Migration, Labor, and Women’s Empowerment: Evidence From an Agricultural Value Chain in Bangladesh

As a substantial portion of the rural labor force migrates to urban areas, it is commonly assumed that women could take over traditionally male tasks in agricultural production, with potentially empowering outcomes for women. We study how changes in the supply of labor may influence female labor participation and empowerment outcomes. Using a detailed panel dataset on jute producers in Bangladesh, we test whether out-migration of household members and perceived labor shortages are associated with the share of household and hired labor performed by women, and women’s empowerment. When a household experiences reduced household or hired labor supply, we observe a relatively larger use of female household labor but not of female hired labor. We find that reduced male household labor supply is associated with improved wages mainly for male laborers, whereas reduced female household labor is associated with improved wages for male laborers and enhanced empowerment of other women in the household. These findings suggest that given existing gender norms, male and female labor are not perfect substitutes for one another, and as a result, male outmigration is not associated with improved outcomes for women in cash crop production. Our results demonstrate a need for better understanding of the role of gender in rural labor markets, particularly in contexts of rapid urbanization.


Introduction
Gender has historically played an important role in shaping the division of labor in agricultural production, and this, in turn, has been shown to have persistent effects on contemporary gender norms in industrialized societies (Boserup 1970;Alesina et al. 2013). Globally, women have increasingly taken on 'visible' public roles in the economy, but without a corresponding increase in the extent to which men participate in care work within households (Evans 2016). As urbanization occurs, and a significant share of the rural labor force migrates to urban areas, it is commonly assumed that women could take over traditionally male tasks in agricultural production (Slavchevska et al. 2019). Such changes may lead individuals to update their beliefs around women's abilities to perform certain tasks and empower women by transforming gender norms. Conversely, an increased role in agricultural production may simply increase women's workload, without necessarily empowering women or enhancing their wellbeing (Pattnaik et al. 2018). The relationship between migration of the rural labor force into cities, increased female participation in agricultural activities, and women's empowerment status is therefore unclear, at least in the short run.
In this paper, we explore the question of how changes in labor supply may influence female labor participation and empowerment outcomes. To answer this question, we analyze data from a context where migration and gendered norms around agricultural production are both prevalent: smallholder farmers in contemporary Bangladesh. Using a panel survey of jute producers in the southern delta region, we study how changes in gender gaps in labor participation, wages and empowerment are associated with two variables that could be indicative of changes in labor supply: whether a household reports an adult member migrating during the survey period (as an indication of reduced labor supply) and whether the household reports having faced increased difficulties finding enough labor for their jute production during that period.
As our empowerment indicator, we use the project-level Women's Empowerment in Agriculture Index (pro-WEAI) (Malapit, et al., 2019), which was administered to the same male and female respondents within the household during two follow-up survey rounds.
Electronic copy available at: https://ssrn.com/abstract=3677436 Analyzing detailed gender-disaggregated data on labor used for jute production and post-harvest activities, we first establish that the division of labor in jute in our study area follows strong gender norms. Whereas male labor is used for a wide range of tasks, female labor is used mainly for post-harvest processing activities that are typically carried out at the homestead. For households reporting migration, we observe an overall increase in the share of labor performed by household females and a reduction in the share by hired females. Female laborers are paid lower wages than men, even when focusing on activities where the use of female labor is prevalent, and men and women perform the same tasks. This gender wage gap does not decrease when households have likely faced a reduction in labor supply. These findings suggest that in the presence of strong gender roles in the production of a primary crop, female labor is not treated as a substitute for male labor, and as a result, changes in the composition of the household labor force do not necessarily create opportunities to empower women in agriculture.
This paper contributes to the literature in several ways. First, we provide empirical evidence on the relationship between available labor supply and changes in how agricultural producer households allocate resources. While the determinants of rural-urban migration and the effects on communities receiving migrants have received considerable attention in the quantitative literature, less attention has been paid to how producer households adjust their input allocations in response to migration. Böhme (2015) uses statelevel U.S. GDP growth rates to instrument for the relative returns to migration among households with migrants in rural Mexico and finds an increase in agricultural investment among households with migrants.
Similarly, using an instrumental variables approach, de Brauw (2010) finds a shift away from labor intensive crops among rice producing households with migrants in Vietnam. Our dataset contains household reports both on migration by individuals within the household, as well as the principal agricultural decisionmaker's perception of local labor availability.
Second, we are able to explore the role of migration in relation to household labor allocation in a novel setting, since current evidence on this topic is predominantly from studies in China. The closest such study to our approach is by Wang et al. (2014), who use a similar approach to test how households allocate inputs in response to out-migration (though their data does not distinguish between male and female labor). Studies on China which look at gender roles are summarized by Mueller et al. (2015), with the common finding of an increased role in agricultural production by women and children in households where the adult male has migrated (Chang et al. 2011;Mu and van de Walle 2011;de Brauw and Mu 2011). Understanding how gender plays a role in this process requires evidence from varied contexts, since norms about women's role in production activities may vary in different cultural settings. Our analysis uses data from jute producers in Bangladesh, but we expect the results to generalize to other agricultural value chains-particularly in South Asia-with similar gender norms restricting women's mobility (Malapit, et al., 2019).
Third, a key contribution of our paper is to demonstrate the value of combining gender disaggregated data on agricultural activities with measures of female (and male) empowerment within households. We collect both data on labor use by male and female, hired and household workers for each stage of agricultural production for their primary crop, as well as detailed individual-level measures of empowerment within agriculture using the pro-WEAI, a new, internationally-validated, measure of women's empowerment designed for project use. This allows us to combine traditional economic measures of welfare for men and women-in terms of their time worked and compensation received-with data on their perceptions of their own individual and collective agency. We hope to encourage future research to incorporate measures of agency within their theory of change, as a complement to measures of economic wellbeing.

