Predictors of persistent moderate and severe food insecurity in a longitudinal survey in Mexico during the COVID-19 pandemic.

Background: Household food insecurity (HFI) increased in Latin America by 9% between 2019 and 2020. Scant evidence shows who was unable to recover from the COVID-19 pandemic. Our aim was to use a Machine Learning (ML) approach to identify consistent and influential predictors of persistent moderate or severe HFI over 2 years. Methods: We use a three-wave longitudinal telephone survey with a probabilistic sample representative of the Mexican population. With a response rate of 51.3 and 60.8% for the second and third waves, the final sample size consisted of 1,074 individuals. The primary outcome was persistent HFI, i.e., respondents who reported moderate or severe HFI in 2021 and 2022. Twelve income-related predictors were measured in 2020, including baseline HFI. We employed 6 supervised ML algorithms to cross-validate findings in models, examined its precision with 4 standard performance indicators to assess precision, and used SHAP values (Shapley Additive exPlanations) to identify influential predictors in each model. Results: Prevalence of persistent moderate/severe HFI in 2021 and 2022 was 8.8%. Models with only a HFI 2020 baseline measure were used as a reference for comparisons; they had an Discussion: Persistent HFI can be a relevant indicator to identify households that are less responsive to food security policies. These households should be prioritized for innovative government support and monitored to assess changes. Forecasting systems of HFI can be improved with longitudinal designs including baseline measures of HFI and socioeconomic predictors.