Introduction

The concept of “nutrition sensitive agriculture” assumes that agricultural production practices have the potential to positively affect the underlying determinants of nutrition (Ruel et al. 2013a, b). Although this assumption is intuitively a sensible one, especially if the focus is narrowed to food crop production, empirically it has proven difficult to support, not least because the causal pathways hypothesized to run between agriculture and nutrition are long and winding. Nevertheless, understanding the capacity of farming systems to contribute to improved nutrition outcomes is gaining ground as an objective among economists and other development professionals (Carletto et al. 2015). The need for more information on this topic is considered especially great when seeking strategies to improve food security in poor countries triply burdened by stunting, micronutrient deficiencies, and increased prevalence of obesity and diet-related chronic disease.

However, surveys which capture all the information required to test the association between farm level production practices, individual level dietary intake, and nutrition outcomes are few. A more common, albeit imperfect, option, are household surveys which capture information on both crop production diversity and household food consumption (Carletto et al. 2013). The latter can be used to construct indicators of household level diet diversity, including the widely used Household Diet Diversity Score (HDDS), which measures the number of food groups (out of a total of 12) consumed by one or more household members over a given reference period; usually 24 h or 7 days (Swindale and Bilinsky 2006; Kennedy et al. 2013).

While measures of household level food diversity should not be used to predict nutrient adequacy of individual level dietary intake (Kennedy et al. 2013), they are well-established indicators of what foods households can afford to eat, not only in terms of diversity but also in terms of quality of food groups consumed (Hoddinot and Yohannes 2002). As such, although HDDSs are not good indicators of individual nutrition outcomes, they can provide information regarding what households are eating as a unit, thus providing important clues about the options available at individual level.Footnote 1

Following this logic, we used household survey data to investigate associations between farm level production diversity and household level diet diversity in Kenya. Our sample consisted of ultra-poor and labour constrained families who were surveyed during an economic evaluation of Kenya’s Cash Transfer for Orphans and Vulnerable Children Programme (CT-OVC), Kenya’s flagship social protection programme. The geographical context of these data is relevant to our objective of understanding the capacity of farming systems to contribute to improved nutrition outcomes given (i) the widespread food insecurity and pervasive undernutrition present in East Africa (Black et al. 2008; FAO 2013), (ii) the kinds of agricultural policies currently promoted sub-continent-wide (i.e. market-based commercial agriculture), and (iii) the fact that agriculture is the primary livelihood base for most Africans (Pinstrup-Andersen 2010). The fact that the sample was ultra-poor and labour constrained is important, given the need to focus especially on the most vulnerable populations when applying a “nutrition lens” to empirical analysis. Around the world, the burden of undernutrition tends to fall disproportionately on the lowest-income groups. This is certainly the case in Kenya, where many more children are stunted and/or wasted in Kenya’s lowest wealth quintile relative to its highest.Footnote 2

Theoretical background

Multiple pathways have been proposed for the various ways through which agriculture may plausibly improve nutrition outcomes, and there is now general consensus on a model which includes 1) agriculture as a direct (via production for own-consumption) and indirect (via income effect) source of food at household level, 2) agriculture (with trade policy) as a driver of national and sub-national food prices; and 3) agriculture as an entry-point for enhancing women’s control over community and household resources, knowledge and status (Gillespie et al. 2012; Meeker and Haddad 2013; Ruel et al. 2013a, b; Herforth and Harris 2014; Webb 2013; Improving Nutrition Through Multisectoral Approaches 2013; Jones et al. 2014; Kadiyala et al. 2014).

These pathways appear in a number of recent agriculture-nutrition conceptual frameworks (Herforth and Harris 2013, 2014; Kadiyala et al. 2014; Kanter et al. 2015) most of which include explicitly labelled outcome boxes on agriculture as a source of food and agriculture as a source of income.

