INEFFICIENCY OF MALE AND FEMALE LABOR SUPPLY IN AGRICULTURAL HOUSEHOLDS: EVIDENCE FROM UGANDA MARTYN J. ANDREWS, JENNIFER GOLAN, AND JANN LAY This article analyzes the efficiency of the intra-household allocation of female and male labor inputs in agricultural production. In a collective household model, spouses’ optimal on-farm labor supply is such that the marginal rate of technical substitution between male and female labor is equated over different crops. Using the Uganda National Household Survey 2005/06, we test whether this condition holds by estimating production functions and controlling for endogeneity using a method proposed by Gandhi, Navarro, and Rivers (2009). We find that women are less productive than men, that there is more female labor input on low productivity parcels, and that men are relatively more productive on female-controlled plots compared with male-controlled plots. Total farm output could be higher and Pareto improvements could be possible if male labor was reallocated to female-controlled plots and/or female labor was reallocated to male-controlled plots. Key words: Intra-household allocation, labor supply, gender, Uganda. JEL codes: D13, J16, O12, Q12. In many developing countries, smallholder production remains the main source of livelihood for poor rural households. The literature on agricultural households has stressed the role of missing insurance, credit, and labor markets in explaining the moderate supply response to market reforms of smallholders (Key, Sadoulet, and De Janvry 2000; Taylor and Adelman 2003). In addition to these market constraints, smallholder labor allocation decisions may be affected by norms and rules that may cause suboptimal outcomes. Specifically, gender roles in agriculture have been previously shown to potentially cause inefficiencies in agricultural Martyn Andrews is a Professor of Economics and Jennifer Golan is a Lecturer in Development Economics, both at the University of Manchester, UK. Jann Lay is head of the Research Program “Development and Globalization” at the GIGA German Institute of Global and Area Studies, Hamburg, Germany, and Assistant Professor at the University of Göttingen, Germany. Jennifer Golan gratefully acknowledges financial support of the Poverty Reduction, Equity, and Growth Network, and the Economics Department of the University of Manchester for funding her PhD. The authors are thankful for comments received by participants of the Royal Economic Society Conference in London, seminar participants at Purdue University, the Econometric Society World Congress in Shanghai, the Bolivian Conference on Development Economics in La Paz, and the German Economic Association Meeting in Göttingen. The authors also thank Mette Christensen, David Colman, Mika Kortelainen, Andy McKay, Ed Taylor, Bernard Walters, and various referees for useful comments that substantially improved this paper. production, though rigorous evidence on this is scarce. The so-called collective model has been established as the “workhorse” for analyzing intra-household resource allocation. Like all other cooperative models, a key assumption of the model is that household decision outcomes are Pareto efficient. The predictions of the collective model for consumption behavior have been tested in different contexts. While studies using data from developed countries tend to support the collective model (Bourguignon et al. 1994; Browning and Chiappori 1998; Thomas and Chen 1994), evidence from developing countries often does not support efficiency (Udry 1996; Duflo and Udry 2004; Goldstein and Udry 2008; Quisumbing and Maluccio 2003). In particular, production decisions have been found to be sub-optimal when households have to adjust to new agricultural conditions such as deeper market integration or changes in relative prices (Jones 1983; McPeak and Doss 2006). In this article, we use a version of the collective model that is characterized by an efficient allocation of factor inputs to evaluate the optimality of on-farm labor allocations in agricultural households in Uganda. The model is based on the original model of Chiappori (1992), which has been extended for household production by Udry (1996), Amer. J. Agr. Econ. 97(3): 998–1019; doi: 10.1093/ajae/aau091 Published online October 27, 2014 © The Author 2014. Published by Oxford University Press on behalf of the Agricultural and Applied Economics Association. All rights reserved. For permissions, please e-mail: [email protected] Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households Apps and Rees (1997), Chiappori (1997), and Donni (2008). For consumption choices, the model predicts certain conditions for domestic production decisions. Domesticallyproduced goods can be divided into those that can be sold or bought in the market (e.g., crop production) and non-marketable goods (e.g., children’s welfare). Chiappori (1997) and Donni (2008) show that if the good has a market substitute, then the allocation of resources in production can be analyzed independently from consumption decisions; essentially this is the same as the standard unitary agricultural household model, where the household has no restrictions in accessing product markets (Singh, Squire, and Strauss 1986). In other words, factors of production are allocated to home production as if the household maximizes domestic profits. Apart from this marketability assumption, a further challenge for collective models with domestic production is that data on inputs and outputs are seldom available in standard datasets. However, in developing countries, domestic production of marketable crops is a main source of household income and standard survey data often also collects input, output, and price information of farm production. There is an increasing body of evidence that female and male labor inputs are allocated sub-optimally in smallholder households. This evidence suggests that intrahousehold farm behavior may be better described by alternative household models in which conflicts of interest between the partners may produce inefficient outcomes. Udry’s seminal (1996) study concludes that productive resources are allocated inefficiently in rural Burkina Faso. Efficiency requires that two equal plots should have identical yield and input allocations, and so any deviation should be only a function of plot characteristics. Controlling for plot characteristics and household-year-crop fixed effects, Udry finds that yields per acre of identical crops are lower for plots controlled by women than men by approximately 30%. This can be explained by fertilizer and labor inputs being more intensely used on maletended plots. In other words, were Burkinabe households to allocate resources efficiently, given diminishing returns, they would reallocate resources to female-tended plots. Using CES production function estimates, Udry shows that male and female labor inputs are equally productive, while all labor inputs are highly substitutable on identical crop-plots. 999 Also, the results indicate no differences in production technology between the plots. This means that, amongst other factors, labor could be reallocated so as to increase output (see Akresh (2005) for further evidence for Burkina Faso). Preceding Udry (1996), Jacoby (1992) studies the division of labor of men and women in livestock and crop production in Peru; he finds that women are more productive in livestock and men in crop production, which means that the efficiency requirement would be violated if labor inputs were allocated in equal proportions. However, the data used does not disaggregate male and female labor inputs by activity, which undermines this test. Nonetheless, Jacoby’s results illustrate that the efficiency of allocating male and female labor not only depends on relative factor productivity, but also on the production technology employed. Udry (1996) and Jacoby (1992) are consistent with the earlier findings of Jones (1983), who explains why married women, compared to widowed women, allocate “insufficient labor” to paddy rice plots in North Cameroon. Incomes from rice production are controlled by the husband, but cultivation relies on both husband’s and wife’s labor inputs. Jones estimates the compensation rate to female labor input in rice production; her results suggest that the suboptimal labor allocation can be explained by the low rate of compensation that married women receive relative to the opportunity costs faced; that is, income from sorghum production, agricultural wage labor, and other incomes. In other words, if households were to allocate resources so as to maximize incomes, compensation to married women would need to increase. In this article, we analyze the efficiency of intra-household allocation of female and male labor inputs in the farm production of various crops in the context of Uganda. Our contribution to the literature is as follows. First, we derive a test of the efficiency of male and female labor allocation over various crops using the collective model derived by Browning and Gørtz (2012). Second, our empirical strategy is based on theoretical conditions that allow us to identify all relevant parameters for testing efficiency, but without suffering from the inherent endogeneity problems of estimating production functions. Third, we use the Uganda National Household Survey (UNHS) 2005/06, which 1000 April 2015 contains information on who controls the output from a given piece of land. Our study is related to but distinct from Udry’s. While Udry analyzes efficiency within crops across different plots, we analyze labor allocation both between plots as well as between crops (within the same household), partly using a different theoretical restriction. Further, this dataset records information on farmers who are much more integrated into markets than those analyzed by Udry. The next section analyzes the literature on traditional gender roles and social norms in agricultural production in Sub-Saharan Africa (SSA) and Uganda. In the following section, we derive a test of efficiency considering different scenarios of labor supply. We then describe the data and discuss econometric issues, particularly relating to the possible endogeneity of labor inputs. The penultimate section discusses the results and the final section concludes. Gender Roles and On-farm Labor Allocation Anthropological, sociological, and anecdotal evidence points to a “gendered division” of crops and tasks in SSA that might explain intra-household labor allocation. In the SSA context, traditional gender roles not only reflect and reinforce comparative advantages in agriculture, but are in many countries a mechanism for the division of income within the household. If established structures become incompatible with new incentives, such as through price increments for a certain type of crop, this might cause difficulties when adjusting the distribution of benefits. While this does not necessarily impact on the effort and amount of labor applied, achieving a compensation rule compatible with profit maximizing production behavior may be fairly complicated under rigid traditions and gender roles. For instance, in many SSA countries, women grow food crops mainly to guarantee domestic food consumption, while men tend to control output from cash crop production while still relying on the female and family labor inputs (FAO, IFAD, and ILO 2010). As noted, Jones’s (1983) study shows that if women do not feel adequately compensated, they tend to supply insufficient labor. Dolan (2001) provides evidence of struggles over resources in horticulture-producing households in Meru, Kenya. French beans, Amer. J. Agr. Econ. a crop originally within the female domain, were expanded for export purposes. The resulting shift in the productive focus of the household increased the labor burden on females. At the same time, men tried to extend their control to the proceeds of French bean production, but thereby adversely impacted the female position within the household, and presumably negatively affected female labor effort in the production process. Traditional gender roles have also been named as a reason for men being reluctant to contribute to food crop production in Africa (Boserup 1976), despite the possibility that they might be more efficient at performing certain tasks. Both men and more often women may be excluded from certain tasks as well as entire income-generating spheres, thereby illustrating the rigidity and persistence of traditional gender roles, in particular with regard to the allocation of labor within the household (Dey 1981; Jones 1983; and von Braun and Webb 1989). Still, whether a certain cropping pattern follows gendered lines has been found to depend on the cultural context as well as agro-ecological context (Doss 2002). We analyze data from Uganda. Here, there is also evidence that cash crop production— particularly coffee production—relies on female labor inputs in the production process, while marketing and income control lie in male hands (Kyomuhendo and McIntosh 2006; Elson and Evers 1996; Evers and Walters 2001; Kasente 1997). In addition to an amount of time-consuming home duties such as fetching water and collecting firewood, producing food crops and specific tasks (such as weeding) required to produce other crops are typically performed by women (Kasente et al. 2000; Dolan 2001). Social norms have been named as a factor constraining women’s ability to mobilize male labor within the household (Hill & Vigneri 2011). Doss and Morris (2001) note that in many African countries men can command women’s labor while the reverse is not true. This is also true for Uganda, and Evers and Walters (2001) argue that this pattern is reinforced by social norms such as obligatory dowry payments to the spouse’s family. Traditional gender roles may be challenged by changes in the structure of the economy and the corresponding changes in economic incentives. One important feature of structural change in SSA is the increasing Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households market integration of rural households (Hazell et al. 2010) that causes a less clear-cut distinction between cash and food crops. For instance, increased prices for leafy vegetables in Kampala, Uganda, shifted men’s focus on its production (World Bank, FAO, and IFAD 2009). Increased marketed production of traditional food crops may cause conflicts of interest between the partners over crops, for example in the case of banana production in Uganda (Ministry of Gender, Labor, and Social Development 2005). These conflicts may result in sub-optimal resource allocations, as previous gender roles and associated compensation rules do not necessarily lead to optimal outcomes under new market conditions. The evidence surveyed in this section hence suggests that sub-optimal production outcomes may be caused inter alia by an inefficient allocation of labor, which in turn may be linked to rigid gender roles in agricultural production. In addition, intra-household compensation for labor inputs in domestic farm production typically depends on control over proceeds from certain crops. Again, control over crop output is determined by traditions and social norms and the evidence above suggests that these may not provide adequate incentives for efficient labor allocation. The Collective Farm Model and Testable Implications The evidence surveyed in the previous section suggests that intra-household struggles over labor input and income control may give rise to production inefficiencies. To test whether this is the case, we derive a condition for the optimal allocation of male and female on-farm labor inputs using a collective farm model based on Browning and Gørtz (2012); the model also draws on Singh, Squire, and Strauss (1986), Udry (1996), and Apps and Rees (1997). The Collective Farm Model The household consists of two partners, labeled A and B. Partner A’s utility depends on the consumption of a good bought in the market qm A , hours of leisure lA , a f domestically-produced subsistence good qA , a non-rival household public good Q, 1001 and a vector of preference factors aA 1 : f uA = uA (qm A , lA , qA , Q; aA ). Partner B’s utility B u depends on exactly the same arguments: f uB = uB (qm B , lB , qB , Q; aB ). Partner A cares about B in the sense that her total utility comprises her own uA plus a proportion λA of B’s. In other words, her welfare function is defined as A = uA + λA uB , with λA ∈ [0, 1]. If λA = 0, A is said to be egoistic, that is, she does not care about B’s utility. If λA = 1, A is completely indifferent between her own utility and her partner’s utility. We rule out the possibility that λA < 0, namely that she dislikes her partner, or λA > 1, that she cares more about her partner than herself. The same considerations apply for B: B = uB + λB uA , with λB ∈ [0, 1]. Further, h is the household’s welfare function, that is, a weighted average of the individual welfare functions: h = μ̃A + (1 − μ̃)B , with μ̃ ∈ [0, 1]. The weight μ̃ reflects the so-called “balance of power” within the household. If μ̃ = 1, then A is all-powerful. Economic theory does not provide a framework explaining the content of μ̃, but it is hypothesized to be a function of so-called distribution factors. These are exogenous factors that impact the allocation of power within the household but not on preferences, like aA and aB , or the budget constraint (Browning and Chiappori 1998). Substituting the individual utility functions into household welfare, we can redefine h as μuA + uB . The so-called Pareto weight μ is a composite index of μ̃, λA , and λB , and captures the relative weight of partner A in the decision process. Each partner has a total time endowment of T hours a day. For A, T can be allocated to leisure lA , food crop fA , cash crop cA , public good production hA , or offfarm employment activities (self- or wage employment), mA . In maximizing household welfare, the partners face the following constraints on their allocation of time, namely T = lA + mA + fA + cA + hA and T = lB + mB + fB + cB + hB . Total food crop production qf can be allof f cated to subsistence consumption qA + qB , f or traded in the local market, qm , at price pf . The food crop production function F f uses domestic labor inputs fA and fB , and land L: 1 For the sake of simplicity, we assume that children can be treated as public goods. 1002 April 2015 f f Amer. J. Agr. Econ. f qf = qA + qB + qm = F f (fA , fB , L). The variable L is assumed to be fixed and exogenous, which allows the analysis to focus on the short-run intra-household allocation of labor inputs in which labor can be more flexibly reallocated between crops than land. The same applies for cash crop production, except that this crop is entirely traded in the market at the cash crop price pc : qc = F c (cA , cB , L). The production of the household public good Q depends on each partners’ time input, hA , hB , and material inputs bought in the market, qH : Q = F Q (hA , hB , qH ). All three production functions, namely food crop F f (.), cash crop F c (.), and public good F Q (.) are increasing in all inputs, and satisfy standard concavity conditions. In addition, domestic on-farm labor inputs and off-farm employment activities are nonnegative. For simplicity, we assume that both partners supply a positive amount of labor to domestic cash crop production. Also, we consider that A’s optimal off-farm labor supply, m∗A , could be limited by social norms to a sub-optimal level m̄A . This means that she may be forced to allocate her time between leisure, public goods, and domestic farm production beyond the desired level. Further, there is an income constraint stating that expenses on consumption goods and material inputs into public good production cannot exceed the total of on-farm profits, nonwage income, y, and labor incomes, mA wA and mB wB . This constraint is written as follows: (1) m H H pm (qm A + qB ) + p q f f = pf [F f (fA , fB , L) − (qA + qB )] + pc F c (cA , cB , L) + y + mA wA + m B wB . The household’s problem is to maximize total household welfare μuA + uB subject to the 5 constraints and 3 production functions. The two relevant first-order conditions are as follows: (2) pf ∂F f ∂F c = pc = wA ∂fA ∂cA (3) pf ∂F f ∂F c = pc = wB . ∂fB ∂cB If both partners supply labor to both crops, equations (2) and (3) are interpreted as if the farm is maximizing the profit of jointly producing the food crop f and the cash crop c. For partner A, the marginal revenue products of labor for both crops are equal to each other and equal to her marginal cost wA . As this is also true for partner B, the marginal rate of technical substitution (MRTS) between male and female labor is equated over the two crops: (4) ∂F f ∂fA ∂F f ∂F c = ∂fB ∂cA ∂F c = wA /wB . ∂cB The left-hand equality is what we test in the rest of this paper; hereafter, we refer to this as our test of allocative efficiency. We also define the two MRTSs as θf and θc , respectively. It can also be shown that if partner A supplies no labor to off-farm activities and she is constrained to supply all the desired labor to the local labor market, this condition still holds. In other words, within the same household, male and female labor inputs cannot be reallocated so as to increase output if the marginal rates of technical substitution are equated over crops, that is, θc = θf . This condition does not depend on the distribution of power within the household. Equation (4) also holds in the presence of market imperfections; more specifically, it does not depend on the existence of a functioning labor market. In the rest of this section, we show how to apply our test of allocative efficiency when estimating production functions for food and cash crops. These different kinds of crops should be seen as a stylized separation of different agricultural production activities. In our empirical analysis we distinguish between plots “controlled” by males or females, respectively. Testing Allocative Efficiency To implement our test, we estimate the four marginal products in equations (2) and (3), which means estimating the parameters of F f (fA , fB , L) and F c (cA , cB , L). However, first we need to address a well-known problem, that is, the labor inputs are likely to be endogenous. There are two possible reasons, both of which are gender-specific. This exacerbates the problem because the biases in estimating the effects of fA and fB will not Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households cancel out when computing θ. The first reason is that there may be idiosyncratic shocks caused by the weather, crop pests, diseases, and so on. These affect men and women differently because