Estimating fixed effects : perfect prediction and bias in binary response panel models, with an application to the hospital readmissions reduction program

The maximum likelihood estimator for the regression coefficients, β, in a panel binary response model with fixed effects can be severely biased if N is large and T is small, a consequence of the incidental parameters problem. This has led to the development of conditional maximum likelihood estimators and, more recently, to estimators that remove the O(T–1) bias in β^. We add to this literature in two important ways. First, we focus on estimation of the fixed effects proper, as these have become increasingly important in applied work. Second, we build on a bias-reduction approach originally developed by Kosmidis and Firth (2009) for cross-section data, and show that in contrast to other proposals, the new estimator ensures finiteness of the fixed effects even in the absence of within-unit variation in the outcome. Results from a simulation study document favourable small sample properties. In an application to hospital data on patient readmission rates under the 2010 Affordable Care Act, we find that hospital fixed effects are strongly correlated across different treatment categories and on average higher for privately owned hospitals.