We adapted one of these recent conceptual frameworks on agriculture-nutrition pathways (Herforth and Harris 2014), to better reflect our focus on the links between production practices and household level dietary diversity, as shown in Fig. 1. Boxes highlighted in yellow reflect the inputs and outputs included in our analysis. Beginning with agricultural production practices (food crops produced and livestock ownership), we used food expenditure data (a proxy for food consumption) as well as data on non-agricultural income to assess what foods were consumed by surveyed households. We used a common measure of household level diet diversity - Household Dietary Diversity Scores - as well as two commonly used indicators of the “depth” of diversity (Simpson and Shannon Indices).

Fig. 1
figure 1

Conceptual framework for agriculture to nutrition pathways, adapted to CT-OVC data analysis

Full size image

As such, our analysis tested the pathways which frame agriculture as a direct and indirect source of food. The former assumes i) that production-for-own-consumption increases food availability and access at household level, and (ii) that increased food availability and access will lead to increased quantity and quality of intake at individual level. The latter assumes that an increase in income (e.g. via sale of agricultural products) will result in the purchase and consumption of not only more food, but of higher quality, nutrient-dense food (Meeker and Haddad 2013; Ruel et al. 2013a, b; Webb 2013; Kadiyala et al. 2014). Hypotheses corresponding to these pathways were:

  1. 1)

    On-farm production diversification correlates positively with household diet diversification in poor and rural settings.

  2. 2)

    Individual production activities are associated with diet diversification through income pathways or through production for own-consumption pathways.

Additionally, to examine whether agriculture is an entry-point for enhancing women’s control over household food resources, we interacted diversity of agricultural production practices with gender of household head. We also analysed the relationship between diversity of agricultural production practices and level of education of household head. While not explicitly included in our conceptual framework, education levels are well-known drivers of nutrition outcomes (Clausen et al. 2005; Thorne-Lyman et al. 2010; Jones et al. 2014). Research hypotheses corresponding to these pathways were:

  1. 3)

    Female household headship correlates positively with diet diversity at household level.

  2. 4)

    Education level of household head correlates positively with diet diversity at household level.

Finally, we tested the association between non-agricultural income sources and diet diversification, as follows:

  1. 5)

    Participation in the CT-OVC programme is associated with household diet diversity.

  2. 6)

    Off-farm income sources are associated with household diet diversity.

Data

We used data collected during the final wave of an evaluation of the welfare and economic impacts of Kenya’s Cash Transfer for Orphans and Vulnerable Children (CT-OVC). This programme targets families representing the poorest 20 % of the population of Kenya and as such these survey data cannot be considered nationally representative. Rather they represent a sample of ultra-poorFootnote 3 and labour constrained families highly affected by the HIV/AIDS pandemic, mostly relying on subsistence farming practices to meet daily basic needs. Furthermore, the data are representative of a large scale cash transfer programme intervention designed to mitigate poverty amongst ultra-poor and vulnerable households.

The impact evaluation used a cluster randomized longitudinal design, with a baseline household survey conducted in 2007 and subsequent follow-ups to the same households in 2009 and 2011. As shown in Map 1, the impact evaluation was carried out in seven districts located in four provinces of Kenya. The evaluation used a delayed entry approach since the government was not planning to saturate districts but rather expand the transfer coverage gradually across a number of districts. Consequently, within each of the seven districts across the country, it was possible to construct an experimental control by first randomly selecting four Locations to enter the study (from all Locations in a district), and then randomly assigning locations to treatment or delayed-entry control status. Note that a Location is the lowest level of programme operation, and represents the unit of public administration below province and district, consisting of a dozen or so communities. Given the budget allocation and scale-up plans at the time, four Locations that were similar in terms of HIV prevalence and overall socioeconomic status were selected for the study, of which two were randomly assigned to treatment status and two to control; households in the former were enrolled in the programme after completing the baseline survey.

Map 1
figure 2

Kenya CT-OVC evaluation sample locations

Full size image

Each survey consisted of a basic questionnaire on household and individual standard of living, consumption expenditure on food and non-food items, and demographic information. The 2011 survey also included an additional, detailed module which collected information on agricultural practices, namely crop production and cattle ownership. Data from this final follow-up survey provided the base for our analysis.