they grow and control different crops. Accordingly, male and female marginal products will be systematically different from each other and these differences would be incorrectly interpreted as inefficiencies. The second reason is that men are more likely to self-select themselves into farming plots and crops that yield higher returns. The issue of endogenous labor inputs was noted in the developing country context by, for example, Fafchamps (1993), who analyzed data for Burkina Faso. Similarly, Udry (1996) was cautious in interpreting his results because it is difficult to find valid instruments that affect factor inputs but are not correlated with production shocks. Like Udry, we are unable to pursue an instrumentalvariables type strategy, and so we follow closely an approach proposed by Gandhi, Navarro, and Rivers (2009). We can view the household’s problem as choosing its inputs fA and fB by maximizing its profits for the food crop: pf F f (fA , fB , L)ev − wA fA − wB fB , where v captures these shock and selection issues, and is observed by the household. Not only is v correlated with fA and fB , the two correlations might well be different. At the same time, the household maximizes its profits for the cash crop, but here we focus solely on the food crop. The first-order conditions equations (2) and (3) are rewritten as follows: (5a) pf ∂F f (fA , fB , L) v e = wA ∂fA (5b) pf ∂F f (fA , fB , L) v e = wB . ∂fB Once fA and fB are chosen, actual food crop output is given by: (6) yf = F f (fA , fB , L)ev e ≡ qf e equations (6) and (5b), so that, for partner B: p f yf log(1/sB ) ≡ log w B fB fB ∂F f ( ) = − log + . F f ( ) ∂fB (7) There is a symmetric equation for Partner A. We emphasize that wB fB is hypothetical in the sense that it is the cost that would be incurred by supplying fB hours at a market wage rate wB . The same applies to wA fA . Both wA and wB reflect the opportunity cost of an hour of labor, while wA fA and wB fB are different from the labor incomes earned in the market, wA mA and wB mB , as in equation (1). Hence, the dependent variable 1/sB , when logged, is the (inverse) share of revenue earned in producing yf that would be paid to partner B for supplying fB hours. Lastly, is the only unobserved variable left in the model. The right-hand side of equation (7) depends on the functional form of the production function. We choose the Constant Elasticity of Substitution (CES) production function because the elasticity of substitution between the labor inputs for partners A and B can be estimated unrestrictedly. Because we want to focus on the elasticity of substitution between fA and fB only, we write the production function for the food crop as: (8) log F f (fA , fB , L) νf −ρ −ρ = − log[δAf fA f + δBf fB f ] + γf Lf ρf where νf is the returns to labor parameter, ρf is the substitution parameter, and δAf and δBf are the distribution parameters that capture the marginal products of partners A and B. Note that 0 ≤ δAf , δBf < 1 and, because of land Lf and other possible labor inputs, δAf + δBf < 1. Equations (6), (5a), and (5b), respectively, become: (9) (10) where is another variable, this time not observed by the household nor the “econometrician.” The key insight of Gandhi, Navarro, and Rivers (2009) is to recognize that v can be substituted out between 1003 log yf = log F f + v + log wA /pf = log F f − (1 + ρf ) log fA −ρ −ρ − log δAf fA f + δBf fB f + log(νf δAf ) + v 1004 April 2015 Amer. J. Agr. Econ. and (11) regression parameters as follows: log wB /pf (14) f θf ≡ MRTSAB ≡ = log F f − (1 + ρf ) log fB −ρ −ρ − log δAf fA f + δBf fB f + log(νf δBf ) + v. −ρf 1+ρf . (15) δAf δBf fA fB −(1+ρf ) δAc = δBc cA cB −(1+ρc ) . −ρ + δBf fB f ] + ρf log fB − log(νf δBf ) + ≈ − ρf λAf log fA + ρf (1 − λBf ) log fB − ρ2f λBf λAf g(fB , fA ) − log(νf δBf ) + where λAf ≡ δAf /(δAf + δBf ) and λBf ≡ δBf /(δAf + δBf ). The last line of equation (12) is a linear approximation that comes from a Taylor series expansion (Kmenta 1967). We note that (13) fB fA This is a non-linear restriction involving six parameters: δAf , δBf , ρf , δAc , δBc , and ρc . The four labor inputs are evaluated at their sample means. log(1/sB ) f f py ≡ log w B fB = log[δAf fA Using the equivalent expression for the cash crop, our test is written as It is v that causes the labor inputs to be endogenous; differencing out v between equations (9) and (11) results in: (12) δAf = δBf ∂F f ∂F f / ∂fA ∂fB g(fA , fB ) ≡ −(log fA − log fB )2 /2. Unique estimates of δAf , δBf , and ρf are identified from equation (12). By symmetry, there is another equation for log(1/sA ), but we do not estimate this equation, as there are issues relating to the quality of the data we have for female wages wA , discussed later. Similarly, we identify δAc , δBc , and ρc from the analogous equation for the cash crop. In what follows, we refer to output as the “dependent variable” because we seek to estimate the parameters of production functions, even though the left-hand side of our regression model is log(1/sB ). Our test of the theory is whether the marginal rate of technical substitution between male and female labor inputs is the same across both crops (see equation (4) above). For the food crop, the marginal rate of substitution is written as a function of the Data In this section we describe the Uganda National Household Survey (UNHS) and the variables we construct for our empirical analyses. Samples The data we use come from the Uganda National Household Survey 2005/06, and was collected by the Ugandan Bureau of Statistics. The UNHS is nationally and regionally representative and contains detailed information on agricultural activities. The agricultural questionnaire covers two cropping seasons: July to December 2004 and January to June 2005, which are labeled “Season t = 1” and “Season t = 2,” respectively.2 We restrict our analysis to male-headed households that live with one or more spouses; this is because analyzing bargaining between spouses is more straightforward in male-headed households.3 After dropping observations with missing values, this gives a sample of 3,665 male-headed households. Hereafter “spouse” and “female” are used interchangeably. 2 Uganda has two rainy seasons. The first starts around March, can last until June/mid-June, and affects the first harvest (t = 2). The second starts around August, lasts until November, and affects the second harvest (t = 1). In northern regions of Uganda, the two rainy seasons tend to merge into each other. 3 In this sample, 21.2% of households are polygamous, the rest are monogamous. Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households 1005 Table 1. Sample Characteristics Households Parcels Plots of which: Monocropped Intercropped Changersc t=1 t=2 t = 1&2 Unitsa Total obsb 3,268 6,407 9,337 3,380 7,092 13,489 2,983 5,225 6,915 3,665 8,274 15,911 6,648 13,499 22,826 4,025 3,309 2,003 6,260 5,226 2,003 2,421 2,491 2,003 7,864 6,044 2,003 10,285 8,535 4,006 Note: Sample is defined as male-headed households with one or more spouses (3,665 households). a indicates column (4) = (1) + (2) − (3), and provides total number of households/parcels/plots. b indicates column (5) = (1) + (2), and provides the total number of household obs./parcel obs./plot obs. c indicates monocropped in one season, intercropped in the other. Table 1 summarizes the sample characteristics of these 3,665 households. Of these, 2,983 are observed in both seasons, resulting in 6,648 household-season observations. Households have land parcels that comprise either monocropped plots, where only one crop is grown, or intercropped plots, where more than one crop is grown. We observe 15,911 plots on 8,274 parcels for the 3,665 households. A parcel is a group of plots with similar soil characteristics, discussed in more detail below. Because there is more rainfall in the first rainy season, households use more plots (13,489) in Season t = 2 compared with Season t = 1 (9,337). This results in 22,826 plot-seasons. Of the 15,911 plots, 7,864 are monocropped (10,285 plotseasons) and 6,044 are intercropped (8,535 plot-seasons), and 2,003 plots change from monocropped to intercropped, or vice versa, between the two seasons. There are 13,499 parcel-seasons for these 3,665 households. Because some permanent crops such as coffee have a two-season growing cycle, we define the dependent variable yp as the total annual value of plot output. Our empirical analysis is based on this plot-level cross-section comprising 15,911 annual plot observations. We also create a crop-level cross-section, where the unit of observation is the total annual value of crop output for a given household for 7,152 household crops. Here, the variable being modeled is denoted as yhc . Generally, the information we require from the agricultural questionnaire is observed at different “levels” (households, parcels, and plots) depending on the variable of interest. This has implications for our empirical analysis. For a given household-season, a) labor inputs are recorded for each plot, whereas b) output is recorded for each crop at the household-season level, but not for each crop on each plot.4 Unfortunately, a) implies that we cannot tell how much time is spent on each crop on an intercropped plot. However, b) implies that output for each plot (at the household-season level) can be constructed from the output for each crop and the area used.5 To do this, we assume that—for a given crop grown on different plots within a household-season—output per acre is the same for all plots. To illustrate, consider a household with two plots labeled p = 1, 2. If these two plots are monocropped (say, coffee on p = 1 and peas on p = 2), we can match the respective labor inputs for both plots with the household-season information on coffee and peas output. However, if the household produces some coffee on p = 2 as well—so that this plot is intercropped—we use the assumption above to share the output value of coffee between the two plots.6 The econometric implications of this “output sharing” rule are discussed below, and the appendix details the actual procedure. The resulting plot-level dataset comprises information on total output value and labor 4 Households are much more likely to report with accuracy their labor inputs by plot, and their output by crop. This is because they may not know from which plot the output of a given crop came, but are more likely to know how much time they spent working on a given plot. 5 The acreage of a specific crop on an intercropped plot is derived from the following survey questions: “If the plot is intercropped, the total plot area should be entered ... for each crop and then the percentage of the plot area under the component crops.” 6 Suppose there are two acres of coffee on the mono-cropped plot and one acre of coffee on the second intercropped plot. In this case, two-thirds of the household’s coffee output would be “allocated” to plot 1 and one-third to plot 2. 