The effect of gender targeting of food transfers on child nutritional status: Experimental evidence from the Bolivian Amazon

Abstract

Some research suggests women are more likely to allocate additional resources to their children than are men. This perception has influenced policies such as in-kind food transfer programs and cash transfer programs, which often target women recipients. We assess whether targeting in-kind rice transfers to female versus male adult household members has a differential impact on children’s short-run nutritional status. We estimate the impacts of transfers of edible rice and rice seeds, randomly allocated to female or male adults, on three anthropometric indicators: BMI-for-age, arm-muscle area, and triceps skinfold thickness. The trial includes 481 children aged 3–11 years in a horticultural-foraging society of native Amazonians in Bolivia. On average, the gender of the transfer recipient does not influence child anthropometric dimensions, possibly due to norms of cooperation and sharing within and between households. We find limited evidence of heterogeneity in impacts. Transfers to women help children who were growth stunted at baseline to partially catch-up to their better-nourished age-sex peers and help boys (but not girls) and children in higher-income households increase their BMI-for-age. The results of this research point to the importance of considering cultural context in determining if allocating food transfers according to gender are most effective.

1. Introduction

Many development programs target women as participants. Food transfer programs have been abundant for decades (Gentilini, 2007) and food transfers to families often target women (Ahmed, Quisumbing, Nasreen, Hoddinott, & Bryan, 2009; Harris-Fry et al., 2018; Hidrobo, Hoddinott, Peterman, Margolies, & Moreira, 2014). Conditional cash transfer programs, more recent key social programs in more than 60 nations in five continents (Honorati, Gentilini, & Yemtsov, 2015), also often target women (Baird, Ferreira, Özler, & Woolcock, 2013; Fiszbein & Schady, 2009; Molyneux, 2006). Targeting women is widely believed to benefit children more than targeting men, presumably from the socialization of women and from a biological predisposition of women to care more about the human capital of their offspring (Gettler, 2010; Trevathan, 2010). The belief about resources transferred to women benefitting children more than resources transferred to men is consistent with household economic models. These models predict that controlling resources allows women to operationalize their preferences, which favour children more than men’s preferences do (Almås, Armand, Attanasio, & Carneiro, 2018; de Groot, Palermo, Handa, Ragno, & Peterman, 2015; Schultz, 1990). This finds support in numerous observational studies (Haddad, Hoddinott, & Alderman, 1996; Piperata, Schmeer, Hadley, & Ritchie-Ewing, 2013; Schultz, 1990; Smith, Ramakrishnan, Haddad, & Martorell, 2003; Thomas, 1990).

However, there is scant evidence from randomized-controlled trials (RCTs) that exogenous allocation of resources between women and men improves children’s well-being more when women rather than men receive them (Yoong, Rabinovich, & Diepeveen, 2012). Understanding whether there are differences in child well-being after transfers of resources to women or men, and what the conditions are for those differences, has implications for the creation and implementation of culturally relevant and effective development programs. This paper contributes directly to filling this gap by using an RCT to estimate the impact of an in-kind food transfer program on child anthropometrics.

To the best of our knowledge, there are no available randomized controlled trials in which food transfers were allocated randomly between women and men. However, a few recent experimental studies provide evidence that the impacts of transferring income to women rather than to men on children’s well-being differ by setting. Cash income and in-kind food transfers are different. However, we draw on evidence from cash income transfers because of the lack of research on in-kind food transfers and because their impacts on the quantity and quality of food consumed are similar (Hidrobo et al., 2014). Three studies exploit random variation in the gender of the income transfer recipient, with contradictory results. Unconditional cash transfers for primary education in rural Morocco lead to large reductions in children’s dropout rates and increased enrolment. However, the effects do not differ by the sex of the household head who received the transfers (Benhassine, Devoto, Duflo, Dupas, & Pouliquen, 2015). On the contrary, Armand, Attanasio, Carneiro, and Lechene (2016) examine a nationwide conditional cash transfer program in Macedonia that randomized the recipients’ gender and find that transfers to women increase the share of household resources spent on food. Last, a two-year pilot study in rural Burkina Faso gave cash-transfers to mothers or fathers (Akresh, de Walque, & Kazianga, 2016). Compared to the control group, cash transfers to women adults did not improve children’s educational and health outcomes, but the study found some evidence that cash given to fathers improved young children’s health in a year with low rainfall, and lead to higher investments in livestock, cash crops, and housing.

