The main result of an experiment is typically an estimate of the average treatment effect (ATE) and its standard error. In most experiments, the number of covariates that may be moderators is large. One way this issue is typically skirted is by interpreting the ATE as the average effect for "some" population. Cornfield and Tukey (1956) famously explained the process of generalization as involving two bridges. The first bridge, they argued, is statistical, from the sample in hand to "some" putative population like it. The second bridge, then, is a subject-matter span from this putative population to the one truly of interest. The focus of this paper is to develop a method for making this second bridge in generalization a statistical bridge. The goal is to formalize the process of moving from a non-random sample in hand to making inferences about the estimate and standard error of a treatment effect in a particular, policy relevant population. The method the author proposes is an extension of propensity score matching, which is commonly used in observational studies to adjust for the process of selection into treatment. Here she proposes using propensity scores to adjust for the process of selection into the experiment. (Contains 1 table and 1 figure.)