1006 April 2015 Amer. J. Agr. Econ. Table 2. Sample Averages by Plot Type, Parcel Control, and Marketed Share Plot type All Output value per acrea Male labor per acreb Female labor per acre Total labor per acrec Propn zero female labor Propn zero male labor Marketed shared Area (acres) Weather shocke Male wagesf No plots No parcels Percent of total 167.3 Mono 167.9 23.1 27.1 53.1 60.7 100.6 115.8 0.053 0.075 0.035 0.289 1.035 0.331 2.013 15,911 0.050 0.277 0.919 0.255 7,864 Parcel control Inter 166.6 All 174.1 194.0 19.1 45.6 85.7 0.031 24.7 48.0 96.9 0.050 22.2 35.4 73.4 0.075 0.020 0.302 1.147 0.406 0.032 0.315 1.310 0.357 0.015 0.371 1.552 0.351 Spouse 170.2 Marketed share Joint 153.3 50.6 3,496 42.3 None 121.2 10.2 35.8 24.3 59.6 55.9 69.6 92.4 126.4 123.3 0.018 0.039 0.046 0.071 0.214 1.052 0.376 0.031 0.308 1.178 0.352 8,047 8,274 49.4 Head 1,742 21.0 3,036 36.7 0.053 0.000 0.673 0.305 <Half ≥Half 192.7 207.0 19.5 46.6 87.8 0.030 25.6 35.7 81.2 0.090 0.020 0.255 1.094 0.371 0.027 0.777 1.518 0.324 6,514 5,164 4,233 40.9 32.5 26.6 Note: The unit of observation is a plot or a parcel. Superscripts indicate the following. a Output value in ,000’s Uganda Shillings. b Labor inputs are measured in person days per year. c Total labor is the aggregate of domestic, hired, and other labor inputs. d Marketed share is defined as the ratio of revenue to output value. e Dummy for whether at least one crop on the plot was affected by drought or floods. f Median, in ,000’s Uganda Shillings, per day; varies at the household-level only. inputs (see table 2). To estimate production functions by crop, we can only use monocropped plots, where, by definition, the crop coincides with the plot, which means we observe the labor inputs used in producing each crop, and the output produced. From this we can create a crop-level dataset by aggregating the data for each crop at the household level. Note that we still have to apply the above rule to share the output between intercropped and monocropped plots if a crop grown on a monocropped plot is also grown on an intercropped plot for a given household-season. This is the case for 49.9% of the household-crop observations. The resulting dataset comprises 22 different crops grown by 3,219 households, totaling 7,982 observations. Given the asymptotic properties of our test, we further restrict this dataset to crops with a minimum of 100 observations. This leaves a final sample of 11 crops grown by 3,128 households, totaling 7,152 crop observations (see table 3). We use the latter for testing whether the marginal rate of technical substitution is the same across the crops. We now detail the key variables used in our analysis. (table 2) or per crop (table 3). This is the quantity harvested (minus quantity wasted) times the producer price, both of which are observed. We use output value (in thousands of Ugandan shillings) rather than output (in kilograms) because we need to aggregate across crops on a given plot in our plot-level dataset. Notice that we do not use the term revenue, because not all output is sold, a distinction that becomes important below. It may be that the price of a crop depends on the gender of the crop manager. For instance, Hill and Vigneri (2011) find that access to coffee markets in Uganda crucially depends on the ownership of means of transportation, and the quantity sold, which differ by the gender of the household head. Once these differences are accounted for, there is no gender difference in accessing markets. In markets, female heads of households receive lower coffee prices than male household heads, but this difference seems due to the marketing channel chosen, rather than gender-based price discrimination. We discuss the possibility of gender-based price differentials in our empirical analysis. Labor Inputs Output The “dependent variable” is the annual value of the output harvested, either per plot Labor inputs are measured in person-days per year. The second, third, and fourth rows of tables 2 and 3 report domestic adult (i.e., Andrews, Golan, and Lay Table 3. Sample Averages by Crop Coffee Cotton Maize Mtooke Grdnts Beans Sorghm Peas Millet Cassva Pots 146.3 27.1 61.2 116.7 0.054 0.068 0.238 1.070 0.263 2.010 7,152 178.0 29.3 22.9 63.2 0.222 0.044 0.940 1.824 0.153 101.5 51.8 44.4 121.5 0.053 0.053 0.943 1.407 0.228 128.8 19.2 30.7 69.0 0.059 0.041 0.399 1.643 0.373 177.4 15.8 24.8 58.0 0.112 0.046 0.215 1.514 0.192 156.5 25.5 65.0 119.0 0.042 0.087 0.223 0.702 0.363 106.3 15.2 44.9 76.9 0.030 0.103 0.215 0.809 0.369 101.6 17.2 45.0 87.3 0.036 0.058 0.185 1.059 0.341 76.1 13.8 38.8 67.2 0.031 0.145 0.178 0.684 0.313 116.9 30.7 79.9 141.0 0.019 0.178 0.148 0.730 0.245 141.4 66.6 93.7 223.0 0.057 0.048 0.134 1.119 0.122 181.1 17.8 91.2 136.8 0.022 0.068 0.071 0.627 0.247 275 189 761 427 667 414 1109 131 376 1044 Note: The unit of observation is a household crop. This table refers to the 11 crops for which there are at least 100 observations. Table columns are ordered by marketed share. Superscripts indicate the following. a Output value in ,000’s Uganda Shillings. b Labor inputs are measured in person days per year. c Total labor is the aggregate of domestic, hired, and other labor inputs. d Marketed share is defined as the ratio of revenue to output value. e Dummy for whether at least one crop on the plot was affected by drought or floods. f Median, in ,000’s Uganda Shillings, per day; varies at the household-level only. 1759 Inefficiency of Male and Female Labor Supply in Agricultural Households Output value per acrea Male labor per acreb Female labor per acre Total labor per acrec Propn zero female labor Propn zero male labor Marketed shared Area (acres) Weather shocke Male wagesf No. crops All 1007 1008 April 2015 18 years or older) male, female, and total labor inputs, per acre. On average, women supply about half the total household annual hours (53.1%), twice that of men (23.1%). In addition to domestic adult female and male labor, total labor also includes domestic child labor, hired labor, and other labor inputs. The information on child labor is not available by gender. Also, 5.3% of plots have no female labor and 3.5% have no male labor. This very much depends on which crops are being cultivated. In table 3, among those crops with more than 100 observations, cassava stands out as the most labor-intensive crop, while coffee, maize, matooke, and peas require considerably less labor. The table confirms that adult female labor supply is substantial. For most crops, this is about half of total labor input, and is typically about two to three times higher than adult male labor input. This is not true, however, for traditional cash crops such as coffee and cotton, where, on average, the male input exceeds that of females. This division of work between crops is also reflected in the next two rows of table 3, which report the proportion of observations where there is no male or no female labor input. There is no female labor on 22.2% of coffee plots (4.4% for males), whereas only 1.9% of the millet plots have no female input (17.8% for males). To summarize, as is well-known, whilst women supply roughly twice as much labor as men, this varies considerably by crops, as does the proportion of crops wholly farmed by one partner or the other. Wages The left-hand side of our regression models is log(1/sB ) ≡ log(pf yf /wB fB ), the hypothetical (inverse) share of revenue earned in producing yf that would be paid to partner B for supplying fB hours (see equation (12)). The final component of log(1/sB ) that we discuss is the wage of the male in the household, wB . There are two sources of information in the survey. The first is constructed from all individuals who report having worked for pay during the previous 12 months. Using information on the last cash payment and the estimated cash value of in-kind payments, we create a daily earnings measure. In our sample of male household heads and female spouses, only 28% of individuals report the information we require, and the majority of Amer. J. Agr. Econ. these (78%) are men. This is one reason why we do not model partner A’s income share in the value of output log(1/sA ). The second variable is the agricultural community person-day wage, recorded in the community module of the survey, and is observed for everybody in our sample. We use this variable to impute wages for those individuals not reporting a wage using a standard Mincer-style regression, additionally controlling for age, gender, years of schooling, a dummy variable if the individual is responsible for a non-farm enterprise, regional fixedeffects, and a dummy variable for urban areas. We also include interaction terms between the gender dummy, the business dummy, and the community wage variable. Using these imputed daily wages, we find that female median earnings are 38% lower than male median earnings. We also find that this mark-down does not vary much by household, only at the community level. In the absence of household variation in wA /wB , it is meaningless to use both first-order conditions in equation (5); this is the second reason why we do not model log(1/sA ). Adverse Shocks Adverse shocks are endemic in agriculture, and can have sizeable effects on output. Our data allow us to control for this, and so we construct a set of dummy variables for such shocks, using the following question: “Before it was harvested, what was the main cause of crop damage?” The possible responses are drought, flood, disease, insects, animals, birds, theft, or some other cause. In the crop dataset, we create dummies for each cause, and assign the value of one if affected in either season. In the plot dataset, the dummies adopt the value of one if at least one crop on a given plot was affected by the said cause. For example, 31.9% of plots were affected by drought, 14.0% were affected by disease, 7.7% were affected by insects, and 5.0% by animals. Because they refer to the crop, these dummies are included in all our regressions. Including the dummies should affect the male and female labor coefficients if the impact of shocks has an important gender dimension. Finally, using the dummies for drought and flood, we create a dummy for adverse weather, which is included in table 2 and table 3; we do this to illustrate the extent to which plots and crops were affected by adverse weather shocks. Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households Gender Control of Crops and Plots As already discussed, table 3 demonstrates that the intra-household allocation of labor between men and women varies by crop. Such a gender division of labor can be efficient if it is made according to comparative advantage and compatible with the production technology employed. If the division of labor reflects the distribution of crop output between partners, and one partner feels inadequately compensated for the labor supplied for crops that are controlled by their spouse, they may either under-allocate labor or supply less effort. Alternatively, if one partner has the power to dictate intra-household labor allocation, too much labor may be supplied to this partner’s crops. To investigate whether any deviations from an efficient labor allocation system are related to who controls the output, we use the following question asked of households (our emphasis added): “Who mainly manages or controls the output from this parcel among the household members?”