Other experimental studies provide indirect evidence by estimating whether resource transfers increase women’s decision-making power in the household. Decision making is most studies’ main hypothesized pathway by which transfers to women benefit children more than transfers to men. Again, findings are mixed. Gitter and Barham (2008) use a RCT and find that conditional income transfers to women in Nicaragua have higher impacts on child school enrolment if the mother has more schooling, a proxy for women’s decision-making power in the household. On the other hand, Braido, Olinto, and Perrone (2012) exploit a natural experiment in Brazil and find no evidence that changes in consumption arising from cash transfer program participation are due to the gender of the transfer recipient. Similarly, an evaluation of the PROGRESA program in Mexico finds no difference between women and men in the use of cash from the program and of earned income (Handa, Peterman, Davis, & Stampini, 2009). The finding presumably reflects no changes in women’s decision-making power within households from the cash transfer programs, raising doubts about the main pathway by which transfers to women are believed to benefit children more than transfers to men. On balance, this limited evidence suggests that the belief that transferring resources to women produces additional benefits for children remains unproven.

Here we present the results of a RCT in which we randomize the gender of recipients of one-time edible rice and rice seeds transfers. Edible rice was allocated to all households in 13 villages, while rice seeds were allocated to all households in another 14 villages. Villages themselves were randomly assigned to the two transfers. Explicitly randomizing the recipients’ gender goes beyond estimating heterogeneity in the impacts of the program by gender because it eliminates correlations with unobserved traits of children and parents if successfully randomized. For example, statistically estimating heterogeneous responses by the gender of the recipient could confound the effects of the intervention with unobserved factors that determine who was present the day of the transfer. Our focus on transferring to a female or male adult, independent of headship, also sidesteps issues with defining the household head (Beegle & van de Walle, 2019).

The trial took place among Tsimane’, a horticultural-foraging society of native Amazonians in Bolivia (Undurraga, Behrman, Leonard, & Godoy, 2016). At the time of the trial, Tsimane’ were still only moderately engaged in markets, and cash was of limited use in the communities. We transferred rice seeds instead of another placebo because Tsimane’ agreed this was an adequate resource for the group that did not receive rice. Seeds were theoretically edible, but we show that households did not treat rice and seeds equally. We transferred rice because it is both fungible and consumable, and one of the most important staple and cash crops (Vadez et al., 2004; Zycherman, 2013). Additionally, relative to other staple crops of the region, notably manioc and plantains, rice is higher in protein and other nutrients. The large transfer of edible rice (58 kg per household on average) can improve children’s nutritional status through (1) increased energy consumption, (2) greater dietary diversity, (3) less use of children in farm work, and (4) increased food security. In this paper, we focus on whether gender targeting of such transfers affects three anthropometric indicators of the short-run nutritional status of children 3–11 years old (inclusive) about five months after the transfers. The three anthropometric indicators include BMI-for-age Z-scores, arm-muscle area, and triceps skinfold thickness.

2. Context: Tsimane’

Tsimane’ live mostly along the Maniqui and the Apere rivers, in the department of Beni in Bolivia. The 2012 national census puts the population at 16,958 (INE, 2015). Tsimane’ live in approximately 95 villages of at least eight households each (Undurraga, Cruz, & Godoy, 2015). Most Tsimane’ remained highly autarkic during the data collection period. A worldwide comparative study of 15 small-scale non-industrial rural societies ranks the Tsimane’ next-to-last in market exposure, with an average of 7% of total household calories bought in markets (Henrich et al., 2010). Tsimane’ earn monetary income through sale of forest and farm goods and occasional wage labour in logging camps or cattle ranches. Both monetary income and non-monetary savings are low (Undurraga, Zycherman, Yiu, TAPS Bolivia Study Team, & Godoy, 2014). In a longitudinal study of 13 villages (2002–2010) we found that about 40% of adults (age ≥ 16 years) during 2002–2007 had earned no monetary income during the two months before the interview.

