Policy Research Working Paper 9065 The Effects of International Scrutiny on Manufacturing Workers Evidence from the Rana Plaza Collapse in Bangladesh Laurent Bossavie Yoonyoung Cho Rachel Heath Social Protection and Jobs Global Practice November 2019 Policy Research Working Paper 9065 Abstract After the tragic factory collapse of Rana Plaza in 2013, both non-garment workers, in districts containing the vast the direct reforms and indirect responses of retailers have majority of export garment factories versus other districts, plausibly affected workers in the Ready-Made Garment pre versus post Rana Plaza. As intended by the reforms, we (RMG) sector in Bangladesh. These responses included a find that increased international scrutiny improved working minimum wage increase, high profile but voluntary audits, conditions by 0.80 standard deviations. In contrast with and an increased reluctance to subcontract to smaller facto- what the theory of compensating differentials would sug- ries. This paper estimates the net impact of these responses gest, we do not find that workers’ wages were negatively using six rounds of the Labor Force Survey and a triple impacted: instead, the post-Rana Plaza responses increased difference approach that compares garment workers to wages by about 10%. This paper is a product of the Social Protection and Jobs 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 lbossavie@worldbank.org or ycho1@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 Effects of International Scrutiny on Manufacturing Workers: Evidence from the Rana Plaza Collapse in Bangladesh ∗ † Laurent Bossavie, Yoonyoung Cho, Rachel Heath‡ § Originally published in the Policy Research Working Paper Series on November 2019. This version is updated on May 2022. To obtain the originally published version, please email prwp@worldbank.org. Keywords: Garment sector, working conditions, gender, minimum wage JEL classification: F16, J16, J31, J32, J81, O12 ∗ World Bank; lbossavie@worldbank.org † World Bank; ycho1@worldbank.org ‡ Department of Economics, University of Washington; rmheath@uw.edu § We thank S Anukriti, Laura Boudreau, Andrew Foster, Melanie Khamis, Adriana Kugler, Annemie Maertens, Tyler McCormick, David Newhouse, Laura Schechter, Shing-Yi Wang, Chris Woodruff, and attendees of the 2018 World Bank/IZA Jobs & Development Conference, the University of Washington Labor-Development Brownbag, the World Bank Jobs Group Brownbag Lunch Series, the Georgetown Center for Economic Research 2019 Biennial Conference, the Korea Development Institute School of Public Policy Seminar, the University of Washington Forum for Political Economy and Economics, the 2019 Determinants of Productivity in Low Income Countries workshop at LMU/DIW, North Carolina State ARE, Howard, the 2020 Econometric Society World Congress, the University of Georgia, Bentley University, the University of Michigan, the 2020 Center for Effective Global Action (CEGA) Research Retreat, and NYU for helpful comments. Iffat Chowdhury and Adam Visokay provided excellent research assistance. This research was made possible by the generous support from the Korea World Bank Partnership Facility Trust Fund for Bangladesh. 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. 1 Introduction The collapse of the Rana Plaza factory building in Bangladesh in April 2013 killed over a thousand workers and injured about 2,500 others. It is widely considered the worst accident in the history of the global garment industry, and the world’s worst industrial disaster since the 1984 Bhopal gas tragedy in India (Reuters, 2017). In its aftermath, local and international attention on the ready-made garment (RMG) sector intensified, resulting in a series of new policies and changes to the behavior of retailers. Direct reforms included an increase in the minimum wage and high- profile but voluntary audits examining safety. Indirectly, retailers may have pressured factories to improve working conditions or have become hesitant to buy from factories that subcontract or that appear to be skirting labor regulations. While the intended goal was to improve working conditions and raise wages, a concern arose that reforms which raise costs among garment factories could ultimately hurt workers. In this paper, we evaluate the net effects of these post-Rana Plaza changes on workers in the RMG sector in Bangladesh. We use six rounds of the Bangladesh Labor Force Survey, spanning from 2003 to 2016, and employ a triple-difference estimation strategy that compares workers in the garment sector to workers in other manufacturing industries, in four “treated” districts whose garment factories comprise 99% of garment exports versus the rest of Bangladesh, pre versus post Rana Plaza. We examine four key outcomes – wages, an index of working conditions, hours of work, and whether the worker receives a written contract – and test parallel trends for each outcome. While we generally fail to reject the null hypothesis of divergent trends among garment workers in treated districts (compared to non-treated districts) pre Rana Plaza, we caution that the results for contracts should be viewed as suggestive because the test there is lower-powered and some of the point estimates are large. Each survey wave was collected in all four quarters of the year, allowing us to separately examine short-term effects (in the remainder of 2013) and medium-term effects (in 2015 and 2016) of the reforms that followed the event. While we believe these medium-term results are important, we acknowledge that by several years after Rana Plaza, it is more likely that there has been a shock that differentially affected garment workers in treated districts, and so the medium-term results should be viewed as somewhat more suggestive than the 2 initial results. We begin by examining wages. Theoretically, wages could have increased due to the minimum wage increase implemented after the Rana Plaza collapse or because of the pressure from buyers to raise wages. Alternatively, wages could fall as firms shift spending on total compensation more towards working conditions. We find that wages increased by 10% across genders and time periods. The results are larger in magnitude for female employees, whose wages increased by 19%, though the difference between men and women is not statistically significant at conventional levels (P = 0.177). We go on to examine measures of working conditions (including non-wage benefits and health/safety), working hours, and job security. As intended by the reforms, we find a large positive effects on working conditions: the working conditions index constructed from the data increased by 0.80 standard deviations after Rana Plaza. In contrast, we find no evidence of average changes in hours of work or contracts as a result of Rana Plaza. Together, these effects paint an optimistic picture of the net effect on workers’ welfare. The fact that working conditions improved is important, given high baseline levels of reports of harassment and abuse. For example, 37% of female garment workers reported abuse in the first quarter of 2013 (i.e., immediately before the Rana Plaza collapse). While a standard compensating differentials model within a competitive labor market would predict that improvements in working conditions come at the expense of wages, we find that if anything, wages increased. This is consistent with earlier evidence that garment exporting firms have some degree of monopsony power in the labor market (Harrison and Scorse, 2010), and thus have scope to raise both wages and working conditions simultaneously. These changes affect many workers and their families, given that the RMG sector in Bangladesh is widely considered to be a key contributor to the country’s robust economic growth, poverty reduction, and women’s empowerment. The sector’s share of total exports increased from 53 percent in 1995 to 83 percent in 2017, with its exports reaching about USD 28.1 billion (Farole et al., 2017). The export share of GDP tripled between 2003 and 2017, and Bangladesh is now the world’s second largest garment exporter after China. The sector provides jobs and earnings 3 opportunities to over 4 million low and semi-skilled workers, and represents over 40 percent of the total industrial employment in the country. Moreover, the large expansion of the garment industry has provided employment opportunities to women, changing their decisions on schooling, marriage, and childbearing (Heath and Mobarak, 2015). There is evidence that women workers particularly value improvements in working conditions (Khosla, 2009, Begum et al., 2010, Gibbs et al., 2019, Subramanian, 2019), showing that international pressure can help women workers reap the benefits of garment jobs with less risk to their personal well-being. We contribute to a small literature on the effects of international pressure on workers’ outcomes. Harrison and Scorse (2010) find that international attention to the wages and working conditions in textiles and footwear factories in Indonesia in the 1990s raised wages. Amengual and Distelhorst (2019) document that a management change at Gap Inc. increased labor compliance in sourcing garment factories. We build on these results to study whether there are sector-wide changes in working conditions as a result of a large industrial tragedy, and whether improvements to working conditions come at the expense of other job quality measures in the context of a highly competitive international garment market in which Bangladeshi factories face pressures to produce at low costs. At the same time, in the same Bangladeshi context as this paper, Boudreau (2019) finds that when provisions in the post Rana Plaza Alliance agreement (as described in section 2.2) that mandate the existence of worker-manager safety committees are randomly enforced, compliance with labor laws increases. While our paper finds that the Rana Plaza reforms had an impact, even absent the external enforcement treatment in Boudreau (2019), her results suggest that the improvements in working conditions we document could be even higher when combined with stronger external enforcement. Our results also add a positive note to a literature that points out that regulation intended to help workers may raise the cost of hiring workers and thus have unintended negative consequences on the workers it is meant to help. For instance, Besley and Burgess (2004) find that pro-worker labor regulations in India decreased productivity and ultimately increased urban poverty. Botero et al. (2004) show that similar results hold in a sample of 85 countries; heavier regulation of labor is associated with lower employment rates. Parker, Foltz and Elsea (2016) find that policies 4 targeting human rights abuses can have similar effects: legislation aimed at stopping the export of conflict minerals hurt infant health in the Democratic Republic of the Congo.1 By contrast, we find that in sectors that have scope to raise both wages and working conditions, pushes to help workers in some job dimensions may not always backfire in other dimensions. Our paper also contributes to a literature on the relationship between wages and working conditions. Standard theories of compensating differentials posit that firms choose wages and working conditions to equalize the marginal utility of an additional dollar investment in wages versus working conditions for the marginal worker. Correspondingly, Summers (1989), Mitchell (1990) and Almeida and Carneiro (2012) find that wages fall when firms are compelled by external forces to improve their working conditions. We find, instead, that improvements to working conditions do not always prompt wage decreases, in an environment where firms plausibly have some market power in the domestic labor market. Finally, our work is related to a literature in trade and development on the ways in which des- tination countries affect workers in export industries in low-income countries. The existence of a wage premium for working in an export factory suggests that access to international markets helps workers in low income countries, as their firms make increased profits from selling to wealthier ıas, Kaplan and Verhoogen, 2009, Amiti consumers in high-income countries (Verhoogen, 2008, Fr´ and Davis, 2011). The sudden opportunity to export that came with Myanmar’s trade liberaliza- tion also improved working conditions in garment factories (Tanaka, 2020). However, exposure to trade can also hurt workers in import-competing industries, and Topalova (2010) and Kovak (2013) find evidence that reductions in trade barriers increased poverty in India and Brazil, re- spectively. Other work finds that consumer sentiment also matters. Dragusanu and Nunn (2018) find that fair trade certification in Costa Rica helps coffee farmers, though there are distributional consequences: unskilled farmers are unaffected. Like Dragusanu and Nunn (2018), we emphasize that it is not just exporting that matters for workers’ outcomes: buyer pressure is also important. Unlike Dragusanu and Nunn (2018), however, we find some evidence that less advantaged workers 1 Adhvaryu, Nyshadham and Xu (2018) and Boudreau (2019) highlight a complementary mechanism in Indian and Bangladeshi garment factories, respectively. These papers argue that reforms that raise workers’ expectations for job quality can decrease job satisfaction if these expectations are not met. 5 (namely, less educated men and younger women) benefit more. The remainder of the paper is organized as follows. Section 2 presents the context of the post-Rana Plaza reforms and international scrutiny that affected the RMG sector following the collapse. Section 6 explains the theoretical mechanisms through which these changes might affect garment workers. In section 3, we describe our data and our identification strategy, and provide evidence for its validity in section 5. Section 4 reports and discusses our main results. Section 7 concludes. 