Agricultural and empowerment pathways from land ownership to women's nutrition in India

Abstract Land size is an important equity concern for the design of ‘nutrition‐sensitive’ agricultural interventions. We unpack some of the pathways between land and nutrition using a cross‐sectional baseline survey data set of 4,480 women from 148 clusters from the ‘Upscaling Participatory Action and Videos for Agriculture and Nutrition’ trial in Keonjhar district in Odisha, India. Variables used are household ln‐land size owned (exposure) and maternal dietary diversity score out of 10 food groups and body mass index (BMI; kg/m2) (outcomes); and mediators investigated are production diversity score, value of agricultural production, and indicators for women's empowerment (decision‐making in agriculture, group participation, work‐free time and land ownership). We assessed mediation using a non‐parametric potential outcomes framework method. Land size positively affects maternal dietary diversity scores [β 0.047; 95% confidence interval (CI) (0.011, 0.082)] but not BMI. Production diversity, but not value of production, accounts for 17.6% of total effect mediated. We observe suppression of the effect of land size on BMI, with no evidence of a direct effect for either of the agricultural mediators but indirect effects of β −0.031 [95% CI (−0.048, −0.017)] through production diversity and β −0.047 [95% CI (−0.075, −0.021)] through value of production. An increase in land size positively affects women's decision‐making, which in turn negatively affects maternal BMI. The positive effect of work‐free time on maternal BMI is suppressed by the negative effect of household land size on work‐free time. Agriculture interventions must consider land quality, women's decision‐making and implications for women's workload in their design.


| INTRODUCTION
There is increasing evidence that agricultural interventions can be designed to improve diets in undernourished, low-income communities (Ruel, Quisumbing, & Balagamwala, 2018). Some of these 'nutrition-sensitive' agricultural interventions have improved dietary quality through crop biofortification (De Brauw et al., 2018) or diversifying crop and livestock production (Darrouzet-Nardi et al., 2016;Olney et al., 2016;Schreinemachers, Patalagsa, & Uddin, 2016), whereas others have used value-chain approaches and agricultural technology to increase household incomes (Alaofè, Burney, Naylor, & Taren, 2016;Le Port et al., 2017). Some interventions, such as sustainable intensification (Pretty & Hine, 2001), rely on participants' ownership of sizeable landholdings, whereas others, such as small livestock interventions and kitchen gardens (Darrouzet-Nardi et al., 2016;Olney et al., 2016;Schreinemachers et al., 2016), require smaller parcels of land. A common theme is that most, if not all, agricultural interventions rely on households to have access to some cultivable land.
Consequently, land size is an important equity concern for the design of agricultural interventions. Some evidence indicates that nutrition-sensitive agriculture interventions improve health outcomes to a greater extent in better-off groups (Jones & de Brauw, 2015;Le Port et al., 2017), suggesting that interventions need to be intentionally designed to be pro-poor. Historically, in many agrarian societies across the world, wealthier landowners have employed poorer, lower class or caste, landless groups to work on their land in a patron-client relationship that entrenched socio-economic inequalities (Cameron, 1995;Lawry, 1990;Scott, 1972). A wealth gradient in nutritional status is commonly observed in national surveys, and determinants analyses have commonly identified wealth as the strongest predictor of dietary quality (Aemro, Mesele, Birhanu, & Atenafu, 2013;Harris-Fry et al., 2015), indicating that these poorer, landless groups have poorer nutritional status (Arimond & Ruel, 2004).
Taken together, this suggests that nutrition-sensitive agriculture interventions may be both less relevant but also more needed in poor households with little or no land.
This heterogeneity may be because the linkages from agricultural production to consumption are highly mixed (Ruel et al., 2018).
Differences in market access and food storage facilities may explain this, with limited market access cornering households into consuming their own production and strengthening the production-consumption pathway in some places more than others (Hoddinott, Headey, & Dereje, 2015). On the other hand, estimates of effects of land on diets may be confounded by wealth, particularly because current evidence is based on observational data so has been unable to isolate causal effects (Shankar et al., 2019). It is likely that wealthier households can afford to cultivate or buy more land and can also afford more adequate diets. Moreover, the land tenure landscape is changing in many places, with land redistribution and titling programmes, migration to urban areas and increasing reliance on non-farm work in rural areas (Holden & Otsuka, 2014;Rigg, 2006). These processes could be weakening the links between land ownership and diets.
Differential effects of land on diets may also be explained by varied roles of women in agriculture. Women have historically had less access to land than men, particularly in patrilineal contexts where sons traditionally inherit land (Doss, Kovarik, Peterman, Quisumbing, & van den Bold, 2015). Perhaps as a result of this, many women also have less control over the use of land across the value chain: from decisionmaking over agricultural processes, processing, marketing and sale and use of agricultural outputs (Meinzen-Dick et al., 2012) • Although land could improve some agriculture and women's empowerment indicators, these may act as suppressors of maternal nutritional outcomes, especially BMI.
• Agriculture programmes aiming to increase household productive assets, such as land transfer programmes, must be designed to consider quality of the transferred land and access to agricultural inputs and their implications for women's and intra-household allocation of labour.
applying a robust causal inference framework to our mediation analysis.