The initial household sample, surveyed in 2007, consisted of 2294 households split between 1540 beneficiaries and 754 delayed beneficiaries (control). However, 4 years after the programme was rolled out, the number of households participating in the evaluation had decreased, leading to a final attrition of 22.32 % with respect to the full sample. The largest attrition occurred between 2007 and 2009, approximately 17 %, while a further 5 % of the households was lost between 2009 and 2011. The final sample size in 2011 consisted of 1782 households (Treatment = 1249; Control = 533). While attrition might seriously undermine statistical inference, in the case of the CT-OVC, mean differences of relevant household characteristics between CT beneficiaries and the counterfactual remained stable over time, at least suggesting that representativeness of the sample remained intact.Footnote 4

Since the present study aimed at exploring the association between on-farm activities - crop production and livestock ownership - and household diet diversification, we excluded from the analysis sample households living in urban areas, as they showed levels of crop production and livestock ownership very close to “0”. We also excluded from the analysis survey data from households located in Garissa. Livelihoods in Garissa are primarily pastoral and crop production is not the main focus (only 1 % of Garissa households surveyed reported growing “at least one crop”). After removing urban households and families residing in Garissa, the remaining sample was 1353 household units.

Measurement of household level diet diversity

HDDSs were calculated by aggregating foods that survey respondents reported consuming in the 7 days prior to the interview into 12 equally weighted groups: (i) cereals, (ii) tubers, (iii) beans and pulses, (iv) fruits, (v) vegetables, (vi) meat, (vii) fish, (viii) eggs, (ix) milk, (x) fats, (xi) sugar and (xii) non-sugar condiments (e.g. salt). The number of groups reported was then summed to obtain an HDDS (0 to 12) for the household as a whole (Swindale and Bilinsky 2006; Kennedy et al. 2013).

Foods included in HDDSs came from one of the following sources: (i) foods purchased outside the home and consumed in the home, (ii) home-produced foods (i.e. production for own consumption), (iii) foods received as gifts, and (iv) foods purchased and eaten outside the home.Footnote 5

The HDDS assesses the presence of various food groups in a household’s meals. However it does not capture differences in the distribution of consumption, as all groups are equally weighted regardless of quantity consumed. The same HDDS score of, say 12, based on a total of twelve food groups, might in reality reflect two very different diet diversity situations, with one representing consumption of relatively large quantities of a very small number of certain food groups but very small quantities from each of the other food items over the recall period, the other reflecting an even distribution of consumption across the twelve groups. As such, higher dietary diversity scores might be more or less meaningful depending on the relative share of each food consumed (Arimond and Ruel 2004).

To mitigate this problem, we used two additional diversity measures – the Simpson index (Simpson 1949) and the Shannon index (Shannon and Weaver 1948) - to estimate the relative concentration or “spread” of food expenditure. (These indices were also used to corroborate HDDS scores for number of food groups consumed). Both indices were constructed making use of food shares calculated based on food expenditure comprising monetary value of food (i) from purchases, (ii) home production and (iii) food received as gifts or eaten out.Footnote 6

The values of each index are calculated as follows:

$$ \mathrm{Simpson}\ \mathrm{index}=1-{\displaystyle \sum_i{\mathrm{w}}_1^2} $$
(1)

Where wi is the expenditure share of food group i. The Simpson index ranges between zero and one; a value of zero implies only one food group is consumed while a value closer to one means a more diversified diet or a more equal distribution of food expenditure by food type is consumed within the sample cases.Footnote 7

$$ \begin{array}{l}\mathrm{Shannon}\ \mathrm{index}\\ {}\kern5em =-{\displaystyle \sum_i{\mathrm{w}}_{\mathrm{i}} \log \left({\mathrm{w}}_{\mathrm{i}}\right)}\end{array} $$
(2)

Where wi is again the expenditure share of food group i. Values for the Shannon index can range from zero to the value of the log of the highest number of food groups. The value of “0” flags consumption of only one food group to a maximum of log n (when all shares equal 1/n).

