Statistical models of text have become increasingly popular in statistics and computer science as a method of exploring large document collections.  Social scientists often want to move beyond exploration, to measurement and experimentation, and make inference about social and political processes that drive discourse and content. In this talk, I will overview our recent work developing a novel topic model which supports this type of substantive research by modeling the relationships between observed covariates and latent topics. Our approach uses a simple generalized linear model framework to allow the analyst to condition on arbitrary structure affecting topic prevalence and content.  I discuss applications from across political science including: the analysis of survey experiments, media reporting in China and Islamic fatwas.  In each case we show how to leverage problem specific structure and perform inference on the resulting covariate relationships. All the methods described are available as part of the open source R package, stm, available at structuraltopicmodel.com.