Setting
We use data collected from smallholder household surveys conducted in four districts of the southern delta region in Bangladesh (Faridpur, Jhenaidah, Madaripur and Narail). Data collection was conducted as part of the Agricultural Value Chains (AVC) program in Bangladesh, which aimed to improve agricultural incomes and food security through a value chains development approach. Details of the project are described in de Brauw et al. (2018). We focus exclusively on households who produce jute as their primary crop, as this was one of the crops targeted by the value chains development program.
As an agricultural commodity, 'jute' refers to the vegetable fibers extracted from the shrub species Corchorus olitorius and Corchorus capsularis 1 which are primarily used as a natural fiber in the production of textiles. Jute has long been produced in Bangladesh, playing a significant role in the globalization of trade in the 19 th century where it was used to produce coarse bags for packaging commodities for shipment. 2 As such, jute is a salient crop for our study. As a crop that has been widely cultivated in the region for a long period of time, it presents an opportunity to examine established gender norms around agricultural production. Since it is a wholly commercial crop, it is unlikely to be affected by changes in household composition through the demand channel, in contrast to horticultural produce and livestock which may be used either for sale or household consumption.

Sample selection
To establish a representative sample of jute producing households, we listed all farming households in fifty jute producing villages prior to the baseline survey. To be eligible for selection, households had to report intending to plant jute as their primary crop in the coming season and cultivate at most 500 decimals (2.02 hectares), though most farmers cultivated substantially less land. 3 For those meeting these sampling criteria, twenty households from each village were randomly selected for an interview in 2016 to create a total sample of 1,000 households. This sample was expanded to include ten additional farmers from each village in 2017, bringing the total sample size to 1500. 4 We exclude households who did not report producing jute throughout the study period, as well as those who could not be located or refused to participate in one or more interviews. As a result, our analysis sample consists of 1 The former is more prevalent in our context. The latter-known as 'white jute'-is generally regarded as producing finer quality fibers. 2 See Ali (2012) for a comprehensive treatment of the historical and political role of jute in Bangladesh. 3 For the 2017 survey the median area was 84 decimals (0.34 hectares). 4 Slightly more restrictive criteria were used for newly sampled households: they were required to have at least ten years of experience farming jute, and to have a minimum landholding of 66 decimals (0.27 hectares). Respondents were interviewed at their dwelling around the end of the primary agricultural season (March-April) up to three times (2016, 2017, and 2018). Our analysis focuses on data from the interviews conducted in the 2017 and 2018 survey rounds to maximize the available number of observations. Where not pooled, we will refer to households selected in 2016 as the 'original' sample and households selected in 2017 as the 'additional' sample.

