Steered Away from the Fields Short-Term Impacts of Oxen on Agricultural Production and Intra-household Labor Supply

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.


Introduction
Most rural households in lower-income countries rely on farming as their primary source of income, and on labor as the primary input to agricultural production. Mechanization, particularly in the form of animal traction, has the potential to raise agricultural incomes and facilitate the structural transformation of developing country economies. Previous research has identified which conditions favor the adoption of animal traction. Yet observational studies on its potential benefits yield nuanced results, and there is a dearth of rigorous evidence on impacts. In addition to knowledge gaps around how mechanization economically influences agricultural households, we also know little about whether (and why) technology adoption has differential impacts within the household. In this paper, we show that labor-saving technology can have large economic benefits, but that examining intra-household impacts-including gendered impacts-is crucial to understanding longer-term ramifications.
In a two-year randomized phase-in trial conducted with 2,546 cotton farmers in rural Côte d'Ivoire, we test the impacts of a matching grant that covered 50% of the cost and delivery of a pair of oxen. We find high take-up and use of the oxen, and positive impacts on agricultural production. In the first agricultural season, producers increased their cotton production and cotton revenues by 7%, despite only being notified about the program and receiving the oxen after planting decisions were made. Second season agricultural outcomes show that households increased their land area under cultivation by 6%, while also increasing spending on complementary non-labor inputs by about 15%. Moreover, household member time on plots decreased by about 1.5 hours per hectare, or 7%. These results represent a lower-bound of the impacts of oxen, as the phase-in trial resulted in a non-trivial share of control households receiving the oxen in advance of the endline survey.
These effects on household members' time use are systematically related to gender. Assignment to treatment had no significant impacts on the producer's time use or boys' time use, while wives and girls decreased the time spent on household plots by almost 6% and 10%, respectively. In addition to this reduction in time spent working on household plots, girls were less likely to be sick, and the duration of illness over the last month reduced by about a quarter. Boys in treatment households were nearly 30% less likely to have dropped out of school, though we find no evidence of increased literacy or numeracy.
What is driving these gender-differential time use impacts? One answer comes from community gender norms around oxen use. Internationally, gendered social norms and assumptions have discouraged-or outright excluded-women from jobs that are perceived as relatively dangerous due to 'benevolent sexism' (Cuddy et al., 2015;Glick et al., 2000;Padavic and Reskin, 2002). 1 Tending to traction oxen is considered a 1 risky task due to their large size and potentially erratic behavior. According to Alesina, Giuliano and Nunn (2013), the plow requires significant upper body strength, and bursts of power, both to pull the plow and control the animal that pulls it. Women are thus commonly considered less suitable for the task-though there is substantial heterogeneity in adherence to this norm. We find that our observed negative treatment effects on the time women and girls spend on household plots are predominantly driven by households in districts with relatively lower female involvement in oxen tasks. Nationally-representative data shows that matrilineal ethnic groups are less prevalent in these districts, and norms are less gender-egalitarian. In these more conservative communities, women shift to working off-farm once oxen are introduced. These empirical patterns align with predictions from our simple model of time allocation under social prescriptions, which builds on Kevane and Wydick (2001).
There are three main contributions of this paper. First, we provide the first experimental evidence on the impacts of animal traction, and the first experimental evidence on the effects of agricultural mechanization as a whole in Sub-Saharan Africa-building on a rich literature that has examined the adoption and use of labor-saving agricultural technology in the region (Pingali, 2007;Owolabi et al., 2012;Guthiga, Karugia and Nyikal, 2007;Williams, 1997). The existing non-experimental research lays out several key stylized facts about animal-traction mechanization, including minimal yield gains, an increase in land under cultivation and devotion of the expanded land to market crops, as well as a reduction in labor required during the land-preparation period (Pingali et al., 1987). More recently, studies of rice production in Côte d'Ivoire highlight how mechanizing land preparation can contribute to agricultural intensification, thereby raising productivity (Mano, Takahashi and Otsuka, 2020), especially when agro-chemicals and mechanized land preparation are combined (Aihounton and Christiaensen, 2023). Within the limited literature demonstrating experimental evidence in developing country settings, our paper is most closely related to a recent study in Karnataka, India, which demonstrated that subsidized access to rental equipment markets lowered labor demand across all farming stages, and disproportionately so in stages not being mechanized, leading to an increase in non-agricultural income (Caunedo and Kala, 2021).
Our second contribution is to examine how the time-use impacts of mechanization vary within the household, speaking to the broader literature on the intra-household consequences of technology adoption (Wodon and Blackden, 2006;Bryceson, 2019;Doss, 2013;Theis et al., 2018). This analysis is particularly important in contexts where the social norm goes against women using the technology: in our case, plowing with oxen (Fisher et al., 2017;Starkey, 1995;Wanjiku et al., 2007). Modern agricultural technologies can affect women's farm labor negatively, facilitating shifts to new tasks or activities, or positively, depending on the context. Importantly, these shifts have the potential to drive structural transformation if they free up women's time for other market-based activities (Goldin and Sokoloff, 1984;Dinkelman and Ngai, 2022). This tension highlights that 'one norm does not fit all': sexist beliefs or practices limiting women's engagement in one economic domain can spur women's engagement in other economic domains.
In India, Afridi, Bishnu and Mahajan (2020) instrument suitability for mechanization by soil characteristics to estimate the gendered impact of mechanization. In line with our findings, they detect a significantly greater decline in women's labor than men's, explained by the complementarity of men's labor and tilling machines, and by the fact that better quality tillage lowers the demand for weeding labor, which is mainly a female task in their setting. In other instances, new technologies might make women's farm labor even more critical, owing to the sexual division of labor and limited technical innovations for women-led tasks (Wilson, 2003;Bassett, 1988). Although limited work has been done to rigorously measure the impacts of other mechanization equipment (Muhabaw and Muhabaw, 2014), to our knowledge, there are no randomized control trials testing the theory underlying improved productivity through animal traction, or its gendered impacts.
Lastly, our study also speaks to the literature on the interaction between crop choice and agricultural technology on the one hand, and education and child labor on the other. Part of this literature draws a causal link between child labor-intensive crops (such as cotton) and negative schooling outcomes (Baker, 2015), while others find that the income effect dominates and schooling outcomes improve (Kazianga and Makamu, 2017;Cogneau and Jedwab, 2012). However, there is little evidence on the relationship between the introduction of mechanization, which can replace tasks done by young children and potentially encourage schooling, beyond cross-sectional studies (Levy, 1985).
From a policy perspective, our findings reaffirm the broad potential for mechanization and productive animal traction to intensify agriculture: the significant short-run impacts suggest that recipient households shifted towards a larger, more capital intensive, and less labor intensive production approach. While receiving oxen shifted women off-farm in districts where the descriptive norm is that women engage minimally with oxen, it is unclear whether this facilitates structural transformation in the longer-term. With this in mind, policy makers seeking to implement agricultural mechanization interventions should consider complementary interventions that may help women retain some control over agricultural income in the face of their reduced participation, or support women in making their shift to off-farm labor market opportunities a profitable one.
The paper proceeds as follows: In Section 2, we present a simple model that highlights the links between agricultural norms, technology adoption and gender norms, and generates testable predictions for our empirical analysis. Section 3 provides information on the local context as well as the experimental design. We present the experimental results in Section 4, while Section 5 discusses the implication of our results and concludes.

