On targeting: Predictors of Consumer Welfare from Catastrophic Insurance

Abstract

This study investigates to what extent machine learning (ML) can be used to predict welfare outcomes using household and consumer characteristics in the context of IBLI insurance in Southern Ethiopia. By focusing on welfare outcomes rather than simple insurance uptake rates and thereby allowing for varying decision qualities, this study closes an important gap in the literature. The aim is to identify relevant variables and subgroups for which targeted interventions can be designed by policymakers. Administrative insurance and survey panel data collected over three rounds in a Randomized Control Trial from households of goat and cattle herders in the rural Borena zone in Southern Ethiopia is used, and four ML models (i.e., Elastic Net, Random Forest, CatBoost, and SVR) are run with a focus on prediction accuracy and predictor stability across models. Results show extremely weak prediction accuracy and an explained variances of no more than 1.5% in the best performing model despite using a large set of predictors. Effect sizes are very weak, but in relative terms trust in the insurance, the non purchase reason, and the main information source are consistently the strongest predictors across models, both for goat and cattle herders. More specifically, external constraints, low trust, and interpersonal contacts as the main information source tend to have welfare reducing effects. Policymakers are advised to explore those three factors in more depth and to identify other variables such as cognitive differences that are more closely related to welfare.