Depending on their age, 30%−36% of Tsimane’ children are growth-stunted (sex-age standardized Z-score (HAZ)<−2) (Foster et al., 2005; Zhang et al., 2016), similar to other native Amazonian societies (Godoy et al., 2010). The high levels of child stunting probably reflect high infectious disease loads, diets limited in critical micronutrients (e.g., zinc iron), and food insecurity, rather than limited total food/energy availability (Blackwell et al., 2011; Brabec et al., 2018; McDade et al., 2007; Puentes et al., 2016; Tanner et al., 2009). Other anthropometric dimensions suggest that Tsimane’ children may not be experiencing low-calorie intake. Specifically, weight-for-height, body fatness, and indices of muscularity among Tsimane’ children all more closely approximate age-specific and sex-specific World Health Organization (WHO) reference values than height-for-age (Blackwell et al., 2016; Foster et al., 2005; Godoy, Reyes-García, Byron, Leonard, & Vadez, 2005).

Assessing the overall socioeconomic status of Tsimane’ women is complex because it requires assessing indicators from a society spanning a broad autarky-to-market continuum. Women lag behind men in monetary income, modern physical assets ownership, schooling, academic skills, and fluency in Spanish, the national language (Godoy et al., 2006). The modern asset values (for example, metal tools) of husbands were 2.5 times larger than that of wives in 2004, and cash earnings were six times larger (Godoy et al., 2006, p. 1525). After controlling for age, adult women had 1.3 fewer grades of schooling, 30 percentage points lower probabilities of speaking fluent Spanish, and lower formal math skills than adult men on average.

However, these disadvantages are partially offset by other indicators of status. Qualitative data suggest that women maintain social status equal to men within community and home life. Although women and men within a household own and keep physical assets separately, they pool resources in consumption, with entire families often eating directly from common pots at mealtimes (Zycherman, 2013). Several aspects of Tsimane’ life suggests gender equality in running households. For example, women and men pool labour resources for farming, and both have an equal say in food production. Besides, matrilocal post-marital residences mean that families settle in villages near the women’s close kin, who provide support and enhance their empowerment (Daillant, 2003). Women also produce shocdye’, a locally-fermented beverage that contributes to social gatherings and, thus, to households’ overall social status (Zycherman, 2015).

Tsimane’ have various norms for sharing, and frequently borrow each other’s assets and do communal work. However, research shows that sharing of food is not common enough that it would likey contaminate the random assignment and impact the results (Ellis, 1996; Undurraga, Zycherman, Yiu, TAPS Bolivia Study Team, & Godoy, 2014; Zycherman, 2013). This is borne out in our data, which show that only 3.7% of the amounts transferred were given to other households and 1.6% were bartered (89% of the rice was eaten by the household, and 87% of the seeds were planted; the rest was stolen/lost, was sold, was left with the household at the time of the follow-up survey, or seeds died).

3. Methods

3.1. Experimental design

To select villages for the trial, we exclude villages participating in other studies, too costly to reach, too small or unsafe, or that contained high proportions of other ethnic groups. This leaves 65 villages, of which we select 27 based on accessibility (Figure 1). At baseline, 8% of households in the sample included only one adult; these are excluded from the gender randomization and this paper’s analyses.

Figure 1.

Tsimane’ villages in the panel study and randomized controlled trial 2008–2009, department of Beni, Bolivia

The colors of the territory denote elevation; mamsl denotes meters above mean sea level. The square symbols and letters in each town are approximately proportional in size to town population. Several villages are located outside of the Tsimane’ territory, an administrative division that does not reflect the lands inhabited by the Tsimane’ (Reyes-Garcia et al., 2014).

The random assignment of rice and seeds transfers to female versus male adult household members is part of a trial to estimate the effects of unconditional in-kind transfers on adult and child health (Undurraga et al., 2016). The 27 villages selected for the trial were first randomly assigned into rice or seeds transfer groups. In the rice group, each of the 13 villages received 782 kg of edible rice divided equally among all village households. Hence, amounts of rice per person are inversely related to the number of households in a village and household size (mean: 58 kg/household; median: 52 kg; SD: 23 kg; range 30–131 kg). In the 14 villages forming the seeds group, each household received 5.9 kg of improved rice seeds. Seeds were edible but included the hull and were labelled as “seeds” when given to households. For most households, seeds were transferred after the planting season. Ultimately, almost all of the rice transferred was eaten, and almost all of the seeds transferred were planted. Consistent with the large transfers of edible rice, households in the rice group report in the follow-up survey having consumed more rice in the last seven days than households in the seeds group (12.0 kg and 8.3 kg, p<0.001).