Taken together, the Simpson and the Shannon indices clarify not only whether particular farm and household characteristics correlate with increased number of foods consumed by households, but also the distribution or “evenness of consumption” of those foods, at least when looking at their expenditure. In so doing, these indices add granularity to the HDDS, which captures “crude diversity” of diets.

Measurement of farm diversification and off-farm activities

A wide variety of indicators for production diversity have been used in recent years to research the association between crop diversity and nutrition, with many studies using some iteration of a crop count of specific species cultivated and livestock species raised (Powell et al. 2015). Given the constraints of our data, we chose instead to construct a unique metric, the “agriculture enterprise score (AES)” for each sampled household by summing the following crop and livestock categoriesFootnote 8:

(i) cereals, (ii) potatoes, (iii) beans and pulses, (iv) vegetables and fruits, (v) cattle, (vi) poultry, (vii)goats and sheep, and (viii) pigs.

We chose to construct our production diversity metric in this way for two reasons: First, because in some cases - e.g. “fruits” - the CT-OVC questionnaire itself collected information on crop groups as opposed to specific species; and second, because only a very small percentage of surveyed households reported certain practices (e.g. cultivating vegetables). As such, we chose to construct broader categories based on taxonomy (e.g. leguminous, grain, tuber or horticultural; small ruminant, large ruminant, poultry, or monogastric) in order to ensure explanatory power.

In sum, the AES aims to reflect basic diversity of food group production. Designed in response to the unique constraints of the CT-OVC dataset, this metric’s drawbacks are discussed under “Limitations of the Study”.

As rural off-farm activities represent an income stream which might affect household diet diversity, we also constructed an “off-farm activities” variable, comprised of the following: (i) wage employment, (ii) annual private transfer income (remittances) and (iii) non-agricultural business. Households which received income from all three sources were scored “3”; households which received income from none of these sources were scored “0”.Footnote 9

Analytical methodology

To estimate the link between farming activities and household diet diversity, we used OLS multivariate regression analysis, constructing our models as follows:

$$ \mathrm{Y}=a+{\beta}_1AES+{\displaystyle {\sum}_{j=2}^n{\beta}_j\mathrm{X}+\upupsilon} $$
(3)

Here, Y is a diet diversification measure constructed using the HDDS count, the Simpson or the Shannon index and AES is the agriculture enterprise count reflecting basic diversity in food group production as explained above. In theory, the more a household diversifies its production practices, the stronger and more significant the positive association with household level food diversification should be. As such, the coefficient β 1 should provide empirical evidence on our first hypothesis.

X consists of a vector of household and community characteristics (e.g. gender of household head and household size) which might plausibly have impacted diet diversification in the sample. Selection criteria for these characteristics were based on empirical research on the drivers of household level dietary diversity (Torheim et al. 2004; Clausen et al. 2005; Thorne-Lyman et al. 2010; Jones et al. 2014). As transfer programmes can increase food expenditure on more nutritious foods, X also includes participation in the CT-OVC programme, so as to purge the association between farming diversification and diet diversification from this potential confounder.

As we wished to test which production activities were most strongly associated with diet diversity, we also disaggregated the AES to investigate how single practices correlated with diet diversity:

$$ \mathrm{Y}=a+{\displaystyle \sum_{k=1}^n{\beta}_kAG+{\displaystyle \sum_{o=n+1}^m{\beta}_oX+\upupsilon}} $$
(4)

Here, AG is a set of dummy variables flagging whether or not a household engaged in each of the production practices captured in the AES, where k is the production practice in question. The coefficient β k tells us if and to what extent k farm practice correlates with dietary diversification. X is the same set of control variables as explained above.

For both regression models, we clustered standard errors at community level to count for intra-correlation in our estimation strategy.

We tested whether the three measures of diet diversity were correlated by using Pearson product-moment, and we used ANOVA analysis to assess whether statistically significant differences in household characteristics and outcome indicators existed across districts.