2 Context 2.1 Employment in the garment sector The garment sector currently employs approximately 4 million workers in Bangladesh in over 5,000 factories. Employment in the sector has grown at an explosive yearly rate of 15.5 percent since its inception in the early 1980s, as indicated by data from the Bangladesh Garment Manufacturers Exporting Association (BGMEA) in Figure 1. This growth has been fueled by migration from rural areas, and rural migrants’ lack of information appears to contribute to poor working conditions (Boudreau, Heath and McCormick, 2017). Workers are fairly mobile; by 12 months after the time of hiring, for instance, only 62 percent of all hired workers who are still working in the garment industry remain in their original factory (Heath, 2018). 2.2 Safety standards in the RMG sector, before and after Rana Plaza Before the Rana Plaza collapse, workers had protested, and some safety measures had been put in place, such as the Fire Prevention and Fire Fighting Law of 2003, the Fatal Accidents Act of 1855, the Bangladesh National Building Code of 2006, and the Labour Act of 2006. However, the effectiveness of these acts was limited; they were overall vague, narrowly defined, and lacked focus (Rahman and Moazzem, 2017). Moreover, authorities lacked technical expertise, and enforcement was weak. For instance, factory visits were pre-announced, and factories with violations found only needed to be revisited within a span of three years. 6 The fast growth of the industry led large factories to subcontract to small factories in order to remain cost competitive and meet deadlines. Labowitz and Baumann-Pauly (2015) estimate that in 2015, 54% of garment factories in Bangladesh are such unregistered, indirect sourcing factories. As illustrated in Figure 2, the immediate aftermath of Rana Plaza was marked by a noticeable decline in the number of factories registered with the Bangladesh Garment Manufacturers and Exporters Association (BGMEA), consistent with the shutdown of non-compliant factories. This drop in the registered number of factories, combined with no noticeable decline in total garment employment post Rana Plaza (Figure 1), indeed suggests that some industrial reorganization took place in the garment industry. While the principal factories are typically compliant with worker and safety norms/codes, their subcontractor factories are small, often huddled into buildings that are not designed for industrial purposes, but house five or more small factories.2 The general consensus among industry experts is thus that working conditions in these factories are worse than those in the principal factories that sell to retailers (Rahman and Moazzem 2017; Naved et al. 2018). Following the Rana Plaza incident, the representatives of the Government of Bangladesh, the European Union (EU) represented by the European Commission, and the International Labour Organization (ILO) supported the Bangladesh Sustainability Compact to promote improved labor standards and responsible business conduct in the RMG sector in Bangladesh. Representatives from industry (including brands, retailers and small and medium enterprises), employers, trade unions and other key stakeholders also supported the initiative. They put in place the National Tripartite Action Plan on Fire Safety and Structural Integrity, which consisted of a set of detailed measures to improve safety in RMG factories. In parallel, although major retailers initially discussed a single coordinated industry response, two different industry agreements emerged. European buyers joined to create the Accord on Fire and Building Safety in Bangladesh (AFBSB), and a group of American buyers joined to create the Alliance for Bangladesh Worker Safety (ABWS) in 2013. Both initiatives were voluntary measures meant to assess RMG factories’ compliance to building and fire safety. The Accord is 2 The Rana Plaza factory was also housed in a building that was permitted for commercial purposes, but it was illegally renovated to have more units for industrial purposes and hosted a total of five garment factories. 7 a legally-binding agreement that commits signatories to a five-year program of safety audits and remediation investments in their Bangladeshi supplier bases. The Alliance has similar stated goals as the Accord but with more limited binding commitments, including for financing remediation. As a result of these measures, the majority of export-oriented RMG factories have been as- sessed and those that did not qualify have been flagged for remediation. Focus groups conducted by Kabeer, Haq and Sulaiman (2019) in 2017 indicated that many workers had witnessed im- provements such as the widening of exit staircases or formation of health and safety committees. Moreover, workers felt newly empowered to speak up about any safety violations they witness; one focus group member gave the example that “if [he] sees a crack on the wall, [he] can speak up.” While the Accord and Alliance have provided some help with remediation financing to make these improvements, the reforms prescribed by the Accord and Alliance have been primarily financed by the factories themselves, under buyer pressure. This led to an important rise in compliance-related expenditures, and therefore to an increase in the cost of doing business for RMG entrepreneurs while facing increased pressure from buyers to bring down their prices (Reuters, 2019). As a result, some factories had to increase workers’ production targets to make up for additional expenditures on compliance requirements, while some could simply not cope with the additional expenses, leading to business closures. 2.3 Minimum wage increase The minimum wage is set at the industry level in Bangladesh, with no formal overall minimum wage. The garment industry is one of the few sectors that has statutory minimum wages; these have been in place since 1985. One minimum is set for each of the seven grade levels, corresponding to different levels of skills and experience in the industry. After subsequent periodic revisions to reflect inflation and cost of living, it was set at BDT 3,000 (USD 38) a month as of 2010 for low-skilled workers. As a response to the Rana Plaza collapse, workers protested in favor of a minimum wage increase, seeking a “living wage” of BDT 9,500 (USD 120). Factory owners argued that such a large increase would hamper competitiveness, and in December 2013, the garment 8 sector’s minimum wage was raised, but only to a compromise of BDT 5,300 (USD 68) a month for low-skilled workers. This represents an increase by 77% compared to the previous minimum wage.3 Figure 3 provides a histogram of the monthly earnings of workers in the RMG sector in 2013 quarter 1 (before the Rana Plaza collapse) and 2015 (after the minimum increase had been in effect for over a year) using the Bangladesh Labor Force Survey data described in section 3.1. Note that while our data on earnings includes all take home pay (inclusive of payments like bonuses and overtime pay) the minimum wage applies only to base pay, so that more workers earn below the new minimum wage than would be expected if they reported base pay separately. Overtime pay is likely to be important: it is legally required for every extra daily hour beyond 8 hours for each of the six working days (Menzel and Woodruff, 2021), and 62% of workers in 2013 quarter 1 worked strictly more than 48 hours per week. Indeed, an estimated 20% of workers’ take-home pay is overtime (Reuters, 2018). Even so, 1.3% of workers earned total take-home pay below the new minimum wage in 2013, while none did by 2015. 3 Empirical strategy 3.1 Data and outcome variables We use the Labor Force Surveys (LFS) conducted in 2003, 2005, 2010, 2013, 2015, and 2016 by the Bangladesh Bureau of Statistics. The Bangladesh LFS was designed to be representative at the national level – for both rural and urban areas – as well as for each of Bangladesh’s eight divisions. All working-age (age 15 to 64) individuals from the selected households are interviewed for detailed labor market outcomes including labor market status, type of employment, monthly wages for those that are paid employees, hours worked, detailed occupation and sector of activity at the four-digit level, as well as working conditions. The sample sizes are large, with over 100,000 individuals in working age interviewed in years prior to 2015, and above 300,000 in 2015 and 2016. The 2003-2013 surveys took place over the course of a year, while the 2015 and 2016 surveys 3 In December 2018, after the period studied by this paper, the minimum wage was further raised to BDT 8,000. 9 represented four distinct quarterly surveys that we merged together. Thus, any seasonality in wages (bonuses or raises) should be averaged out over the course of each survey. In each survey round, workers were asked the number of hours they had worked over the past seven days. By contrast, the measures of earnings varied from year to year,4 so to construct our measure of hourly wages: • In 2003, workers were asked monthly earnings from their main activity in the past year. We calculate each worker’s average hourly earnings by dividing this amount by the hours worked in the past week multiplied by 4.33 (the average number of weeks in a month, across the year). • In 2005, 2010 and 2013, workers were asked about earnings from the previous week in their primary job, so we divide this by the hours worked in the past week. • In 2015 and 2016, workers were asked take home earnings for the past one month. We calculate average hourly earnings by dividing this amount by the hours worked in the past week multiplied by 4.33. In each round, earnings were asked to paid employees: both regular wage workers and well as casual/piece rate workers. While, prior to 2013, self-employed workers were asked about their earnings, for consistency over our period of study, we restrict our sample to paid employees. We deflate all wages to 2010 taka using the consumer price index.5 There may be differences in the wages calculated across the different survey rounds. The fact that the 2003 survey asked about monthly earnings over the past year presumably means that respondents reported their average monthly income over the year, which reduces variance between workers. Similarly, the average weekly pay over the past month in 2015 and 2016 could display lower variance than in rounds 2005-2013, when workers reported weekly pay directly. However, any level differences between years would be absorbed in the year fixed effects in the regression 4 Because these different measurements may have different amounts of noise, it is possible that estimation results are affected by heteroskedasticity. Following the recommendation of Ferman and Pinto (2019) on differences-in- differences settings with few treated and many control groups in the presence of heteroskedasticity, we reestimate treatment effects using FGLS in Table A2. Results are very similar to our main estimates, providing reassurance that those main estimates are not significantly affected by heteroskedasticity. 