Endogenous Units of Analysis

Presenter: Tara Slough

Plan

Empirical political science research increasingly emphasizes experimental and observational methods for causal inference. However, a surprising number research designs seek to make causal inferences on the basis of “endogenous units of analysis.” Units are endogenous when the identity, survival, or composition of the units upon which outcomes are being measured is plausibly endogenous to the treatment being studied. I show that when units are endogenous to treatment, standard causal estimands are undefined. This finding holds for both individual- and cluster-assigned treatments. Using principal stratification, I describe the Survivor Average Causal Effect (SACE) as a relevant estimand for a broad set of cases in which units of analysis may be endogenous to treatment. However, our ability to bound informative estimates of the SACE depends critically on the research design and the process through which treatment affects the composition of units. In the presentation, I will discuss the implications of these findings for works seeking to make causal inferences using long-run historical data and works drawing upon downstream analysis of experiments.