2 Rural households operating under social prescriptions
Our context is that of households in an agriculture-based rural economy, in which members maximize their utility over consumption and leisure. Households' utility depends, inter alia, on how their members conform (or do not conform) to behaviors that are considered socially appropriate. The purpose of the framework is to i) illustrate the links between traction power, agricultural production and labor supply in this setting, and ii) predict what happens to household members' labor supply in the face of gendered social prescriptions related to traction power.

Setup
Our framework incorporates Akerlof and Kranton (2000)'s formalization of the insight that identity-including gender identity and the social prescriptions that accompany it-is a motivation for behavior, and draws on Kevane and Wydick (2001)'s model of time allocation under social norms.
In our simple setup, farm households maximize utility U subject to income, production technology and time constraints. Men own farm capital K, including traction oxen, and female household members divide T units of work time between farm work, T f , off-farm work T o , and domestic work T d . The household has a production function A(K, T f ) that corresponds to output on the husband's farm, and a function H(K, T d ) that corresponds to domestic production, the value of which also depends on the household's farm capital.
Off-farm production, T o , is valued by the market at price p. We assume that A T T <0, A KK <0, and A T K > 0, and similarly for H(). 2 The labor supply of the husband is assumed inelastic and its allocation is fixed across activities for the purposes of this simplified model, where we are interested in examining female household members' labor supply.
As in Akerlof and Kranton (2000), an individual's utility is a function of their social category and the extent to which their own and others' actions correspond to behavior indicated by social prescriptions P. In our rural setting of northern Côte d'Ivoire, one example of such a 'benevolently sexist' social prescription is that women should not undertake activities perceived as risky, including handling large livestock (as discussed in Section 1). Since this social prescription does not regard women's farm work overall, but specifically farm work when oxen are used (K=1), we modify Kevane and Wydick (2001)'s formalization of this norm's impact on behavior by focusing on women's deviation from average oxen use by other women in their community: (1) where j = f , o, d denotes the set of activities available to the woman. The benefit from her behavior conforming to social prescriptions is given by P. The parameter a j represents the intensity of social penalties for violating norms regarding risky work in each of the different activities. K j is the mean intensity of oxen use in activity j for women in a particular community. Social penalties depend on the intensity parameter a j , whether the woman performs relatively more tasks with oxen than women in her community (i.e., K > K j ), 3 and how much time she spends on these tasks (T j ). For simplicity and given our setting, we focus on social penalties related to farm work, setting a o = a d = 0. 4 The household allocates the woman's labor to maximize: subject to the time constraint: (3) This setup yields a set of tractable comparative statics for the relationship between a change in capital (oxen) and labor across the different categories. 5