Survey instruments
The survey consisted of two separate forms. The respondent for the main form was the primary agricultural decision-maker in the household (who was generally male) and focused on agriculture, with detailed modules on jute production, production of other crops, livestock holdings, input and labor allocation, as well as other household level outcomes not related to agriculture. The labor module is of particular interest, as it is a primary source of data used in our analysis. It was designed around the specific stages of jute production, comprising eleven stages from land preparation through transportation and marketing of jute fibers. For each stage, we collected information on the type and duration of labor use for both household and hired laborers, disaggregated by gender within each category, as well as data on wages paid to hired laborers, and perceived challenges in finding labor.
A secondary form was administered at the individual level within each household. It was administered to the spouse of the primary agricultural decision-maker (typically female) and contained questions on household members, including migration, and household consumption. For the additional sample, both the main form and this secondary form also included a series of modules comprising the project-level Women's Empowerment in Agriculture Index (pro-WEAI), which collected detailed data on a range of domains of empowerment (Malapit, et al., 2019). Hence, for the additional sample, we have responses to pro-WEAI modules from both male and female household respondents. Interviews were administered in private by an Electronic copy available at: https://ssrn.com/abstract=3677436 enumerator of the same gender as the respondent, to ensure that respondents could share information on potentially sensitive topics while preserving their privacy. Table 1 provides an overview of the data collected for both samples.

Variable definitions
Our analysis focuses on two explanatory variables that measure changes in the amount of labor available to households in our sample. The first is a variable indicating whether a household reported that one or more of its members had migrated at the time of a given survey. This variable is defined based on the responses given by the person (generally the female spouse of the primary agricultural decision-maker) who completed the secondary interview form. In completing the household roster module, they were asked about the status of any members who they had reported as being part of the household in the interview the previous year. If they responded that one or more individual had left the household because they had migrated from the household (whether temporarily or permanently) this variable was coded as one, and zero otherwise.
We additionally disaggregate this variable by gender, to create two variables employing the same definition: whether the household reported one or more male (female) members having migrated. Note that for the additional sample we do not have this information for the 2017 round (since they did not complete an interview in 2016) hence these households are always assigned zero for the first round. To ensure that this assignment is not driving results, as a robustness check, we run our main specifications using an alternative Electronic copy available at: https://ssrn.com/abstract=3677436 definition of the migration variable for which we have complete data, and find similar results (Appendix Tables 2-5). 5 The second key explanatory variable is an indicator for reported availability of hired labor. As part of the main survey labor module, after providing estimates of wages paid to male and female laborers, the primary decision-maker was asked whether they had experienced difficulty finding enough labor for a given jute production stage in the preceding agricultural season. We aggregate this variable across activities, so the indicator takes the value one if the farmer reports having faced difficulties finding labor for jute production during at least one stage of production, and zero otherwise.
For outcomes, we consider two groups of variables: labor-related outcomes, which are available for the full sample, and empowerment measures constructed using responses to the Pro-WEAI modules, which are only available for the additional sample. All labor outcomes are disaggregated by gender and reported separately for each activity in jute production. 6 We construct indicators of female labor participation as the amount of female household labor as a share of the total amount of household labor, the amount of female hired labor as a share of the total amount of hired labor, and the amount of female labor as a share of the total amount of labor used. We aggregate these labor shares by summing across activities, whereas wages are aggregated across activities by taking the mean wage across activities for which a male (female) worker was hired. Aggregated expenditures are divided by the area (in decimals) on which the household cultivated land to facilitate comparisons between farmers with differently sized farms.
For the pro-WEAI outcomes, the survey modules capture information on three domains of empowerment (3DE). The domains are intrinsic agency (one's own power); instrumental agency (power to do); and collective agency (power as part of a group). Within these three primary domains, questions are divided 5 The alternative measure uses retrospective reports on migration collected for all households in the final survey. While this provides complete data on migration, it is reliant upon respondents' recall of the timing of a migration in order to assign whether migration occurred before or after a given survey, hence we prefer to use the contemporaneous measure. 6 To prevent reporting errors from influencing results, we winsorize all continuous variables at the 1 st /99 th percentile.
into twelve sub-domains. 7 For each of these sub-domains an adequacy score is computed to determine an individual's level of empowerment within that sub-domain. 8 The individual is then assigned a binary score for each sub-domain (which takes the value one if the adequacy score for that sub-domain is sufficiently high, and zero otherwise). If the mean value of these scores is greater than or equal to 0.75, or in other words if their level of empowerment is adequate in at least nine of the twelve sub-domains, a respondent is considered empowered.
For our empowerment outcomes we use both an indicator variable for whether the respondent is considered empowered overall (i.e. whether the respondent's level of empowerment is adequate in at least 75 percent of the sub-domains), and a continuous variable that represents the proportion of the twelve sub-domains in which the respondent is considered adequately empowered. Combined, these two variables allow us to compare both changes in the share of those meeting the overall empowerment threshold, and changes in the share of sub-domains in which respondents are considered empowered.

