Abstract
In quasi-experimental and observational studies, the causal effect of treatment cannot be estimated without controlling for the systematic differences between treated and control subjects. In a sequence of two papers, we suggest an approach for these studies that first, without having access to outcome data, balances these differences and defines appropriate models for the key outcome analyses. Then, when access to outcome data is gained, these models can be fitted without repeated attempts. In this document, we design a hypothetical randomized experiment, studying peer effects related to smoking, that is close to an actual quasi-experimental study. We find that the balance of the observed covariates in the designed study is better than would be expected in a randomized experiment. Furthermore, we define the models of the key outcornes concerning smoking behavior that we commit to run in our next document. Finally, we gain an understanding of which treatment effects we can expect to be real and important in the real outcome analyses by conducting practice analyses on intermediate outcomes using our design.