5 https://ycharts.com/indicators/bangladesh_consumer_price_index_wdi 10 analysis, and would not be a leading concern unless they systematically vary between garment and non-garment sectors, conditional on the control variables in our main specification (equation 1). The data also contain several measures of non-wage benefits and other aspects of job quality. Some of these variables are only available in later rounds, as summarized in the table below: Measure Years available Details Sick Leave all Paid sick leave provided by employer Maternity Leave 2010 onwards Maternity leave provided by employer (does not specify paid) Contract 2005 onwards Written contract (oral contract coded as zero) Injured 2013 onwards Hurt in any accident at work Dangerous 2013 onwards At least one reported danger (see below) Abused 2013 onwards At least one report of abuse (see below) Specifically, the reported dangers could include dust, fumes, noise or vibration; fire, gas, flames; extreme cold or heat; dangerous tools; work underground or at heights; work in water/pond/river; workplace too dark or confined; chemicals/explosives; other things (specify). Reported abuse could include: constantly shouted/insulted; beaten /physically hurt; sexually abused; others. Table A1 gives the frequency of specific abuses and dangerous conditions. The most common abusive behavior faced by garment workers was being constantly shouted at or insulted (13% of women and 12% of men), though 5% of women reported experiencing sexual abuse. The most common dangerous conditions were dust, fumes, noise and vibrations (19% of women and 17% of men) and dangerous tools (15% of women and 14% of men). We combine these five binary measures (sick leave, maternity leave, reports of injuries on the job, having been abused, and dangerous conditions) into a single index of working conditions by standardizing each variable (within the estimation sample) to have a standard deviation of one. We then sum the standardized variables available for each observation (depending on the round from which the observation came) and then re-standardize that sum so that the final working conditions measure has a standard deviation of one across the estimation sample. 11 We examine written contracts as a separate outcome from working conditions. Even if it is very unlikely in practice that a worker would go to court to enforce a contract, a contract is useful for workers who wish to prove that they have formal employment or seek governmental social benefits (Afrin, 2014). We thus argue that contracts provide an important signal of job security to workers, even if lacking a formal way to enforce them. A further caveat to our results on contracts is that the test for parallel trends for contracts is lower-powered than other outcomes and has some large point estimates (even though we formally fail to reject parallel trends), as shown later on in section 5.1. Overall, while we believe contracts are worthy of examination, these caveats suggest that readers should give less weight to them than our other three outcomes in evaluating the overall impacts of Rana Plaza on workers’ welfare. 3.2 Timing of the treatment The collapse of Rana Plaza occurred on April 24, 2013. Rescue efforts continued until May 13, including the high profile rescue of a woman on May 10. Workers began to riot two days after the collapse and were joined the next day by some political parties, including the Bangladesh National Party. Rioters demanded the arrest of those whose negligence they believed contributed to the collapse and an independent commission to identify safety threats. Protests continued off and on for many months. Response from retailers also began shortly after the collapse. Specifically, retailers and NGO’s met one week after the collapse and created the Accord on Factory and Building Safety. The Alliance came in place in July 2013. Activists also pushed retailers in the months following the collapse, with some calling for boycotts, but others calling for brands to stay and “clean up the industry” (Judy Gearheart, executive director of the International Labor Rights Forum on April 29, 2013). Overall, wage and working conditions responses to the Rana Plaza incident and subsequent reforms likely unfolded gradually across the months following the incident. Some responses – such as compositional changes in the industry due to factory shutdowns – could have taken even longer. We consider the first quarter of 2013 to be pre-treatment, drop the second quarter (as it is a mix 12 of pre-treatment survey periods and those surveyed in the immediate aftermath), and consider the third and fourth quarters to be post-treatment. Thus, while our baseline specification averages the effects on workers surveyed from 5 weeks to 3 years after the Rana plaza collapse, we also run specifications where we explicitly consider treatment effect heterogeneity by time duration since the collapse. 3.3 Identification strategy Our main identification strategy is a triple-difference that tests for differential outcomes in apparel workers (vs non-apparel manufacturing workers), in the four districts of Bangladesh that constitute the vast majority of export apparel factories (vs the rest of the country), pre-vs post Rana Plaza. We define the apparel industry according to the ISIC-3 classification (ISIC-3 two-digit code 18), which we also refer to as the garment industry. Our treated districts are Dhaka, Gazipur, Narayanganj, and Chittagong. As these districts represent 3229 of the 3251 (99%) export-oriented ready-made garment factories mapped as of March 2021 (Mapped in Bangladesh, 2021), we argue that they were most closely affected by the Rana Plaza reforms, given the role of international buyers described in section 2.2. Although garment employment in Bangladesh is disproportionately concentrated in the four treated districts, control districts still host 46% of garment workers in the sample, giving us sufficient control observations in our triple-difference estimation. Figure 4 shows the distribution of garment workers across treatment and control districts. We thus estimate the following specification for an outcome Y for worker i in district j at time 13 t: Yijt =β Garmentijt × P ostRanaP lazat × T reatedDistrictj (1) + δ1 Garmentijt + δ2 Garmentijt × F emaleijt + δ3 Garmentijt × T reatedDistrictj + δ4 Garmentijt × T reatedDistrictj × F emaleijt + λt + λt × T reatedDistrictj + λt × F emaleijt + T reatedDistrictj × λt × F emaleijt + λt × Garmentijt + λt × Garmentijt × F emaleijt + γt F emaleijt × Xijt × λt + εijt The effect of Rana Plaza is estimated by the triple difference term Garmentijt × P ostRanaP lazat × ˆ.6 Our identification assumption is then that any differential T reatedDistrictj and thus given by β trends affecting garment (vs non-garment) workers in treated districts versus non-treated districts cannot vary pre vs post Rana Plaza. We also include year fixed effects interacted with gender and a vector of control variables Xijt – namely age and education. These controls flexibly account for changes in labor outcomes over time that may vary by location or a worker’s human capital, and possibly differently so by gender. We discuss possible violations of this identification assumption in sections 5. In particular, we conduct a test for parallel pretrends in section 5.1, and address the possibility of shocks that are contemporaneous to Rana Plaza in section 5.2. Another possible threat to identification is that garment workers in control districts or non-garment workers in garment-exporting (i.e. treated) districts might have been affected by Rana Plaza; we discuss these possibilities in section 5.3. 3.4 Summary statistics Table 1 reports summary statistics by gender, garment industry (vs not), and location in treatment district (vs not) for the estimation sample. There are level differences in personal characteristics and labor outcomes between garment workers by area, and garment vs non-garment workers within area. While these level-differences are not necessarily a concern for identification if the control 6 Note that the double-difference terms Garmentijt × P ostRanaP lazat and T reatedDistrictj × P ostRanaP lazat are captured by the inclusion of time fixed effects interacted with Garmentijt and T reatedDistrictj (and each variable’s interaction with gender). 14 group plausibly displays a parallel trend in each key outcome (a question which we discuss in section 5), they are useful for interpreting the type of workers in treatment and control indus- tries. For instance, both garment and non-garment workers are better educated and earn more in treated districts, compared to non-treated districts. Interestingly, given evidence from Myanmar that exporting causally improves working conditions (Tanaka, 2020), working conditions are not uniformly better in garment factories in treated districts: workers in treated districts are more likely to have sick leave and maternity leave, but women report more abuse and both genders report higher rates of dangerous conditions. Within treated districts, earnings do not significantly differ between male garment and non- garment workers, but female garment workers do earn more than non-garment workers (34.5 taka per hour for garment workers, vs 31.6 for non-garment workers). Working conditions between garment and non-garment workers in treated districts present mixed results: garment workers are more likely to have sick leave and maternity leave and less likely to report dangerous conditions or injuries, but are more likely to report abuse. 4 Results Table 2 reports the results of our triple difference specification (equation 1) on our key outcomes: hourly wages, working conditions, contracts, and hours of work. Starting with hourly wages, the estimate in column 1 shows that the post-Rana Plaza reforms increased by 10 percent, pooled across genders and years. In panel B, we find that this wage increase was qualitatively driven by women’s wages; women’s wages rose by 19% after Rana Plaza, while men’s rose by 6.7% (though the difference between men’s and women’s wages is not statistically significant at traditional levels; P = 0.177). In Figure 5, we break down treatment effects for each variable by both gender and year and find that the point estimates for both men’s and women’s wages were more strongly positive in 2013 and 2015 (for women) and 2013 (for men). This is consistent with some of the wage effect being driven by the minimum wage increase, though, given that other mechanisms discussed in section 6 may vary by year, our estimates are net impacts of the channels of Rana Plaza in each year and 15 we cannot cleanly isolate the minimum wage channel. We also acknowledge that allowing a triple difference to vary by both gender and time period is demanding on the data, and the estimates are accordingly noisy, so they should be viewed as suggestive. We further caution that, as discussed in the introduction, the more time that passed after Rana Plaza, the greater the likelihood of a shock to garment workers in treated districts that would violate the parallel trends assumption needed for a causal interpretation of the triple difference in equation 1. So the medium-term results should be viewed as somewhat more suggestive, though it’s broadly reassuring for policy-makers interested in improving the welfare of garment workers that positive point estimates for wages do not dissipate completely in the medium term. The stated objective of much of the international scrutiny post Rana Plaza reforms was to improve safety and working conditions inside garment factories. In column 2 of Table 2, we report estimates of the effects of the Rana Plaza reforms on the index of working conditions described in section 3.1. Overall, we find that Rana Plaza increased the working conditions faced by garment workers by 0.80 standard deviations, with similar orders of magnitude between men and women. From Figure 5, we see that the effects are particularly strong in 2016 (especially for women), though again the estimates are fairly noisy and this pattern should be viewed as suggestive. By contrast, we do not find qualitatively large or statistically significant effects on hours of work or the probability of receiving a contract. Table 3 examines other sources of potential heterogeneity behind the average effects given in panel A of Table 2. Specifically, we consider four binary dimensions of heterogeneity: at or below median age and education in estimation sample (7 years of education and 28 years of age), whether the respondent is married, and whether there is at least one child in the respondent’s house7 . These variables were chosen because they either reflect human capital (lacking direct information on total garment sector experience, age is a close proxy) or increased demand for nonwage benefits among those with care-taking responsibilities. Because there may be important interactions between these variables and gender, we include each variable’s interaction with both 7 This measure may miss respondents’ biological children if they are living with relatives in a rural area, which anecdotally is not uncommon among garment workers, but not available in the LFS surveys. Still, however, the measure captures garment workers who are likely to have increased care-taking responsibilities. 