Alternatives to calorie-based indicators of food security: An application of machine learning methods

Identifying food insecure households in an accurate and cost-effective way is important for targeted food policy interventions. Since predictive accuracy depends partly on which indicators are used to identify food insecure households, it is important to assess the performance of indicators that are relatively easy and inexpensive to collect yet can proxy for the “gold standard” food security indicator, calorie intake. We study the effectiveness of different variable combinations and methods in predicting calorie-based food security among poor households and communities in rural Bangladesh. We use basic household information as a benchmark set for predicting calorie-based food security. We then assess the gain in predictive power obtained by adding subjective food security indicators (e.g., self-reported days without sufficient food), the dietary diversity score (DDS), and the combination of both sets to our model of calorie-based food security. We apply machine learning as well as traditional econometric methods in estimation. We find that the overall predictive accuracy rises from 63% to 69% when we add the subjective and DDS sets to the benchmark set. Our study demonstrates that while alternative indicators and methods are not always accurate in predicting calorie intake, DDS related indicators do improve accuracy compared to a simple benchmark set.