Estimating a counter-factual in which a treatment did not occur allows political science researchers to better understand the effect of an intervention. To date, the most prominent attempt in the literature has been the introduction of the synthetic control method. However, this method has important and related drawbacks. Perhaps most importantly, the synthetic control method does not lend well to estimates of uncertainty, making traditional hypothesis testing ad hoc. We develop a new method, Gaussian process projection (GPP), that circumvents the issues associated with the synthetic control method. By comparing the projected unit to the true unit, we can estimate the causal effect of an intervention, and by using the measure of uncertainty, we can determine how unlikely the difference would be due to chance. After demonstrating the usefulness of the method in traditional settings, we generalize the method to allow for multiple treated units, treated at different times, and allow for link functions for applications with discrete outcomes.

Speaker: Dr. David Carlson, Assistant Professor of International Relations at Koç University