Building Empirical Configurational Typologies with Regularized Loglinear Models

Comparative researchers commonly build typologies that characterize the types by feature configurations. In some settings, the types are not defined ex ante, but rather chosen so that they describe a population of interest well. Existing approaches--cluster analysis, finite mixture models, or QCA--each have different disadvantages. Unlike QCA, finite mixture models handle uncertainty probabilistically, but assume that a set of restrictive types describes the population. This is problematic if only some observations belong to configurational types. For example, if such types result from lock-ins and some configurations are thus substantially more persistent. This paper introduces to political science an approach free of these disadvantages and capable of probabilistic feature selection, and applies it to the V-Dem dataset to identify persistent regime types and evaluate possible feature lock-ins. The approach is based on Bayesian regularized loglinear models and closely related to psychometric Configural Frequency Analysis.