[In the context of regression models with stochastic explanatory variables, the exact sampling distribution of the omitted variable (OV) estimator is derive. We show that stochastic regressors present the problem of larger variance for the OV estimator than in the nonstochastic case, and derive conditions under which omission may be better under a minimum mean square error criterion. Since the errors-in-variables problem can be interpreted as a specification error problem with stochastic regressors, we also consider the issue of MSE dominance of the proxy-variable estimator over the OV estimator.]

We estimate a multistage sequential logit model reflecting the structure of the disability determination process of the Social Security Administration (SSA). The model is estimated using household survey information exact-matched to SSA records on disability adjudications from 1989 to 1993. Under program provisions, different criteria dictate outcomes at different steps of the determination process. We find that, without the multistaged structural approach, effects of many important health, disability, and vocational factors are not readily discernible. As a result, split-sample predictions of overall allowance rates from the sequential model perform considerably better than do those for the conventional allowed/denied logit regression.