violence, such as physical or emotional IPV were missing. We also apply Lee’s bounds to determine the range of possible estimates from the missing responses being either all yes or all no. For each individual question, we imputed missing responses to be all affirmative in the treatment group and all negative in the control group. The resulting index yielded the upper bound. We did the reverse to calculate the lower bound. Among IPV variables, the upper bound for controlling behavior is significant, as is exposure to sexual violence in the household by a non-partner (Tables A13 and A14). When we apply Kling & Liebman bounds, statistically significant results for the upper bounds suggest that IPV increases; 12 Available upon request. 12 however, only controlling behavior has a lower bound that is also positive; the bounds of the other IPV variables include zero (Table A15). The upper and lower bounds for the GBV estimates are all significant, but they also include zero (Table A16). Overall, the results from the bounding exercises using extreme assumptions regarding non-response do not generate strong evidence that calls into question conclusion of our core specification. Our pre-analysis plan included heterogeneity analyses along several dimensions that have been linked to women’s status in the home and labor opportunities: ethnicity, income, age, education, education relative to the spouse’s, empowerment index, household decision making, and social contact. We only found one significant result (the most educated women who participated experienced less non-IPV GBV) that did not stand up to multiple hypothesis testing when examining the two outcomes ever-IPV and ever non-IPV GBV (results available upon request). 7. Conclusion PWPs, like cash transfer programs, are an important tool in creating social protection and increasing the well-being of vulnerable populations that are typically poorer and have less access to the traditional labor market. Unlike cash-transfer programs, public work programs require participants to be employed, often in labor intensive jobs, in order to receive a transfer. This requirement may challenge gender norms. While a consensus has emerged that cash transfers are likely to reduce IPV, at least on average (Buller et al., 2019), there is a concern that PWPs may have unintended negative consequences on the incidence of GBV both inside and outside the home. To the best of our knowledge, we are the first to investigate the causal relationship of PWPs that target women, and GBV (both IPV, domestic violence and GBV outside the household). The randomized control trial design of the Lao PDR Road Management Group program confirms that treatment increases women’s employment in wage work and their earnings. We do not find that participation in the program changes their exposure to GBV. There are a variety of possible explanations for the lack of impact on GBV with respect to specific program components. First, though women were bringing in more income, the PWP wage was lower than the male wage. Thus, men may not have felt like their economic power was being challenged. GBV may not have increased because the work was done in groups of women; harassment is easier if there are no bystanders. Additionally, men were not part of the work groups, so harassment from co-workers was not a possibility. The communities may be very aware of the benefits of the roads, which were only recently improved; this social commitment to maintaining a public good may also have protected the women. Although our findings are reassuring, future PWPs should still analyze how their program design may influence gender divisions both inside and outside the home before and after implementation. 13 Our results are reassuring for those concerned with the potential negative effects of PWPs on GBV: we do not find evidence that participation in PWPs may endanger women through increased risk of violence either from their partners and other household members, or the public at large in the position of authority. On the other hand, unlike in the literature on cash transfers, we did not find that PWPs lowered IPV. Considering evidence that add-on interventions to cash transfer programs, such as behavioral change components, are likely to improve their potential to lower GBV (Roy et al., 2017), policy makers may consider integrating additional components into the design of future PWP programs. 