Predictions
We start with the first equation. The marginal effect of capital (in our context, introducing traction oxen) on women's farm work will depend on the sign of A T K H T T − a f (K − K)H T T , since |H| is negative as shown in Appendix A. Recall that A T K is positive by assumption, while H T T is negative by assumption. This implies that if a f (K − K) > A T K -that is, the social penalty she incurs from spending time on-farm with oxen is greater than the return of her labor to farm production when capital increases marginally-then the intro-duction of oxen will decrease the time she spends on the farm.
In the presence of social prescriptions that it is unsuitable for women to handle traction oxen, a f (K − K) is large and positive such that: Note that for lower values of the social penalty in less conservative communities, where the norm is for women to already be relatively more involved in handling oxen, introducing oxen will naturally have a smaller negative effect on women's time spent on farm work.
Turning to domestic work, as explained above, we assume that H T K is positive but only weakly so. Since A T T is negative and |H| is negative, this implies that the marginal effect of capital on domestic work is positive but close to zero: Lastly, we turn to the case of women's off-farm work. Since H T K is near-zero, and H T T is negative, the sign will again depend on the relative size of a f (K − K) and A T K . If a f (K − K) > A T K as above, then the first term of the numerator will be negative. Since we know that the second term of the numerator is positive and close to zero, this implies that in households where women would go against the norm by working with oxen on-farm and thus face a social penalty, ceteris paribus: Another way to see this result is that Note also that as the social penalty shrinks, the increase in women's off-farm work resulting from an increase in capital will also get smaller.
Next, we turn to describing our experiment in more detail before presenting household-level impacts and bringing our model to the data in Section 4.

6
3 Experimental design and implementation

Experimental design
We conducted a randomized evaluation of a traction oxen matching grant for cotton farmers working under formal cotton societies in four northern regions of Côte d'Ivoire under the World Bank's Agricultural Support Program (PSAC). 6 Specifically, the project provided a matching grant of 50% of the cost of receiving two traction oxen and related equipment, with producers financing the remainder through their own funds or credit from cotton ginning companies and agricultural cooperatives. 7 In economic terms, this is a large treatment. Previous studies have found that an ox provides farm 'labor' equivalent to five to eight men, and that traction oxen are at least twice as powerful as smaller bovines: 400-500 watts of power versus less than 200 watts for small bovines (Smil, 2004;Starkey and Faye, 1990).
The participant cotton societies formed the farmer sampling frame using their member lists and a mix of objective criteria (such as being married, having been a member of a cotton society for at least 3 years, and belonging to a single cotton society) and more subjective criteria determined by cotton society staff (being credit-worthy, trust-worthy, and able to increase their cotton farmland). Few farmers in the sampling frame owned traction oxen before the intervention, although many owned other types of farm equipment (Appendix Table B1).
We used a randomized phase-in approach for our experiment: from an initial sample frame of 2,546 eligible farmers, we randomly assigned 1,273 producers to receive the oxen matching grant offer in the fall of 2016 (treatment group) and 1,273 to receive the offer starting in April 2018 (control group). We stratified the randomization by cotton zone, whether the cotton societies' administrative data indicated that they already owned any oxen, whether they requested one or two oxen, and whether they owned a cart. We measure impacts using detailed household surveys conducted in July and August 2018. 8 As cotton cultivation is a key outcome of interest in this experiment, Figure 1 presents the timing of the distribution of the oxen under the experiment together with the key activities of the 2017 and 2018 cotton 6 The cotton component of the project operated in Bere, Poro, Tchologo and Worodougou, and partnered with Intercoton-the cotton value-chain interprofessional organization in Côte d'Ivoire-and its five constituent cotton societies (CIDT, COIC, IVOIRE COTON, SECO and URECOSCI) to implement the matching grant intervention under the broader goal of increasing agricultural productivity in the country. 7 The market value of two traction oxen at the time of delivery was FCFA 480,000, implying a matching grant value of FCFA 240,000 (US $408). In practice, there are small variations in treatment across farmers. While 96% of farmers requested two oxen, 4% instead requested one ox: households assigned to the treatment group were provided with the opportunity to purchase their requested number of oxen. Moreover, 7% of farmers in the treatment group requested and received a multicultivator, 6.4% a seed drill, 5.1% a plow, and less than 3.5% a received tillers or carts. Appendix Table B1 provides a full list, and shows that households own an average of 2 pieces of equipment, irrespective of their treatment status. The most commonly owned tools are multicultivators (52% households), followed by carts (46%) and plows (43%). 8 We only collected information on agricultural inputs use from a random half of the sample, due to budgetary reasons. The data is balanced across this split.