16 the main garment treatment effect (Garment × Post × Treated District) and its interaction with gender. We find that less-educated and married men particularly benefit from the Rana Plaza responses. Less educated men enjoy a 5.4 percent wage gain, compared to more-educated men, leading to 12 percent net gain compared to pre-Rana Plaza. They also improve working conditions by 0.20 standard deviations more than less-educated men (0.94 net gain relative to pre Rana Plaza) and are 12 percentage points more likely to receive a contract (net effect 13 percentage points). Married men receive a 6.5 percent wage gain compared to unmarried men (net effect 14 percent) and a 0.19 standard deviation gain in working conditions compared to unmarried men (net effect 0.88 standard deviations). While less educated women and married women do not display larger effects than women with above-median age or education, young women workers do benefit more than older women workers.8 In particular, young female garment workers enjoy a 9.6 percentage point higher wage increase than older female garment workers (net improvement 22%), an 0.18 standard deviation higher improvement in working conditions (net improvement 1.1 standard deviations), and a 4.9 higher percentage point increase in the probability of receiving a contract (though this results in a net impact of roughly zero; older female workers were actually less likely to receive contracts). 5 Threats to identification We acknowledge that identifying the causal impacts of Rana Plaza is difficult, given that the garment industry is unique in Bangladesh, and the fact that it might have had spillover effects on the control groups that we use for identification in the triple difference strategy given in equation 1. To build a case for identification, we first conduct a test for parallel pre-trends (section 5.1) and then discuss the possibility of time-varying shocks that might hit garment factories that export in 8 While, as shown in Table 1, female garment workers tend to be younger and less educated than male garment workers, the difference is not so stark that we lack the power to detect effects among female workers with above median age or education. In particular, 24% of female garment workers have above the median years of education in the estimation sample (i.e, the sample across both genders and including garment workers and non-garment manufacturing workers) – this median is 7 years – and 34% of female garment workers are above the estimation sample median age (28 years). 17 section 5.2. Then, in section 5.3, we address the possibility that control groups (namely, garment workers in non-treated districts, and non-garment manufacturing workers in treated districts) are impacted by Rana Plaza, either directly or through worker flows. As part of this discussion, we test for changes in observable characteristics of garment workers in treated districts and discuss the potential implications of such compositional changes in the interpretation of our estimated treatment effects. 5.1 Test for parallel pre-trends Equation 1 estimates the causal impact of the Rana Plaza reforms on garment workers under the assumption that the difference in outcomes between garment and non-garment workers, pre versus post Rana Plaza, does not vary between treatment and control districts. As a first piece of evidence supporting this assumption, we conduct a parallel trends test to assess whether this triple difference was significant before Rana Plaza. Specifically, we examine a Garmentijt × T reatedDistrictj dummy interacted with year dum- mies over the pretreatment time periods (2003, 2005, 2010 and 2013 quarter 1), to test whether there is a non-linear pretreatment trend: f Yijt =βt Garmentijt × T reatedDistrictj × λt + βt Garmentijt × T reatedDistrictj × F emaleijt × λt + δ1 Garmentijt + δ2 Garmentijt × T reatedDistrictj + δ3 Garmentijt × T reatedDistrictj × F emaleijt + λt + T reatedDistrictj × λt + F emaleijt × λt + T reatedDistrictj × F emaleijt × λt + γt F emaleijt × Xijt × λt + ijt (2) ˆt ˆt and β We then test whether the estimated β f coefficients are jointly significant, which would indicate that the difference between garment versus non-garment workers in treated versus non- treated districts (among men or women or both) varies from year to year in the pre-treatment period. The estimation results are given in Table 4 and visually in Figure 6. We consider tests of joint 18 significance for four potential departures from the parallel trends assumption: joint significance ˆt ˆt and β of the β f ˆt ˆt coefficients; joint significance of the β coefficients; joint significance of the β f ˆt ˆt + β coefficients; and the net effect on women (test of joint significance of the sum β f across years). We fail to reject parallel pretrends in each case, with the exception of a marginally significant ˆt result for working conditions for the β f coefficients. (Given that we do examine four outcomes and conduct four tests per outcome, such a result is not unexpected.) Note, however, that the test for pre-trends in contracts is lower-powered (given that data on contracts is available only ˆt ˆt and β 2005 onwards) and there are still some large point estimates among the β f coefficients in column 4. So, while we do display results for contracts given their potential importance for welfare, we do highlight the caveat that they may be impacted by non-parallel trends (in addition to the interpretation considerations raised in section 3.1). 5.2 Contemporaneous shocks to garment factories in treated districts One may also be concerned that the policy changes that took place after Rana Plaza are con- temporaneous to other shocks affecting the garment industry, which could also affect workers’ outcomes. Our triple difference identification strategy allows for garment-sector specific shocks, but we are cognizant that some shocks may particularly fall on export factories in our treated districts. For example, the garment industry in Bangladesh may have been affected by increased global competition or other shifts in the global economy in the post Rana Plaza period. While we include fixed effects interacted with gender and other control variables to capture macroeco- nomic shocks, it is still possible that the macroeconomic variables affect the garment industry, say because it more extensively relies on exports, as discussed in section 2.1. To investigate this possibility descriptively, we look at changes in garment exports from Bangladesh and its main competitors over time before and after Rana Plaza. As shown in Figure 7, we do not observe a noticeable change in export growth from Bangladesh after the Rana Plaza collapse. Garment exports from Bangladesh continued to increase after 2013 at a steady pace, and its global market share actually increased after Rana Plaza: the share of Bangladesh in global clothing exports had risen from 2.5% in 2005 to 4.2% in 2010, and continued to rise to 6.4% in 2018 19 (WTO, 2019). Combined with the evidence from Figure 8 that there are also no large changes in the growth rate of garment sector employment after Rana Plaza, these figures provide evidence against a potential channel where firms respond to Rana Plaza by changing the capital/labor ratio. We also do not observe a noticeable break in exports for the other main exporters of garment post Rana Plaza, except for China, where the value of exports declined after 2013.9 Although this descriptive evidence is not a formal proof of the absence of shocks to the gar- ment industry contemporaneous to Rana Plaza, it alleviates concerns that our estimates primarily capture the influence of contemporaneous shocks. It also indicates that international demand for garment products from Bangladesh was not substantially affected by the Rana Plaza event, possibly due to the rapid safety and regulation responses documented in section 2.2. Therefore, while reduced demand could have been a mechanism behind the treatment effects we estimate, they appear instead to be primarily driven by firms’ responses to increased international scrutiny. 5.3 The extent to which control groups are treated by Rana Plaza responses Even if trends are parallel leading up to Rana Plaza, a violation of the parallel trend assumption required for our triple difference identification strategy would occur if a control group was indirectly affected by the treatment. In the context of the triple difference strategy, this could occur either if garment workers in non-treated (i.e, not exporting) districts are affected, or if non-garment manufacturing workers in treated districts are affected. Table 5 summarizes the potential ways in which both groups could potentially be affected; we discuss each channel in detail in turn. 5.3.1 Garment workers in control districts Garment workers in control districts might have been affected by Rana Plaza, either because the Rana Plaza reforms affected them directly or because of indirect effects through subcontracting channels or workers flows (in particular, if workers leave garment factories in treated districts for 9 The decline in Chinese export value has been mostly attributed to a willingness to move away from low-value apparel up to the value chain, and to an increase in labor costs. 20 factories in control districts). First, consider the minimum wage increase. While the minimum wage increase arose out of the post Rana Plaza protests (as described in section 2.3) it applied to the whole industry, and thus any effects that apply equally in treated and non-treated districts are netted out of the treatment effects given in equation 1. However, it is possible that international scrutiny leads to increased enforcement of existing labor laws given the role of multinationals in enforcing labor law (Boudreau, 2019). In this case, our identification strategy would correctly attribute this aspect of the minimum wage increase to international scrutiny. Worker flows might affect garment workers in control districts if workers either voluntarily left (or were fired from or laid off) garment sector jobs in treated districts and went to work in garment factories in control districts. While we cannot use the panel component of our labor force survey data to speak to the empirical frequency of this event (as workers are not tracked if they move), a separate survey conducted by Heath (and collaborators Laura Boudreau and Khandker Wahedur Rahman) speaks to the phenomenon. In that survey, workers surveyed in 2017 by Kabeer, Haq and Sulaiman (2019) were recontacted in either November 2020 or March-April 2021. The workers all initially worked in treated districts. Of the workers who were found and still worked in the garment industry, 94% (266 of 283) were still in the same district. Among those workers who had moved, only 1 (i.e., .003 %) had moved to a control district. Thus, this survey presents one piece of evidence that flows of garment workers between treated and control districts are relatively small. Next, we consider indirect effects through subcontracting, or spillover effects of international scrutiny. In particular, garment workers in control districts might be impacted if they had pre- viously been in factories that subcontracted with export factories, and this subcontracting was discouraged by the international scrutiny following Rana Plaza. Or their factories might fear worker reprisal if they don’t make some of the same improvements to working conditions or wages that factories in treated districts do. If so, the identification assumption for the triple difference might be violated. To alleviate this concern, we estimate a double-difference specification that compares garment workers to non- 21 garment manufacturing workers within treated districts. While this estimation doesn’t allow for the possibility of differential trends within the garment industry overall, it does assess the extent to which the estimated effects on garment workers in treated districts are not primarily driven by changes in the outcomes of garment workers in non-treated districts. These results are given in Table 6. We find that female garment workers in non-treated districts do receive lower wages (a point estimate of 22% lower, conditional on the human capital controls in equation 1), compared to other manufacturing workers, after Rana Plaza. There are broadly two ways to interpret this change. The first is that these workers capture overall trends in the garment industry (unrelated to Rana Plaza), which, if ignored, would bias our estimations of the effects of Rana Plaza. There is indeed reason to think that there are broader sectoral trends affecting garment workers, absent Rana Plaza, such as increasing global competition (International Labour Organization, 2019). Alternatively, the wage decreases among female workers could represent a causal spillover effect of Rana Plaza, potentially because of reduced demand from export factories to factories that subcontract out. Recall that Figure 8 showed the potential for somewhat reduced employment among women workers in non-treated districts after Rana Plaza, compared to the first quarter of 2013, which is consistent with some demand decrease. If Rana Plaza causally lowered wages of female garment workers in non-treated districts, then our triple difference strategy (equation 1) – which compares wages in treated versus non-treated districts – could overestimate the wage increases among female workers. Still, it is important to note that the double-difference among only workers in treated districts gives a zero (rather than a negative) effect on female workers’ wages, and still shows a clear increase in working conditions of roughly the same magnitude as the triple difference (0.69 standard deviations). That is, even in the most extreme assumption that all the effects on female workers in non-treated districts are causally due to Rana Plaza, female workers in treated districts still benefit from better working conditions without suffering lower wages. 