A Least Squares Correction for Selectivity Bias

Abstract

WHEN ESTIMATING REGRESSION MODELS it is very nearly always assumed that the sample is random. The recent literature has begun to deal with the problems which arise when estimating a regression model with samples which may not be random. The most general case in which one only has access to a single nonrandom sample has not been addressed since it is a very imposing problem. The case which has been addressed starts with a random sample but considers the problem of missing values for the dependent variable of a regression. If the determination of which values are to be observed is related to the unobservable error term in the regression, then methods such as ordinary least squares are in general inappropriate. By constructing a joint model which represents both the regression model to be estimated and the process determining when the dependent variable is to be observed, some progress can be made towards taking into account nonrandomness for the observed values of the dependent variable. The actual techniques employed fall into two rough groups, full information maximum likelihood models, and limited information methods which are more easily estimated. In the full information category are two methods. One model combines the probit and the normal regression models, and the other combines the Tobit or limited dependent variable model with the normal regression model. The form of the probit regression model is