Within both groups, a separate randomization determined whether the rice or seeds were given to an adult female or male in the household. We exploit this randomization in this paper. Table 1 shows that households and children assigned to the two groups were similar at baseline along a series of key characteristics (Appendix Table A1 also shows that the age composition of households was similar in both groups).

Table 1.

Randomization check: Comparison of child, household and village characteristics at baseline (2008).

Dependent variable:	1 if female adult received transfer; 0 if male adult received transfer
BMI-for-age Z-score	−0.011 (0.043)
Arm muscle area (cm2)	0.005 (0.009)
Triceps skinfold thickness (mm)	0.000 (0.014)
Boy (binary variable)	−0.041 (0.051)
Child’s age (years)	−0.008 (0.018)
Child’s schooling (grades)	−0.011 (0.033)
Household size	0.026 (0.017)
Rice received (kg)	−0.002 (0.002)
Father’s schooling (grades)	−0.016 (0.014)
Mother’s schooling (grades)	0.040 (0.025)
Area deforested by household (ha)	−0.044 (0.041)
Household monetary income in last 7 days (1000 Bolivianos)	−0.011 (0.060)
Village size (# households)	−0.005 (0.007)
Distance to nearest town or road (hours walking in dry season)	0.004 (0.008)
Constant	0.454* (0.233)
Observations	441
R-squared	0.057
F test that all variables are jointly = 0	0.88
Prob > F	0.581

Standard errors clustered by household in parentheses.

^***p<0.01,

^**p<0.05,

^*p<0.1.

BAZ=BMI-for-age Z-score; AMA=arm-muscle area (cm²); TST=triceps skinfold thickness (mm). Walking time is a village-level variable (N=27). Tsimane’ deforest for agricultural cultivation, so area deforested is a proxy for income from crop production. Household monetary income comes from wage labour and from the sale of goods.

The baseline survey happened during February-May 2008, the transfers during October 2008-January 2009, and the follow-up survey during February-May 2009.

3.2. Significance of the transfers

We transferred rice or rice seeds because of the importance of rice as both a staple and cash crop (Zycherman, 2013). With a dual purpose as food and as a tradeable item, rice or rice seeds transfers should improve children’s nutritional status through several paths. Total monetary expenditures on all foods do not change due to the rice or seeds transfers, suggesting that the transfers do not displace other foods. The p-value of a t-test of the difference in mean household expenditures on all foods in the last seven days, for rice and seeds households in the follow-up year (2009) is 0.189.

The rice transferred has a value of ~US$11/person, the equivalent of income earned over 12.4 days for an average Tsimane’. Annual panel data from 2002 to 2007 (Leonard & Godoy, 2008) suggest that mean weekly rice consumption per person is ~1 kg, so a person in an average household (6 people) would have enough rice for ten weeks. The monetary value of the improved rice seeds given to households in the seeds group was US$1.70/person. However, the perceived value of the improved rice seeds at delivery time may have been lower, as there was no market for improved rice seeds in the study area. In focus groups and interviews in February 2011, after the study had ended, Tsimane’ reported they did not like to eat the rice produced from improved rice seeds and did not perceive that the cash earning potential of harvested rice from improved seeds was large.

The edible rice transfers represent a substantial infusion of energy and protein. On average, the amounts transferred provide households with the equivalent of ten weeks’ worth of their baseline rice consumption. Assuming 10% wastage from the time of the transfer until the time of the follow-up survey (150 days), the average 58 kg transfers add 1,249 kcal/day and 25 g/day of protein to average household availabilities. Considering the nutritional needs of an average Tsimane’ child, even a small allocation of the household rice could improve child nutrition. We do not have data on the rice consumed by each household member, but for the average 58 kg rice transfers, a 10% allocation to five-year-old children would meet 10% of children’s daily energy requirements and 16–17% of daily protein needs. Households had an average of six members; even with an uneven distribution of rice between adults and children, a 10% allocation to each child is a reasonable estimate. Studies find that increasing early-life energy intakes, particularly in the form of proteins, has significant and substantial effects on child anthropometric measures such as weight and height (Adair et al., 2010; Habicht, Martorell, & Rivera, 1995; Puentes et al., 2016).

