Sample selection bias correction for missing response observations

We propose a new simpler estimation technique to handle missing response fo r categorical dependent variables, which we apply to multinomial legit mode ls. Survey data often have a significant number of incomplete or missing re sponses. If such data an systematically missing (i.e., not missing at rando m) and if such observations are deleted from the analysis, biased sample se lection results. The standard approach to correct sample selection bias inv olves the calculation of complicated joint or conditional probabilities. Ou r approach requires only marginal choice probability specification and may be used with any appropriate probability structure such as legit or probit. We apply our new method to the empirical analysis of work status determina tion.