14 References Aizer, Anna. 2010. “The Gender Wage Gap and Domestic Violence.” The American Economic Review 100 (4): 1847–59. https://doi.org/10.1257/aer.100.4.1847. Akerlof, George A., and Rachel E. Kranton. 2000. “Economics and Identity.” The Quarterly Journal of Economics 115 (3): 715–53. https://doi.org/10.1162/003355300554881. Amaral, Sofia. 2019. “Do Improved Property Rights Decrease Violence against Women? Evidence from India.” manuscript. Amaral, Sofia, Siddhartha Bandyopadhyay, and Rudra Sensarma. 2015. “Employment Programmes for the Poor and Female Empowerment: The Effect of NREGS on Gender-Based Violence in India.” Journal of Interdisciplinary Economics 27 (2): 199–218. https://doi.org/10.1177/0260107915582295. Amarante, Verónica, and Martin Brun. 2018. “Cash Transfers in Latin America: Effects on Poverty and Redistribution.” Economia 19 (1): 1–31. Bloch, F., and V. Rao. 2002. “Terror as a Bargaining Instrument: A Case Study of Dowry Violence in Rural India.” American Economic Review 92 (4): 1029–43. Bobonis, Gustavo J., Melissa González-Brenes, and Roberto Castro. 2013. “Public Transfers and Domestic Violence: The Roles of Private Information and Spousal Control.” American Economic Journal: Economic Policy 5 (1): 179–205. https://doi.org/10.2307/23358341. Buller, Ana Maria, Amber Peterman, Meghna Ranganathan, Alexandra Bleile, Melissa Hidrobo, and Lori Heise. 2018. “A Mixed-Method Review of Cash Transfers and Intimate Partner Violence in Low- and Middle-Income Countries.” The World Bank Research Observer 33 (2): 218–58. https://doi.org/10.1093/wbro/lky002. Bulte, Erwin, and Robert Lensink. 2019. “Women’s Empowerment and Domestic Abuse: Experimental Evidence from Vietnam.” European Economic Review 115 (C): 172–91. Chin, Yoo-Mi. 2012. “Male Backlash, Bargaining, or Exposure Reduction?: Women’s Working Status and Physical Spousal Violence in India.” Journal of Population Economics 25 (1): 175–200. https://doi.org/10.1007/s00148-011-0382-8. Corradi, Consuelo, Chaime Marcuello-Servós, Santiago Boira, and Shalva Weil. 2016. “Theories of Femicide and Their Significance for Social Research.” Current Sociology 64 (7): 975–95. https://doi.org/10.1177/0011392115622256. Cullen, Claire. 2020. “Method Matters: Underreporting of Intimate Partner Violence in Nigeria and Rwanda.” SSRN Scholarly Paper ID 3624515. Rochester, NY: Social Science Research Network. https://papers.ssrn.com/abstract=3624515. Davila, Manuel Alejandro Estefan. 2019. “Essays in Development Economics.” Doctoral Thesis, UCL (University College London). Doctoral, University College London. https://discovery.ucl.ac.uk/id/eprint/10076088/. Dervisevic, Ervin, Seth Garz, Aneesh Mannava, and Elizaveta Perova. 2020. In Light of What They Know: How Do Local Leaders Make Targeting Decisions? Policy Research Working Papers. The World Bank. https://doi.org/10.1596/1813-9450-9465. East Asia & Pacific Gender Innovation Lab. 2020. “Impact Evaluation of Laos Road Maintenance Groups Program.” The World Bank. Kling, Jeffrey R., Jeffrey B. Liebman, and Lawrence F. Katz. 2007. “Experimental Analysis of Neighborhood Effects.” Econometrica 75 (1): 83–119. https://doi.org/10.1111/j.1468-0262.2007.00733.x. 15 Kotsadam, Andreas, and Espen Villanger. 2020. “Jobs and Intimate Partner Violence - Evidence from a Field Experiment in Ethiopia.” SSRN Scholarly Paper ID 3541457. Rochester, NY: Social Science Research Network. https://papers.ssrn.com/abstract=3541457. Macmillan, Ross, and Rosemary Gartner. 1999. “When She Brings Home the Bacon: Labor-Force Participation and the Risk of Spousal Violence against Women.” Journal of Marriage and Family 61 (4): 947–58. https://doi.org/10.2307/354015. Manser, Marilyn, and Murray Brown. 1980. “Marriage and Household Decision-Making: A Bargaining Analysis.” International Economic Review 21 (1): 31–44. https://doi.org/10.2307/2526238. Pillinger, Jane. 2017. “Violence and Harassment against Women and Men in the World of Work.” International Labour Organization. Roy, Shalini, Melissa Hidrobo, John Hoddinott, Bastien Koch, and Akhter Ahmed. 