. From the replies to this question we construct three categories to record whether output is controlled by a) the head only, b) the spouse only, or c) by both and/or other household members and/or other people (see the middle panel of table 2). Just because one partner (typically the man) controls output on some or all of the parcels on a farm does not mean that the allocation of factor inputs on the farm is inefficient, as the theory allows for all-powerful men or women. However, the literature for SSA agriculture cited above suggests that inefficiencies may arise when partners feel inadequately compensated for their labor inputs. A total of 36.7% of the parcels are jointly managed or managed by someone else; of the rest, men exert control over twice the number of parcels as women (our sample only consists of male-headed households with at least one female spouse). Table 2 also shows that much less male labor per acre is used on their spouse-controlled parcels, and vice versa; in fact, men spend about twice as much time on their own parcels compared with their spouses’ parcels, whereas their spouses spend six times more time on their own than on their husbands’ parcels. The table also shows that female spousecontrolled parcels are much smaller (by 32%) and have a lower output value than male-controlled parcels (by 12%). For similar 1009 findings for Malawi, see Kilic, Palacios-Lopez, and Goldstein (2013). We accordingly stratify the sample to determine whether our test of inefficiencies is related to gender roles in the household. In addition, we are interested in examining whether the role of market integration possibly causes such inefficiencies, as conflicts over labor allocation are likely to arise once a certain proportion of the household’s output is traded in the market. As we observe the revenue from a given plot when its output is sold on the market, we define a marketed share variable, r, as the ratio of revenue to output value. Table 2 shows that this proportion does not vary between the monocropped and intercropped plots, but table 3 shows that it varies considerably over crops in the crop-level dataset. As expected, the share is higher for cash crops: 94% of the output value of all coffee produced is sold, whereas this figure is 7.1% for potatoes. The 11 crops we analyze are listed in order of marketed share in table 3. To capture the idea that men’s interests in controlling traditionally female-dominated crops will increase with the amount of output that is sold, we define the following three categories: no output is sold in the market (r = 0); “some” output is traded on the market (0 < r ≤ 0.50); and “lots” of output is traded on the market (r > 0.50). Table 2 shows that the average time that females spend per acre is about twice as high (69.6 days per year, per acre) when the market share is zero compared with when more than 50% of the output is sold in the market (35.7 days per year per acre). For men, there are no substantial differences. The same pattern occurs when we examine single food crops that traditionally fall under female control. Accordingly, table 2 shows that the marketed share and control variables are correlated, as expected. The proportion of output sold, r, is much larger for male-controlled parcels (37.1%) than female-controlled parcels (21.2%). When we regress r on two control dummies and a constant, this difference in means of 0.152 is highly significant with a household-cluster robust standard error of 0.0109. When parcels are jointly controlled, the marketed share is 30.8%. While these are interesting correlations, both endogeneity of the marketed share variable (higher output causes a higher marketed share) and simultaneity with labor allocation decisions preclude using it as an explanatory April 2015 variable in a production function. Nor can we stratify the sample by marketed shares. Nonetheless, these descriptive statistics and correlations are a strong indication that market integration and gender roles are interrelated. Households integrate into markets not only by selling more of a given crop portfolio, but also by adopting more marketable crops. We have shown that the marketed shares differ considerably between crops. This is also why we are interested in investigating the equality of the MRTS between male and female labor across crops within the household. We have pointed out that labor input choices and the marketed share of a specific crop are likely to be simultaneously determined. This is less of a problem if we compare the MRTS between crops, as one can reasonably assume that labor allocation decisions are being made once a specific crop choice has been made. Parcels A unique feature of the data is that it records information at the parcel level as well as for plots. A parcel is a contiguous piece of land with identical (uniform) tenure and physical characteristics and is entirely surrounded by land with other tenure and/or physical characteristics or infrastructure, for example, water, a road, forest, etc., that does not form part of the holding.7 This implies that a parcel is part of a holding that is physically separate from other parts of the holding; a holding is made up of one or more parcels. To assess the extent to which parcels differ from plots, we note that the number of plots per parcel is Poisson distributed, with a mean of 1.72. First note that the distribution of the number of crops per parcel-season is very similar to that of the number of crops per plot-season. Similarly, when we compare the distribution of the number of crops per plot with the distribution of the average number of crops per plot over the 13,499 parcelseasons, we see that the standard deviation of the former—the total variation—is very similar to that of the latter—the betweenparcel variation—which means that there is very little within-parcel variation. Figure 1 illustrates why there is so much between variation: of 13,499 parcel-seasons, 5,734 have 7 A holding is a household in our terminology. Amer. J. Agr. Econ. 6000 4000 Frequency 1010 2000 0 0 2 4 6 8 avge no crops per parcel-season 10 Figure 1. Distribution of average number of crops per plot by parcel exactly one crop on every plot (these are the monocropped plots), whereas another 3,494 parcel-seasons have an average of exactly two crops per plot, and 1,210 parcel-seasons have an average of exactly 3 crops per plot. In other words, there is considerable variation in how households use the plots within a given parcel in terms of how many crops they grow.8 This information at the parcel-level allows us to do two things. First, we observe who controls each parcel, namely the male, the spouse, or whether the parcel is jointly controlled. We can then investigate which control regimes are more productive, and then test whether the MRTS between female and male labor is equated across plots under different control regimes (head, spouse, joint control). Second, in both the plot and crop datasets, we have repeated observations on parcels and are able to control for parcel fixed-effects in our regressions. In other words, using parcel fixed-effects allows us to control for soil fertility characteristics that potentially determine which plots and crops are farmed by men rather than women in the same household. Specification and Econometric Issues In this section we specify the empirical model. Each household h owns a number 8 Crop diversification is a typical feature of poor rural households. On the one hand, diversification results from optimal portfolio choices of crops given differences in expected returns and risks. On the other hand, it is well-established that agricultural households simultaneously decide on production and consumption in the presence of different combinations of market imperfections. This implies that crop choice is determined by shadow prices that reflect the household’s factor endowments, technology, and preferences. Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households of parcels, denoted a, that are sub-divided into plots, denoted p. For the plot dataset, our estimating equation is written: (16) yp = x p β + η a + v p where implicitly p = p(a(h)). This is standard notation for a nested data structure such as ours, and indicates that p varies as parcels vary in the data, which themselves vary as households vary. Comparing this with equation (12), it is clear that yp replaces log(1/sB ) ≡ log(pf yf /wB fB ). Similarly, the vector xp contains log fA and log fB , the labor inputs of partners A and B, g(fA , fB ), the CES variable defined in equation (13) above, other labor inputs (including domestic children and hired labor), and any remaining control variables. Even though there is a γf Lf term in equation (8), because we assume that the production function is additively separable in Lf , it does not appear in the first-order conditions, and hence the model being estimated. This is clearly true of the plot acreage. However, we do not drop all the other labor inputs because they could have been specified as part of a larger aggregated labor input in the CES production function. We choose not to do this because we are not interested in estimating these other substitution elasticities, but they should be added as controls nonetheless. Further, ηa is a parcel-level fixed-effect that controls for soil-quality or the slope/steepness of the cultivation area, while vp is the idiosyncratic error term. In some regressions, we replace ηa by ηh , a household fixed-effect. From the estimates on log fA , log fB , and g, we compute δA , δB , and ρ, and then, using the sample averages for fA and fB , we compute the MRTS parameter θ (see equation (14) above). To test for the equality of the MRTS across plots, we then stratify the sample using our gender control variable. In the crop dataset, we focus on monocropped plots only, and then compute household-crop-level aggregates for all variables in the data. We then examine each of the 11 crops—coffee, cotton, maize, matooke, groundnuts, beans, sorghum, peas, millet, cassava, and potatoes—separately. In effect, we replace p by hc: (17) yhc = x hc β + ηa + vhc , c = 1, . . . , 11. 