3.3. Anthropometric measures of short-run nutritional status

Because we measure the impacts of the transfers on children’s nutritional status about five months after the transfer, we focus on anthropometric indicators of short-run nutritional status. We collect children’s anthropometric data following the protocols of Lohman, Roche, and Martorell (1988) and Antón, Snodgrass, and Bones and Behavior Working Group (2009). We measure linear growth (stature/length; cm) using portable Seca 213 stadiometers. Bodyweight is measured to the nearest 0.1 kg using standing scales (Tanita Corp.). Mid upper-arm circumference (MUAC), used to derive arm muscle area (AMA), is measured to the nearest 0.1 cm using plastic tape measures, and triceps skinfold thickness (TST) is measured to the nearest mm with Lange skinfold callipers.

BMI-for-age Z-scores (BAZ) are used to assess thinness, overweight, and obesity (de Onis et al., 2012; Flegal, Wei, & Ogden, 2002), and calculated relative to WHO reference values (de Onis et al., 2007; World Health Organization, 2009). For under-nutrition, thinness is defined as BAZ<−2, while severe thinness is defined as BAZ<−3.

Arm-Muscle Area (AMA, cm²) captures muscular development and protein reserves (Saito et al., 2010). Arm muscle area is preferred over MUAC because it more accurately reflects changes in muscle tissue and is more strongly correlated with biochemical measures of protein status (e.g., creatinine excretion) (Gibson, 2005). AMA was calculated from MUAC and triceps skinfold thickness following the procedures in Frisancho (2008). As with low BAZ, low AMA is indicative of protein-energy malnutrition.

Triceps Skinfold Thickness (TST, mm) measures the thickness of subcutaneous adipose tissue and captures total body fat and energy reserves (Jebb, Murgatroyd, Goldberg, Prentice, & Coward, 1993; Pecoraro et al., 2003). High fat content is associated with high calorie intake or low energy expenditure (Frisancho, 1990). Muscle and fat measures are of primary interest because these dimensions are affected by nutritional disorders and can change in the short run (Briend, Garenne, Maire, Fontaine, & Dieng, 1989; Holliday, 1978).

3.4. Sample and analysis strategy

We limit analyses to children aged 3 to 11 years. We set the lower age at three years because Tsimane’ mothers breastfeed their children on demand for about two years (Brabec et al., 2018; Veile, Martin, McAllister, & Gurven, 2014). Thus, including children younger than three in the 2009 survey (two in the 2008 survey) would have increased age-related heterogeneity regarding possible children’s consumption of rice or other foods. We set the upper age to limit the effects of puberty on growth rates (Proos & Gustafsson, 2012; Walker et al., 2006). The final sample includes 27 villages, 196 households, and 481 children (girls=230, boys=251) (Table 2).

Table 2.

Descriptive statistics of children and villages at baseline (2008).

	N	Mean	Std. Dev.	Median	Min.	Max.
BAZ	481	0.44	0.76	0.42	−2.69	3.15
AMA	481	17.6	4.4	17.3	3.1	42.7
TST	481	7.2	2.3	7.0	3.0	20.0
Boy (%)	481	52	50	100	0	100
Age (years)	481	6.5	2.2	7.0	3.0	10.0
Walking time from village centre to nearest town or to road, dry season (hours)	27	5.0	6.0	2.4	0.1	24.0

BAZ=BMI-for-age Z-score; AMA=arm-muscle area (cm²); TST=triceps skinfold thickness (mm). Walking time is a village-level variable. The sample includes all children in households that received rice and seeds.