22 5.3.2 Other manufacturing workers in treated districts We argue that the possibility that other manufacturing industries in treated districts were impacted by international scrutiny was likely small; the responses we describe in section 2 were closely targeted at the garment industry. Control workers may still have been affected by labor supply shifts, if workers fled the garment industry for other manufacturing sectors (corresponding to the inward labor supply shift in theoretical model we posit in section 6). 5.4 Evidence on labor supply shifts We can use our data to provide some additional evidence on the worker flows described in the prior subsection. We first consider the potential for changes in the extensive margin, i.e, the total number of garment workers in treated versus control districts over time, in in figure 8. It shows a short-term dip in garment sector employment (though more strongly appearing in non-treated districts10 ), which appears to dissipate by 2015. Thus, we conclude that extensive margin effects are not a large part of the overall story. Even if there are not large effects on the size of the garment sector workforce, there could be a composition change in the types of workers that remain in the garment industry post Rana Plaza (compared to those in control industries). Indeed, in section 6, we will argue that an inward shift in the labor supply curve is likely important in explaining how total compensation increased without an increase in the quantity of labor supplied. Of course, a shift in labor supply doesn’t automatically imply a change in worker characteristics, it could, if workers who are less likely to work for any given compensation after different than those still willing to work. While we can partially address this issue by controlling for observable worker demographics interacted with year and gender dummies in equation 1, we acknowledge that there might be changes in unobservable worker characteristics (such as ability) that are correlated with labor market outcomes. If so, treatment effects estimated by equation 1 will capture both compositional changes and causal 10 Note that the drop off in garment employment in non-treated districts between 2010 and 2013 does not appear to be driven by Rana Plaza, because the decrease also appears in the first quarter of 2013, which is before Rana Plaza. The 2010 survey was at the peak period of Bangladesh’s low unit cost production and subcontracting, which likely affected non-treated districts particularly strongly (Farole et al., 2017). 23 effects on a given type of worker. While we of course cannot test for changes in unobservable characteristics directly, we examine composition changes along key observables (workers’ years of education, gender, age, marital status, whether a child lives in their household, urban location, and Dhaka area location), under the premise that changes in unobservables are likely correlated with changes in observables (Altonji, Elder and Taber, 2005, Oster, 2019). Figure 9 estimates compositional changes by gender, both pooled across all post-treatment time periods and separately by period. We do find some evidence of compositional change. Female garment sector workers were approximately 0.4 standard deviations less likely to be in urban areas (and also less likely to be in Dhaka), post Rana Plaza: this effect is driven by a short term change in 2013 which has gone away by 2015 and 2016. There were increases in education of approximately 0.2 standard deviations for both male and female workers, though the pattern for female workers again appears to be stronger in the short term. These education increases suggest that pushes to improve working conditions appear to have pulled in workers with better outside options to the garment industry, and this appears to be a stronger force that any supply shifts inward (as we discuss when we discuss the theoretical underpinnings for our results in section 6). Given evidence of compositional changes in the garment sector workforce after Rana Plaza, a natural question is whether our estimated coefficients are driven entirely by these changes, or also reflect causal changes on a given type of worker. We first note that our main empirical specification (equation 1) contains controls for observable human capital (age and education), interacted with gender and year fixed effects, so if the key changes are on these observable dimensions, they are already accounted for. We acknowledge, however, that there may be changes among unobservables as well. To address this possibility, we follow Oster (2019) in assessing the degree of selection on unobservables that would be needed to explain the entire estimated treatment effects. Table AA3 shows that it would take high level of selection to explain the wage results. For instance, if the degree of selection is the same between observables and unobservables, the treatment effect drops to 5.7%, but it is still statistically significant and economically relevant. We think this is a relatively conservative upper bound for selection on unobservables, given that selection on observables is 24 important for wages in the garment sector: there are high returns to education and experience (Heath and Mobarak, 2015, Heath, 2018). Meanwhile, because observables don’t explain much of the results on working conditions, even a high degree of selection (twice as much selection on unobservables compared to unobservables), leaves the working conditions results practically unchanged. Further, Table A4 looks for differences in gender composition. There is an overall increase of 2.7 percentage points that a garment sector worker is female, post Rana Plaza. However, this effect is not statistically significant, and also is driven by 2013. All together, we argue that our results are not entirely driven by changes in the types of workers in garments. Rather, they also represent causal effects on a fixed type of worker. 6 Theoretical mechanisms through which international scrutiny may affect garment workers Section 2 describes several of the different mechanisms through which international scrutiny post Rana Plaza may have affected garment workers in exporting factories, leading to the net effects seen in the empirical results in section 4. These mechanisms have different theoretical implications on wages and labor supply, depending on whether the labor market is perfectly competitive or monopsonistic. Accordingly, in this section we first discuss theoretical predictions on working conditions and contracts, and then go on to discuss theoretical predictions on wages and labor supply separately in the cases of perfect competition and monopsony. We then argue in section 6.4 that overall, our empirical results are more consistent with a model in which firms have monopsony power in the labor market. 6.1 Did firms respond to international scrutiny by improving working conditions? Buyers and other international observers pushed for better working conditions after Rana Plaza. While these efforts may have succeeded, it is also possible that garment factories improved official 25 conditions, but these improvements did not actually impact workers’ day to day experiences (as captured in their answers to the labor force survey).11 For instance, firms could instruct lower- level managers not to abuse workers, but the managers may not follow these directions. It is also possible that working conditions may improve on some dimensions – in particular, those that are easier to scrutinize – and get worse on others, leading to a net zero change in an aggregate index of working conditions. Thus, whether labor markets are perfectly competitive or monopsonistic, we anticipate that working conditions either increased, or did not change if the push toward better conditions was subverted by firms. 6.2 Contracts The international scrutiny that workers faced after Rana Plaza may have increased the uncertainty that firms face about the future. For instance, firms may perceive an increased likelihood that orders will be cancelled if audits find any issues in their factories. In this environment, firms would place increased value on flexibility, such as the ability to quickly scale up or down the size of their labor force, and they may be less likely to offer workers formal contracts. At the same time, if international scrutiny involved direct pressure related to contracts, firms might be more likely to offer formal contracts. Thus, the theoretical prediction on contracts is ambiguous, whether labor markets are perfectly competitive or monopsonistic. 6.3 Wages and labor supply in perfectly competitive labor markets If firms did respond to international scrutiny by improving working conditions, then this response could impact their wage in a world in which firms divide total compensation between wages and non-wage benefits. In the context of a perfectly competitive labor market – where total compensation was set at workers’ marginal product, pre Rana Plaza – this would mean that wages decreased a greater fraction of total compensation was shifted towards non-wage benefits (Summers, 1989, Mitchell, 1990, Almeida and Carneiro, 2012). 11 This possibility evokes a strand of the labor markets literature that highlights the potential for firms to evade labor market regulation, e.g. Ashenfelter and Smith (1979). 26 However, there is another potential channel affecting wages, which occurs if workers now per- ceive jobs as less safe than pre-Rana Plaza. This change could lead to a shift inward in labor supply, which means that total compensation could rise. If the rise in total compensation is greater than the tendency to shift a larger fraction of total compensation toward working con- ditions, then wages could rise.12 Note, however, that in this case, total quantity of labor would decrease due to the inward shift in labor supply; see figure 10. We note also that there might be additional channels for predictions on hours of work per worker (that are particularly salient in a perfectly competitive labor market), in addition to the overall labor supply results discussed above. For instance, firms might want to reduce hourly wages (as proposed above), but they find it easier to ask workers for unpaid overtime than to decrease base pay, because it is more effective to compete for workers on base pay. This is particularly likely in a competitive labor market, where firms may compete more heavily on base pay – which is more easily observable to workers at the time of hiring – than whether overtime is paid in a timely fashion (Boudreau, Heath and McCormick, 2017).13 6.4 Wages and labor supply in monopsonistic labor markets If firms have some monopsony power in the labor market, then they set workers’ wage equal to the marginal revenue they get from offering a worker that wage, which reflects the fact that they have to pay existing workers a higher wage. This results in the well-known result that firms higher fewer workers under monopsony than they would under perfect competition, in order to keep wages low. In this world, we consider the Rana Plaza reforms to shift the marginal revenue curve closer to the demand curve for workers in a perfectly competitive market, because firms now face external pressure (and the possibility of lost orders or other consequences) from paying compensation that 12 Another plausible factor that might affect total compensation is that the damaged reputation of the Bangladeshi name might lower demand for Bangladeshi clothing (if buyers now perceive an increased risk of bad publicity). But this would presumably be reflected in lower export volume, which we don’t see in figure 7. Moreover, this would lower wages, which is inconsistent with our empirical findings, so we do not focus on that channel in the section. 