Sample characteristics
At endline, we tracked and administered a survey to 91% of farmers in the sampling frame, resulting in a sample size of 2,314 producers: 1,146 producers in the control group and 1,168 producers in the treatment group. 9 We focus on the 2,113 producers who reported comprehensive agricultural production, sales, and time use data. 10 Table 1 reports the socio-demographic characteristics of the sample at endline, together with the difference between treatment and control, and the level of significance of this difference. Sample households have large families with an average of over 8 members, including just under two children younger than 5 years old. Educational attainment is low, with 81% of household heads having received no formal schooling.
The surveyed households are predominantly male-headed, with less than 1% being female-headed (only 17 households). Of the households in the sample, 42.3% are polygamous. There are a few imbalances between treatment and control: treatment households have fewer boys aged 6-16, are slightly more likely to have a household head that completed at least primary school, and are slightly less likely to have experienced a shock in the last year. We are limited on the range of variables we can test balance on given that we do not have baseline data in this experiment. However, this table also contains pre-program administrative data for a sub-sample of 86% of the cotton producers. Reassuringly, we do not find imbalances in the pre-program cotton area, which is around 3.65 ha in both experimental groups, nor in the cotton production, which is around 3770 kg.
The survey data provide a detailed picture of household agricultural tasks and cotton production in Côte d'Ivoire. While cotton producers are almost exclusively men, both men and women in these farm households are engaged in cotton production. Contrary to conventional wisdom detailing stark gender divisions in the performance of agricultural tasks, with men being more responsible for physically-intensive tasks like plowing and land clearing and women more responsible for weeding and sowing, Table 2 shows that there is little gender specialization in cotton-related tasks in our setting. The starkest gender difference is that household men are generally more involved than women in all tasks: over 80% of male producers perform each type of task and producers are over 10 percentage points more likely to conduct any given task than any other household member. Still, spouses are broadly engaged in agricultural tasks: 74% of spouses perform transport-related tasks while 68% plow (though consistent with the social prescriptions outlined in Section 2, rates are higher for plowing without oxen than with oxen). Similarly, while 55% of spouses sow and weed, 85% of male producers also perform these tasks. The only two tasks in which women (producer's spouse and other women in the household) are relatively less involved are land clearing (40% for spouses) and marketing (9% for spouses).
Our data suggests systematic gender differences in perceived skill in handling oxen, which are endogenous to treatment. Figure 2 presents the distribution of different household members' perceived skill levels, as reported by the cotton producer. Over 85% of girls and women are reported to have no oxen skills, about 30 and 60 percentage points above the equivalent rates for boys and for the male producer, respectively.
The sizable difference in skill rates between boys and girls persists even for children under 10 years old, suggesting that social norms attribute higher perceived skills for boys and men, relative to girls and women. The presence of social norms regarding handling oxen is reinforced by the existence of district-level patterns in the likelihood that women work with oxen. In our survey data, almost a third of households reside in districts where other households' spouses performed fewer than one in three agricultural oxen tasks (sowing/weed, plow, or transport) over the prior year, while another third reside in districts where women performed at least 1.33 tasks. The intra-household labor impacts from the distribution of oxen are likely to critically depend on norms around the appropriateness of women working with oxen.
Women's likelihood of working with oxen is highly correlated with broader district-level descriptive norms regarding women's decision-making over their own health care, large household purchases, and visits to family and relatives (correlation coefficient of 0.75). Districts in which women are less likely to work with oxen are also those where women have less overall decision-making power. Regional patterns in these norms are shown in Figure 3. 11 Underlying these patterns in women's decision-making is variation in kin- ship structure, particularly in matrilineal vs. patrilineal systems, which has been shown to be a predictor of modern-day gender norms (Lowes, 2022). According to DHS data, districts with less restrictive gender norms are 70% Gour (traditionally matrilineal) and 16.5% Mande (traditionally patrilineal), while more restrictive areas are 41% Gour and 38% Mande. 12 In line with our conceptual framework, in Section 4.3 we explore social prescriptions regarding gender norms as a potential driver of whether introducing oxen to agricultural tasks will concentrate farming among men, shifting women away from agricultural work. Table 3 summarizes the implementation and compliance with the study design. In line with the phase-in design, we find a low but significant difference of 9.1 percentage points in the likelihood that the treatment group received a traction oxen by the endline survey. This relatively small difference between the oxen distribution rates for the treatment and control groups is largely due to our experimental design, as 46% of the control group had already received an ox by the time of the survey. However, the earlier distribution of the oxen is also apparent: the likelihood that the treatment group received an oxen before September 2017 is 18.6 percentage points, or 77%, higher than in the control group. 12 Both regions are about 5% Akan and less than 1% Krou.