| Study setting and population
The survey was conducted between November 2016 and January We sampled households containing at least one child aged 0-23 completed months and interviewed the female primary caregivers (aged 15-49 years) and their husbands (or other adult males, if the husband was not available). We aimed to sample 32 households per cluster, giving an intended sample size of 4736 mother-child dyads.
The sample size was calculated to detect differences in the two primary outcomes of the trial (percentage of children aged 6-23 months consuming at least four out of seven food groups per day and maternal mean BMI). We excluded any households where the mother or child had a discernible disability affecting their anthropometric measurements or ability to respond to the questionnaires. Data on household food expenditures were collected on a randomly selected 50% of the sample, using a Household Expenditure and Consumption Survey.
Informed consent was obtained from all respondents.

| Data collection
Nine teams of local, trained interviewers interviewed the respondents using a pretested questionnaire translated into Odia language. We trained interviewers for 3 weeks before the start of data collection.
Interviewers measured women's height using Seca 213 Stadiometers, weight using PLAX-Cruzer scales, and mid-upper arm circumference (MUAC) using Seca circumference tapes. We standardised anthropometric measurements by comparing against a 'gold standard' measurer and calculating inter-and intra-technical error of measurement. We also standardised dietary diversity assessments, by asking each interviewer and a gold standard measurer to question the same woman using a set probing technique. Weaker interviewers were given additional training.
We collected data on paper questionnaires, and a quality assurance team checked them for plausibility and logic at the field site before double entry into a database in the nearest city (Bhubaneswar).
The data management team observed 10% of the interviews to ensure data quality and adherence to procedures, and took repeat measurements for a subset of questions in 20% of households.
The variables used in this study are described below and in more detail in the Supporting Information: • -Value of agriculture production: ln-value of total agricultural production in the last three agricultural seasons (in the last 12 months), in 1,000 Indian Rupees. Production from all cultivated land (owned, rented, shared or other arrangements such cultivating on an extend family member's land or community land) included.
• -Agriculture production diversity: Count of 10 food groups produced, regardless of land ownership status, in the last three agricultural seasons (in the last 12 months) by households in any quantity.
• -Women's decision-making: Women's self-reported involvement • -Women's land ownership: Women's self-reported land ownership, in two categories: none versus any (joint or sole) ownership.
In addition to the variables of interest on the pathway from land to maternal nutrition, we also used the following variables that we identified well known a priori as confounders: caste group (Coelho & Belden, 2016), years of maternal education (Subramanian & Smith, 2006), count of household assets (Subramanian & Smith, 2006), household size (Rashid, Smith, & Rahman, 2011), female-only households (Rashid et al., 2011) and maternal age ( Harris-Fry et al., 2015).
The asset score includes the following 15 assets: high cost consumer durables, low cost consumer durables, jewellery, mobile phone, electricity, computer, internet, motorbike, mechanised agricultural assets, business assets, high-quality fuel type, finished flooring, finished roofing, finished walls and toilet. This score excludes assets that are highly correlated with land ownership: house, bicycle, small livestock, large livestock and non-mechanised equipment. 2010), we assess mediation using the 'potential outcomes framework' .

