In Social Sciences, many experiments rely on responses to information treatments. Experimental subjects in the treatment group receive some information that subjects in the control group don't. Often, the proportion of people in the treatment and control groups who were pre-exposed to the information is unknown and uncontrolled by the researchers. If that pre-exposure is ignored, it can bias the treatment effect estimation and lead to incorrect conclusions. I propose two estimation procedures for latent pre-exposure in this paper. One combines designed-based and model-based methods and relies on ancillary sampling and predictive probability of pre-exposure. The other is data-driven and relies on exploring latent effect heterogeneity using unsupervised learning methods and Dirichlet Process models. I compare both approaches using a real application to investigate public attitudes toward government tax policies to reduce economic inequality.

Dr. Diogo Ferrari