The analyses proceed in two steps. First, we use the following double-difference estimator to assess the effects of the rice transfers recipients’ gender on children’s nutritional status:

$$Y_{\text{aihvt}}=\mathrm{\alpha }+\mathrm{\beta }\text{*}Female{T}_{\text{hvt}}+\mathrm{\gamma }\text{*}Afte{r}_{\text{ihvt}}+\mathrm{\delta }\text{*}FemaleT*Afte{r}_{\text{ihvt}}+{\epsilon }_{\text{aihvt}}.$$ (1)

The subscripts stand for anthropometric indicators (a), child (i), household (h), village (v), and year (t). FemaleT is a binary variable equal to one if transfer recipients are women, and zero if they are men, After is an indicator for time (0=2008, baseline; 1=2009, follow-up), FemaleT*After is an interaction term, and ε is the error term. The coefficient δ is the difference-in-difference (DID) estimator for differential increases in anthropometric measures over time when female rather than male adults receive the transfers. We use ordinary least square regressions and cluster standard errors at the household level, as the random assignment of transfer recipients’ gender was at that level (Glennerster & Takavarasha, 2013). We also analyse our data with an augmented version of Equation (1) including a vector of control variables to increase precision: point estimates are unchanged (Appendix Tables A2-A3 mirror Tables 3–4), consistent with the successful random assignment of the transfer recipients’ gender (Table 1, Panel B).

Table 3.

Gender of transfer recipient and impacts of the rice transfers.

	(1)	(2)	(3)
Dependent variable:	BAZ	AMA	TST
Female recipient * After	−0.048 (0.090)	0.143 (0.673)	0.038 (0.372)
Female recipient	0.059 (0.123)	0.122 (0.553)	−0.351 (0.387)
After	−0.177*** (0.064)	1.199*** (0.351)	−0.209 (0.228)
Constant	0.490*** (0.087)	17.387*** (0.310)	7.484*** (0.255)
Observations	442	442	442
R-squared	0.016	0.020	0.007
Number of children at baseline	221	221	221
Mean of dep. var. at baseline	0.44	17.6	7.2
Effect size (SD)	−0.063	0.033	0.017

Standard errors clustered by household in parentheses.

^***p<0.01,

^**p<0.05,

^*p<0.1.

BAZ=BMI-for-age Z-score; AMA=arm-muscle area (cm²); TST=triceps skinfold thickness (mm). The sample includes only children in households that received rice. The effect size is indicated in units of standard deviations (SD) of each outcome, calculated by dividing the regression coefficient on the interaction term by the standard deviation of each outcome at baseline.

Table 4.

Gender of transfer recipient and impacts of the seeds transfers.

	(1)	(2)	(3)
Dependent variable:	BAZ	AMA	TST
Female recipient * After	0.012 (0.101)	0.009 (0.555)	−0.363 (0.351)
Female recipient	−0.004 (0.113)	−0.077 (0.613)	0.258 (0.329)
After	−0.239*** (0.063)	0.962** (0.388)	0.257 (0.249)
Constant	0.385*** (0.090)	17.788*** (0.461)	6.933*** (0.230)
Observations	520	520	520
R-squared	0.021	0.010	0.002
Number of children at baseline	260	260	260
Mean of dep. var. at baseline	0.44	17.6	7.2
Effect size (SD)	0.016	0.002	−0.160

Standard errors clustered by household in parentheses.

^***p<0.01,

^**p<0.05,

^*p<0.1.

BAZ=BMI-for-age Z-score; AMA=arm-muscle area (cm²); TST=triceps skinfold thickness (mm). The sample includes only children in households that received seeds. The effect size is indicated in units of standard deviations (SD) of each outcome, calculated by dividing the regression coefficient on the interaction term by the standard deviation of each outcome at baseline.

Second, we use the same equation to analyse anthropometric indicators of children living in households that receive improved seeds. The rationale for this analysis is that seeds may lead to changes in farm investments of parents, which, in turn, could affect children’s nutrition. For example, parents may focus on growing complementary crops with high nutritional value or increase total harvests by using fertilizer.

4. Results

We describe in turn the main results (section 4.1), robustness checks (section 4.2), and heterogeneous treatment effects (section 4.3).

4.1. Main results

Tables 3 and 4 show the main results. The tables contain parameter estimates for Equation [1] applied to children whose households received rice (Table 3) or seeds (Table 4). The difference-in-difference coefficients (δ) are very small and not statistically significant for all outcomes, whether we analyse the groups of rice or seeds recipients. When converted into units of standard deviations of each outcome at baseline (shown in the bottom row), regression coefficients indicate that transferring rice to and adult woman rather than man leads to 0.06 standard deviations lower BMI-for-age Z-scores, 0.03 standard deviations higher arm muscle area, and 0.02 standard deviations higher triceps skinfold thickness in children.

