Dynamic variable selection in dynamic logistic regression: an application to Internet subscription

We extend the dynamic model averaging framework for dynamic logistic regression proposed by McCormick et al. (Biometrics 68(1):23–30, 2012) to incorporate variable selection. This method of accommodating uncertainty regarding predictors is particularly appealing in scenarios where relevant predictors change through time, and there are potentially many of them, as a consequence, the computational burden is high. Simulation experiments demonstrate that our greedy variable selection strategy works well in identifying the relevant regressors. We apply our algorithm to uncover the determinants of Internet subscription in Medellín (Colombia) among 18 potential factors, and thus 262,144 potential models. Our results suggest that subscription to pay TV, household members studying, years of education and number of household members are positively associated with Internet subscription.