We examine how pension rule changes affected teacher retirement by estimating an option-value retirement model on a large cohort of late career Missouri public school teachers from 1994 to 2008. In so doing we offer potential solutions to several statistical challenges that arise in estimating structural models of retirement on large panel data sets. The first concerns modelling the formation of teacher expectations of future pension rules. The second is bias induced by baseline sample selection: in baseline cohorts we only observe teachers who are still working. This bias also evolves with pension rule changes. A third challenge arises from maximum-likelihood estimation using large panels of micro-data on individual teachers. The teacher-level data can be difficult to obtain and the likelihood of teacher-data is costly to compute in large panels. We address these challenges by incorporating policy expecta- tions and sample selection directly into estimation of the likelihood. We also show that the likelihood can be efficiently estimated by using teacher data grouped by age and experience cells, which permits:  a) estimating structural models of teacher retirement with data that are more widely available, and b) dramatic reductions in computation cost. Counterfactual simulations of the estimated structural model suggest that Missouri’s pension enhancements led to earlier retirement by about 0.4 years on average for the 1994 cohort and by more than one year in a steady state. Enhancements increased steady state pension liabilities by 16 percent for senior teachers.