The estimates of impacts of transferring seeds to women rather than men are more varied, ranging from children’s lower triceps skinfold thickness (−0.16 standard deviations) to no change in arm muscle area (0.002 standard deviations) to higher BMI-for-age Z-scores (0.02 standard deviations). In effect, most point estimates are practically indistinguishable from zero in the magnitude of the effects, and lack statistical significance.

To help address issues with our small sample size (to which we return below), we also analyse the impacts of allocating transfers to female versus male adult in the household on the combined group of rice and seeds recipients. Results are similar to those of the separate analyses of rice and seeds transfers, with estimated coefficients ranging from -0.06 standard deviations to +0.01 standard deviations; none of these coefficients are statistically significant (Appendix Table A4).

4.2. Robustness tests

Here we show that our results are robust to three changes in the analysis. First, we re-estimate the Table 3 regressions with standard errors clustered at the village level (chosen because of the randomization of rice and seeds at the village level). In the main analyses we cluster at the household level because the exogenous variation we exploit is that of the gender of the adult household member who received the transfers (Glennerster & Takavarasha, 2013). As shown in Appendix Table A5, clustering standard errors by village tends to increase standard errors, and coefficients remain not statistically significant.¹ Second, we re-estimate our main regressions using hierarchical linear models allowing treatment to vary randomly among households in the village. All coefficients of interest remain very small and not statistically significant (Appendix Table A6). Finally, instead of a difference-in-difference estimator, we use an ANCOVA model and regress the outcome at follow-up against the treatment (FemaleT) and the outcome at baseline (McKenzie, 2012). Like all other analyses, coefficient estimates are very small in magnitude and not statistically significant (Appendix Table A7).

4.3. Heterogeneity in impacts

We assess whether the small magnitudes of estimated effects of edible rice transfers in the full sample hide larger but opposing effects in sub-samples (Table 5). We examine heterogeneity in impacts by children’s gender, age at baseline, nutrition status at baseline, and household income at baseline by including triple difference-in-difference estimates. For example, we interact all independent variables in Equation 1 with Boy to assess variation by child’s gender (itself given by the coefficient on the triple interaction term FemaleT*After*Boy). We find limited evidence that transferring income to female adult household members has heterogeneous impacts. The strongest result is that transfers to female adults help already-stunted children catch-up on their better-nourished peers. The coefficient estimates on the triple interaction terms FemaleT*After*Stunted are positive for all outcomes, large (they represent 0.22 to 0.64 units of standard deviations of the three outcomes at baseline), and statistically significant at the 5 percent level for BMI-for-age Z-scores and triceps skinfold thickness (p=0.014 and 0.023).

Table 5.

Heterogeneity in impacts of the gender of rice transfer recipient.

	(1)	(2)	(3)
Dependent variable:	BAZ	AMA	TST
Female recipient * After * Boy	0.494*** (0.161)	0.708 (0.989)	−0.708 (0.530)
Female recipient * After * Child age	−0.054* (0.031)	0.039 (0.298)	−0.184 (0.124)
Female recipient * After * Stunted	0.426** (0.171)	0.967 (1.182)	1.457** (0.626)
Female recipient * After * Bottom 20%	−0.579*** (0.155)	0.407 (1.334)	−1.245* (0.740)
Observations	442	442	442
Number of children at baseline	221	221	221
Mean of dep. var. at baseline	0.44	17.6	7.2
Effect size (SD) of interaction with
Boy	0.647	0.162	−0.312
Child age	−0.071	0.009	−0.081
Stunted	0.558	0.222	0.642
Bottom 20%	−0.759	0.093	−0.549

Standard errors clustered by household in parentheses.

^***p<0.01,

^**p<0.05,

^*p<0.1.

Each row reports coefficients from separate regressions. All combinations of the three interacted variables are included in the regressions but not shown for clarity (all coefficients are shown in Appendix Tables A9-A12). Child age is a continuous variable. Boy, Stunted, and Bottom 20% are binary variables. Bottom 20% is an indicator of living in a household at the bottom 20% of the village income distribution at baseline. BAZ=BMI-for-age Z-score; AMA=arm-muscle area (cm²); TST=triceps skinfold thickness (mm). The effect size is indicated in units of standard deviations (SD) of each outcome, calculated by dividing the regression coefficient on the triple interaction term by the standard deviation of each outcome at baseline.

