The paper proposes a technique for the estimation of possibly nonlinear panel data models in the presence of heterogeneous unit nonresponse. Attrition or unit nonresponse in panel data usually renders parameter estimators inconsistent unless the unavailable infor-mation is missing completely at random. For moment based estimators this problem can be expressed in terms of the impossibility to construct the sample equivalents of the population moments of interest. However, if the attrition process is conditionally mean independent of the variables of interest then the sample equivalents of the population moments can be recov-ered by weighting the moment functions with the conditional response probability (or propen-sity score). The latter is usually unknown and has to be estimated. In the presence of nonre-sponse heterogeneity the propensity score can be estimated by conventional parametric esti-mation methods like the multinomial logit or probit model. The technique proposed in this paper leads to a moment estimator which simultaneously exploits the weighted moment func-tions of interest and the score function of the multinomial choice model. The use of simulated moments is discussed for applications with many nonresponse reasons. An applications of the estimator to firm level data is presented where the variables of interest are R&D investments related to product and process innovations.

published