2019. Can Transfers and Behavior Change Communication Reduce Intimate Partner Violence Four Years Post-Program? Experimental Evidence from Bangladesh. Internatinal Food Policy Research Institute. https://doi.org/10.2499/p15738coll2.133421. UN Department of Economic and Social Affairs Statistical Office. 2014. Guidelines for Producing Statistics on Violence Against Women: Statistical Surveys. United Nations. 16 Table 1 - Attrition Attrition Attrition (Lottery -> Baseline) (Baseline -> Endline) RMG -0.013 (0.048) 0.245 (0.161) Age -0.001 (0.001) -0.001* (0.001) Poverty Ranking -0.006 (0.007) RMG x Age 0.001 (0.001) -0.000 (0.001) RMG x Poverty Ranking -0.009 (0.014) Married -0.048** (0.021) Ever Attended School 0.017 (0.015) Ethnicity: Mon-Khmer 0.050** (0.020) Ethnicity: Tibeto-Burman 0.013 (0.040) Ethnicity: Hmong-Mien & Other 0.094*** (0.035) Household Size -0.002 (0.003) Sqm of residential land -0.000 (0.000) House has brick wall 0.028 (0.033) House has solid floor -0.010 (0.027) Number of rooms in house 0.002 (0.007) House has piped water -0.053*** (0.019) House has toilet 0.017 (0.017) House has outdoor roofed kitchen -0.009 (0.015) House has electricity -0.033* (0.019) Number of cell phones in HH 0.007 (0.006) HH owns a vehicle -0.061 (0.047) HH owns a motorcycle -0.015 (0.017) HH owns a refrigerator/freezer 0.004 (0.031) HH owns a steam rice cooker 0.005 (0.033) HH owns an electric rice cooker 0.079 (0.102) HH owns a tractor -0.022 (0.020) HH owns a rice mill 0.025 (0.017) HH owns a TV 0.008 (0.018) Table continues in next page 17 Attrition Attrition (Lottery -> Baseline) (Baseline -> Endline) Table continues from previous page RMG x Married 0.100** (0.042) RMG x Ever Attended School -0.013 (0.029) RMG x Ethnicity: Mon-Khmer -0.052 (0.037) RMG x Ethnicity: Tibeto-Burman -0.031 (0.069) RMG x Ethnicity: Hmong-Mien & Other -0.100 (0.062) RMG x Household Size -0.006 (0.005) RMG x Sqm of residential land 0.000 (0.000) RMG x House has brick wall -0.047 (0.059) RMG x House has solid floor 0.010 (0.049) RMG x Number of rooms in house 0.016 (0.014) RMG x House has piped water -0.008 (0.037) RMG x House has toilet -0.013 (0.032) RMG x House has outdoor roofed 0.027 (0.028) kitchen RMG x House has electricity -0.002 (0.037) RMG x Number of cell phones in HH -0.004 (0.012) RMG x HH owns a vehicle 0.101 (0.087) RMG x HH owns a motorcycle -0.014 (0.031) RMG x HH owns a refrigerator/freezer -0.019 (0.056) RMG x HH owns a steam rice cooker -0.008 (0.062) RMG x HH owns an electric rice cooker -0.268* (0.139) RMG x HH owns a tractor -0.022 (0.037) RMG x HH owns a rice mill -0.018 (0.034) RMG x HH owns a TV -0.003 (0.033) Constant 0.069*** (0.026) 0.041 (0.109) Adj R-sq 0.001 0.018 p-val. (RMG) 0.788 0.128 F-stat (controls) 0.834 1.945 p-val. 0.435 0.004*** F-stat (interaction) 0.358 0.880 p-val. 0.699 0.631 N 1182 1146 18 Table 2 - The effect of treatment on women's labor market activities ITT (OLS) Regular earner Average Regular Income, monthly (HN), Winsor RMG 0.774*** 173,558.246*** (0.023) (7392.307) Control Group Mean 0.050 16,249.033 N (Obs.) 1,099 1,099 Adj R-sq 0.634 0.350 p-val 0.000 0.000 ToT (2SLS) Regular earner Average Regular Income, monthly (HN), Winsor RMG 0.807*** 181,147.937*** (0.022) (7384.186) Control Group Mean 0.048 16,859.697 N (Obs.) 1,099 1,099 Adj R-sq 0.676 0.374 p-val 0.000 0.000 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao- Tai, Mon-Khmer, Sino-Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. Outcomes are (1) binary indicator of whether she is a regular earner, and (2) her average regular monthly income in Laotian Kip. USD 1 = LAK 9, 350 at time of writing. * p < 0.1, ** p < .05, *** p < 0.01. 19 Table 3 - Summary Statistics for Gender-Based Violence, exposure in the last 12 months Difference in Control Treatment Non-response Type of Intimate Partner Violence No No Yes No N Yes No N Resp. Resp. Diff. p-val. (%) (%) (Obs.) (%) (%) (Obs.) (%) (%) EXPOSURE TO CONTROLLING BEHAVIOR 49.29 46.05 4.66 708 54.64 39.52 5.84 291 -1.181 0.437 Ever experienced jealousy 18.22 49.01 32.77 708 20.62 47.42 31.96 291 0.810 0.804 Ever accused of unfaithfulness 14.12 51.41 34.46 708 16.49 50.17 33.33 291 1.130 0.732 Ever socialisation restricted 19.63 48.59 31.78 708 19.93 43.30 36.77 291 -4.990 0.128 Ever family contact restricted 14.69 42.94 42.37 708 14.43 39.52 46.05 291 -3.675 0.287 Ever husband must know location 23.16 38.98 37.85 708 26.46 31.27 42.27 291 -4.415 0.194 Ever household spending