PNNL

Abstract

We consider the problem of Bayesian variable selection for the dynamic logistic regression model and propose using penalized credible regions to select parameters of the updated state vector. This method avoids the need for shrinkage priors, is scalable to high-dimensional dynamic data, and allows the importance of variables to vary in time as new information becomes available. To fit the dynamic logistic regression model and improve computation time for posterior simulation, we utilize the Polya-Gamma latent variable approach. This technique alters the joint posterior distribution, providing pseudo-Gaussian data which allows the utilization of the Forward Filtering Backward Sampling algorithm, a Kalman filter based technique, to sequentially update states. A substantial improvement in both precision and F1-score using this approach is demonstrated by means of simulation. Finally, we apply the proposed model fitting and variable selection methodology to the problem of forecasting civil unrest in Latin America using daily protest-related terms scraped from Twitter as model features and show improved accuracy compared to the current static approach.