1011 As above, we can control for ηa or ηh , or neither. To ensure that we estimate the same ηa or ηh across all crops, we pool the data and use crop-dummy interactions. Finally, we test the null hypothesis that the MRTS parameter is the same for all 11 crops. A potential problem arises because output is recorded for each crop, but not each plot. As described in the appendix, for a given crop grown across different plots in the same household, we assume that the output per acre is the same for all plots cultivating the same crop. This is a possible concern if this assumption is violated because some plots are more productive than others and it is these plots that males choose to control. We control for this potential endogeneity in three different ways: using the method of Gandhi, Navarro, and Rivers (2009), specifying parcel fixed-effects, and using information on who actually controls each plot. This means there are no bias implications for our estimates because all that is left is standard measurement error in the dependent variable. Finally, for the same reason, we do not need to be concerned about the fact that the male wage rate is measured with error (this variable is derived from the community wage and other observable characteristics for some males in the sample). Results We first discuss our estimates of the underlying production function using the plot-level dataset; in the following sub-section, we examine the crop-level dataset. Evidence from Plot Data Following our discussion of inefficiency above, we first test whether the MRTS between female and male labor is equated across plots under different control regimes (head, spouse, joint control). In table 4, columns 2 and 3 report what happens when equation (16) is estimated controlling for household fixed effects ηh , and parcel fixedeffects, ηa , respectively. Column 1 reports ordinary least squares (OLS) estimates. There are three important features of table 4. In all our regressions, the elasticity of substitution parameter ρ is estimated slightly out of its legal range of ρ < −1. It is wrong to interpret the production functions as having linear isoquants (ρ = −1) because there are 1012 April 2015 Table 4. Plots Amer. J. Agr. Econ. Fixed Effects Estimates for All (1) OLS Female labor log fA Male labor log fB CES variable g(fA , fB ) Log child labor Log hired labor Log other labor Drought dummy Disease dummy Head control dummy (parcel variable) Spouse control dummy (parcel variable) 0.1142 (0.0113) −0.7233 (0.0108) −0.0526 (0.0037) 0.0423 (0.0068) 0.1335 (0.0080) 0.0719 (0.0088) −0.2162 (0.0279) −0.2452 (0.0383) −0.0123 (0.0373) (2) (3) Household Parcel 0.1375 (0.0110) −0.7329 (0.0098) −0.0494 (0.0033) 0.0426 (0.0071) 0.1090 (0.0060) 0.0396 (0.0071) −0.0585 (0.0249) −0.0706 (0.0325) 0.2198 (0.0696) −0.0857 −0.0640 (0.0520) (0.0691) 0.1692 (0.0132) −0.7406 (0.0124) −0.0505 (0.0041) 0.0479 (0.0092) 0.0982 (0.0081) 0.0401 (0.0100) −0.0613 (0.0311) −0.0288 (0.0396) · · · · Female share λA 0.0964 0.1258 0.1629 (0.0132) (0.0134) (0.0165) Male share λB 0.3891 0.3290 0.2872 (0.0297) (0.0246) (0.0259) Elast. param. ρ −1.1841 −1.0923 −1.0389 (0.0499) (0.0328) (0.0283) MRTSAB param.θ 0.2799 0.4064 0.5819 (0.0484) (0.0616) (0.0944) No. Household dummies No. Parcel dummies 3,665 8,274 Notes: Estimates of equation (12); household cluster standard errors appear in parentheses. Column (3) includes parcel fixed-effects; column (2) restricts these to household fixed-effects. All specifications include crop and shock dummies. The sample size is 15,911 plots. The parameters λA , λB , ρ, and θ are derived from the estimates above. See equation (14) of main text for how the MRTS parameter θ is estimated. very few observations where either the male or female exclusively works on the plot or crop (see table 3). It is better to interpret these estimates as meaning that one day of male labor input is being easily substituted by one day of female labor input; this is consistent with ρ being, for example, −0.9 (which lies inside the 95% confidence interval in most regressions). Rather than drop observations where either fA = 0 or fB = 0, we follow standard practice and add a small number of hours to fA or fB . Second, because the MRTS parameter θ is estimated less than unity, the data imply that men are more productive than women. In all three regressions, the parameter is significantly lower than unity. The third feature is that the estimate of θ increases from 0.280 when there are no fixed-effects, to 0.406 when we control for ηh , and to 0.582 when we control for ηa . This is because, as we move from left to right, the estimated female share parameter, λA —which is closely related to the marginal productivity of women—almost doubles, whilst that of men, λB , drops by one quarter. Said differently, the more we control for, the more productive are women, as the following correlations also illustrate: ηh ηa log fA log fB g(fA , fB ) ηh 1.0000 0.8731 1.0000 ηa −0.0441 −0.0980 1.0000 log fA −0.0194 0.0190 0.2086 1.0000 log fB g(fA , fB ) −0.0149 0.0248 0.0425 0.7166 1.0000 . The correlation between the parcel fixedeffect η̂a from column (3) and log fA is −0.098, but is 0.019 for men. When we compute the correlation between η̂h and log fA , it is smaller at −0.044.9 This is evidence that the parcel fixed-effects indeed capture parcel characteristics that matter for productivity, and so we are not purely picking up a household effect when taking parcel effects into account. Whilst it is often suspected that women cultivate less productive land, we believe that concrete econometric evidence for this finding is scarce. For example, the World Bank, FAO, and IFAD (2009) note that “Women—especially if they are the main providers of staple food crops—are particularly affected by declining soil fertility. Men often control the best land with the best soil to produce commercial crops, and women more often farm marginal land.” Two other features of our results are noteworthy. First, almost all of the adverse shock variables have a negative effect on output 9 Obviously η̂a and η̂h are highly correlated. Hausman tests in the two fixed-effects regressions confirm that the observed labor inputs are indeed strongly correlated with the estimates of ηa and ηh . Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households in the OLS regression. This is because the shocks are correlated with output in the pf yf component of sB , and not with labor inputs or wages. Table 4 reports that the estimate for drought is −0.216 and for disease it is −0.245, and both are significant and substantial. The table does not report the rest, but in fact, insects (−0.181), animals (−0.159), and stealing (−0.294) are also significant. In the other two regressions, the effects are much weaker, suggesting limited variation in shocks across plots within the same household or parcel. However, when we drop these dummies, in none of the three regressions do the estimates on log fA , log fB , or g(fA , fB ) alter at all, and so θ is unaffected too. This means that the shocks are not correlated with labor inputs, which in turn implies that there is no gender dimension to these observed shock adjustments. To see that there is no correlation between the weather shock dummy—either drought or flood—and who controls the parcel, see table 2. Second, we also include the parcel-level control variables discussed above. The variables are small and insignificant in column (1), but when we control for household fixed-effects, we find that when the parcel is controlled by males, such plots are 22% more productive than all other plots. This is consistent with our findings just reported. We now investigate this further; in table 5, we report what happens when column (3) of table 4 is stratified by who controls the parcel (see also table 2). The key finding is that θ is significantly smaller, at 0.212, when the spouse (female) controls the parcel compared with the head (male) controlling the parcel, where θ is estimated as 0.591. On jointly-controlled parcels the estimate is even higher at 0.708. This reinforces the finding above: not only is more female labor applied on low-productivity parcels, the marginal productivity of female labor is much lower and the marginal productivity of male labor is much higher on a female-controlled parcel. Recalling that there is much less labor input when females control parcels compared with male- and joint-controlled parcels, these results mean that households fail to allocate labor optimally across different parcels when these are under the control of different household members. In other words, total output could be higher—a Pareto improvement—if male labor were reallocated from male- to female-controlled parcels or vice versa. Using the estimates for the 1013 Table 5. Parcel Fixed-effects Estimates for All Plots, Stratified by Who Controls the Parcel (1) Head (2) Spouse (3) Joint 0.1824 (0.0204) −0.7306 (0.0192) −0.0574 (0.0069) 0.0414 (0.0157) 0.1026 (0.0138) 0.0433 (0.0168) −0.0419 (0.0497) −0.0778 (0.0623) 0.0780 (0.0377) −0.7098 (0.0327) −0.0689 (0.0094) 0.0741 (0.0197) 0.0809 (0.0190) 0.0597 (0.0246) −0.0080 (0.0748) −0.0815 (0.0841) 0.1664 (0.0196) −0.7574 (0.0179) −0.0389 (0.0061) 0.0458 (0.0138) 0.0935 (0.0110) 0.0350 (0.0139) −0.1020 (0.0469) 0.0569 (0.0623) 0.1745 (0.0249) Male share λB 0.3009 (0.0367) Elast. param. ρ −1.0451 (0.0403) MRTSAB param. θ 0.5909 (0.1356) No. of Parcel 3,496 dummies No. of Plots 6,752 0.0490 (0.0381) 0.5544 (0.1432) −1.5929 (0.4744) 0.2118 (0.0714) 1,742 0.1679 (0.0247) 0.2358 (0.0392) −0.9912 (0.0379) 0.7080 (0.1891) 3,036 3,148 6,011 Female labor log fA Male labor log fB CES variable g(fA , fB ) Log child labor Log hired labor Log other labor Drought dummy Disease dummy Female share λA Notes: See table 4 column (3), but re-estimated for head, spouse, and joint/other controlled parcels separately. “Head” and “Spouse” columns in table 5, if all the female labor (the 59.6 figure in table 2) were moved from spouse-controlled to head-controlled plots, total output from spouse-controlled and head-controlled plots would increase by 0.092 log-points. Similarly, if all the male labor (the 22.2 figure in table 2) were moved in the opposite direction, total output would increase by 0.189 log-points. These are sizeable and significant effects. The finding that the MRTS is highest on jointly-controlled parcels is in line with a recent study by Kazianga and Wahhaj (2013), who argue that in Burkina Faso jointlycontrolled plots serve for the household provision of public goods which, given social norms, enables joint plot managers to draw 1014 April 2015 more household resources towards its cultivation than individually-controlled plots can command. It may be that jointly-controlled plots overcome individual incentive problems in terms of labor supply and effort application. In addition, our results suggest strong productivity differences between head- and spouse-controlled plots. One explanation for our finding that θ varies by who controls the output is that men and women face different prices for trading identical crops; if so, this would explain why marginal products are not equated over the crops. To investigate this further, for each plot we construct a price-index for the output sold by dividing revenue per kilogram sold by the actual output sold, in kilograms. We then regress this variable on the three control dummies, controlling for crop and shock dummies and household fixed-effects. It turns out that we do not find any price differentials between spouse- and male- or jointly-controlled parcels. In fact, our estimated price differential between head-controlled and spouse-controlled is 25.8 Ugandan shillings in favor of spouses, with a robust standard error of 27.8 (the average kilogram price is 342.86 shillings). Unfortunately, we cannot explore the determinants of potential price differentials further because our data does not contain disaggregated information on the ownership of means of transport, access to networks, or other factors that may explain why and if gender-based price differentials exist. See also Hill & Vigneri (2011), who examine this issue for a sample of female heads of households. Before we turn to the crop-level dataset and test the equality of the MRTS across crops, we examine whether we can detect any systematic differences between monocropped and intercropped plots. This allows us to assess whether the crop-level dataset that is constructed using only monocropped plots is a representative sub-sample of the plot dataset. We therefore stratify the data into monocropped and intercropped sub-samples to examine whether any sample selection issues are likely to affect our results. For example, it is possible that intercropped plots are more likely to be farmed by women and to be food crops instead of cash crops like coffee or cotton. In fact, table 2 shows that monocropped plots have slightly