restricted 14.69 42.51 42.80 708 13.40 41.92 44.67 291 -1.877 0.587 Ever not trusted with money 20.06 50.71 29.24 708 21.65 44.67 33.68 291 -4.440 0.167 EXPOSURE TO EMOTIONAL VIOLENCE 32.06 60.73 7.20 708 31.62 60.82 7.56 291 -0.357 0.844 Ever publicly embarrassed 17.66 62.85 19.49 708 20.27 57.73 21.99 291 -2.502 0.372 Ever threatened 17.80 60.45 21.75 708 19.59 58.08 22.34 291 -0.585 0.839 Ever insulted/degraded 21.75 59.32 18.93 708 19.93 58.08 21.99 291 -3.067 0.270 EXPOSURE TO PHYSICAL VIOLENCE 31.78 65.11 3.11 708 33.33 62.20 4.47 291 -1.360 0.289 Ever pushed/shaken 17.23 63.70 19.07 708 20.27 62.89 16.84 291 2.229 0.409 Ever slapped 14.41 63.84 21.75 708 17.18 60.48 22.34 291 -0.585 0.839 Ever arm twisted/hair pulled 12.71 65.68 21.61 708 13.75 70.10 16.15 291 5.459 0.050 Ever punched/hit 14.41 68.22 17.37 708 15.46 63.23 21.31 291 -3.933 0.146 Ever kicked/beaten 13.14 69.07 17.80 708 16.15 64.60 19.24 291 -1.447 0.591 Ever choked/burned 11.86 71.33 16.81 708 12.37 71.13 16.49 291 0.313 0.904 Ever threatened with weapon 11.44 71.89 16.67 708 13.40 70.45 16.15 291 0.515 0.842 EXPOSURE TO SEXUAL VIOLENCE 19.35 72.60 8.05 708 23.37 70.45 6.19 291 1.865 0.310 Ever forced sex 12.15 67.66 20.20 708 12.71 71.13 16.15 291 4.047 0.139 Ever forced non-sex sexual acts 10.88 68.36 20.76 708 13.40 64.95 21.65 291 -0.887 0.755 Ever threatened into sex 10.45 70.20 19.35 708 16.49 67.01 16.49 291 2.855 0.292 20 Difference in Control Treatment Non-response Type of Non-Intimate Partner Violence No No Yes No N Yes No N Resp. Resp. Diff. p-val. (%) (%) (Obs.) (%) (%) (Obs.) (%) (%) PHYSICAL VIOLENCE BY OTHER HH MEMBER 10.14 69.86 20.00 720 8.67 70.67 20.67 300 -0.667 0.809 EXPOSURE TO PHYSICAL VIOLENCE, OUTSIDE HH 12.50 69.03 18.47 720 11.33 69.00 19.67 300 -1.194 0.657 Physical violence by friend/family 8.47 72.22 19.31 720 7.67 71.67 20.67 300 -1.361 0.619 Physical violence by authority 9.17 71.25 19.58 720 7.33 71.67 21.00 300 -1.417 0.607 Physical violence by random stranger 8.33 71.39 20.28 720 8.67 70.33 21.00 300 -0.722 0.795 Physical violence by other type of individual 8.19 71.94 19.86 720 8.67 71.00 20.33 300 -0.472 0.864 FORCED SEX BY OTHER HH MEMBER 7.64 75.00 17.36 720 8.67 76.67 14.67 300 2.694 0.292 EXPOSURE TO SEXUAL VIOLENCE, OUTSIDE HH 12.36 71.67 15.97 720 13.00 73.67 13.33 300 2.639 0.285 Forced sex by friend/family 8.47 74.03 17.50 720 9.00 75.33 15.67 300 1.833 0.478 Forced sex by authority 9.31 73.47 17.22 720 9.00 76.33 14.67 300 2.556 0.317 Forced sex by random stranger 8.06 74.58 17.36 720 8.33 76.00 15.67 300 1.694 0.511 Forced sex by other type of individual 8.06 74.58 17.36 720 8.00 76.33 15.67 300 1.694 0.511 21 Table 4 - Effect of treatment on GBV: IPV in the past 12 months ITT (OLS) Exposure to Exposure to Exposure to Exposure to Sexual Controlling Emotional Violence Physical Violence Violence Behavior RMG 0.057 -0.000 0.020 0.033 (0.036) (0.035) (0.033) (0.030) Control Group Mean 0.517 0.346 0.328 0.210 N (Obs.) 949 926 964 924 Adj R-sq 0.006 -0.003 0.038 0.039 p-val 0.111 0.991 0.551 0.285 ToT (2SLS) Exposure to Exposure to Exposure to Exposure to Sexual Controlling Emotional Violence Physical Violence Violence Behavior RMG 0.057 -0.000 0.020 0.033 (0.036) (0.034) (0.033) (0.030) Control Group Mean 0.521 0.343 0.324 0.211 N (Obs.) 949 926 964 924 Adj R-sq 0.006 -0.003 0.038 0.039 p-val 0.109 0.991 0.549 0.282 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary outcomes where the outcome equals 1 if she has experienced this type of violence in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01. 22 Table 5 - Effect of treatment on GBV: Non-IPV GBV in the past 12 months ITT (OLS) Exposure to Phys. Exposure to Sexual Exposure to Phys. Exposure to Sexual Violence, in HH Violence, in HH Violence, outside Violence, outside non partner non partner HH HH RMG -0.016 0.007 -0.012 -0.003 (0.024) (0.022) (0.027) (0.026) Control Group Mean 0.124 0.091 0.153 0.147 N (Obs.) 828 865 828 865 Adj R-sq 0.025 0.023 0.058 0.051 p-val 0.508 0.737 0.644 0.896 ToT (2SLS) Exposure to Phys. Exposure to Sexual Exposure to Phys. Exposure to Sexual Violence, in HH Violence, in HH Violence, outside Violence, outside non partner non partner HH HH RMG -0.016 0.007 -0.012 -0.003 (0.024) (0.022) (0.026) (0.026) Control Group Mean 0.122 0.094 0.151 0.149 N (Obs.) 828 865 828 865 Adj R-sq 0.025 0.023 0.058 0.051 p-val 0.506 0.736 0.641 0.895 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary outcomes where the outcome equals 1 if she has experienced this type of violence in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01. 