Estimation
For our primary specifications of interest, we estimate two regressions with household fixed effects. 9 We first estimate the change in a given outcome Y associated with a change in our two main explanatory variables: Migrant, an indicator variable which takes the value 1 if the household reports at least migrant at survey period t and 0 otherwise, which we interpret as a proxy for reduced household labor availability; and Scarcity, an indicator variable which takes the value 1 if a household reports difficulty finding labor for period t and 0 otherwise, which we interpret as a proxy for reduced supply of hired labor. 10 The model is estimated as a pooled regression, with household and survey round fixed effects ( and ) and the error term .
(1) = + 1 + 2 + + + We additionally estimate a variation of this specification, disaggregating Migrant by sex, so that MaleMigrant (FemaleMigrant) takes the value 1 if the household reports that at least one male (female) household member migrated out of the household between survey period t -1 and t, and 0 otherwise, giving: For individuals in our additional sample, we do not observe migration in the first survey round. As a result, we assign these observations zero for migration indicators for our main specification. This also means that we cannot estimate our pooled specification for empowerment outcomes, since these are observed only for the additional sample. Instead, for this set of outcomes, we specify a static model, where we regress an outcome in the second period (t=1) on reported migration in the second period, and control for the lag of the outcome reported in the first period. Since we observe reported labor scarcity in both periods, we take the difference between the second and first period, giving: (1), we additionally disaggregate the migration indicator variable by gender: We cluster standard errors at the village level across all specifications. We control for the area planted with jute, as this will be an important time-varying determinant of our outcome variables, which will not be captured by our household fixed effects.
An important qualifier to our results is that they are by nature descriptive, rather than causal. Since the decision by a household member (or members) to migrate is not plausibly exogenous, we are not able to construct the counterfactual effect of what would have happened had an individual remained (or conversely migrated from a household without migrants). Our aim is rather to describe observable changes in empowerment status within households with migrants versus those without, as these changes have important implications for thinking about policies around gender, migration and women's empowerment in agriculture. Table 2 presents summary statistics on the sample households and their agricultural production. The person identified as the main respondent was almost exclusively male-only 4% of household identified a woman as the primary decision-maker on jute production. Almost all (95%) are married. They are typically middle aged, with an average age of 48 years, compared to migrants who are generally young, with an average age of 21. Migration in our sample is therefore generally a change in the available pool of potential household laborers available to a decision-maker, rather than a change in the individual making the decisions. These available laborers are roughly balanced in terms of gender, with a slightly higher mean number of available women than men. In 2017, the average household in our sample cultivated an area of 99 decimals of jute (0.99 acres). While the farmers in our sample are all smallholders, there is nonetheless considerable variation in the reported area planted (between 10 and 460 decimals). In addition to growing jute, most households produced other crops, principally rice, wheat and pulses, both for sale and consumption. The mean revenue from jute sales in 2017 was $501.90, slightly less than the total mean amount of revenue received from all other crops ($520.62), and for a majority of households (57%), revenue from jute exceeds that from all other crops combined.

Respondent and household characteristics
All households employ household labor for jute production, typically of both genders: 98% of households report using male household labor, while 86% of households report using female household labor. Use of hired labor is notably high in the sample, with 96% of households reporting hiring laborers at some point during the season. However, while hiring male labor is common (96% of households), only 53% of the sample reports hiring female laborers.

