Conditional Topic Allocations for Open-Ended Survey Responses

This paper develops a novel topic model for text data by conditioning on observables, named the "Conditional Topic Allocation" (CTA). This data-driven method allows identification of latent topics that explain other observable variables. It is particularly suited for small-scale text data, such as open-ended survey responses. First, CTA is used to extract topics from open-ended text answers that explain single observable variables. Then, in empirical models where text is a control variable for unobservable characteristics, CTA's flexible scope of applications allows uncovering latent variables from the text. As a proof-of-concept, this approach is used to analyze the sentimental value homeowners place on their homes and how this relates to energy-efficiency valuation. Specifically, responses from open-ended survey questions are used as control variables in a hedonic regression, and CTA serves to identify latent preferences associated with nearly 50% of the valuation of energy efficiency for single-family houses