A common concern when using multivariate regression analysis to analyze correlation is multicollinearity between the independent variables used to generate the model. That is, magnitude of some coefficient estimates might be increased because of associations between predictor variables, resulting in misleading measurements of the strength of the association in question. To test if this was an issue in our model, we observed variance inflation factors that ranged between 2.1 to 2.5, well below the suggested cut-off value of 10 provided by Kutner et al. (2004). We therefore concluded that multicollinearity was not an issue in our model.

To examine the robustness of our findings we re-ran the model with a stepwise exclusion of the control variables.

Finally, while we controlled for confounding variables – gender and education of household head, participation in the CT-OVC transfer programme, wealth quintile, land ownership, and district fixed effects – it should be noted that unobserved characteristics might still be of concern with respect to estimated strength of association between farming diversity and diet diversity.

Sample household characteristics

Table 1 presents descriptive characteristics of sample households disaggregated by district. Approximately 67 % of households received cash transfers through the CT-OVC programme. Mean land holdings were 3.1 acres. Most households were headed by women over 60 (64 %). Mean years of education was less than 4 years, and average household size was approximately 5.

Table 1 CT-OVC household sample characteristics by district
Full size table

Households spent more money on food than anything else, an average of approximately 2124 Ksh (US$21) per capita, per month. For contrast, monthly non-food expenditures averaged around 515 Ksh (US$5). The mean for off-farm activities was 1.14, indicating that, on average, sample households were benefitting from at least one off-farm income source other than agricultural production.Footnote 10 Although 81 % of surveyed households reported road access and 84 % reported access to potable water, only 29 % of households reported living in a village with a local market.

Taken together, these results imply a semi-autarkic food environment among surveyed households, comprised of some combination of production-for-own consumption and purchased foods.

Agricultural diversification and household level dietary diversity

Per Table 2, surveyed households reported harvesting approximately 2 crops per production cycle. The most frequently grown crops were cereals (96 %), followed by beans and pulses.

Table 2 Household agriculture practices
Full size table

Eighty-eight percent of surveyed households reported owning at least one animal. Seventy-four percent of households reported raising poultry, 57 % reported owning cattle, and 46 % reported owning goats or sheep. A very small number - 2 % of households reported raising pigs. Again, data from Kwale indicate that families living in that district were relatively worse-off, most notably with respect to cattle and poultry holdings.

In terms of incidence of consumption – i.e. foods grown for home consumption and purchased (Table 3A) - all households reported consuming cereals. In terms of micronutrient rich foods, almost all households reported vegetables, 91 % reported fish, and 78 % reported pulses. Fewer households reported fruit or milk, and well under half reported meat or eggs (Table 3A).

Table 3 Food expenditure and diet diversity indicators by district
Full size table

However, in terms of consumption shares (Table 3B), households spent 44 % of their food budgets on cereals, with far less allocated to micronutrient rich foods. Given that the latter are more expensive than cereals, it is reasonable to assume that only minimal amounts of micronutrient rich foods were actually being eaten.

In line with this assumption, although mean HDDS was 9.15 (Table 3C), indicating a relatively high level of intake (i.e. approximately 9 out of 12 food groups per week), mean Simpson index level was 0.73 and mean Shannon index level was 1.61, indicating that a less than “even” distribution of food groups were being eaten, (1 and 2.48 indicate perfect “evenness” of distribution for the Simpson and Shannon indices, respectively).

There was strong and significant correlation between the three indicators (p < 0.00): The Pearson cross-product correlation coefficient was equal to 0.62 between the HDDS and the Simpson index and 0.79 between the HDDS and the Shannon index. The correlation between the Simpson and the Shannon index was 0.90 (data not shown).

Determinants of household diet diversity

Per Table 4, the AES was strongly and positively correlated with all three household diet diversity indices (p < 0.01). All else equal, this result is important given that the sample in question was of ultra-poor families who were likely relying heavily on starchy staples to meet their caloric needs. The positive and significant association across all three diversity indices implies not only that diversified farming practices were associated with higher levels of diet diversity but also that farm diversification positively correlates with a more even distribution of consumption expenditure across all food groups, including more nutrient dense foods.