13 Other explanations related to hours of work per worker would affect the monopsony and perfectly competitive cases roughly similar. For instance, when there is reduced need for coordination between workers, the same amount of labor can be more productive when provided by fewer workers (Goldin, 2014). This might apply to firms’ responses to Rana Plaza if they reduced subcontracting, and thus need to rely more on fewer skilled workers who can do more parts of a garment. 27 is too low. We depict this case in figure 11. Similar to the case of perfect competition case depicted in figure 10, total compensation rises after a supply shift inward as workers’ perceptions of safety go down, resulting in increases to both wage and non-wage benefits (even as firms shift a greater share of compensation to non-wage benefits). In contrast to the perfect competition case, however, quantity of labor may not necessarily decrease in this case, due to the shift in firms’ marginal revenue curve. Thus, while either the monopsony or perfect competition models can explain the fact that both wages and working conditions improved, the fact that the quantity of labor did not change (either on the extensive margin in figure 8 or in hours of work per worker in table 2) is more consistent with the monopsony case just described than the perfect competition case described in section 6.3.14 This conclusion coincides with empirical evidence elsewhere that large manufacturing firms – especially those that are export-oriented – engage in profit-sharing with employees (Kruse, 1991, FitzRoy and Kraft, 1995, Black and Brainerd, 2004, Amisano and Del Boca, 2004, Heywood and Jirjahn, 2009, Harrison and Scorse, 2010, Brooks et al., 2021). We can then interpret our heterogeneous treatment effects in tables 2 and 3 in the context a model of monopsonistic labor markets. The two key channels for predictions on total compensation in the monopsony model are an inward labor supply shift (which also drives predictions in the case of perfectly competitive labor markets) and a shift outward in the marginal revenue curve. If workers are heterogeneous, this shift will be particularly impactful for workers whose labor supply is inelastic, which allowed the firms to lower wages particularly effectively, pre Rana Plaza. Recall that table 2 found qualitatively higher wage impacts of women (19%, versus 6.7% for men, though this difference is not statistically significant at conventional levels), and we believe it is indeed plausible that women’s labor supply is more inelastic, given that there are fewer alternative 14 We acknowledge that additional shocks or constraints could generate an increased labor supply in a perfectly competitive model. The need to keep production constant, given a fixed set of existing contracts, is one possibility. But contracts are generally set only months in advance, so production constraints could primarily only explain short-run results. Another possibility is a positive shock to output demand or a reduction in input costs which are complementary to labor. We do not think the former is likely, given that it seems unlikely that Rana Plaza would improve demand for Bangladeshi garment exports, and figure 7 shows no trend break in exports around 2013. While the latter is possible, we argue that our triple difference would absorb garment sector-wide shocks, and the production process is very similar in exporting and non-exporting factories. 28 employment opportunities for them, outside the garment sector. Column 1 of table 3 suggests that the same is true for low-skilled men; men with below median education increase their wages by 5.4% more than those with above-median education, and their working conditions also particularly increase (column 5). We also find that wages of men who are older, married, and have a child particularly increase. It is either plausible that they also have more inelastic labor supply if their demographics make it harder to move to find other jobs15 , or that they have an stronger inward shift in labor supply (which also leads to particular increases in wages if workers are not perfect substitutes) if safety concerns are particularly salient given that they have families to support. 7 Conclusion The Rana Plaza collapse of April 2013 killed over 1000 Bangladesh garment workers and injured many more. In response, international scrutiny on the garment sector increased, leading to a push by retailers to improve working conditions, an increased hesitancy for export factories to subcontract to smaller factories, and a minimum wage increase. This paper studies the net effects of these post-Rana Plaza responses on labor market outcomes of garment workers, using six rounds of the Bangladeshi Labor Force Survey and a triple difference identification strategy that compares garment workers to non-garment workers, in garment-exporting districts versus non-exporting districts, pre versus post Rana Plaza. We find that increased international scrutiny achieved its goal of improving working conditions, while also leading to average wage increases for garment workers in treated factories. We did not find average impacts on hours of work or the receipt of written contracts. These results suggest that garment manufacturers had scope to increase total compensation, compared to a model of compensating differentials in a perfectly competitive labor market in which improvements to working conditions come at the cost of reduced wages. While a commonly voiced concern is that, in a highly competitive global garment sector market, pushes to improve compensation will lead to firms to leave one country for one with even cheaper labor costs (Haq, 2018), we still find the 15 Menzel and Woodruff (2021) find that mobility is an important component of wage increases in the Bangladeshi garment sector. 29 potential for external pressure to improve workers’ welfare. 30 References Adhvaryu, Achyuta, Anant Nyshadham, and Huayu Xu. 2018. “Hostel takeover: Living condi- tions, reference dependence, and the well-being of migrant workers.” Afrin, Samina. 2014. “Labour Condition in the Apparel Industry of Bangladesh: Is Bangladesh Labour Law 2006 Enough?” Development Country Studies, 4(11). Almeida, Rita, and Pedro Carneiro. 2012. “Enforcement of labor regulation and informality.” Amer- ican Economic Journal: Applied Economics, 4(3): 64–89. Altonji, Joseph G, Todd E Elder, and Christopher R Taber. 2005. “Selection on observed and unobserved variables: Assessing the effectiveness of Catholic schools.” Journal of political economy, 113(1): 151–184. Amengual, Matthew, and Greg Distelhorst. 2019. “Can Sourcing Help Enforce Global Labor Stan- dards? Evidence from the Gap Inc Supply Chain.” Amisano, Gianni, and Alessandra Del Boca. 2004. “Profit related pay in Italy: A microeconometric analysis.” International Journal of Manpower. Amiti, Mary, and Donald R Davis. 2011. “Trade, firms, and wages: Theory and evidence.” The Review of Economic Studies, 79(1): 1–36. Ashenfelter, Orley, and Robert S Smith. 1979. “Compliance with the minimum wage law.” Journal of Political Economy, 87(2): 333–350. Begum, F, RN Ali, MA Hossain, and Sonia B Shahid. 2010. “Harassment of women garment workers in Bangladesh.” Journal of the Bangladesh Agricultural University, 8(452-2016-35685). Besley, Timothy, and Robin Burgess. 2004. “Can labor regulation hinder economic performance? Evidence from India.” The Quarterly Journal of Economics, 119(1): 91–134. Black, Sandra E, and Elizabeth Brainerd. 2004. “Importing equality? The impact of globalization on gender discrimination.” ILR Review, 57(4): 540–559. Botero, Juan C, Simeon Djankov, Rafael La Porta, Florencio Lopez-de Silanes, and Andrei Shleifer. 2004. “The regulation of labor.” The Quarterly Journal of Economics, 119(4): 1339–1382. Boudreau, Laura. 2019. “Multinational enforcement of labor law: Experimental evidence from Bangladeshs apparel sector.” mimeo. Boudreau, Laura, Rachel Heath, and Tyler McCormick. 2017. “Migrants, information, and work- ing conditions in Bangladeshi garment factories.” mimeo. Brooks, Wyatt J, Joseph P Kaboski, Illenin O Kondo, Yao Amber Li, and Wei Qian. 2021. “Infrastructure Investment and Labor Monopsony Power.” IMF Economic Review, 69(3): 470–504. Dragusanu, Raluca, and Nathan Nunn. 2018. “The effects of Fair Trade certification: evidence from coffee producers in Costa Rica.” National Bureau of Economic Research Working Paper 24260. Farole, Thomas, Yoonyoung Cho, Laurent Bossavie, and Reyes Aterido. 2017. “Bangladesh jobs diagnostic.” Jobs working paper 9. Washington, D.C. : World Bank Group. 31 Ferman, Bruno, and Cristine Pinto. 2019. “Inference in differences-in-differences with few treated groups and heteroskedasticity.” Review of Economics and Statistics, 101(3): 452–467. FitzRoy, Felix R, and Kornelius Kraft. 1995. “On the choice of incentives in firms.” Journal of Economic Behavior & Organization, 26(1): 145–160. ıas, Judith A, David S Kaplan, and Eric A Verhoogen. 2009. “Exports and wage premia: Fr´ Evidence from Mexican employer-employee data.” Columbia University mimeo. Gibbs, Andrew, Rachel Jewkes, Samantha Willan, Mahfuz Al Mamun, Kausar Parvin, Marat Yu, and Ruchira Naved. 2019. “Workplace violence in Bangladesh’s garment industry.” Social Science & Medicine, 235: 112383. Goldin, Claudia. 2014. “A grand gender convergence: Its last chapter.” American Economic Review, 104(4): 1091–1119. Haq, Naimul. 2018. “Bangladeshs Garment Industry Boom Leaving Workers Behind.” Inter Press Service, February 9. Harrison, Ann, and Jason Scorse. 2010. “Multinationals and anti-sweatshop activism.” The American Economic Review, 100(1): 247–273. Heath, Rachel. 2018. “Why do firms hire using referrals? Evidence from Bangladeshi garment factories.” Journal of Political Economy, 126(4): 1691–1747. Heath, Rachel, and A Mushfiq Mobarak. 2015. “Manufacturing growth and the lives of Bangladeshi women.” Journal of Development Economics, 115: 1–15. Heywood, John S, and Uwe Jirjahn. 2009. “Profit sharing and firm size: The role of team produc- tion.” Journal of Economic Behavior & Organization, 71(2): 246–258. International Labour Organization. 2019. “The future of work in textiles, clothing, leather and footwear.” Kabeer, Naila, Lopita Haq, and Munshi Sulaiman. 2019. “Paradigm shift or business as usual? Workers perspectives on multi–stakeholder initiatives in Bangladesh after Rana Plaza.” Development and Change. Khosla, Nidhi. 2009. “The ready-made garments industry in Bangladesh: A means to reducing gender- based social exclusion of women?” Journal of International Women’s Studies, 11(1): 289–303. Kovak, Brian K. 2013. “Regional effects of trade reform: What is the correct measure of liberalization?” American Economic Review, 103(5): 1960–76. Kruse, Douglas L. 1991. “Profit-sharing and employment variability: Microeconomic evidence on the Weitzman theory.” ILR Review, 44(3): 437–453. Labowitz, Sarah, and Dorothee Baumann-Pauly. 2015. “Beyond the Tip of the Iceberg: Bangladesh’s Forgotten Apparel Workers.” NYU Stern, Center for Business and Human Rights. Mapped in Bangladesh. 2021. “Mapped in Bangladesh: Vision for an Efficient, Transparent, and Accountable RMG Industry.” Menzel, Andreas, and Christopher Woodruff. 2021. “Gender Wage Gaps and Worker Mobility: Evidence from the Garment Sector in Bangladesh.” Labour Economics, 102000. 32 Mitchell, Olivia S. 1990. “The effects of mandating benefits packages.” National Bureau of Economic Research Working Paper 3260. Naved, Ruchira, Tabassum Rahman, Samantha Willan, Rachel Jewkes, and Andrew Gibbs. 2018. “Female garment workers experiences of violence in their homes and workplaces in Bangladesh: A qualitative study.” Social Science & Medicine, 196: 150–157. Oster, Emily. 2019. “Unobservable selection and coefficient stability: Theory and evidence.” Journal of Business & Economic Statistics, 37(2): 187–204. Parker, Dominic P, Jeremy D Foltz, and David Elsea. 2016. “Unintended consequences of sanc- tions for human rights: Conflict minerals and infant mortality.” The Journal of Law and Economics, 59(4): 731–774. Rahman, Mustafizur, and Khondaker Golam Moazzem. 2017. “The legacy of Rana Plaza: Im- proving labour and social standards in Bangladeshs apparel industry.” In Governing Corporate Social Responsibility in the Apparel Industry after Rana Plaza. 81–109. Springer. Reuters. 2017. “Factbox: Major industrial accidents in Bangladesh in recent years.” July 3. Reuters. 