Gendered labor usage
We next consider how households use male versus female labor during different stages of jute production.
We organized jute production activities into eleven distinct tasks: pre-harvest activities (ploughing, seeding and weeding); harvesting; processing (drying, curing, bundling, stripping, and bailing); and marketing (sorting and transporting to market). Figure 1 summarizes the types of labor engaged in each of these tasks. Whereas male household members are involved across all production and post-production activities, the hiring of male laborers varies by task, with high rates of use reported for labor intensive (and time sensitive) activities such as weeding and harvesting, and for some post-harvest activities (bundling and stripping). In contrast, households rarely report using labor by female household members prior to or during harvesting.
Female household members are involved primarily in processing activities (stripping and bailing) and in sorting. Hiring of female laborers is essentially constrained to two tasks: weeding and stripping. Female workers, whether household or hired, are almost never involved in transporting and marketing jute.
Qualitative research suggests that labor patterns are driven by strong gender-based norms regarding the type of work that women can do, including norms around mobility; post-harvest tasks can be performed in the homestead, in contrast to work in the fields or in transporting jute to market (Rubin et al. 2018).
We next consider gender differences in the outcomes for those providing labor. We first consider wage payments to hired laborers. Figure 2 shows median wages for male and female hired workers for both survey rounds. We show the two stages for which women are predominantly hired (weeding and stripping) as well as aggregate categories for field activities (ploughing, seeding, weeding and harvesting) and postharvest activities (drying, curing, bundling, stripping, bailing, sorting and transporting to market).

Figure 2 -Median hourly wage by production stage and gender
Across all four categories in both survey rounds, median wages are significantly higher for male than for female workers. In 2017, for weeding, the hourly wage for women was 25% lower than for men, and for fiber stripping, the median wage paid to women was half of what men were paid for the same task. While productivity for field activities could potentially vary by gender, which could help explain a wage gap in field activities, variation in productivity is unlikely for stripping fibers, a fine motor task for which women are typically preferred in the survey area (Rubin et al. 2018). A more likely explanation is that female workers receive less than male workers for equivalent tasks because of lower bargaining power.
Since household members do not earn wages when working on the family farm, we consider an alternative measure of rewards for household laborers: survey respondents' empowerment within sample households.
Focusing on our additional sample, where the pro-WEAI survey instrument was administered, Figure 3 summarizes the mean 3DE scores and the share of empowered respondents for male and female respondents Male Female across the two survey rounds. While empowerment scores increase across the two survey rounds, there is a persistent gap between male and female scores: 14% of women were considered empowered based on the 2017 survey, relative to 32% of men; and 21% based on the 2018 survey, relative to 45% of men.

Figure 3 -Pro-WEAI status by gender and survey round, additional sample
Overall, we observe strong gender differences in agricultural production in our setting. Gender is predictive of the types of tasks that workers undertake in growing, harvesting and processing jute, the level of compensation that wage workers receive, and the extent to which household members-often providing household labor for jute production-are considered empowered. In the next section, we explore the extent to which these gender gaps change when households report a change in the supply of labor available to them, and whether such changes may represent normative gains for women.