Table 4 Regression analysis of determinants of household dietary diversity
Full size table

Participation in the CT-OVC programme also had a positive and significant association with diet diversity, confirming previous studies (Asfaw et al. 2014; The Kenya CT-OVC Evaluation Team 2012) which found that receipt of the transfer increased consumption of nutrient-dense foods. To determine whether a production-for-own consumption effect was occurring in addition to the obvious income effect, we tested the association between participation in CT-OVC and the AES (through simple regression with AES as a dependent variable). The results (not shown) were negligible, indicating that the transfer’s impact on diet diversification was channeled primarily through foods purchased, rather than via increased production diversification. These results are in line with findings from Asfaw et al. (2014).

Gender of household head was not associated with diet diversity. Age of household head had a negative and significant association with the HDDS though no significant relationship was detected with respect to the Simpson and Shannon indices.

Not surprisingly, wealth was a key determinant of diet diversification. Households in the bottom quintileFootnote 11 of the consumption expenditure distribution consumed on average 1.8 food groups (p < 0.01) fewer compared to families in the wealthiest quintile. They also displayed significantly (p < 0.01) lower values of the Simpson and Shannon indices.

Off-farm income was also positively associated with household diet diversity. These results are in line with theories which frame income diversification as an ex-ante risk management strategy for food insecure poor families (Barrett et al. 2001a).

Association between individual agricultural practices and household diet diversity

In an attempt to assess which practices might play a greater role in shaping household dietary patterns in the sample population, we decomposed the AES to explore the association between individual farming practices and the three household diet diversity indicators (Table 5).

Table 5 Regression analysis of individual crop and livestock production practices on household diet diversity
Full size table

Anticipating that disaggregation of the AES would likely result in confounding effects or intra-correlation, we also ran a series of robustness checks (data not shown). Results indicated the following clean correlations:

  • Cultivation of pulses was associated with a significant (p < 0.01) increase in the number of food groups consumed (i.e. HDDS). However, no significant relationship was detected with respect to diet distribution (i.e. Simpson and Shannon indices).

  • Poultry ownership was significantly (p < 0.01) and positively correlated with all three diet diversification outcome variables.

Should these associations be attributed to increased availability of foods due to production-for-own-consumption, or to income effects resulting from the sale of agricultural products? As described in “Theoretical background” section, both are principle pathways through which agricultural production is hypothesized to impact household dietary diversity and, eventually, nutrition outcomes. However, since most farmers practice some mix of subsistence and market-oriented production (Jones et al. 2014), disentangling these pathways can be a complex task.

To assess the extent to which bean production and poultry ownership as well as other production practices were associated with household diet diversity through income effects versus production-for-own-consumption effects, we tested whether individual crop and livestock practices correlated with incidence of HDDS food groups (Table 6).

Table 6 Individual crop and livestock practices by incidence of HDDS food groups
Full size table

If cultivation of a certain crop or ownership of a certain livestock type significantly increased the likelihood of consuming a variety of food groups, the conclusion was that the activity might plausibly increase diet diversity primarily via income effects. Conversely, if a particular production practice was significantly associated with only those few foods which could be the result of that specific activity (e.g. cow ownership and milk) the own-consumption pathway was assumed to be more likely.

Pulses were strongly associated with their own consumption and milk’s (p < 0.01), as well as more weakly with tubers, fish, fat (p < 0.05), and meat (p < 0.1). Poultry displayed a strong and significant association (p < 0.01) with pulses, fruit, and meat, as well as a weaker but still significant association with eggs (p < 0.05) and milk (p < 0.1).

These results are intuitive, suggesting that as well as being eaten on-farm, beans were being used to purchase, inter alia, other foods. Similarly, poultry require few inputs, mature quickly, and are affordable relative to larger livestock. Previous studies have found that poultry is frequently sold in order to purchase other types of foods as well as non-food items (Azzarri et al. 2014; Robinson et al. 2007). Poultry can therefore be considered “liquid assets” which are attractive and accessible to extremely poor households facing chronic, severe income constraints.

