This paper utilizes recent simulation techniques in a two-stage estimation method which is applicable for a wide range of statistical models in the presence of missing data. The first stage of the method provides a way to estimate (and simulate from) the joint distribution of missing variables when the missing variables are continuous, binary, or ordered discrete. The second stage uses the first-stage estimates to “integrate” out the effects of the missing variables and obtain model estimates. The implementation of the method in this paper allows theoretically important, partially missing wage and school characteristic variables-which are not necessarily independently determined-to be included in a proportional hazard model of teacher attrition.

Using novel data from the Berea Panel Study, we show that the beauty wage premium for college graduates exists only in jobs where attractiveness is plausibly a productive characteristic. A large premium exists in jobs with substantial amounts of interpersonal interaction but not in jobs that require working with information. This finding is inconsistent with employer taste-based discrimination, which would favor attractive workers in all jobs. Unique task data address concerns that measurement error in the importance of interpersonal tasks may bias empirical work toward finding employer discrimination. Our conclusions are in stark contrast to the findings of existing research.