JOINT CENSORING OF REGRESSORS AND OUTCOMES:SURVEY NONRESPONSE AND ATTRITION

Censoring of outcomes (selection) is a common consequence of survey nonresponse and attrition in panels, and has received much attention. Joint censoring of regressors and outcomes is also common, but it has remained unexplored. This paper shows that the problem of identification when regressors and outcomes are jointly censored is equivalent to the problem of identification that would arise if only outcomes were censored but the censoring rate were higher than the actual one. As in the selection problem, one can identify informative bounds on conditional probability distributions, though the distributions themselves are not identified. These findings are illustrated using CPS income data and PSID data on welfare transitions. The identifying power of increasing the response rate to a survey is investigated.