Goats and sheep, which may also be considered “liquid” relative to cattle, were significantly associated with increased consumption of fish (p < 0.01) as well as pulses and meat (p < 0.05). Cereals were associated with increased consumption of meat and milk (p < 0.05) and more weakly with beans (p < 0.1).

Conversely, tubers were strongly associated with their own consumption (p < 0.01) and nothing else, and cattle holdings were markedly associated with milk (p < 0.01). These results imply a production-for-own consumption effect.

Cattle are less “liquid” than poultry and smaller livestock. They require more inputs, mature less quickly and are less affordable. Within the CT-OVC sample, cattle ownership might have contributed substantially to milk consumption via production-for-own consumption, but not to overall diet diversity via income effect. This conclusion is in line with the data’s reflection of thin markets, implying high perishability risk and consequent reduced incentive to sell.

Interaction terms and robustness checks

We interacted the AES with a number of control variables - gender of household head, level of education of household head, proximity to a local market, and participation in the CT-OVC programme - to further test associations between farming diversification, household characteristics and diet diversification (data not shown except for gender, Table 7).

Table 7 Interaction between gender of the head and agriculture enterprise score
Full size table

While the interaction terms were always positive, we found a significant result (at 10 %) only when interacting the AES with female headed households in the HDDS model, implying a stronger association between farm diversification and diet diversity for female than male headed households.

Finally, as a robustness check, we reran the model with a stepwise exclusion of control variables deemed to be relevant in the model specification (Table 8). With these covariates removed from the model, we found that the association between the AES and household diet diversity increased consistently and remained significant, thus providing evidence of the overall robustness of our model.

Table 8 Agriculture enterprise count on food diversity, robustness checks
Full size table

Limitations of the study

The cross sectional nature of the CT-OVC data we used (2011 survey only) removed the possibility of establishing causality. Despite a thorough exploration of the data, we were unable to construct an instrumental variable which would have been a good candidate on theoretical grounds to eliminate endogeneity concerns. Nor was it possible to model the adoption of different farming practices using propensity score matching, as production diversification is the result of individual choices endogenously made by farmers.

While we controlled for confounding variables - gender and education of household head, participation in the CT-OVC transfer programme, wealth quintile, land ownership and district fixed effects - unobserved characteristics might still be of concern with respect to estimated strength of association between farming diversity and diet diversity.

Food consumption was estimated using expenditure data collected during a 7 day recall period, rather than a 24 or 48 h time frame. While a longer recall period might have captured a wider variety of foods consumed by a household, it would also add some level of “noise” to the estimates by reducing their accuracy.

As previously mentioned, measures of household level food diversity should not be used to predict nutrient adequacy of individual level dietary intake and consequently we were not able to include individual diet quality as an explicit output of our analysis.

While using an unweighted score of crop and livestock categories permitted identification which of those categories contributed the most to diet diversification, the AES had no capacity to identify the roles of specific crop or livestock species within each group in determining household diet diversification. We attempted to address this issue later in our analysis by disaggregating the AES into individual crop practices for which data were available. Theoretically, the use of this summary measure based on unweighted categories also runs the risk of “masking” the nutritional implications of production practices. For example, farms raising a single cereal variety and cultivating a single fruit tree would receive a higher AES than farms growing a wide variety of vegetables. However as mentioned above, one rationale for our need to use the AES - as opposed to a more conventional unique crop count - was the very fact that few CT-OVC farmers were producing vegetables in the first place.

Finally, our findings are based on a social protection programme targeting the very poor households and as such they are not nationally representative.

Conclusions and policy message

In many contexts, the burden of malnourishment is concentrated disproportionately among poor farming families. Children from rural low-income groups often show significantly worse nutritional outcomes. This is certainly the case in Kenya where extremely poor farming households with one or more orphans and vulnerable children struggle to meet their daily subsistence requirements. While historically, limited attention has been given to linking agriculture to nutrition, especially in the context of subsistence producers, growing interest in the feedback loops between agriculture, food systems, and nutrition is now building the evidence base for this type of information, with a growing number of studies looking explicitly at the links between household diet diversity and various measures of production diversification (Powell et al. 2015; Carletto et al. 2015).

