A Bayesian networks approach for analyzing voting behavior

The problem of finding the factors influencing voting behavior is of crucial interest in political science and is frequently analyzed in books and articles. But there are not so many studies whose supporting information comes from official registers. This work uses official vote records in Spain matched to other files containing the values of some determinants of voting behavior at a previously unexplored level of disaggregation. The statistical relationships among the participation, the vote for parties and some socio-economic variables are analyzed by means of Gaussian Bayesian Networks. These networks, developed by the machine learning community, are built from data including only the dependencies among the variables needed to explain the data by maximizing the likelihood of the underlying probabilistic Gaussian model. The results are simple, sparse, and non-redundant graph representations encoding the complex structure of the data. The generated structure of dependencies confirms many previously studied influences, but it can also discover unreported ones such as the proportion of foreign population on all vote variables.