查看文章

We compare experimental and nonexperimental estimates from a social and informational messaging experiment. Our results show that applying a fixed effects estimator in conjunction with matching to pre-process nonexperimental comparison groups cannot replicate an experimental benchmark, despite parallel pre-intervention trends and good covariate balance. The results are a stark reminder about the role of untestable assumptions–in our case, conditional bias stability–in drawing causal inferences from observational data, and the dangers of relying on single studies to justify program scaling-up or canceling.