Editing and multiply imputing German establishment panel data to estimate stochastic production frontier models

This paper illustrates the effects of item-nonresponse in surveys on the results of multivariate statistical analysis when estimation of productivity is the task. To multiply impute the missing data a data augmentation algorithm based on a normal/Wishart model is applied. Data of the German IAB Establishment Panel from waves 2000 and 2001 are used to estimate the establishment's productivity. The processes of constructing, editing, and transforming the variables needed for the analyst's as well as the imputer's models are described. It is shown that standard multiple imputation techniques can be used to estimate sophisticated econometric models from large-scale panel data exposed to item-nonresponse. Basis of the empirical analysis is a stochastic production frontier model with labour and capital as input factors. The results show that a model of technical inefficiency is favoured compared to a case where we assume different production functions in East and West Germany. Also we see that the effect of regional setting on technical inefficiency increases when inference is based on multiply imputed data sets. This could have influence on the economic and regional policies in Germany in the future.