Relating Seasonal Hunger and Prevention and Coping Strategies: A Panel Analysis of Malawian Farm Households

Abstract

Relative to chronic hunger, seasonal hunger in rural and urban areas of Africa is poorly understood. This paper examines the extent and potential correlates of seasonal hunger in Malawi using panel data from 2011–2013. We find that both urban and rural households report seasonal hunger in the pre-harvest months. Certain strategies to smooth consumption, including crop storage and livestock ownership, are associated with fewer months of hunger. In addition, we find that Malawian households that experience seasonal hunger harvest their crops earlier than average – a short-term coping mechanism that can reduce the crop’s yield and nutritional value, possibly perpetuating hunger.

Acknowledgements

We thank Katie Panhorst Harris and Margaret Beetstra for excellent research assistance. We are grateful to two anonymous referees for providing comments that greatly improved this paper. We also thank the Bill and Melinda Gates Foundation for supporting the research leading to this paper. The findings and conclusions presented here are those of the authors and do not necessarily reflect the positions or policies of the foundation. All remaining errors are our own. Stata .do files used to prepare the data and conduct the analysis are available on our website at https://evans.uw.edu/policy-impact/epar.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. While definitions differ, in this paper we refer to hunger as insufficient caloric intake (following the United States Department of Agriculture [USDA], 2015).

2. The median number of acres owned is identical in both survey waves, suggesting that the increase in acres owned is driven primarily by increased land holding for households in the top half of the distribution.

3. The qualitative results are unchanged using the current four month seasonal hunger period, the three months prior to but not including the month of first harvest by the household, or the three months prior to and including the month of first harvest.

4. Only agricultural households are included in these regressions, as the models include various farm characteristics.

5. We test for heteroskedasticity in our EA models using the White test. The test strongly rejects the null hypothesis of homoskedasticity. As such, we cluster our standard errors at the EA level, which results in standard errors that are robust to both general heteroskedasticity as well as autocorrelation (at the EA level) (Bertrand, Duflo, & Mullainathan, 2004). To test for multicollinearity, we first constructed a correlation matrix of the main independent variables. The highest pairwise correlation coefficient between independent variables is just 0.3816. In addition, we construct variance inflation factors (VIF) following estimation; the highest VIF is 1.27, which is well below the commonly used cut-off of 10 (Wooldridge, 2009).

6. This diversity index is calculated as: $Simpson=1-\sum \frac{a_i}{A}$, where $a_i$ is the area devoted to crop i and A is the total area devoted to crops in the household. Constructed in this way, the variable equals zero when there is no diversity (that is, when there is only one crop planted). An increasing index thus represents increasing diversity (Meng, Smale, Ruffa, Brennan, & Godden, 1999). We do not include area planted to cash crops including tobacco, cotton, and paprika in these calculations.

7. In other words, households in Wave 1 of the survey were asked about the 2008–2009 cropping season, while households in Wave 2 were asked about the 2011–2012 cropping season.

8. Similarly, households in Wave 1 were asked about the 2009–2010 cropping season, while households in Wave 2 were asked about the 2012–2013 cropping season.

9. In the household fixed effects models, we cluster at the household level, which again allows for arbitrary correlation over time and constructs estimates for the standard errors that are valid in the presence of heteroskedasticity, autocorrelation, or both (Bertrand et al., 2004). We also compute the correlation of differences at the household level. In other words, for each independent variable we use in the household fixed effects estimates, we construct the difference in that variable (value in Wave 2 minus value in Wave 1) at the household level. The highest pairwise correlation is only 0.1527. Based on these statistics, we conclude that multicollinearity is not likely to be a problem. Households that split off from Wave 1 panel households and were interviewed only in Wave 2 are not included in these regressions.

10. Household size and count of non-poultry livestock are associated with earlier harvests of any crop in the ordered logit models (Table A1), but not in the OLS models.

Additional information

Funding

This work was supported by the Bill and Melinda Gates Foundation [OPP1135685].