A new estimation technique is proposed to deal with missing response variab
les in the context of a nested multinomial legit model. Survey data often h
ave a significant number of incomplete or missing responses. If such data a
re systematically missing (i.e. not missing at random) and if such observat
ions are deleted from the analysis, biased sample selection results. The ne
w method is applied to the empirical analysis of determining job loss statu
s.