2018. “Pay more for your clothes, Bangladesh workers tell global fashion brands.” September 27. Reuters. 2019. “Six years after Bangladesh’s Rana Plaza disaster, fashion brands urged to pay more.” April 24. Subramanian, Nivedhitha. 2019. “Workplace Attributes and Womens Labor Supply Decisions: Evi- dence from a Randomized Experiment.” mimeo. Summers, Lawrence H. 1989. “Some simple economics of mandated benefits.” The American Economic Review, 79(2): 177–183. Tanaka, Mari. 2020. “Exporting Sweatshops? Evidence from Myanmar.” Review of Economics and Statistics, 1–44. Topalova, Petia. 2010. “Factor immobility and regional impacts of trade liberalization: Evidence on poverty from India.” American Economic Journal: Applied Economics, 2(4): 1–41. Verhoogen, Eric A. 2008. “Trade, quality upgrading, and wage inequality in the Mexican manufacturing sector.” The Quarterly Journal of Economics, 123(2): 489–530. WTO. 2019. “World Trade Statistical Review 2019.” Geneva: World Trade Organization. 33 Figure 1: Total employment in the garment industry in Bangladesh Nation-wide Employment in the Garment Industry 4 3 millions of workers 21 0 1985 1990 1995 2000 2005 2010 2015 Year Source: Bangladesh Garment Manufacturers' Export Association (BGMEA) report of total garment sector employment by year. 18 34 Figure 2: Total Number of registered garment factories Source: Bangladesh Garment Manufacturers and Exporters Association (BGMEA) 35 Figure 3: Wages in the garment industry before and after Rana Plaza, by gender Quarter 1 of 2013 Quarter 1 of 2015 Male Female Male Female 50 50 40 40 30 30 Percent Percent 20 20 10 10 0 0 50 0 10 00 15 00 20 00 25 00 30 000 0+ 50 0 10 00 15 00 20 00 25 00 30 000 0+ 50 0 10 00 15 00 20 00 25 00 30 000 0+ 50 0 10 00 15 00 20 00 25 00 30 000 0+ 0 0 0 0 0 0 0 0 0 0 0 0 00 00 00 00 nominal monthly wage in taka nominal monthly wage in taka Notes: The vertical red line represents the new nominal minimum wage implemented in the garment industry from January 2014 onwards after the Rana Plaza collapse. Source: Bangladesh National Labor Force Surveys. 36 Figure 4: Distribution of garment workers across treated and non-treated districts of Bangladesh 37 Figure 5: Marginal effects by year Log(hourly wage) Females Males .4 .4 .3 .3 Estimated marginal effect Estimated marginal effect .2 .2 .1 .1 0 0 -.1 -.1 2013 2015 2016 2013 2015 2016 Working conditions index Females Males 3 3 2 2 Estimated marginal effect Estimated marginal effect 1 1 0 0 -1 -1 -2 -2 2013 2015 2016 2013 2015 2016 38 Marginal effects by year (continued) Hours last week Females Males 5 5 Estimated marginal effect Estimated marginal effect 0 0 -5 -5 -10 -10 2013 2015 2016 2013 2015 2016 1(Contract) Females Males .6 .6 .4 .4 Estimated marginal effect Estimated marginal effect .2 .2 0 0 -.2 -.2 -.4 -.4 2013 2015 2016 2013 2015 2016 Notes: Depicted are marginal effects by gender and year, from estimates of equation 1. These regressions include controls for worker’s age, gender, level of schooling, region dummies, and dummy for urban location, interacted with year dummies and gender dummies, and a triple interaction with year and gender. Sampling weights included. Standard errors clustered at the primary sampling unit: *** p<0.01, ** p<0.05, * p<0.1 39 Figure 6: Trends over time Log(hourly wage) Female Male Test of joint sig of parallel trends: Test of joint sig of parallel trends: F = .9350000000000001; P = .423 F = .058; P = .982 4 4 3.5 3.5 log(wage) log(wage) 3 3 2.5 2.5 2 2 03 05 10 13 20 5 16 03 05 10 13 20 5 16 1 1 20 20 20 20 20 20 20 20 20 20 year year Non-Garm Non-Treated District Non-Garm Treated District Garment Non-Treated District Garment Treated District Working conditions index Female Male Test of joint sig of parallel trends: Test of joint sig of parallel trends: F = 1.221; P = .3 F = .787; P = .501 4 4 2 2 standard deviations standard deviations 0 0 -2 -2 -4 -4 03 05 10 13 20 5 16 03 05 10 13 20 5 16 1 1 20 20 20 20 20 20 20 20 20 20 year year Non-Garm Non-Treated District Non-Garm Treated District Garment Non-Treated District Garment Treated District 40 Trends over time (continued) Hours last week Female Male Test of joint sig of parallel trends: Test of joint sig of parallel trends: F = .885; P = .448 F = .465; P = .706 60 60 50 50 hours hours 40 40 30 30 03 05 10 13 20 5 16 03 05 10 13 20 5 16 1 1 20 20 20 20 20 20 20 20 20 20 year year Non-Garm Non-Treated District Non-Garm Treated District Garment Non-Treated District Garment Treated District 1(Contract) Female Male Test of joint sig of parallel trends: Test of joint sig of parallel trends: F = .997; P = .393 F = 2.475; P = .06 1 1 .8 .8 percent percent .6 .6 .4 .4 .2 .2 03 05 10 13 20 5 16 03 05 10 13 20 5 16 1 1 20 20 20 20 20 20 20 20 20 20 year year Non-Garm Non-Treated District Non-Garm Treated District Garment Non-Treated District Garment Treated District Notes: The test of joint significance of parallel trends includes the following three coefficients for Log(wage), Working Conditions, and Hours: Year = 2005 × Garment × Treated Dist , Year = 2010 × Garment × Treated Dist, Year = 2013 × Garment × Treated Dist × Female. For Contract, it includes the following two coefficients: Year = 2010 × Garment × Treated Dist, Year = 2013 × Garment × Treated Dist × Female. 41 Figure 7: Total value of apparel exports by exporting country (in 1,000 USD) Yearly apparel exports In millions of USD 40000 30000 00 00 20 00 20000 00 15 00 00 10 10000 0 00 50 0 2000 2005 2010 2015 2020 year 0 2000 2005 2010 2015 2020 China year Bangladesh Vietnam Turkey India Cambodia Hong Kong Source: UN COMTRADE and ITC statistics. 42 Figure 8: Percentage of workers in the garment industry Treated Districts Women Men .8 .8 .6 .6 fraction in garments fraction in garments .4 .4 .2 .2 0 0 03 05 10 3 20 5 16 03 05 10 3 20 5 16 1 1 1 1 20 20 20 20 20 20 20 20 20 20 year year Population of Employed Workers Pop. of Wage Workers Population of Manufacturing Workers Un-Treated Districts Women Men .8 .8 .6 .6 fraction in garments fraction in garments .4 .4 .2 .2 0 0 03 05 10 13 20 5 16 03 05 10 13 20 5 16 1 1 20 20 20 20 20 20 20 20 20 20 year year Population of Employed Workers Pop. of Wage Workers Population of Manufacturing Workers 43 Figure 9: Compositional changes in garment workers, post Rana Plaza Female: Overall ge 2013 2015 A 2016 Male: Overall 2013 2015 2016 Female: Overall n tio 2013 2015 2016 a uc Male: Overall Ed 2013 2015 2016 Female: Overall d rie 2013 2015 2016 ar M Male: Overall 2013 2015 2016 Female: Overall H 2013 H 2015 2016 in ld Male: Overall hi 2013 C 2015 2016 Female: Overall an 2013 2015 rb 2016 U Male: Overall 2013 2015 2016 Female: Overall ka 2013 ha 2015 2016 Male: Overall D 2013 2015 2016 -1 -.5 0 .5 marginal effect, in standard deviations Notes: Depicted are marginal effects by gender and year estimated from regressions (separately for each gender) of the bold variables on the left on an indicator for Garment × P ost × T reatedDist, as well as the main effect of Garment (and its interaction with T reatedDist) and year dummies interacted with Garment and T reatedDist. Sampling weights included. Standard errors clustered at the primary sampling unit: *** p<0.01, ** p<0.05, * p<0.1 44 Figure 10: Compensation and quantity of labor after a supply shift: Perfect competition case Value of total compensation (C) S’ S C’ nwb' C nwb w' w D Quantity of Q’ Q labor 45 Figure 11: Compensation and quantity of labor after a supply shift: Monopsony case Value of total compensation (C) S’ S C’ nwb' C nwb D w' w MR’ MR Quantity of labor Q = Q’ 46 Table 1: Summary statistics in estimation sample Treated District Non-Treated District P-values for test of equality Garment Non-Garm Garment Non-Garm Garment vs NonGarment TreatedDist vs Non TreatDist NonTrDist Garm NonGarm Sex F M F M F M F M F M F M F M F M Age 25.76 29.92 27.55 31.62 25.93 27.42 31.42 31.82 0.000 0.000 0.000 0.000 0.000 0.454 0.257 0.000 Educ (years) 5.37 7.31 5.07 6.82 4.62 7.09 3.48 5.48 0.000 0.000 0.000 0.000 0.000 0.008 0.000 0.000 Married 0.684 0.712 0.659 0.690 0.652 0.541 0.683 0.672 0.015 0.005 0.023 0.000 0.007 0.000 0.047 0.007 Children in hh 0.66 0.63 0.65 0.73 0.88 0.82 0.87 0.91 0.683 0.000 0.944 0.000 0.000 0.000 0.000 0.000 Urban 0.706 0.742 0.732 0.695 0.227 0.120 0.276 0.282 0.009 0.000 0.000 0.000 0.000 0.000 0.000 0.000 Dhaka district 0.748 0.815 0.893 0.796 0.319 0.337 0.274 0.298 0.000 0.004 0.001 0.000 0.000 0.000 0.000 0.000 Hourly wage 34.5 46.5 31.6 46.5 30.8 43.9 23.7 34.8 0.000 0.926 0.000 0.000 0.000 0.000 0.000 0.000 Hours last week 54.2 56.7 53.2 56.2 51.5 55.0 43.8 52.5 0.000 0.003 0.000 0.000 0.000 0.000 0.000 0.000 1(Sick leave) 0.404 0.470 0.254 0.357 0.235 0.296 0.133 0.193 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 47 1(Mat leave) 0.538 0.386 0.298 0.182 0.218 0.163 0.174 0.088 0.000 0.000 0.002 0.000 0.000 0.000 0.000 0.000 1(Contract) 0.611 0.653 0.492 0.486 0.579 0.687 0.446 0.380 0.000 0.000 0.000 0.000 0.024 0.000 0.002 0.000 1(Injured) 0.014 0.026 0.013 0.043 0.026 0.012 0.011 0.053 0.829 0.000 0.008 0.000 0.011 0.001 0.552 0.028 1(Abused) 0.190 0.120 0.139 0.103 0.138 0.150 0.185 0.112 0.000 0.006 0.002 0.000 0.000 0.000 0.000 0.108 1(Dangerous) 0.279 0.329 0.445 0.428 0.204 0.217 0.469 0.376 0.000 0.000 0.000 0.000 0.000 0.000 0.154 0.000 Firm size 10+ 0.693 0.770 0.726 0.628 0.649 0.737 0.594 0.469 0.001 0.000 0.002 0.000 0.001 0.000 0.000 0.000 Year = 2003 0.025 0.022 0.062 0.085 0.269 0.040 0.273 0.193 0.000 0.000 0.736 0.000 0.000 0.000 0.000 0.000 Year = 2005 0.077 0.102 0.102 0.162 0.060 0.119 0.196 0.213 0.000 0.000 0.000 0.000 0.010 0.005 0.000 0.000 Year = 2010 0.117 0.171 0.091 0.119 0.186 0.210 0.054 0.080 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 Year = 2013 0.344 0.243 0.323 0.208 0.163 0.159 0.262 0.156 0.044 0.000 0.000 0.581 0.000 0.000 0.000 0.000 Year = 2015 0.211 0.226 0.228 0.202 0.169 0.243 0.109 0.175 0.057 0.000 0.000 0.000 0.000 0.041 0.000 0.000 Year = 2016 0.225 0.235 0.193 0.223 0.153 0.229 0.105 0.183 0.000 0.093 0.000 0.000 0.000 0.456 0.000 0.000 N 8434 7047 2721 6849 1786 4140 2895 14111 Notes: Wages deflated to 2010 taka; the average exchange rate in 2010 was 69 taka/1 USD. Firm size available only 2010 onwards. Contract available 2005 onwards. Injured, Dangerous, and Abused available only 2013 onwards. Table 2: Treatment effects of Rana Plaza on garment workers Dependent Variable Log(wage) Working Hours 1(Contract) Conditions last week Panel A: Overall Garment × Post × Treated District 0.100** 0.803*** -0.777 -0.00221 (0.0494) (0.205) (1.047) (0.0576) Garment -0.0294 0.000448 4.205*** 0.544*** (0.0510) (0.0901) (0.992) (0.0365) Garment × Female 0.202*** 0.0861 10.78*** -0.158** (0.0666) (0.0832) (1.311) (0.0797) Treated District 0.142*** 0.104 4.295*** 0.0938 (0.0447) (0.0929) (1.157) (0.0573) Treated District × Garment -0.108** -0.160 -0.691 -0.104** (0.0460) (0.153) (0.881) (0.0461) Treated District × Garment × Female -0.00566 0.0317 -1.744** 0.107** (0.0423) (0.150) (0.885) (0.0425) R2 0.421 0.277 0.178 0.157 N 46638 46638 46638 42439 Panel B: Interactions with Gender Garment × Post × Treated Dist 0.0673 0.731*** -0.639 0.00764 (0.0526) (0.208) (1.072) (0.0548) Garment × Post × Treated Dist × Female 0.121 0.266 -0.503 -0.0393 (0.0897) (0.301) (1.811) (0.0937) Garment -0.0377 -0.0178 4.240*** 0.548*** (0.0513) (0.0913) (0.996) (0.0353) Garment × Female 0.221*** 0.128 10.70*** -0.174** (0.0696) (0.0935) (1.346) (0.0867) Treated District 0.139*** 0.0980 4.307*** 0.0955* (0.0448) (0.0933) (1.157) (0.0567) Treated District × Garment -0.0878* -0.117 -0.773 -0.111** (0.0486) (0.156) (0.903) (0.0441) Treated District × Garment × Female -0.0727 -0.115 -1.466 0.131 (0.0798) (0.220) (1.440) (0.0808) Net effect on females 0.188 0.996 -1.142 -0.0320 P-value 0.0290 0.00200 0.532 0.761 R2 0.421 0.277 0.178 0.157 N 46638 46638 46638 42439 Notes: All regressions include controls for time fixed effects interacted with Garmentijt and T reatedDistrictj (and each variable’s interaction with gender); T reatedDistrictj interacted with Garmentijt and a triple interactions with gender; worker’s age and level of schooling, interacted with year dummies and gender dummies, and a triple interaction with year and gender. Sampling weights included. Standard errors clustered at the primary sampling unit: *** p<0.01, ** p<0.05, * p<0.1 48 Table 3: Heterogeneous treatment effects of Rana Plaza on garment workers Dependent Variables Log(wage) Working Conditions Interaction LowEduc Young Married Child LowEduc Young Married Child Garm × Post × TreatedD 0.0683 0.0707 0.0700 0.0668 0.735*** 0.733*** 0.737*** 0.731*** (0.0524) (0.0528) (0.0525) (0.0527) (0.208) (0.208) (0.208) (0.208) Garm × Post × TreatedD × Female 0.120 0.120 0.108 0.121 0.262 0.274 0.215 0.268 (0.0897) (0.0900) (0.0898) (0.0899) (0.301) (0.301) (0.302) (0.301) Garm × Post × TreatedD × INT 0.0544*** -0.0678*** 0.0651*** 0.0299* 0.207*** -0.0536 0.146** -0.0334 (0.0163) (0.0146) (0.0147) (0.0154) (0.0623) (0.0591) (0.0628) (0.0651) Garm × Post × TreatedD × Female × INT -0.0265 0.0964*** 0.0130 -0.0193 -0.233* 0.184* 0.194* -0.00384 (0.0250) (0.0220) (0.0200) (0.0212) (0.121) (0.103) (0.102) (0.0995) Net effect, females, INT = 0 0.189 0.191 0.178 0.188 0.996 1.008 0.952 0.999 P-value 0.0280 0.0270 0.0390 0.0290 0.00200 0.00200 0.00300 0.00200 Net effect, males, INT = 1 0.123 0.00300 0.135 0.0970 0.942 0.680 0.883 0.698 49 P-value 0.0280 0.956 0.0120 0.0790 0 0.00200 0 0.00100 Net effect, females, INT = 1 0.216 0.220 0.256 0.198 0.970 1.138 1.292 0.962 P-value 0.0140 0.0150 0.00300 0.0220 0.00500 0.00100 0 0.00400 INT = 1 effect - INT = 0 effect, females 0.0280 0.0290 0.0780 0.0110 -0.0260 0.130 0.340 -0.0370 P-value 0.173 0.102 0 0.494 0.815 0.118 0 0.658 R2 0.421 0.421 0.422 0.421 0.277 0.277 0.278 0.277 N 46638 46638 46638 46638 46638 46638 46638 46638 Table 3 (continued): Heterogeneous treatment effects of Rana Plaza on garment workers Dependent Variable Hours Contract Interaction LowEduc Young Married Child LowEduc Young Married Child Garm × Post × TreatedD -0.649 -0.619 -0.599 -0.639 0.0106 0.00844 0.00889 0.00780 (1.071) (1.070) (1.069) (1.071) (0.0545) (0.0547) (0.0547) (0.0548) Garm × Post × TreatedD × Female -0.493 -0.580 -0.651 -0.474 -0.0416 -0.0375 -0.0420 -0.0367 (1.812) (1.811) (1.816) (1.811) (0.0938) (0.0936) (0.0940) (0.0936) Garm × Post × TreatedD × INT -0.535 -0.412 0.991*** -0.0142 0.116*** -0.0166 0.0271* -0.00761 (0.328) (0.292) (0.306) (0.301) (0.0156) (0.0141) (0.0149) (0.0151) Garm × Post × TreatedD × Female × INT 0.640 -0.240 -0.171 -0.433 -0.0873*** 0.0485* -0.00756 -0.0328 (0.486) (0.461) (0.405) (0.391) (0.0256) (0.0251) (0.0212) (0.0214) Net effect, females, INT = 0 -1.142 -1.198 -1.250 -1.113 -0.0310 -0.0290 -0.0330 -0.0290 P-value 0.532 0.512 0.496 0.542 0.766 0.780 0.751 0.781 Net effect, males, INT = 1 -1.185 -1.030 0.392 -0.653 0.126 -0.00800 0.0360 0 50 P-value 0.294 0.365 0.715 0.565 0.0290 0.886 0.522 0.997 Net effect, females, INT = 1 -1.037 -1.850 -0.430 -1.561 -0.00200 0.00300 -0.0140 -0.0690 P-value 0.590 0.330 0.812 0.400 0.981 0.978 0.896 0.516 INT = 1 effect - INT = 0 effect, females 0.105 -0.651 0.820 -0.447 0.0280 0.0320 0.0200 -0.0400 P-value 0.796 0.0770 0.0150 0.134 0.184 0.119 0.318 0.0240 R2 0.178 0.178 0.179 0.178 0.160 0.157 0.157 0.157 N 46638 46638 46638 46638 42439 42439 42439 42439 LowEduc and Young are a binary variable that equals one if observation is at or below the median of the estimation sample (7 years of education; 28 years of age). Child = 1 if there is a child living in the respondent’s household. INT denotes the interaction with the dependent variable reported in the corresponding column. All regressions include controls for time fixed effects interacted with Garmentijt and T reatedDistrictj (and each variable’s interaction with gender); T reatedDistrictj interacted with Garmentijt and a triple interactions with gender; worker’s age, gender, level of schooling, region dummies, and dummy for urban location, interacted with year dummies and gender dummies, and a triple interaction with year and gender. Sampling weights included. Standard errors clustered at the primary sampling unit: *** p<0.01, ** p<0.05, * p<0.1. Table 4: Pre-trends Dependent Variable Log(wage) Working Hours 1(Contract) Conditions last week Garment -0.0423 0.0241 5.036*** 0.528*** (0.0597) (0.0817) (1.078) (0.0361) Garment × Female 0.227*** 0.00144 9.653*** -0.0955 (0.0784) (0.0817) (1.489) (0.110) Treated District 0.137*** 0.113 4.587*** 0.0847 (0.0470) (0.100) (1.208) (0.0661) Treated District × Garment -0.0770 -0.216* -2.661 -0.0723 (0.0966) (0.130) (1.975) (0.0701) Treated District × Garment × Female -0.0912 0.352** 1.506 0.00682 (0.133) (0.157) (2.844) (0.151) Year = 2005 × Garment × Treated Dist -0.00395 -0.187 1.687 (0.140) (0.236) (2.769) Year = 2010 × Garment × Treated Dist -0.0281 0.192 2.691 -0.0474 (0.122) (0.253) (2.305) (0.0963) Year = 2013 × Garment × Treated Dist 0.0144 0.684 1.640 -0.118 (0.121) (0.846) (2.759) (0.129) Year = 2005 × Garment × Treated Dist × Female -0.227 0.00635 -5.892 (0.250) (0.339) (4.477) Year = 2010 × Garment × Treated Dist × Female 0.191 -1.072** -0.838 0.311 (0.193) (0.476) (3.769) (0.200) Year = 2013 × Garment × Treated Dist × Female 0.0893 -1.096 -4.623 0.0480 (0.181) (0.922) (3.879) (0.210) F-stat for joint sig of garment × year dummies 0.556 1.338 0.642 1.168 P-value 0.766 0.236 0.697 0.323 F-stat – male dummies 0.0575 0.787 0.465 0.426 P-value 0.982 0.501 0.706 0.653 F-stat – female dummies 0.994 2.133 0.945 1.478 P-value 0.394 0.0940 0.418 0.228 F-stat – net effect on females 0.935 1.221 0.885 1.490 P-value 0.423 0.300 0.448 0.226 R2 0.358 0.334 0.240 0.231 N 12873 12873 12873 8726 Notes: Contract available 2005 onwards, so 2005 dummies (and their interactions with gender, Garmentijt and T reatedDistrictj ) omitted in column 4. The F test for joint significance of garment × year dummies involves six coefficients in columns 1-3 and four coefficients in column 4. The F test for male and female dummies involves three coefficients in columns 1-3 and two coefficients in column 4. The F test for the net effect on females involves three coefficients (Year = 2005 × Garment × Treated Dist × Female, Year = 2010 × Garment × Treated Dist × Female, Year = 2013 × Garment × Treated Dist × Female) in columns 1-3 and two coefficients (Year = 2005 × Garment × Treated Dist × Female, Year = 2010 × Garment × Treated Dist × Female) in column 4. All regressions include controls for time fixed effects interacted with Garmentijt and T reatedDistrictj (and each variable’s interaction with gender); T reatedDistrictj interacted with Garmentijt and a triple interactions with gender; worker’s age and level of schooling, interacted with year dummies and gender dummies, and a triple interaction with year and gender. Sampling weights included. Standard errors clustered at the primary sampling 51 unit: *** p<0.01, ** p<0.05, * p<0.1 Table 5: Hypothesized impacts on workers in control groups Treated Industry? Yes No Treated Yes Min wage increase Within-district, cross-industry worker flows District? (possibly with stronger enforcement) Minimal international scrutiny International scrutiny No Min wage increase Likely minimal Indirect effects of international scrutiny (directed at non-exporting firms) Reduced subcontracting Within-industry, cross-district worker flows 52 Table 6: Treatment effects of Rana Plaza on garment workers in treated versus non-treated districts (1) (2) (3) (4) (5) (6) (7) (8) Non-treated districts Treated districts Log(wage) Working Hours 1(Contract) Log(wage) Working Hours 1(Contract) Conditions last week Conditions last week Garment × Post -0.0488 -0.0463 -1.349** -0.190*** -0.00291 0.682*** -1.397 -0.183*** (0.0315) (0.136) (0.610) (0.0315) (0.0416) (0.168) (0.880) (0.0469) Garment × Post × Female -0.175*** 0.0205 -5.981*** 0.175** -0.0489 0.0117 -2.494** 0.115* (0.0611) (0.199) (1.318) (0.0748) (0.0672) (0.245) (1.242) (0.0626) Net effect on females -0.224 -0.0260 -7.330 -0.0150 -0.0520 0.693 -3.892 -0.0680 P-value 0 0.897 0 0.848 0.425 0.00900 0.00300 0.371 R2 0.400 0.194 0.176 0.161 0.431 0.278 0.109 0.134 N 22371 22371 22371 19281 24267 24267 24267 23158 Notes: All regressions include controls for time fixed effects interacted with Garmentijt and its variable interaction with gender; the main effect of Garmentijt and its interaction with gender; worker’s age, gender, level of schooling, region dummies, and dummy for urban location, interacted with year 53 dummies and gender dummies, and a triple interaction with year and gender. Sampling weights included. Standard errors clustered at the primary sampling unit: *** p<0.01, ** p<0.05, * p<0.1 Appendix 54 Table A1: Summary statistics on specific abusive behaviors and dangerous conditions Garments Other P-values for test of equality Manufacturing F vs M Gar vs Other Female Male Female Male Gar Other F M Abusive behavior constantly shouted/insulted 0.126 0.118 0.158 0.100 0.088 0.000 0.000 0.000 beaten /physically hurt 0.006 0.003 0.008 0.006 0.002 0.213 0.327 0.004 sexually abused 0.047 0.001 0.011 0.001 0.000 0.000 0.000 0.758 others 0.003 0.002 0.000 0.004 0.020 0.001 0.008 0.001 Dangerous conditions dust, fumes,noise or vibration 0.185 0.172 0.234 0.185 0.011 0.000 0.000 0.008 fire, gas, flames 0.007 0.019 0.049 0.118 0.000 0.000 0.000 0.000 extreme cold or heat 0.029 0.040 0.058 0.079 0.000 0.000 0.000 0.000 dangerous tools 0.153 0.137 0.178 0.187 0.001 0.266 0.004 0.000 55 work underground or at heights 0.002 0.004 0.011 0.028 0.083 0.000 0.000 0.000 work in water/pond/river 0.003 0.003 0.001 0.009 0.827 0.000 0.172 0.000 workplace too dark or confined 0.020 0.021 0.031 0.028 0.566 0.409 0.002 0.001 chemicals/explosives 0.005 0.013 0.036 0.057 0.000 0.000 0.000 0.000 other things (specify) 0.001 0.005 0.006 0.013 0.000 0.001 0.000 0.000 N 10,386 11,363 2229 11,527 All variables available only 2013 onwards. Table A2: FGLS Treatment effects of Rana Plaza on garment workers Dependent Variable Log(wage) Working Hours 1(Contract) Conditions last week Panel A: Overall Garment × Post × Treated Dist 0.0836*** 0.646*** -0.344 -0.0154 (0.0218) (0.0800) (0.435) (0.0208) Garment -0.0546* 0.0392 4.560*** 0.242*** (0.0288) (0.0994) (0.574) (0.0196) Garment × Female 0.178*** 0.0150 10.39*** -0.132*** (0.0374) (0.124) (0.743) (0.0428) Treated District 0.111*** 0.0523 3.867*** 0.326*** (0.0188) (0.0777) (0.380) (0.0252) Treated District × Garment -0.0914*** -0.0968 -1.241*** -0.132*** (0.0181) (0.0667) (0.358) (0.0175) Treated District × Garment × Female 0.000129 -0.0202 -1.221** 0.0826*** (0.0245) (0.0871) (0.489) (0.0230) Observations 46638 46638 46638 42439 Panel B: Interactions with Gender Garment × Post × Treated Dist 0.0612** 0.573*** -0.116 -0.0273 (0.0255) (0.0949) (0.509) (0.0240) Garment × Post × Treated Dist × Female 0.0834* 0.253 -0.849 0.0469 (0.0492) (0.177) (0.981) (0.0478) Garment -0.0608** 0.0269 4.619*** 0.245*** (0.0291) (0.0998) (0.578) (0.0198) Garment × Female 0.193*** 0.0413 10.24*** -0.186*** (0.0385) (0.126) (0.762) (0.0351) Treated District 0.109*** 0.0457 3.886*** 0.328*** (0.0189) (0.0779) (0.380) (0.0253) Treated District × Garment -0.0780*** -0.0532 -1.376*** -0.125*** (0.0197) (0.0733) (0.391) (0.0190) Treated District × Garment × Female -0.0460 -0.166 -0.761 0.0526 (0.0365) (0.135) (0.723) (0.0382) Observations 46638 46638 46638 42439 Notes: All regressions include controls for time fixed effects interacted with Garmentijt and T reated Districtj (and each variable’s interaction with gender); T reated Districtj interacted with Garmentijt and a triple interactions with gender; worker’s age, gender, level of schooling, region dummies, and dummy for urban location, interacted with year dummies and gender dummies, and a triple interaction with year and gender. Sampling weights included. Standard errors clustered at the primary sampling unit: *** p<0.01, ** p<0.05, * p<0.1 56 Table A3: Estimated treatment effects under various assumptions about selection on unobserv- ables, relative to observables δ 0.5 1 1.5 2 Log(wage) 0.079 0.057 0.036 0.014 [ 0.028] [ 0.030] [ 0.032] [ 0.036] Working conditions 0.792 0.780 0.768 0.756 [ 0.085] [ 0.088] [ 0.093] [ 0.098] Hours -0.725 -0.674 -0.623 -0.572 [ 0.462] [ 0.471] [ 0.483] [ 0.497] Contract -0.021 -0.041 -0.060 -0.079 [ 0.022] [ 0.024] [ 0.025] [ 0.028] Notes: δ gives the relative amount of selection on unobserveables, compared to observables, as described in Oster (2019). We assume that the additional r-squared explained by unobserveables is equal to the amount explained by observables (education and age interacted with gender and year fixed effects and gender interacted with treated district and period and a triple interaction of gender × treated district × period), i.e., Rmax = R∼ + (R∼ − R0 ). Sampling weights included. Standard errors calculated by bootstrapping, taking repeated samples stratified and the psu level and using 1000 repetitions: *** p<0.01, ** p<0.05, * p<0.1 57 Table A4: Changes in gender composition, post-Rana Plaza Dependent Variable Female Garment × Post × Treated District 0.0269 0.0444 (0.0283) (0.0360) Garment × Post × Treated District × 2015 -0.0770 (0.0519) Garment × Post × Treated District × 2016 -0.0189 (0.0501) Observations 46776 46776 R2 0.128 0.128 Notes: Control variables include the main effect of Garment (and its interaction with T reatedDistrict) and year dummies interacted with Garment and T reatedDistrict. Sampling weights included. Standard errors clustered at the primary sampling unit: *** p<0.01, ** p<0.05, * p<0.1 58