*4.2. Determinants of Farm Production Diversity*

In the analysis of factors determining the observed farm production diversity, we present results based on crop-livestock count and the number of food groups produced—our primary indicators of farm production diversity—as dependent variables. Despite a few differences, the results from the two indicators of diversity provided a similar picture. Here we interpret the Poisson regression results based on crop-livestock count for both regions and the pooled sample (Table 4).

Results showed that farm production diversity is positively and significantly influenced by age of household head, availability of labor in the household and access to credit, for both Kilosa and Chamwino districts. For Kilosa, column (1), education of the household head and access to non-farm self-employment were also significantly and positively associated with increased farm production diversity. Interestingly, increased distance to nearest paved road had a significant positive influence on production diversity only for Kilosa with better market access suggesting an increased role of self-sufficiency for households far from market opportunities. However, for Kilosa and the pooled sample, agricultural shocks were negatively associated with farm production diversity. This could suggest that resource-constrained households may opt for a few highly resistant crops and livestock—or even venture into non-agricultural activities—after the experience of agricultural shock. In addition, the onset of an agricultural shock (such as drought, crop pests or unusually heavy rainfall) may have severe and negative impacts which may further reduce their agricultural production including its diversity. For Chamwino, the preparedness of a household to undertake risk, availability of land and other assets were significant in raising farm production diversity. Locational dummies also confirm the pattern observed in descriptive analysis, where residing in villages in Kilosa was negatively related to farm production diversity, unlike in Chamwino.


#### **Table 4.** Determinants of farm production diversity.


#### **Table 4.** *Cont*.

All models are estimated with Poisson regressions; \*\*\*, \*\* and \* indicate a significance level of 1%, 5%, and 10%, respectively; Values shown in parentheses are standard errors.

#### *4.3. The Role of Farm Production Diversity on Household Food Consumption Diversity*

In the analysis of the role of farm production diversity on food consumption diversity of households, we used several regression models. As pointed out earlier, the aim was to assess this relationship based on the two regions with distinct agro-ecological and market access characteristics as well as to ascertain whether farm production diversity plays a role in influencing seasonal food consumption diversity. For farm production diversity, we used crop-livestock count and the number of food groups. To get insights on food consumption diversity and its seasonal nature, the dependent variables were HDDS and FVS; and HDDS (planting), HDDS (pre-harvest) and HDDS (post-harvest) respectively. All regression models were estimated with Poisson regression except for FVS which were estimated with negative binomial regressions. In the latter regressions, the test of the over-dispersion parameter indicated that alpha is significantly different from zero, suggesting inappropriateness of Poisson regression. Table 5 presents these results showing the determinants of food consumption diversity.

Taking the case of crop-livestock count, results showed that farm production diversity had an overall positive and significant influence on household food consumption diversity. Going beyond farm production diversity, results also showed that household food consumption diversity was also influenced by market access characteristics. Access to market information and income from non-farm self-employment was significantly associated with increased food consumption diversity. Similarly, per capita food expenditure per month was positively related to food consumption diversity indicating that sourcing of different varieties of food from markets seems to be a relevant factor. Distance to nearest paved road was negatively related to food consumption diversity suggesting that market access plays an important role. Specifically, residing far from markets lowers the level of food consumption diversity in the households. A largely similar pattern of influences was observed for results of regressions using the number of food groups produced as an indicator of farm production diversity (see Table 6).

While results for district-specific regressions (presented in Tables A1–A4) showed almost consistent positive effects of farm production diversity on household food consumption diversity for Chamwino district, the same effects were not observed for Kilosa, except for HDDS (planting). The magnitudes of effects are also consistently higher for the former than the latter. The results suggest that the role of farm production diversity is more pronounced in Chamwino, which has relatively poor market access and agricultural potential as compared to Kilosa district with better market access. Additionally, the crop-livestock indicator showed that farm production diversity had a positive effect on seasonal food consumption diversity. However, the role of market access was less pronounced for Chamwino district. Despite a significant influence of access to market information on food consumption diversity, distance to nearest paved road and access to income from non-farm self-employment (except for HDDS for post-harvest) were insignificant. However, there was still a

significant positive association between per capita food expenditure per month and household food consumption diversity.

**Table 5.** Effects of farm production diversity on household food consumption diversity—Pooled sample (Farm production diversity indicator: Crop-livestock count).


\*\*\*, \*\* and \* indicate a significance level of 1%, 5%, and 10%, respectively; Values shown in parentheses are standard errors; Negative binomial model used for FVS regression: Likelihood-ratio test of alpha = 0; chibar2 (01) = 258.20; Prop > = chibar2 = 0.000. This suggests that alpha is non-zero rendering Poisson model less appropriate.



\*\*\*, \*\* and \* indicate a significance level of 1%, 5%, and 10%, respectively; Values shown in parentheses are standard errors; Negative binomial model used for FVS regression: Likelihood-ratio test of alpha = 0; chibar2 (01) = 197.16; Prop > = chibar2 = 0.000. The estimated alpha coefficient for the Negative binomial model is significant suggesting absence of equi-dispersion which would favor the use of a Poisson model.