The data also suggest that transferring income to female adults, compared to male adults, leads to two other large effects: 0.65 standard deviations higher BMI-for-age Z-scores for boys (p=0.003), and 0.76 standard deviations lower BMI-for-age Z-scores for poorer children (p<0.001). We do not find heterogeneity in impacts on other anthropometric indicators. In sum, the only consistent evidence of heterogeneous impacts is by child stunting at baseline (which transfers to women helped redress). Additional evidence points to possible benefits of transferring income to women on the BMI-for-age Z-scores of boys and children in wealthier households.

5. Discussion and conclusion

On average, in our setting, added food resources in the hands of female adult household members do not improve anthropometric indicators of children’s nutritional status more than added food resources in the hands of male adult members. In the full sample, impacts are small and not statistically significant. The results hold up to different analyses. We find limited evidence that impacts vary within the sample; in particular, we find stunted children grow faster when women received the transfer.

Although our sample size is small, our main estimated effect sizes are nearly zero, and standard errors indicate that possible effect sizes remain modest. 95% confidence intervals for the effect from our largest sample range from -0.28 to 0.19 standard deviations (calculated from coefficients in Appendix Table A4, and shown in Appendix Table A8). Minimum detectable effect sizes, shown in Appendix Table A8, range from 0.22 to 0.26 standard deviations for the impacts of rice transfers, and from 0.20 to 0.24 standard deviations for the impacts of seeds transfers. While these sizes are small by the standards of Cohen (1988), our estimated effects are much smaller. Even though we lack sample size to detect whether our estimated effect sizes are indeed statistically significantly different from zero, the very small size of the coefficients renders the point moot. Even if statistically significant, these small effects would lack socioeconomic and nutritional importance.

We provide two explanations for the lack of impacts and three policy implications. First, social norms among Tsimane’ support intra-household cooperation, resource pooling, and equal decision-making between female and male adult household members. Together, these features of Tsimane’ society may lead to similar decisions on the use of transfers by female or male adults. A similar result arises in Gitter and Barham (2008), who estimate that the benefit of directing conditional cash transfers to women diminishes as their years of schooling far exceeded those of their male partners, an indication that their decision-making power is high before the transfers. In our setting, women make household decisions on an equal footing with men. Tsimane’ horticulture requires cooperation. Men cut large trees for farming, but thereafter women and men work jointly, clearing underbrush, burning, planting, weeding, and harvesting. Both have equal rights to farm plots and equal say on end uses of the harvests, including bartering or selling crops. Besides, Godoy et al. (2006) find that Tsimane’ women see themselves as the primary decision-makers and tie-breakers when spouses cannot reach an agreement in various domains of the household economy, such as buying or selling goods or child schooling.

Second, Tsimane’ households may share many personal preferences. Tsimane’ are a tightly-knit, small-scale endogamic society. Evidence also suggests that Tsimane’ practice positive assortative mating for age, schooling, body size and type, ethnobotanical knowledge, and psychological traits (Godoy et al., 2008). These practices probably increase harmony between male and female adults in the household and decrease the instances of gender-oriented differences in determining how food is utilized in the household. As noted above, there is understanding between adults as to how much seed needs to be stored for sowing next year, what the current needs are for fungible rice, and how much is kept in the kitchen to be consumed. Rice is the second most common food in the Tsimane’ diet, and there is an appreciation for the preparation of rice in common dishes like Jona (or Arrosh/Arroz), which involves boiling rice with fish or meat (Zycherman 2013).

Our results have at least three policy implications. First, the very small impacts on child nutritional status from directing rice transfers to women rather than to men raise questions about the universal need to target women in transfer programs to improve child well-being. Targeting transfers to women may unnecessarily raise implementation costs or even create gender-based conflicts (Hidrobo & Fernald, 2013). Second, norms of reciprocity, family structure, marriage rules, and gender roles can erase or increase the impact of gender-targeted transfers on child well-being. Third, the absence of effects could reflect the size and type of transfer we made. Unlike our transfers, larger, more frequent, or conditional transfers could tip the balance of food allocation within households and produce visible impacts on child well-being.