Replication Data for: "Estimating Substantive Effects in Binary Outcome Panel Models: A Comparison"

Dummy variable maximum likelihood (ML) estimation for binary response panel models struggles to estimate coefficients or substantive quantities in either short or rare event panels. The standard response is a conditional ML that consistently estimates the coefficients on time-varying covariates but makes substantive effects impossible to compute. In light of this problem, multiple suggestions have appeared for computing these effects, but there is little-to-no guidance as to when one solution may be preferred to another. I address this question by comparing one of these approaches, a correlated random effects (CRE) estimator, to the maximum likelihood dummy variable estimator (MLDV). I find that when the number within-group observations is small or events are rare, the CRE is preferred, but as panels get longer the differences between approaches fades. NOTE: The data used in the replication of Goldman (2018) comes from the 2008 National Annenberg Election Survey Internet Panel. While this data is freely available online (https://www.annenbergpublicpolicycenter.org/data-access/), its terms and conditions forbid posting it online.