Abstract
Two-way fixed-effects (TWFE) designs are a popular identification strategy in the non-experimental program evaluation literature. The approach produces what are considered to be convincing causal effect estimates where the identifying assumptions hold. Unfortunately, the “parallel paths” assumption can be vulnerable to confounding that may not be detectable by pre-treatment trends tests and may not be controlled for through observable covariate adjustment. This dissertation proposes a testable condition that might replace the assumption of parallel paths in the context of location based confounders. It provides an outline of the circumstances where such a condition might be met; demonstrates how this condition may be used to recover the (joint) distribution of “unobservable” confounding variables; and quantifies the reduction in omitted variable bias.
Using the framework of matching as non-parametric control, this dissertation applies a machine learner (specifically a genetic algorithm) to the problem of estimating the (spatial) distribution of an “unobserved” confounder. In the first chapter, the performance of the method is evaluated in a series of simulations.
The latter chapters apply this method in two empirical contexts, estimating the size of natural hazard effects in residential real estate markets. These applications provide opportunities to compare results against conventional estimation procedures. They also provide opportunities to test the limitations of the method and discuss its value as a supplementary diagnostic tool.
The first empirical chapter finds that conventional estimation methods substantially underestimate the “true” effect of interest. Conventional estimates are anywhere from 12% to 70% smaller in size than the unconfounded effect depending on how those models are specified. This chapter suggests (and shows) that this underestimation may be explained by post-treatment confounding arising in the control observations. Controlling for the estimated distribution of the “unobserved” (spatial) confounder ensures that the control group is restricted to those observations with similar “location” characteristics.
The second finds that, despite having the appearance of an ideal natural experiment, compelling unconfounded location controls cannot be easily found. Consequently, estimates of the effect of interest should be interpreted with some additional caveats. The failure to converge on a convincing spatially confounding distribution is in itself a valuable piece of diagnostic information and is also discussed.