This study attempted to contribute to this growing evidence base by testing whether the livelihood systems of ultra-poor, labor-constrained, subsistence-oriented farmers are associated with nutritional outcomes. Our sample consisted of 1353 households surveyed in 5 districts, namelyy, Kisumu, Migori, Suba and Kwale, during the final wave (2011) of an evaluation of Kenya’s Cash Transfer for Orphans and Vulnerable Children Programme (CT-OVC). The results of an OLS multivariate regression analysis showed that diversification in agricultural production practices was significantly and positively correlated with household diet diversification. We also found a significant and positive association between certain farming practices such as livestock ownership and household food diversification.

Of all the on-farm activities included in the analysis, poultry had the most compelling correlation with household diet diversification, followed by pulses. In both cases, the association was most plausibly attributed to an income effect. There was also a significant association between cattle holdings and milk consumption, likely attributable to a production-for-own-consumption effect, and significant findings linking small livestock to a variety of food groups. In addition, factors such as participation in the CT-OVC programme, level of education of the household head and participation in off-farm income sources are positively and significantly associated with household diet diversification. These findings are in line with previous studies such as those of Rawlins et al. (2014); Hoddinott et al. (2015), Azzarri et al. (2014) and Nicholson et al. (2003) among others.

From a “nutrition-sensitive” policy perspective, our findings are thus indicative of the potential value of four broad intervention areas: (i) Support to diversified farming systems and diversified income sources; (ii) promotion and support of poultry and small livestock holdings; (iii) promotion and support of cow ownership; and (iv) “pro poor” attention to districts with limited agricultural potential and labor constraints.

In a semi-autarkic smallholder context, a diversification strategy which integrates crop and livestock production adds value directly via increasing diet diversity and quality, and indirectly via income effects. In addition to improving diets, diversifying into the production of pulses, vegetables and fruits (currently not widely practiced) and livestock serves as a risk management instrument, protecting against weather and market shocks. Pulse production in particular is a sound investment strategy given their nutrient value, low water footprint, and low carbon-to-nitrogen ratio, the latter especially important given current soil depletion challenges facing many smallholders.

That said, given that our findings are not nationally representative and as such must be applied exclusively to extremely poor, rural populations, it may be that sub-district or even community-based promotion of and support for poultry enterprises is an especially important intervention to emphasize. Scavenging family poultry provide much-needed protein and income at very low investment and operating costs. Chicken meat and eggs are sources of not only high-quality protein, but also important vitamins and minerals. And while increased milk consumption is a valuable consideration when the ultimate objective is improved nutrition outcomes in small children, the start-up and maintenance costs of cow ownership may put this type of intervention out of reach for very low income farmers. In contrast, poultry require few inputs, mature quickly, and are affordable even for extremely poor households facing chronic and severe income constraints.

An additional, related consideration concerns the fact that many districts in Kenya suffer from water stresses and over-pumping of boreholes. The chances of poultry production efforts attaining success is thus increased if complementary measures to establish adequate and sustainable water supplies are in place. Such measures are of course also incentives for diversification into small livestock holdings and horticulture.

In conclusion, it is important to note the shrinking size of African smallholder farms (Jayne et al. 2014). Farm families are having to do more with less, as in many cases area expansion is not an option. And while increasing yield per hectare of one or two heavily promoted and often subsidized cereal crops has been the de facto response for decades, climate change concerns, land degradation, loss of biodiversity and other sustainability issues - not to mention stubborn and deadly rates of undernutrition - point to an increasingly pressing need to do things differently (FAO 2013; Pingali 2015; World Bank 2016; Global Panel on Agriculture and Food Systems for Nutrition 2014). Government policies in agriculture, especially those directed towards small and marginal farms, need to support diversified farming systems, giving greater attention to poultry, pulses, fruits and vegetables than hitherto.