IV-Schätzung eines linearen Panelmodells mit anonymisierten Betriebs- und Unternehmensdaten

Recently researchers in Germany were given the opportunity to use some of the business data provided by the Federal Statistical Office at their own workplace as Scientific Use Files (SUF). One of the anonymisation procedures used for the generation of SUF is multiplicative stochastic noise. Since this method leads to inconsistent estimates of linear panel models the use of correction procedures is necessary. This article analyses the instrument variables estimation as a correction method. As instrument variables we test a) a lagged variable, b) the difference between two lagged variables, and c) an anonymised variable from an additional anonymised data set. We conclude that only the last instrument leads to consistent estimators. We also study numerical problems which arise if the denominator of the estimator tends towards zero.