4 Analysis
In this section, we report intent-to-treat (ITT) estimates of the impacts of the mechanization treatment on households assigned to the treatment group. Treatment assignment was random within strata, so the impacts of the intervention on a given outcome Y i can be measured using the following regression: where Y i represents outcome y for household i, Treat i is an indicator variable equal to 1 if household i was assigned to receive oxen in the first wave, X i is a vector of unbalanced demographic variables, and δ stratum is a series of strata fixed effects. As treatment was assigned at the individual/household level, we report heteroskedasticity-robust standard errors. 13 In addition, we examine impact heterogeneity by whether households live in districts where women are more or less involved in oxen tasks. Specifically, we run the following regression: where High j is an indicator variable equal to one for households living in districts where the share of oxen tasks conducted by women within the district exceeded the median across all districts. We restrict to control households for this calculation, since our goal is get as close as we can to capturing baseline-level norms.
As a robustness check, we-run this analysis using pre-program Demographic and Health Survey (DHS) data to divide districts into ones that are more conservative (below-median women's decision-making power over their health care, making large household purchases and visiting family and friends) versus less conservative, as described in Section 3.2. We confirm that results are unchanged. Table 4 presents intent-to-treat impacts on agricultural yields and area cultivated, showing that traction oxen increase yields and lead farmers to bring a larger area under cultivation. 14 Panel A of Table 4 focuses on treatment effects for first season agricultural outcomes, showing that assignment to the oxen treatment arm led to significant increases in agricultural output, even though the oxen were distributed mid-season and recipient producers were not notified that they would receive oxen before they made their planting decisions. Producers increased cotton production by 7%, yielding an equivalent increase in cotton revenues. 15

Household-level impacts
Notably, the total market value of all crops did not increase, with the increase in cotton production partially 13 For individual-level outcomes with more than one observation per household, such as children's education and health, and wives' initial labor outcomes, we cluster standard errors at the household level.
14 As noted in Section 3.3, above, non-compliance within both the treatment and control groups suggests that our estimates represent a lower-bound for the treatment effects of the matching grant for the oxen. 15 The cotton society-dominated market structure yields little variation in the price farmers receive for their cotton, with the median reported sales prices within communities in the sample varying by less than 2%. 12 offset by a negative but not statistically significant impact on other agricultural production. 16 We also find limited impact on total household income, which almost exclusively comprises agricultural sales. The fact that this increased cotton output stems from the early-season, unanticipated receipt of oxen after planting demonstrates that producers were able to increase output through increased productivity by using oxen for crop maintenance and harvesting. Table 4 presents the impacts on early second season agricultural outcomes. In contrast to the first season when households received oxen unexpectedly and after several key input decisions had been made, households were able to take the oxen into account for planning and land preparations in the second season.

