Abstract
With this document, we conclude our demonstration of how Rubin's Causal Model (RCM) can be used to draw causal inferences in a two-step procedure. In the first step, we designed a study to evaluate if Harvard freshmen were more prone to start smoking when sharing a suite with at least one smoker than they would have been when sharing a suite with only non-smokers. Treated students were matched with control students, and models for the outcome analyses were specified. In this second step, we fit these models and evaluate the treatment effects. We also discuss how robust the effects are to various assumptions, as demonstrated by the variation in the effects across the different models. Our main result is that our effect of treatment is small and insignificant when we fit our statistical models on a well-balanced study. Also, this result is robust to the assumptions we make both with regard to the missing potential outcomes and to the various covariate adjustments. Our secondary result is that we would have found peer effects had we instead fitted a model on a less balanced sample, as has been done previously in the peer effect literature, using the traditional approach of causal inferences. However, this secondary result is not robust to the covariate adjustments we make. This exercise illustrates that it is difficult to replicate the results we find when we evaluate peer effects using a well-balanced study (RCM) when we evaluate peer effects using a less-balanced study (traditional approach). The result is reminiscent of the classic results of LaLonde (1986).