higher factor input levels per acre than intercropped plots. However, the proportion of male labor out of total labor is Amer. J. Agr. Econ. almost the same for both sub-samples, as it is for female labor. Also, the average output value per acre is the same, on average, for both sub-samples. As parcel fixed-effects should deal with such potential selection effects, we simply re-estimate the model in table 4, column 3, to see whether the same results are obtained for the two plot-types (not reported). It turns out that the MRTS parameter θ is the same, estimated as 0.518 for monocropped plots and 0.682 for intercropped plots. The F-statistic for testing their equality is 1.53, with a p-value of 0.22. In other words, we conclude that sample selection is not driving our subsequent results using our crop-level dataset. Evidence from Crop Data We now analyze the 7,152 crop observations for 3,128 households by estimating equation (17) above (see table 6). The three blocks report estimates of the MRTS parameter, θ, by crop. At the top, we report OLS estimates; we also control for household fixed-effects (middle block) and for parcel fixed-effects (bottom block). For the OLS regression, the estimates of θ across the 11 crops are robust, considering the relatively low number of observations for each crop. Eight of the estimates lie between 0.28 and 1.07, that is, they are consistent with all earlier estimates of θ reported in the previous subsection. For three crops—beans, sorghum, and millet—θ is estimated above unity (the peas and millet estimates have particularly large standard errors). In general, the standard errors are much larger when we stratify by crops, which reflects small sample sizes and the non-linear nature of how θ is computed. This means that we are unable to reject the null hypothesis that θ is unity for nine crops. Our test of efficiency is whether these 11 estimates are the same as each other; the F-statistic is 0.84 (p-value = 0.46). Because the estimates for peas and millet have very high standard errors, we could exclude them from the test, but nothing changes: the F-statistic is 0.99 (p-value = 0.44). This constitutes evidence in favor of efficiency in the allocation of labor across the crops. The household and parcel fixed-effects estimates are similar, albeit slightly more variable, and usually higher. For cotton, maize, matooke, groundnuts, millet, and cassava, the results are fairly robust. For the rest, the standard Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households 1015 Table 6. Estimates of MRTS Parameter θ, by Crop Coffee Cotton Maize Mtooke Grdnts Beans Sorghm Peas Millet Cassva Pots F-stata OLS 0.297 0.925 0.667 1.067 0.501 1.544 1.302 0.280 1.568 0.825 0.873 0.84 (0.200) (0.565) (0.153) (0.290) (0.414) (0.724) (0.886) (2.476) (1.904) (0.230) (0.312) [0.59] Household FEb 0.687 1.327 0.804 1.352 0.457 1.233 6.597 1.288 0.512 0.835 1.965 0.88 (0.200) (1.231) (0.199) (0.385) (0.207) (0.890) (13.195) (1.452) (0.287) (0.232) (0.714) [0.55] Parcel FEc 1.371 0.607 0.690 1.106 0.460 4.185 3.613 1.892 0.795 1.379 4.583 0.67 (0.475) (0.903) (0.263) (0.432) (0.413) (6.066) (3.603) (2.118) (0.418) (0.501) (4.142) [0.75] 275 189 1109 761 427 667 414 131 376 1044 1759 Notes: This table refers to the 11 crops for which there are at least 100 observations. Fixed effects estimates of equation (17), household-cluster robust standard errors in parentheses. All models estimated as one equation, so that fixed-effects constrained equal across the columns. In all 33 cases, we cannot reject the hypothesis that ρ = −0.9 (that is, the two factor inputs are highly substitutable for each other); θ is computed imposing this restriction. Adverse shock dummies are included. There are 7,152 household-crop observations. Table columns are ordered by “marketed share.” Superscript a indicates F-statistics in the first block have a F(10, 3127) distribution under the null; the second block they have a F(10, 3127) distribution; and in the third block they have a F(10, 5065) distribution; P-values in brackets. b indicates 3,128 household fixed-effects. c indicates 5,066 parcel fixed-effects. errors increase appreciably. The F-statistics are now 0.88 (p-value = 0.55) for householdfixed effects and 0.67 (p-value = 0.75) for parcel-fixed effects, and so lead to the same outcome. The OLS test statistic is the most reliable, and OLS provides the most precise estimates of the three blocks. Using the OLS estimates, we now see what happens when we explicitly control for who controls the output. For this purpose, we construct a household-level version of the parcel-level control variable analyzed earlier, and use interactions in the regression to generate estimates for “head,” “spouse,” and “joint.”10 We do this for each crop; the estimates are reported in table 7. We have already established that labor inputs are correlated with who controls the plot, and that θ falls when it is the female spouse in control. Recall that our estimates of θ from table 5 are 0.59 if the parcel is head-controlled, 0.21 if spouse-controlled, and 0.71 for joint control. In addition to corroborating our previous finding of inefficiencies, the aim here is to also determine whether there are any variations across crops. We would, for example, be able to see whether cash crops—traditionally controlled my males—are more prone to inefficiencies than female-controlled food crops. It is worth reporting that the proportion of household-crop observations that are “joint” 10 The household-level control variable classifies households as head-, spouse-, or jointly-controlled if all the parcels for given household are either controlled by the head or the spouse; otherwise, a household is categorized as jointly-controlled. is constant over the 11 crops, but that some crops are less likely to be spouse-controlled than others. For example, for coffee and cotton, the proportions are 14% and 11%; these are double for millet, beans, and peas. Note that the rows of the table are ordered by the marketed share variable, which is why the traditional cash crops of coffee and cotton are in the left-most columns. There are only three crops with significant differences in male and female labor input coefficients across the three control regimes: coffee, cassava, and potatoes. For coffee, θ varies considerably across head-controlled, spouse-controlled, and jointly-controlled crops. Table 3 shows that coffee is typically controlled by the head of the household and female labor input is relatively low. For headcontrolled coffee, θ is negative, which is due to λA being negative. It is interesting that on female-controlled coffee crops, female labor input is relatively productive, but there are very few such cases. Cassava and potatoes are traditional food crops that are mostly consumed within the household. Still, the market share for cassava is higher than for potatoes and there are relatively few spouse-controlled cassava crops (see table 3). Also, male labor input is higher on cassava relative to potato crops. Interestingly, for both crops, θ is highest on jointly-controlled crops, which might reflect that most of the proceeds from cassava and potato production are used for the provision of household consumption. The low MRTS on spouse-controlled compared to headcontrolled cassava is surprising; it is driven 1016 April 2015 Amer. J. Agr. Econ. Table 7. OLS Estimates of MRTS Parameter θ, by Crop and by Who Controls the Parcel Coffee Cotton Maize Mtooke Grdnts Beans Sorghm Head control −0.083 (0.275) 0.51 Spouse control 1.336 (0.842) 0.14 Joint control 0.050 (0.231) 0.36 Obs 275 p−valuea [0.03] Peas Millet Cassva Pots 0.768 0.796 0.711 0.432 1.920 1.786 −1.296 1.189 1.143 (0.843) (0.268) (0.289) (0.416) (1.177) (1.810) (1.813) (1.379) (0.563) 0.57 0.50 0.41 0.36 0.33 0.39 0.33 0.34 0.44 0.929 (0.327) 0.37 1.183 0.369 0.891 0.423 1.062 0.669 −0.435 0.719 0.157 (3.996) (0.199) (0.435) (0.342) (0.707) (0.763) (6.119) (0.755) (0.130) 0.11 0.13 0.20 0.22 0.26 0.19 0.24 0.27 0.18 0.434 (0.149) 0.24 1.030 0.745 2.124 0.762 1.659 1.156 0.456 2.313 1.211 1.415 (0.682) (0.249) (1.125) (0.930) (0.779) (1.108) (1.485) (4.301) (0.470) (0.753) 0.32 0.37 0.39 0.42 0.41 0.43 0.43 0.39 0.38 0.39 189 1109 761 427 667 414 131 376 1044 1759 [0.53] [0.45] [0.46] [0.55] [0.09] [0.96] [0.41] [0.58] [0.0001] [0.0003] Notes: See tablenotes to table 6. The estimates on log fA and log fB are interacted with the 3 control dummies, thereby generating 3 different estimates of θ. The third row in each block is the proportion of observations controlled. Superscript a denotes 4 degrees of freedom, testing whether the 6 parameters can be restricted to 2, associated with the variables log fA , log fB . by λA being higher on head-controlled than on spouse-controlled cassava. An explanation for this could be that women tend to oversupply labor to the crop, hence yielding low returns. In contrast, for potatoes, men seem to under-supply labor to spouse-controlled crops, with θ being higher under head control, and λA being similar under both control regimes, while λB is higher under spouse control. The results by crop are hence partially consistent with inefficient allocations across crops. We observe inefficient labor allocations, in particular for coffee, cassava, and potatoes. These again imply that output could be increased by labor reallocations and/or changes in control regimes. Conclusion This article analyzes the efficiency of female and male labor allocation in agricultural households using data from the Uganda National Household Survey 2005/06. In particular, in Sub-Saharan Africa, previous empirical evidence hints at the possibility of sub-optimal labor allocations due to traditional gender roles in agriculture. These gender roles may limit the flexibility of labor allocation in agricultural production. In addition, gender roles also govern the control of the proceeds of production, which constitutes an element of intra-household compensation mechanisms. We provide evidence that farm households indeed fail to efficiently allocate female and male labor in agricultural production. We test whether the marginal rate of technical substitution between female and male labor inputs is equated over different agricultural production activities. This optimality condition does not hold when we compare male- and female-controlled plots. The findings imply that total farm output could be increased by reallocating male labor to female-controlled plots, or vice versa. An alternative explanation for our finding is that men and women face different prices for identical crops, which would also imply that the marginal products are not equated over the crops. Partners may, for example, face different (transaction) costs for trading identical crops. This might be particularly relevant once an individual controls or manages the output of a certain crop. While we do not find gender-based price differentials to explain our results, we find that women operate on less productive parcels of the farm. In addition, the inefficient intra-household allocation of labor results in a severe lack of male labor on female-controlled parcels. These findings are confirmed by production function estimates for single crops. For crops with significant labor input differences across the different control regimes, we also find that households do not equate