Abstract: Applied studies frequently involve variables that cannot be measured directly. Such variables are usually incorporated through several observable variables, or proxies, designed to capture the nature of the underlying concept. For example, the rurality of a county is a social construct that cannot be observed directly but is clearly related to, for instance, the proportion of households with access to the public water supply or the proportion of households that use wood for heating. These so-called “latent" variables are widely used in disciplines such as economics, psychology and epidemiology. In this work we address the problem of selecting latent variables proposed as potential regressors for a response of interest. We investigate the notion of grouped Bayes factors: weighted averages of the Bayes factors built from the proxies that generate each latent variable. We discuss the choice of weights and show that standard default choices may lead to unsatisfactory results. This leads us to argue in favor of informed alternatives that account for the correlation structure of the proxies. The resulting method is fully Bayesian and objective, and it can be applied to problems with a large number of predictors, as we illustrate with real datasets.