Labor allocation
To examine how the use of female labor varies in response to local labor availability, in Table 3 we regress the amount of female labor used, expressed as a share of the total amount of labor used, on our indicator variables for household migration, and reported difficulty finding labor during the season (Equations 1 and 2). We focus on the share of total labor performed by women to account for variation in the extent or intensity of jute production activities.
In columns (1) and (2), the dependent variable is defined as the number of days that female household members worked on jute production as a share of the total number of days of household labor used in jute production, and as a share of the total number of labor days used to produce the jute, including hired labor, respectively. Columns (3) and (4) use similar dependent variables, but focusing on hired instead of household labor, while (5) and (6) use the total amount of female labor used (household and hired labor) as a share of the total number of labor days used in the entire jute production process. We find a statistically significant association between households who report at least one member moving out between survey rounds and female labor usage in specifications (1) and (3). The sign of the coefficient varies with the type of labor. Households with migrants report a modest increase in the use of female household members' labor as a share of total labor days (2.6 percentage points). For hired labor, the correlation is negative: households with migrant members report hiring less female labor as a share of total labor days (4.6 percentage points). Specifications (2) and (4) suggest the magnitude of these changes does not vary with the sex of the migrant: we see a statistically similar increase in household female labor share, and decrease in hired female labor share, whether the sex of the migrant is male or female. In other words, in this context, male migration does not create increased labor market opportunities for women. In fact, our findings suggest that male and female out-migration reduce demand for female hired labor.
Our reported measure of difficulty in finding labor is more likely to reflect perceived scarcity of hired labor.
Here, the association between labor scarcity and the share of labor performed by women is positive across specifications, though the coefficient is only statistically significant in specifications (3) and (5); households reporting increased labor scarcity use more labor by female household members. Surprisingly, we observe no shifts in the relative use of female hired labor when households perceive increased difficulties finding labor.
In Table 4, we disaggregate the share of labor performed by women by stage of production. We do so by separating the two activities for which female labor use is most prevalent (weeding and stripping) and aggregating activities performed on plots (field activities, including weeding; Panel A) relative to activities performed following harvest (post-field activities, including stripping; Panel B). For field activities, we do not observe a significant change in the female share of jute labor when a household member moves out, regardless of whether we focus on household or hired labor. Thus, changes in labor allocations for field activities do not account for the finding that households with migrant members replace female hired labor by female household labor. However, specifications (2) and (4) show that this is because households with a female migrant do not change their labor allocations for field activities; when a male household member migrates, households do replace female hired labor by female household labor, consistent with the estimates in Table 3. Thus, households where male members leave (who previously worked in the fields), female members appear to take over some of the work in the field, while reducing the relative use of female hired labor.
For post-field activities, in which women are traditionally more engaged, we observe the same results for both the aggregate and migrant gender-disaggregated specifications. For the latter, we consistently fail to reject the null hypothesis that the coefficients on having at least one male migrant or at least one female migrant are equal. Thus, the gender of the migrant appears to matter for field activities, where female labor is used relatively little. A possible explanation is that households do not face a reduction in available household labor supply for field activities when female household members move out, because women were not working in the field to begin with. For post-harvest activities, female outmigration could have a more noticeable impact, as these tasks are often done by women. This combination of findings suggests that although male and female household labor for homestead activities can be treated as substitutes, and although households could in principle turn to hired labor, the loss in labor supply is disproportionately made up by female household members independent of the gender of the migrant. Turning to the relative returns to working in jute production for men versus women, Table 5 considers associations between wages and migration. When at least one household member moves out, associated with a reduction in the relative use of female hired labor in Table 3, wages paid to male laborers increase significantly (equivalent to 0.08 USD per hour), whereas the change in wages paid to female laborers is not statistically different from zero. For households hiring both male and female laborers, we do not observe a statistically significant change in the wage gap associated with migration. We find similar trends in the even-numbered specifications, which disaggregate changes in outcome variables by the gender of a migrant.
Migration is associated with an increase in wages for men, not for women, and in fact, in Column (6), we observe a marginally significant increase (p < 0.10) in the gender wage gap within households with male migration.
Finally, we explore whether changes in labor availability are associated with changes in empowerment outcomes, using our additional sample, for whom we collected pro-WEAI data in both 2017 and 2018. Table 6 presents estimates of the model in Equations (3) and (4), using both an aggregate indicator for migration (odd-numbered specifications) and an indicator disaggregated by migrant gender (evennumbered specifications). For male respondents, we find no significant changes in empowerment scores when either male or female household members move out. By contrast, for female respondents, we do observe a statistically significant and positive increase in empowerment when a household member moves out. Controlling for their empowerment status in 2017, female respondents in households with migrants are 9.4 percentage points more likely to be empowered in 2018 than respondents in households without migration; and these women's empowerment scores (i.e. the proportion of sub-domains in which they are considered adequately empowered) increase by 3.5 percentage points. Changes are largest in households with female members moving out: when another woman leaves, female respondents are 20 percentage points more likely to be empowered relative to their peers. This is not because with a female member moving out, the respondent changes too. Restricting our sample to households without change in respondent from 2017 to 2018 yields similar results (Appendix Table 1).

Figure 4 -Regression estimates of disaggregated pro-WEAI outcomes on female migrant status
In Figure 4, we analyze which sub-domains are driving this result, using the same specification with genderdisaggregated migrant data, i.e. Equation (4), but using as our outcome variables the indicator for being adequately empowered within each sub-domain of the pro-WEAI. Whereas women face an increased workload when other female household members move out, we find statistically significant improvements in empowerment across three sub-domains: attitudes to intimate-partner violence, asset ownership and respect among household members for households reporting one or more female members migrating. One potential hypothesis is that having multiple adult women within a household may lead to more competition over resources or relative standing, which migration alleviates. This interpretation is consistent with Raghunathan et al (2019), who find the empowerment of different women in the household relative to one another is an important determinant of household outcomes. These findings underscore that analyses on the relationship between intra-household dynamics and empowerment outcomes should move beyond interactions between husband and wife and take into consideration broader household structure.

