Using Multiverse Analysis for Estimating Response Models: Towards an Archive of Informative Features

Abstract

This study investigates which features are informative in predicting non-response for certain target variables in the context of European behavioral- and social-pattern data and how a multiverse approach can guide the process of identifying these predictors, with the long-term goal of building an archive of essential predictors. This will help researchers to design their studies in such a way that the missing at random mechanism can be assumed safely, ensuring valid use of advanced imputation techniques. Within the context of this study, a consensus on the types of variables that are informative can be discerned. That is, the results suggest the importance of variables related to employment, education level, domicile, and household and partner information. Limitations remain in accounting for the researcher degrees of freedom and the missing data in the observed variables, indicating the relevance of conducting similar, additional analyses to get a more robust collection of essential predictors. Nevertheless, this study provides an initial set of important predictors in the context of social science-related data and shows that multiverse analysis can adequately guide the process of identifying predictors of non-response by enabling flexibility in the construction and deployment of a set of models, rendering it easy to implement in di↵erent domains.