| Data analysis
The potential outcomes framework is a non-parametric approach that applies the logic of counterfactuals to identify mediated effects Although only one combination can be observed for each woman and all other possible combinations are counterfactual for that woman, the combinations that are counterfactual for one woman are observed for other women. Therefore, fitting models allows us to predict the unobserved values (counterfactuals) for any woman given her characteristics and characteristics of similar women in the sample.
Here we use non-parametric simulations to estimate these counterfactuals and their uncertainty.
The simulation process follows a four-step algorithm designed by In order to make inferences on any mediated effects, two 'assumptions of sequential ignorability' must be satisfied (Imai, Keele, & Yamamoto, 2010). Broadly speaking, the first assumption relates to the ignorability of the exposure: potential mediators and outcomes must not affect exposure. The second assumption is that the mediator is ignorable given the observed exposure and covariates.
The first assumption is satisfied if the exposure precedes the mediator and outcomes in time and is satisfied here. In our sample, 90.9% of the sample owned land in ancestral name. Only 4.9% reported owning land with a record of rights, and 1.3% reported having owned a piece of land without a record. Because land size owned is predominantly determined by inheritance and government allocations (which mostly occurred between 1960 and 2013), in this context, rather than women's empowerment and agriculture productivity, we assume that this exposure precedes the hypothesised mediators, at least in the short run. As such, given the counterfactual framework and the ignorability of the first assumption, following the methodology, we use the term effect rather than association while presenting our results.

| Respondent characteristics
We visited 5,427 households and interviewed 4,480 households, giving an 83% response rate. Two hundred eighty-two women were pregnant or post-partum (gave birth <42 days before the interview), so they were excluded from analyses with BMI as an outcome.
Descriptive statistics on the exposures, outcomes, confounders, mediators and other sociodemographic characteristics are given in Table 1.
Few households (6%) had no land. Around three quarters own a small plot <2.5 acres, but only 17% of women reported owning any land themselves. Maternal diets were inadequate, with around four fifths not consuming the recommended five or more food groups per day. Diversity of diets and agricultural production were similar (mean: 3.7 and 3.6 food groups, respectively), and the value of agricultural production was also strikingly low (median: 4,469 INR/year). Most women were involved in household decisions on agricultural activities, but their work burdens were high; less than half (40%) worked less than 10.5 h in a day. Figure 1 shows that women from households with the largest landholdings consumed around one more food group per day, compared with households with the smallest landholdings, but no evidence of a gradient with land size for BMI apart from perhaps women with the largest landholdings (Figure 1).

| Agricultural mediators
The results in Figure    The results in Table 2     F I G U R E 2 Pathways from land size to maternal diet diversity: Estimations from potential outcomes framework analysis. All coefficients are from linear or logit regression regressions. All models adjust for caste group, years of maternal education, asset score, household size, female-only households, maternal age and clustered study design. Estimations of associations between mediators and outcome are adjusted for exposure (ln land size) Abbreviations: ACME, average causal mediation effects; CI, confidence interval. In Table 2, we see a similar pattern to that observed for the hypo-