Panel B of
This added flexibility led to large shifts in land holding and cultivation decisions: households assigned to receive traction oxen increased their total land holdings by 8%, land area under cultivation by 6%, and land cultivated with cotton by 10%. This increased focus on cotton, an inedible cash crop, does not come at the expense of food security, as shown in Table 5. Table 6 shows that, for the subsample of households asked about agricultural inputs, the oxen increased producers' use of traction power, increased their use of other non-labor inputs, and decreased plot household labor. Panel A shows that treatment households increased the likelihood that households used traction oxen by 10 percentage points and the number of plots on which they used the traction oxen by almost 30%, while also decreasing the likelihood of renting in oxen, indicating that they are better able to meet their traction needs despite a larger area to manage.
Panel B focuses on non-labor inputs, showing that treatment households increased the share of plots on which they use organic fertilizer by almost 5 percentage points. There is little evidence that the oxen increased use of inorganic fertilizer or pesticides though extensive usage of these inputs was already high, with over 95% of households using at least one. Overall, households increase the value of non-labor inputs used by about 15%. 17 Finally, Panel C shows that households assigned to receive the oxen decreased their aggregate time spent on household plots over the last week without an accompanying increase in hired labor, indicating that labor productivity has increased due to the introduction of oxen. Overall, household member time spent on plots decreased by about 1.5 hours per hectare or 7%.
In the next section, we build on the observed decrease in aggregate household agricultural labor to examine the intra-household labor impacts of the oxen. 16 Appendix Table B2 presents impacts on the production of other crops. Farmers assigned to receive the traction oxen increased rice production by 9%. Rice is primarily produced for household-consumption with only 8% of the rice-producing households selling any amount. 17 Appendix Table B3 presents the estimated impacts of the treatment on the proportion of expenditures across six categories, finding that treatment households increased their spending on agricultural investments as a share of their overall spending.

Intra-household labor allocation
This section unpacks the intra-household composition of the observed reduction in agricultural labor provision as a result of the intervention, as well as its determinants, testing the predictions in Section 2. We restrict attention to the 1,834 married producers with complete time-use data for spouses. 18 Table 7 shows that the oxen had different impacts for different household members, with time on household plots decreasing for women and girls. Columns 1 and 2 show that assignment to treatment had no significant impacts on the producer's time use. Columns 3-8 show that treatment households decreased the time spent on household plots by almost 6% and 9% for wives and girls, respectively. The large decrease in household plot activities for girls represents around two-thirds of a larger decrease in their total reported active time. 19 In this table, we report results for both the first spouse and other spouses for polygamous households. The impact of the treatment on time use, including household plot activities, is not statistically different between first spouse (column 4) and other spouses (column 6). Results on other spouses being noisier but equivalent to results on first spouse, the rest of the paper studies the average impact of treatment on spouses, by taking the average of outcomes across the producer's wives, when applicable. Finally, boys in treatment households increased time spent on household chores by almost 0.6 hours, representing an increase of over 30%. 20 These findings speak directly to our model: the significant decrease in female plot labor supports the presence of sizable social prescriptions dictating that women should not handle traction oxen (a f > 0). Relatedly, the lack of significant impacts on domestic work suggests a near-zero response of the marginal product of labor to a capital shock (H T K ≈ 0). We examine the ambiguous observed impact on non-farm labor in more detail in Section 4.3, below, by exploiting district variation in social prescriptions regarding women working with oxen (a f (K − K)).
Building on the decreased time women and girls spend on household plots, Tables 8 and 9 examine treatment effects for education and health outcomes. Panel A of Table 8 shows that the program did not increase either overall school enrollment for children aged 6-16 years old or student performance on a simple literacy and numeracy assessment. However, this may have been a tall order, since it would have required many older children to enroll in school for the first time or to return to school after having dropped out previously. With this in mind, Panel B examines whether the oxen increased the likelihood that students who had previously attended school were still attending school, showing that boys in treatment households were almost 30% less likely to have dropped out of school. Among children that had attended school, we find no evidence of increased performance on the literacy or numeracy assessment.
Decreased time spent on agricultural plots also has the potential to affect child health by decreasing their exposure to agricultural chemicals and zoonotic diseases. Table 9 presents impacts of treatment on likelihood and duration of illness among the producers' spouse and children. While all six of the coefficients are negative, the impacts are only statistically significant for girls, for whom treatment decreases the likelihood of illness and duration of illness over the last month by about a quarter. The coefficients for boys are about half as large as those for girls and are not statistically significant. This makes sense, since girls are reducing the time they spend working in agriculture at higher rates.