23 Appendix Table A1 - Baseline characteristics by treatment status Mean Mean Difference p-val (Control) (Treated) Age 33.256 32.718 0.538 0.432 Married (Dummy) 0.868 0.889 -0.021 0.342 Ever attended school (Dummy) 0.590 0.553 0.038 0.239 Education level: None 0.410 0.447 -0.038 0.239 Education level: Primary & lower 0.476 0.438 0.038 0.247 Education level: Lower Secondary 0.102 0.093 0.009 0.644 Education level: Upper Secondary + 0.012 0.021 -0.009 0.268 Ethnicity: Lao-Tai 0.225 0.222 0.003 0.916 Ethnicity: Mon-Khmer 0.662 0.619 0.043 0.165 Ethnicity: Tibeto-Burman 0.044 0.057 -0.013 0.359 Ethnicity: Hmong-Mien & Other 0.069 0.102 -0.033 0.058* Household Size 5.846 6.282 -0.436 0.029** Sqm of residential land 231.634 256.694 -25.060 0.107 House has brick wall 0.129 0.129 0.000 0.999 House has solid floor 0.196 0.192 0.003 0.896 Number of rooms in house 1.913 1.928 -0.015 0.833 House has piped water 0.180 0.162 0.017 0.481 House has toilet 0.416 0.420 -0.005 0.884 House has outdoor roofed kitchen 0.486 0.444 0.041 0.203 House has electricity 0.795 0.826 -0.031 0.227 Number of cell phones in HH 1.386 1.483 -0.097 0.284 HH owns a vehicle 0.023 0.027 -0.004 0.716 HH owns a motorcycle 0.588 0.607 -0.019 0.560 HH owns a refrigerator/freezer 0.077 0.087 -0.010 0.588 HH owns a steam rice cooker 0.062 0.054 0.007 0.628 HH owns an electric rice cooker 0.995 0.985 0.010 0.079* HH owns a tractor 0.231 0.231 0.000 1.000 HH owns a rice mill 0.276 0.279 -0.004 0.897 HH owns a TV 0.326 0.297 0.029 0.345 N (Obs.) 813 333 24 Table A2 - Percent of positive responses in each category of GBV: IPV ITT (OLS) Pct Positive Pct Positive Pct Positive Pct Positive Response to Response to Response to Response to Sexual Controlling Emotional Violence Physical Violence Violence Behavior RMG 0.017 0.015 0.022 0.022 (0.024) (0.026) (0.022) (0.023) Control Group Mean 0.261 0.223 0.163 0.139 N (Obs.) 949 926 964 924 Adj R-sq 0.008 0.007 0.033 0.039 p-val 0.462 0.573 0.333 0.334 Bonferroni (α = .05) 0.013 0.013 0.013 0.013 ToT (2SLS) Pct Positive Pct Positive Pct Positive Pct Positive Response to Response to Response to Response to Sexual Controlling Emotional Violence Physical Violence Violence Behavior RMG 0.017 0.015 0.022 0.022 (0.024) (0.026) (0.022) (0.022) Control Group Mean 0.259 0.219 0.161 0.140 N (Obs.) 949 926 964 924 Adj R-sq 0.008 0.007 0.033 0.039 p-val 0.459 0.571 0.330 0.331 Bonferroni (α = .05) 0.013 0.013 0.013 0.013 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are the proportion of questions she answered affirmatively to in the category. * p < 0.1, ** p < .05, *** p < 0.01. 25 Table A3 - Percent of positive responses in each category of GBV: Non-IPV GBV ITT (OLS) Pct Positive Pct Positive Pct Positive Pct Positive Response to Response to Response to Sexual Response to Sexual Physical Violence, Physical Violence, Violence, in HH Violence, outside in HH non partner outside HH non partner HH RMG -0.016 -0.003 0.007 0.004 (0.024) (0.008) (0.022) (0.009) Control Group Mean 0.124 0.043 0.091 0.042 N (Obs.) 828 827 865 865 Adj R-sq 0.025 0.056 0.023 0.044 p-val 0.508 0.750 0.737 0.645 Bonferroni (α = .05) 0.013 0.013 0.013 0.013 ToT (2SLS) Pct Positive Pct Positive Pct Positive Pct Positive Response to Response to Response to Sexual Response to Sexual Physical Violence, Physical Violence, Violence, in HH Violence, outside in HH non partner outside HH non partner HH RMG -0.016 -0.003 0.007 0.004 (0.024) (0.008) (0.022) (0.009) Control Group Mean 0.122 0.042 0.094 0.042 N (Obs.) 828 827 865 865 Adj R-sq 0.025 0.056 0.023 0.044 p-val 0.506 0.749 0.736 0.643 Bonferroni (α = .05) 0.013 0.013 0.013 0.013 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are the proportion of questions she answered affirmatively to in the category. * p < 0.1, ** p < .05, *** p < 0.01. 