| DISCUSSION
To our knowledge, this is the first study to unpack the pathways from land size to maternal nutrition. We find that the pathways between household land size and women's nutrition outcomes are complex, often acting in opposing directions. First, we find small effects between land size and dietary diversity but not maternal BMI. Second, agricultural production indicators appear to partially mediate these effects, by improving diets but compromising BMI. We see agriculture production mediators acting as 'suppressors' of the effect of land on BMI; that is, the total effect of land size on BMI (shown in Figure 3 Odisha requires each homestead-less household to be allocated 10 decimals (~0.01 acres) of homestead land, which may explain the F I G U R E 3 Pathways from land size to maternal body mass index (BMI) diet diversity: Estimations from potential outcomes framework analysis. All coefficients are from linear and logit regressions. All models adjust for caste group, years of maternal education, asset score, household size, female-only households, maternal age and clustered study design. Estimations of associations between mediators and outcome are adjusted for exposure (ln land size) very low percentage of households owning no land (Government of Odisha, 2018). However, these land distribution schemes have been criticised for being ineffective due to the low quality of land provided (Deo, 2011). Variable quality of land, along with other agricultural inputs (technology, labour and climatic factors such as rainfall) may explain some disconnect between land ownership and agricultural production (Rahman, 2010;Zepeda, 2001). Beyond subsistence agriculture, with increasing reliance on migration, wage labour including agricultural wage labour and non-farm businesses, other non-farm sources of income may be more important for improving nutrition outcomes (Shankar et al., 2019). Furthermore, diversifying agricultural production may require relatively little land. Chickens (which can produce both meat and egg food groups) can be kept near the homestead; vegetables can be grown on small kitchen gardens, and people may also collect fruits or other foods from publicly owned forest land.
The relatively weak linkages from agricultural production to dietary diversity are consistent with the growing body of evidence on this relationship (Dillon, McGee, & Oseni, 2015;Jones, Shrinivas, & Bezner-Kerr, 2014;Sibhatu, Krishna, & Qaim, 2015). A disconnect may be explained by households selling their produce (Singh, Squire, & Strauss, 1986) or allocating the food to household members other than women (Harris-Fry et al., 2018). The temporal mismatch in measurement may also explain a weak association, with production being measured over 1 year and diets over 1 day.
The negative association between agricultural production and maternal BMI suggests that improvements in diets may be insufficient to compensate for the women's energy expenditure required to participate in agriculture. Increasing agricultural production may require more physical activity, placing women in negative energy balance. The same could apply for women's decision-making, if women who control decisions about agricultural production also take responsibility for carrying out those decisions. This hypothesis is corroborated by the finding that women with more work-free time have higher BMIs ( Figure 3). Our research questions guiding the paper emerged after the definition of the study design and data collection. In particular, our research question arose due to concern of excluding the landless in our own (and other agricultural) interventions for which these data were collected. Thus, variable specifications, especially of mediators, were guided by the data available. Our application of counterfactuals to identify mediated effects using potential outcomes framework for causal mediation offers a rigorous analytical approach for a crosssectional study, especially given that that our analysis satisfies the first assumption of sequential ignorability. However, our study cannot account for all confounding, especially due to time-varying unobservables. As such, any causal inference should be made cautiously. Future work using experimental or quasi-experimental study designs is needed to further test these pathways.
Our study suggests that land transfer programmes may need to be coupled with other agricultural inputs, such as soil quality, fertiliser, irrigation and labour to improve nutrition status and should support more equitable workload allocation. This is consistent with other studies showing weak nutritional effects of land-titling schemes in India (West Bengal;Santos, Fletschner, Savath, & Peterman, 2014), Vietnam (Menon, Van Der Meulen Rodgers, & Nguyen, 2014) and Ethiopia (Muchomba, 2017). This is especially the case where such land transfer programmes are for homesteads, where poor land quality might mean low potential for kitchen gardens. Nutrition-sensitive agriculture interventions should consider reliance on land and land quality for food production because land size does seem to partially determine agricultural production and, in turn, nutritional outcomes. However, these effect sizes are small, and it is possible that, in the Odisha context where homestead-less households have been provided with some land, agricultural diversification can be achieved on small plots.
Other study contexts, where more people do not have any land, may require more careful programme design. The negative association between agricultural production and BMI indicates that programmes aiming to increase or diversify production would need to carefully consider their implications for women's and intra-household allocation of labour (Johnston, Stevano, Malapit, Hull, & Kadiyala, 2018). This may partially explain the observed discrepancy between the larger, more consistently positive effects of nutrition-sensitive agricultural interventions on dietary intakes compared with anthropometric outcomes (Ruel et al., 2018).
Further research is needed on the trade-offs between improvements in household agricultural productive assets; women's empowerment in agriculture and women's nutritional outcomes; and how to mitigate these trade-offs (women's work-free time and energy expenditure) to optimise women's nutritional outcomes.

ACKNOWLEDGMENTS
We thank the study participants and DCOR's data collection team;