Role of community norms in driving intra-household impacts
As discussed in Sections 2 and 3.2, community norms around women working with oxen are likely to play a key role in determining the intra-household impacts of a mechanization intervention. If the community norm is that women are not involved in oxen tasks, introducing oxen to agricultural work may shift women away from agricultural work. This subsection builds on these theoretical considerations by examining treatment effect heterogeneity for households residing in districts where women perform relatively more or relatively fewer oxen tasks. 21 Table 10 presents heterogeneous impact estimates for second-season time use outcomes. For this heterogeneity analysis, we average spouse outcomes across wives in households where the producer had more than one wife. Column 1 shows the estimated impact of delivering oxen to households residing in districts where women conduct more oxen tasks. In these households, oxen treatment had relatively limited time-use impacts with none of the 12 coefficients being statistically significant. This contrasts with the results presented in Columns 2-3, where we display the marginal impact of treatment within households who reside in districts where women perform fewer oxen-related tasks (Column 2) as well as the p-value of the test comparing these impacts to those in districts where men dominate oxen work (Column 3). In these households, we find large intra-household impacts of the oxen delivery. In particular, in districts where women are typically not engaged in oxen work, we see a significant 1.78 hour/week increase in how much time the spouse spends on off-farm work. This matches qualitatively with the fact that the spouses' (and girls') reduction in time spent working on the producer's plots is concentrated in these same districts. This is consistent with our model: the impacts of introducing capital depend critically on the social penalty to capital intensive 21 We define households as residing in districts with relatively high levels of women's involvement in oxen tasks if the share of oxen tasks conducted by women in control households within the district exceeded the median across all districts. Appendix Table B5 presents summary statistics comparing households residing in districts with relatively high and relatively low rates of women's involvement in oxen tasks, showing significant differences across several household and production characteristics. We include controls for each of the demographic characteristics in our heterogeneity specification. Controlling for production-related differences gives equivalent results. work for different household members, a f (K − K). Taken together, these results show that gender norms around suitability of working with oxen lead to a concentration of household plot activities under men in the household, with women shifting to non-agricultural work.
One concern with the above heterogeneity analysis is the potential endogeneity of women's involvement in oxen tasks. While we define this using endline data from control group households, the phase-in design means that many of these households also received the oxen and may have changed behaviors in response to treatment. With this in mind, we demonstrate the robustness of the heterogeneity results to using districtlevel differences in women's decision-making power, shown in Figure 3, as an alternative construction of gender norms. Appendix Table B6 presents equivalent heterogeneity results using pre-program DHS data to characterize districts in the sample as either more or less conservative based on women's decision-making power. Since, as discussed in Section 3.2, the two measures of district norms are highly correlated (0.75), this alternative way of analyzing heterogeneity correspondingly yields a similar pattern of results.