the MRTS of male and female labor for the same crop when output is controlled by different members of the household. The finding that inefficiencies arise when comparing plots controlled by different individuals in the households implies that the household’s intra-household compensation Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households rules do not yield Pareto-optimal outcomes. Control over output can be thought of as a constituent component of such compensation rules and our results suggest that, in many agricultural households in Uganda, these rules do not seem to generate adequate incentives. It is important to note that our findings are not driven by different opportunity costs of male and female labor. These may be due to male-female earnings differentials on labor markets or different responsibilities within the household, such as caring for children or the elderly. All these factors may affect the relative productivity of males and females. Despite the resulting relative productivity differences, households should still allocate labor between different agricultural activities so as to maximize output. Finally, possible inefficiencies are often considered in the context of increasing market integration of farmers and changes in relative prices. There is anecdotal evidence that traditional norms and gender roles may not respond quickly enough to new incentives, and therefore lead to sub-optimal outcomes in the presence of such changes. Our findings are not generally supportive of such a view. In fact, more market integration does not appear to be related to higher inefficiencies in allocating male and female labor. In light of previous findings from less market-integrated households with large inefficiencies, our results may also be taken as a sign that market integration is not necessarily related to less cooperation in production, given that a considerable fraction of output is jointly controlled for crops that are mainly traded in markets. Yet our evidence for these relationships is at best indicative. In our view, assessing the causal relationship between gender roles, market integration, and productivity therefore constitutes an interesting avenue for future research. References Akresh, R. 2005. Understanding Pareto Inefficient Intrahousehold Allocations. IZA Discussion Paper No. 1858, Institute for the Study of Labor, Bonn. Apps, P., and R. Rees. 1997. Collective Labor Supply and Household Production. The Journal of Political Economy 105 (1): 178–90. 1017 Boserup, E. 1976. The Traditional Division of Work between the Sexes, a Source of Inequality. International Institute for Labor Studies, Geneva. Bourguignon, F., M. Browning, P.A. Chiappori, and V. Lechene. 1994. Income and Outcomes: A Structural Model of Intrahousehold Allocation. Journal of Political Economy 102 (6): 1067–96. Browning, M., and P.A. Chiappori. 1998. Efficient Intra-household Allocations: A General Characterization and Empirical Tests. Econometrica 66 (6): 1241–78. Browning, M., and M. Gørtz. 2012. Spending Time and Money within the Household. Scandinavian Journal of Economics 114 (3): 681–704. Chiappori, P.A. 1992. Collective Labor Supply and Welfare. Journal of Political Economy 100 (3): 437–67. ———. 1997. Introducing Household Production in Collective Models of Labor Supply. Journal of Political Economy 105 (1): 191–209. Dey, J. 1981. The Gambian Women: Unequal Partners in Rice Development Projects? Journal of Development Studies 17 (3): 109–22. Dolan, C. 2001. The “Good Wife”: Struggles over Resources in the Kenyan Horticultural Sector. Journal of Development Studies 37 (3): 39–70. Donni, O. 2008. Labor Supply, Home Production, and Welfare Comparisons. Journal of Public Economics 92 (7): 1720–37. Doss, C. 2002. Men’s Crops? Women’s Crops? Gender Patterns of Cropping in Ghana. World Development 30 (11): 1987–2000. Doss, C., and M. Morris. 2001. How does Gender Affect the Adoption of Agricultural Innovations? The Case of Improved Maize Technology in Ghana. Agricultural Economics 25 (1): 27–39. Duflo, E., and C. Udry. 2004. Intrahousehold Resource Allocation in Côte d’Ivoire: Social Norms, Separate Accounts and Consumption Choices. Working Paper No. 10498, National Bureau of Economic Research. Cambridge, MA. Elson, D., and B. Evers. 1996. Uganda: Gender Aware Country Strategy Report. Unpublished, Mimeo. School of Social Sciences, University of Manchester, U.K. Evers, B., and B. Walters. 2001. The Model of a Gender-segregated Low-income 1018 April 2015 Economy Reconsidered: Evidence from Uganda. Review of Development Economics 5 (1): 76–88. Fafchamps, M. 1993. Sequential Labor Decisions under Uncertainty: An Estimable Household Model of West-African Farmers. Econometrica 61 (5): 1173–97. Food and Agriculture Organization of the United Nations (FAO), International Fund for Agricultural Development (IFAD), and International Labour Organization (ILO). 2010. Gender Dimensions of Agricultural and Rural Employment: Differentiated Pathways Out of Poverty. The Food and Agricultural Organization of the United Nations, the International Fund for Agricultural Development and the International Labor Office, Rome. Gandhi, A., S. Navarro, and D. Rivers. 2009. Identifying Production Functions Using Restrictions from Economic Theory. Unpublished, Mimeo. University of Wisconsin-Madison, Madison, USA. Goldstein, M., and C. Udry. 2008. The Profits of Power: Land Rights and Agricultural Investment in Ghana. The Journal of Political Economy 116 (6): 981–1022. Hazell, P., C. Poulton, S. Wiggins, and A. Dorward. 2010. The Future of Small Farms: Trajectories and Policy Priorities. World Development 38 (10): 1349–61. Jacoby, H. 1992. Productivity of Men and Women and the Sexual Division of Labor in Peasant Agriculture of the Peruvian Sierra. Journal of Development Economics 37 (1–2): 265–87. Jones, C. 1983. The Mobilization of Women’s Labor for Cash Crop Production: A Game Theoretic Approach. American Journal of Agricultural Economics 65 (5): 1049–54. Kasente, D. 1997. Agricultural Intensification Strategies, Women’s Workloads and Well-being in Uganda. Paper prepared for the workshop on Gender, Poverty and Well-being hosted by the United Nations Research Institute for Social Development, the United Nations Development Programme, and the Centre for Development Studies in Kerala, India. Kasente, D., M. Lockwood, J. Vivian, and A. Whitehead. 2000. Gender and the Expansion of Non-traditional Agricultural Exports in Uganda. UNRISD Occasional Paper, United Nations Research Institute for Social Development, Geneva. Amer. J. Agr. Econ. Kazianga, H., and Z. Wahhaj. 2013. Gender, Social Norms, and Household Production in Burkina Faso. Economic Development and Cultural Change 61 (3): 539–76. Key, N., E. Sadoulet, and A. De Janvry. 2000. Transactions Costs and Agricultural Household Supply Response. American Journal of Agricultural Economics 82 (2): 245–59. Kilic, T., A. Palacios-Lopez, and M. Goldstein. 2013. Caught in a Productivity Trap: A Distributional Perspective on Gender Differences in Malawian Agriculture. World Bank Policy Research Working Paper 6381, Washington DC, USA. Kmenta, J. 1967. On the Estimation of the CES Production Function. International Economic Review 8 (2): 180–89. Kyomuhendo, G., and M. McIntosh. 2006. Women, Work and Domestic Virtue in Uganda, 1900-2003. Ohio University Press. McPeak, J., and C. Doss. 2006. Are Household Production Decisions Cooperative? Evidence on Pastoral Migration and Milk Sales from Northern Kenya. American Journal of Agricultural Economics 88 (3): 525–41. Ministry of Gender, Labor, and Social Development. 2005. Key Gender, Poverty and Employment Dimensions of Agricultural and Rural Development in Uganda: A Case Study of Bushenyi and Mpigi Districts. Ministry of Gender, Labor and Social Development, Uganda, December. Quisumbing, A., and J. Maluccio. 2003. Resources at Marriage and Intrahousehold Allocation: Evidence from Bangladesh, Ethiopia, Indonesia, and South Africa. Oxford Bulletin of Economics and Statistics 65 (3): 283–328. Singh, I., L. Squire, and J. Strauss. 1986. Agricultural Household Models: Extensions, Applications, and Policy. Johns Hopkins University Press for the World Bank. Baltimore, Maryland. Taylor, J., and I. Adelman. 2003. Agricultural Household Models: Genesis, Evolution, and Extensions. Review of Economics of the Household 1 (1): 33–58. Thomas, D., and C.L. Chen. 1994. Income Shares and Shares of Income: Empirical Tests of Models of Household Resource Allocation. RAND Labor and Population Program Working Paper Series No. 94-08, RAND, Santa Monica, CA. Andrews, Golan, and Lay Inefficiency of Male and Female Labor Supply in Agricultural Households Udry, C. 1996. Gender, Agricultural Production, and the Theory of the Household. The Journal of Political Economy 104 (5): 1010–46. Vargas Hill, R., and M. Vigneri. 2011. Mainstreaming Gender Sensitivity in Cash Crop Market Supply Chains. Agricultural Development Economics (ESA) Working Paper No. 11-08, Food and Agriculture Organization of the United Nations. von Braun, J., and P. Webb. 1989. The Impact of New Crop Technology on the Agricultural Division of Labor in a West African Setting. Economic Development and Cultural Change 37 (3): 513–34. World Bank, Food and Agriculture Organization of the United Nations (FAO), and International Fund for Agricultural Development (IFAD). 2009. Gender in Agriculture Sourcebook. The World Bank: Washington DC. Appendix Measurement Error when Constructing Crop Output? A potential problem arises because, for a given household–season, output is recorded for each crop, but not each plot. In this appendix, we describe our crop-sharing rule. First write the production function for crop β β c as q = kAfAA fBB . Explicitly subscripting by c, kc is an unobserved crop-specific technical factor that captures the idea that k1 coffee-plants can be grown on an acre of 1019 land (because of the distance between plants etc.), and is different from k2 for potatoes, and so on. Next, we consider crop c grown by household h over a number of plots. For each plot, we observe the area on which the crop is grown, Apc , but we only observe crop output, qc , for the whole household. We assume that output per acre is the same for all plots, which means we can compute q̂p = qc Apc Ac where Ac ≡ p Apc . Note that out of 12,288 observations on crop output, 49.9% were created using this method (note that there are more crop observations than in table 3, precisely because some households grow the same crop on different plots). The potential problem is that kc might not be constant from plot to plot within a given household. Of the 6,129 observations that were imputed, 36.3% come from the same parcel, where it is reasonable to assume that kc is constant (because parcels have the same soil characteristics, etc.). Of the remaining 3,902 observations, suppose that some plots are better than others in the sense that one can produce more coffee output per acre, then it is possible that males may choose to control those plots rather than poorer plots. However, also note that there are only 408 households (out of 3,665) where the control variable varies within a household across different parcels. None of this is an issue once our econometric procedures control for the endogeneity of labor inputs, as discussed in the main text.
© Copyright 2026 Paperzz