26 Table A4 - Individual outcomes: Controlling Behavior ITT (OLS) Ever experienced Ever accused of Ever socialization Ever family Ever husband Ever household Ever not trusted jealousy unfaithfulness restricted contact restricted must know spending with money location restricted RMG 0.029 0.026 0.016 -0.004 0.087* -0.018 0.038 (0.039) (0.036) (0.041) (0.042) (0.045) (0.041) (0.039) Control Group Mean 0.271 0.216 0.288 0.255 0.373 0.257 0.283 N (Obs.) 674 658 667 565 608 566 694 Adj R-sq 0.010 0.014 0.016 0.025 0.019 0.018 0.016 p-val 0.459 0.481 0.688 0.914 0.057 0.656 0.325 ToT (2SLS) Ever experienced Ever accused of Ever socialization Ever family Ever husband Ever household Ever not trusted jealousy unfaithfulness restricted contact restricted must know spending with money location restricted RMG 0.029 0.026 0.016 -0.004 0.087* -0.018 0.038 (0.038) (0.036) (0.040) (0.041) (0.045) (0.040) (0.039) Control Group Mean 0.268 0.211 0.284 0.253 0.369 0.251 0.281 N (Obs.) 674 658 667 565 608 566 694 Adj R-sq 0.010 0.014 0.016 0.025 0.019 0.018 0.016 p-val 0.456 0.477 0.685 0.914 0.054 0.653 0.322 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino-Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary indicators equal to 1 if she has experienced the individual type of GBV in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01 27 Table A5 - Individual outcomes: Emotional Violence ITT (OLS) Ever publicly Ever threatened Ever embarrassed insulted/degraded RMG 0.039 0.024 -0.007 (0.035) (0.034) (0.035) Control Group Mean 0.219 0.227 0.268 N (Obs.) 797 780 801 Adj R-sq 0.016 0.008 0.010 p-val 0.264 0.475 0.848 ToT (2SLS) Ever publicly Ever threatened Ever embarrassed insulted/degraded RMG 0.039 0.024 -0.007 (0.034) (0.034) (0.034) Control Group Mean 0.215 0.221 0.264 N (Obs.) 797 780 801 Adj R-sq 0.016 0.008 0.010 p-val 0.261 0.472 0.847 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino-Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary indicators equal to 1 if she has experienced the individual type of GBV in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01 28 Table A6 - Individual outcomes: Physical Violence ITT (OLS) Ever Ever slapped Ever arm Ever punched/hit Ever Ever Ever threatened pushed/shaken twisted/hair pulled kicked/beaten choked/burned with weapon RMG 0.033 0.040 -0.004 0.022 0.038 0.006 0.022 (0.032) (0.032) (0.029) (0.031) (0.030) (0.027) (0.028) Control Group Mean 0.213 0.184 0.162 0.174 0.160 0.143 0.137 N (Obs.) 815 780 799 814 817 832 834 Adj R-sq 0.023 0.013 0.022 0.033 0.034 0.039 0.033 p-val 0.307 0.206 0.903 0.470 0.210 0.813 0.418 ToT (2SLS) Ever Ever slapped Ever arm Ever punched/hit Ever Ever Ever threatened pushed/shaken twisted/hair pulled kicked/beaten choked/burned with weapon RMG 0.033 0.040 -0.004 0.022 0.038 0.006 0.022 (0.032) (0.032) (0.029) (0.030) (0.030) (0.027) (0.027) Control Group Mean 0.211 0.186 0.160 0.170 0.156 0.141 0.140 N (Obs.) 815 780 799 814 817 832 834 Adj R-sq 0.023 0.013 0.022 0.033 0.034 0.039 0.033 p-val 0.304 0.203 0.903 0.467 0.207 0.812 0.415 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino-Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary indicators equal to 1 if she has experienced the individual type of GBV in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01 29 Table A7 - Individual outcomes: Sexual Violence ITT (OLS) Forced sexual Forced other sexual Ever threatened into intercourse acts sex RMG -0.007 0.029 0.061** (0.028) (0.029) (0.029) Control Group Mean 0.152 0.137 0.130 N (Obs.) 809 789 814 Adj R-sq 0.031 0.022 0.045 p-val 0.794 0.322 0.036 ToT (2SLS) Forced sexual Ever other sexual Ever threatened into intercourse acts sex RMG -0.007 0.029 0.061** (0.028) (0.029) (0.029) Control Group Mean 0.151 0.138 0.134 N (Obs.) 809 789 814 Adj R-sq 0.031 0.022 0.045 p-val 0.793 0.319 0.035 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino-Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary indicators equal to 1 if she has experienced the individual type of GBV in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01 30 Table A8 - Individual outcomes: Non-IPV Physical Violence ITT (OLS) Physical violence Physical violence Physical violence Physical violence Physical violence by other HH by friend/family by authority by random stranger by other type of member individual RMG -0.017 -0.009 -0.023 0.007 0.006 (0.024) (0.023) (0.023) (0.024) (0.023) Control Group Mean 0.127 0.105 0.114 0.105 0.102 N (Obs.) 814 819 816 811 816 Adj R-sq 0.028 0.027 0.032 0.034 0.035 p-val 0.476 0.699 0.321 0.781 0.789 