Conclusion and discussion
In this paper, we provide experimental evidence on the short-term impacts of mechanization on household production and intra-household labor allocation among cotton farmers in Côte d'Ivoire. We find positive impacts on cotton production-a male-dominated cash crop-and intensification of agriculture, along with an increase in the value of complementary agricultural inputs used. Though our results are short-term and we do not detect a significant increase in household income in our survey time-frame, second season outcomes indicate that households expand their overall agricultural cultivation and not merely cotton. Moreover, our data indicates that animal traction can lead to yield-not just production-improvements, as farmers were able to increase their cotton production after land cultivation decisions were made.
Our results shed additional light on four stylized facts related to agricultural mechanization identified in observational studies. First, earlier observational studies analyzed by Pingali et al. (1987) suggested minimal yield gains for animal traction farms relative to hand-hoe farms. The increased cotton and rice production we detect in the first season, together with the fact that the oxen were distributed after land cultivation decisions were made, suggest that the oxen increased yields through either task quality or efficiency improvements during crop maintenance, harvest, and/or crop transport. Second, Pingali et al. (1987) documented a relationship between animal traction use and a larger area under cultivation. Our estimates indicate that this relationship is causal, with households who gain access to traction oxen bringing new land under cultivation.
Third, earlier findings have suggested that additional area brought under cultivation as a result of mechanization was concentrated among market crops including cotton, rice, and groundnuts (Pingali et al., 1987). We find impacts concentrated among both cotton and rice but caution against the market crop characterization: few farmers in this setting report selling rice, suggesting that the impacts may stem from specific traction oxen efficiencies of cultivating rice rather than a market crop designation. Finally, Pingali et al. (1987) note that there is "general agreement that a transition [to animal traction] reduces the amount of labor required during the land-preparation period". We find evidence of broader labor savings from oxen: specifically, we document a reduction in household agricultural labor during crop maintenance, when households may have been expected to need more labor due to their increased land under cultivation. This is in line with results from Afridi, Bishnu and Mahajan (2020), who point to the fact that mechanization leads to higher-quality plowing, which in turn lowers the demand for weeding labor.
Our results show the broad labor-saving potential of animal traction beyond the land preparation period, but also how this labor-saving does not affect everyone in the household equally. Only wives and girls significantly reduce the time they spend working on household plots as a result of mechanization. The human development impacts are similarly gendered: girls appear to benefit in terms of their health, while boys benefit in terms of their schooling (possibly because of an anticipatory income effect by households).
Our findings indicate that the adoption of labor-saving technology in male-dominant activities can have large intra-household effects and welfare impacts. They also underscore the importance of complementary development efforts. For example, while boys are less likely to drop out of school as a result of mechanization, they do not observe an increase in education outcomes-pointing to the potential need for increased investment in school capacity. Moreover, our analysis shows how gender norms are tightly linked to mechanization impacts for women, with women in districts with stronger norms against women working with oxen shifting off-farm. Looking forward, it will be important to better understand how these relatively conservative norms interact with greater off-farm engagement.
In terms of consequences for structural transformation, it is worth noting that norms related to women working with technology and in capital-intensive tasks affect many more domains than simply working with oxen.
Prescriptions regarding the acceptability of different types of work for women are not just relegated to informal norms, but are also enshrined in formal legislation. In 49 of 190 countries worldwide, women are not allowed to work in jobs deemed 'dangerous' in the same way as men (World Bank, 2023). Beyond examining the effects of legal and normative changes regarding what sectors it is acceptable for women to work in, future research should examine whether offering complementary gender-targeted programming alongside mechanization can strengthen women's economic empowerment, and more broadly continue unpacking the relationship between technology, labor and (gendered) power. Note: Robust standard errors reported in parenthesis. *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals. Pre-program cotton area and production come from cotton producer database.  Note: *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals, respectively. indicates variables that have been winsorized at the 5% level. All regressions include controls for whether household head completed primary school, number of boys between 6 and 16 years old, and stratum fixed effects. Note: *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals, respectively. indicates variables that have been winsorized at the 5% level. All regressions control for whether household head completed primary school, number of boys between 6 and 16 years old, and stratum fixed effects.  Note: *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals, respectively. Agricultural input usage only collected for half the sample. indicates variables that have been winsorized at the 5% level. All regressions control for whether household head completed primary school, number of boys between 6 and 16 years old, and stratum fixed effects. Note: *, **, and *** indicate significance at the 90, 95, and 99 % confidence intervals, respectively. indicates variables that have been winsorized at the 5% level. Regression reports impacts of treatment on weekly average household time use for given tasks: for households with multiple wives, daughters, or sons, the time use is averaged and regressions are reported at the household level. All regressions control for whether household head completed primary school, number of boys between 6 and 16 years old, and stratum fixed effects.  Note: *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals, respectively. indicates variables that have been winsorized at the 5% level. All regressions control for whether household head completed primary school, number of boys between 6 and 16 years old, and stratum fixed effects.

Appendixes Appendix A: Derivation of Comparative Statics
The household allocates women's labor across on-farm (f), off-farm (o), and domestic (d) work to maximize utility, subject to a fixed time constraint. This can be represented as the following maximization equation: subject to the time constraint: The associated Lagrangian yields the following first order optimality conditions: where λ is the Lagrange multiplier on the time constraint. As detailed in the body of the paper, we set a o = a d = 0 for ease of exposition. From the first order conditions, we derive a set of reduced form equations for the allocation of time into each of the three activities. Time in each activity will depend on the level of household capital, the price of the market product, the extent of conformity to social norms, and intensity of penalties from deviating or rewards for conforming.
Following Kevane and Wydick (2001), we examine whether the task-induced and capital-dependent social norms influence the time allocation of women using comparative statics showing how time allocation varies with changes in animal traction K. Dropping the subscripts, the matrix of totally differentiated first order conditions is: The Hessian determinant is: While Cramer's rule yields the desired partial derivatives: Appendix B: Supplemental Tables      Note: Robust standard errors reported in parenthesis. *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals, respectively. † indicates hours in last week. All regressions include controls for household composition, household-head education, and stratum fixed effects. Note: *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals, respectively. indicates variables that have been winsorized at the 5% level. All regressions include controls for whether household head completed primary school, number of boys between 6 and 16 years old, and stratum fixed effects. Note: *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals, respectively. indicates variables that have been winsorized at the 5% level. All regressions control for whether household head completed primary school, number of boys between 6 and 16 years old, and stratum fixed effects. Note: *, **, and *** indicate significance at the 90, 95, and 99% confidence intervals, respectively. Agricultural input usage only collected for half the sample. indicates variables that have been winsorized at the 5% level. All regressions control for whether household head completed primary school, number of boys between 6 and 16 years old, and stratum fixed effects.