insecurity. This is especially of concern given the relatively few questions that define the food insecurity measure.
Several questions in the food insecurity supplement are follow-up questions asked only if “yes” is answered to a previous question. This structure should be taken into account in summarizing information about food insecurity. In particular, the responses to the stem item and the frequency follow-up question are not conditionally independent, an assumption made by the IRT model.
Under current practice, estimates of food insecurity at the household level ignore the large amounts of inherent uncertainty that exist due to its measurement by a small number of items. As a result, it is more appropriate to regard each household’s value of as estimated by its posterior distribution of -values rather than by an estimated value that does not reflect its uncertainty. It is not appropriate to ignore this uncertainty when classifying households or estimating the prevalence of food insecurity, either overall or in subsets of households.
Currently USDA collects information on the intensity and frequency of food insecurity in U.S. households. The information does not address the measurement of the duration of spells of food insecurity, either overall or in subsets of households.
These conclusions lead to the following recommendations to improve the categorization of households into food insecurity levels:
Recommendation 5-1: USDA should consider more flexible alternatives to the dichotomous Rasch model, the latent variable model that underlies the current food insecurity classification scheme. The alternatives should reflect the types of data collected in the Food Security Supplement. Alternative models that should be formally compared include:
Modeling ordered polytomous item responses by ordered polytomous rather than dichotomous item response functions.
Treating items with frequency follow-up questions appropriately, for example, as a single ordered polytomous item rather than as two independent questions.
Allowing the item discrimination parameters to differ from item to item when indicated by relevant data.