ToT (2SLS) Physical violence Physical violence Physical violence Physical violence Physical violence by other HH by friend/family by authority by random stranger by other type of member individual RMG -0.017 -0.009 -0.023 0.007 0.006 (0.024) (0.023) (0.023) (0.024) (0.023) Control Group Mean 0.124 0.103 0.112 0.104 0.100 N (Obs.) 814 819 816 811 816 Adj R-sq 0.028 0.027 0.032 0.034 0.035 p-val 0.473 0.698 0.318 0.780 0.787 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino-Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary indicators equal to 1 if she has experienced the individual type of GBV in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01 31 Table A9 - Individual outcomes: Non-IPV Sexual Violence ITT (OLS) Forced sex by other Forced sex by Forced sex by Forced sex by Forced sex by other HH member friend/family authority random stranger type of individual RMG 0.008 -0.002 -0.013 -0.004 -0.006 (0.022) (0.023) (0.023) (0.023) (0.022) Control Group Mean 0.092 0.103 0.112 0.097 0.097 N (Obs.) 851 847 852 848 848 Adj R-sq 0.025 0.034 0.042 0.029 0.040 p-val 0.727 0.940 0.575 0.873 0.800 ToT (2SLS) Forced sex by other Forced sex by Forced sex by Forced sex by Forced sex by other HH member friend/family authority random stranger type of individual RMG 0.008 -0.002 -0.013 -0.004 -0.006 (0.022) (0.023) (0.023) (0.022) (0.022) Control Group Mean 0.096 0.103 0.113 0.102 0.096 N (Obs.) 851 847 852 848 848 Adj R-sq 0.025 0.034 0.042 0.029 0.040 p-val 0.725 0.940 0.573 0.872 0.799 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino-Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary indicators equal to 1 if she has experienced the individual type of GBV in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01 32 Table A10 - Effect of treatment on GBV: KKL Indexes of Exposure to IPV ITT (OLS) KKL Index of KKL Index of KKL Index of KKL Index of Controlling Emotional Violence Physical Violence Sexual Violence Behaviour Experience Experience Experience Experience RMG 0.039 0.037 0.057 0.065 (0.053) (0.063) (0.060) (0.065) Control Group Mean -0.033 -0.029 -0.010 -0.006 N (Obs.) 949 926 964 924 Adj R-sq 0.008 0.007 0.033 0.039 p-val 0.465 0.556 0.344 0.321 Bonferroni (α = .05) 0.013 0.013 0.013 0.013 ToT (2SLS) KKL Index of KKL Index of KKL Index of KKL Index of Controlling Emotional Violence Physical Violence Sexual Violence Behaviour Experience Experience Experience Experience RMG 0.039 0.037 0.057 0.065 (0.053) (0.062) (0.060) (0.065) Control Group Mean -0.037 -0.039 -0.014 -0.003 N (Obs.) 949 926 964 924 Adj R-sq 0.008 0.007 0.033 0.039 p-val 0.462 0.553 0.341 0.318 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are indexes of the form specified by Kling, et.al. (2007) for each category of GBV in the past 12 months. * p < 0.1, ** p < .05, *** p < 0.01 33 Table A11 - Effect of treatment on GBV: KKL Indexes of Exposure to non-IPV GBV ITT (OLS) KKL Index of KKL Index of KKL Index of KKL Index of Physical Violence, Physical Violence, Sexual Violence, in Sexual Violence, in HH non partner outside HH HH non partner outside HH RMG -0.053 -0.006 0.026 -0.000 (0.074) (0.071) (0.075) (0.069) Control Group Mean 0.008 0.023 -0.014 0.016 N (Obs.) 814 827 851 865 Adj R-sq 0.028 0.042 0.025 0.048 p-val 0.476 0.930 0.727 0.998 Bonferroni (α = .05) 0.013 0.013 0.013 0.013 ToT (2SLS) KKL Index of KKL Index of KKL Index of KKL Index of Physical Violence, Physical Violence, Sexual Violence, in Sexual Violence, in HH non partner outside HH HH non partner outside HH RMG -0.053 -0.006 0.026 -0.000 (0.074) (0.071) (0.075) (0.068) Control Group Mean -0.001 0.017 -0.003 0.018 N (Obs.) 814 827 851 865 Adj R-sq 0.028 0.042 0.025 0.048 p-val 0.473 0.930 0.725 0.998 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are indexes of the form specified by Kling, et.al. (2007) for each category of GBV. * p < 0.1, ** p < .05, *** p < 0.01 34 Table A12 - Effect of treatment on GBV: Binary indicators of exposure to any GBV and any IPV ITT (OLS) Experienced any Experienced any GBV IPV RMG 0.025 0.037 (0.033) (0.033) Control Group Mean 0.668 0.663 N (Obs.) 1,003 980 Adj R-sq 0.008 0.000 p-val 0.439 0.265 ToT (2SLS) Experienced any Experienced any GBV IPV RMG 0.025 0.037 (0.032) (0.033) Control Group Mean 0.671 0.668 N (Obs.) 1,003 980 Adj R-sq 0.008 0.000 p-val 0.437 0.262 