Conclusion
In this paper, we describe changes in the gendered allocation of labor among smallholder jute-producing households in a context of widespread rural out-migration in southern Bangladesh. We do not find strong evidence to suggest that out-migration is welfare-enhancing for women who remain in agricultural households. Out-migration of household members and scarcity of agricultural labor is associated with an increase in the share of labor for jute performed by female household members, and-in the case of outmigration-a decrease in the hiring of female workers. Thus, the increased workload associated with reduced labor supply is carried disproportionally by female household members, who are not necessarily rewarded for this added burden: women's empowerment increases only when another female household member moves out, not when a male household member migrates. Moreover, migration is associated with a reduction in the use of female hired labor, and a widening gap in wages paid to male versus female workers, suggesting that out-migration is associated with worse outcomes for female wage laborers.
The development literature has long acknowledged that hired labor is an imperfect substitute for family labor (e.g. Singh, Squire and Strauss, 1986;Benjamin, 1992). We add nuance to this finding in a context where gender norms are closely linked to the division of tasks between male and female workers, and the compensation male and female workers each receive for completing those tasks. Households appear to hire male labor instead of adopting capital that can substitute for labor, and we do observe female household members taking on male laborers' tasks when facing labor shortages, suggesting that hired labor and family labor are substitutes to some extent. That said, gender norms remain a significant influence; we do not observe women taking on men's tasks in the field, regardless of whether labor shortages are associated with out-migration of male family members or with general shortages of hired labor; and when female household members move out, households do not offset the reduced labor supply by hiring additional female workers, suggesting that especially female hired labor is an imperfect substitute for family labor.
These findings could also be important for understanding agricultural technology adoption, as capital can substitute for labor in agricultural production. An important question for future research is how introducing labor-saving technologies influence welfare outcomes for men versus women. In this context, we would anticipate quite different dynamics around a labor-saving device for activities typically undertaken by men, versus one for activities undertaken by women, or both sexes. In promoting inclusive development, it is therefore important to identify the presence of such norms, their potential to persist, and influence on whether intended outcomes can be achieved. Had we not collected data disaggregated by both gender and task, we would have been unable to make these observations. Our findings support arguments by Doss (2002Doss ( ,2017 on the importance of female labor to crops traditionally perceived as 'male'. While we do not have similarly detailed data for other crops, we expect similar norms to be prevalent for grains such as rice and wheat, which are grown on fields outside of the compound, where women in qualitative work reported not being able to work. Overall, however, our analysis underscores the value of collecting gender-and task-disaggregated data on labor in the agricultural production process, and we believe that it is worth extending such detailed data collection to other similar contexts in which men and women collaborate in agricultural production, to learn more about how gendered divisions of different tasks within crop production can be influenced by external factors such as increasing migrant opportunity, and how this, in turn, could influence the relative benefits of labor-saving technologies for men versus women in producer households and agricultural labor markets. Finally, the paper leverages the gender-and task-disaggregated labor modules by linking these findings to panel data that allow documenting changes in women's (and men's) empowerment in agriculture. Our analysis does find that household female respondents report greater intrinsic agency following outmigration of female household members. We do not find this result when male household members move out, even though male and female outmigration are associated with a similar increase in the relative contribution of female family labor to production. This demonstrates the potential importance of intrahousehold dynamics for empowerment across the entire household-including between women-not just between spouses.
In conclusion, while gender roles in agriculture may evolve with increased migration, our findings indicate that this process does not happen rapidly. In traditional societies like Bangladesh, gender segregation of tasks continues to be the norm, and gaps in labor use, wages, and empowerment persist, despite the labor shortages that households face as their members, and laborers from their communities, seek their fortune by migrating to urban settings. Better understanding the nuances of these labor dynamics, and identifying how they relate to gender gaps in empowerment, through analyzing sex-and task-disaggregated data, is key for inclusive agricultural development.

Appendix Tables
Alternative specification (Table 5)      Wages regression restricted to households reporting both male and female labor