A Flexible Model of Food Security : Estimation and Implications for Prediction

We propose a novel Bayesian Graded Response Model (BGRM) for food security measurement. Our BGRM has several attractive features. It produces continuous food security estimates and measures of estimation uncertainty at the household level. Unlike the USDA's official measurement model, the BGRM can be used with binary and polytomous items. We further modify our BGRM to include any combination of binary, ordered polytomous, and continuous variables. With data from the 2017-18 National Health and Nutrition Examination Survey (NHANES), we estimate our BGRM for responses to the 10 adult core Food Security Module (FSM) questions. We find substantial uncertainty in household-level estimates, emphasizing the inherent uncertainty of latent trait estimation. We observe overlap in BGRM estimates across USDA-defined food security categories and significant variation within categories. We estimate our model using Current Population Survey (CPS) data as a robustness check. CPS results are qualitatively similar to those from the NHANES, highlighting possible implications for national USDA food security estimates. We explore the BGRM's ability to explain variations in health outcomes associated with food security and compare results to those produced using standard USDA category definitions. Finally, we demonstrate the BGRM's flexibility by incorporating an additional continuous variable, the Healthy Eating Index (HEI), into the model, capturing nutrition quality and food security information in a novel latent construct. The adaptability of our BGRM positions it as a versatile tool for measuring food security and other latent traits requiring a diverse range of variable types like nutrition security