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon- Khmer, Sino-Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. Outcomes of interest are dummy variable, if a respondent reports to have ever experienced any GBV or any IPV. * p < 0.1, ** p < .05, *** p < 0.01 35 Table A13 - Lee bounds for Exposure to GBV: IPV Exposure to Exposure to Exposure to Exposure to Sexual Controlling Emotional Violence Physical Violence Violence Behavior RMG (Upper Bound) 0.059* 0.004 0.037 0.049 (0.035) (0.034) (0.033) (0.030) Control Group Mean 0.516 0.340 0.314 0.195 N (Obs.) 981 963 995 950 Adj R-sq 0.006 -0.001 0.035 0.039 p-val 0.094 0.896 0.263 0.104 RMG (Lower Bound) 0.049 0.001 0.026 0.028 (0.035) (0.034) (0.033) (0.030) Control Group Mean 0.527 0.344 0.326 0.215 N (Obs.) 982 964 996 951 Adj R-sq 0.007 -0.000 0.035 0.039 p-val 0.170 0.982 0.429 0.349 Controls Controls Controls Controls Controls Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary outcomes where the outcome equals 1 if she has experienced this type of violence in the past 12 months, modified using the formula specified by Lee (2009). * p < 0.1, ** p < .05, *** p < 0.01 36 Table A14 - Lee bounds for Exposure to GBV: Non-IPV GBV Exposure to Exposure to Sexual Exposure to Exposure to Sexual Physical Violence, Violence, in HH Physical Violence, Violence, outside in HH non partner non partner outside HH HH RMG (Upper Bound) -0.006 0.041** -0.004 0.030 (0.024) (0.021) (0.026) (0.025) Control Group Mean 0.116 0.057 0.145 0.115 N (Obs.) 863 879 863 879 Adj R-sq 0.023 0.023 0.057 0.045 p-val 0.785 0.046 0.886 0.243 RMG (Lower Bound) -0.011 0.001 -0.008 -0.011 (0.024) (0.022) (0.026) (0.026) Control Group Mean 0.120 0.098 0.149 0.156 N (Obs.) 864 880 864 880 Adj R-sq 0.023 0.021 0.059 0.052 p-val 0.647 0.975 0.755 0.672 Controls Controls Controls Controls Controls Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary outcomes where the outcome equals 1 if she has experienced this type of violence in the past 12 months, modified using the formula specified by Lee (2009). * p < 0.1, ** p < .05, *** p < 0.01 37 Table A15 - Kling-Liebman bounds for Exposure to GBV (0.25 diff): IPV Exposure to Exposure to Exposure to Exposure to Sexual Controlling Emotional Violence Physical Violence Violence Behavior RMG (Upper Bound) 0.097*** 0.046* 0.066** 0.077*** (0.029) (0.028) (0.028) (0.025) Control Group Mean 0.501 0.320 0.304 0.191 N (Obs.) 1,193 1,193 1,193 1,193 Adj R-sq 0.013 0.002 0.034 0.036 p-val 0.001 0.099 0.019 0.002 RMG (Lower Bound) 0.011 -0.044 -0.010 -0.003 (0.029) (0.028) (0.028) (0.025) Control Group Mean 0.543 0.365 0.341 0.231 N (Obs.) 1,193 1,193 1,193 1,193 Adj R-sq 0.001 0.003 0.030 0.033 p-val 0.702 0.114 0.726 0.919 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary outcomes where the outcome equals 1 if she has experienced this type of violence in the past 12 months, modified using the formula specified by Kling and Liebman (2004). * p < 0.1, ** p < .05, *** p < 0.01 38 Table A16 - Kling-Liebman bounds for Exposure to GBV (0.25 diff): Non-IPV GBV Exposure to Phys. Exposure to Sexual Exposure to Phys. Exposure to Sexual Violence, in HH Violence, in HH Violence, outside Violence, outside non partner non partner HH HH RMG (Upper Bound) 0.031* 0.039** 0.040** 0.041** (0.017) (0.017) (0.019) (0.020) Control Group Mean 0.098 0.076 0.124 0.128 N (Obs.) 1,193 1,193 1,193 1,193 Adj R-sq 0.018 0.020 0.043 0.037 p-val 0.069 0.021 0.035 0.042 RMG (Lower Bound) -0.056*** -0.030* -0.056*** -0.043** (0.017) (0.017) (0.019) (0.020) Control Group Mean 0.142 0.113 0.173 0.173 N (Obs.) 1,193 1,193 1,193 1,193 Adj R-sq 0.025 0.018 0.051 0.043 p-val 0.001 0.074 0.003 0.034 Robust standard errors in parentheses. We control for the following baseline characteristics: age, a binary indicator for married, an indicator for ever attending school, indicators for ethnicity being Lao-Tai, Mon-Khmer, Sino- Tibetan, Hmong-Iumien or other, household size and monthly household income. Data on gender-based violence was not collected at baseline so we cannot control for exposure before the RMG program. All outcomes are binary outcomes where the outcome equals 1 if she has experienced this type of violence in the past 12 months, modified using the formula specified by Kling and Liebman (2004). * p < 0.1, ** p < .05, *** p < 0.01 39