Texto para Discussão, Instituto de Pesquisa Econômica Aplicada (IPEA) 1347
Our objective is to analyze the gravity of panel attrition in the Brazilian Monthly Employment Survey (PME). There are three types of panel attrition: a) random, that does not produce any selection bias; b) function of observables, whose bias can be corrected through an appropriate re-weighing scheme; and c) function of unobservable variables, whose bias can be corrected through a bivariate probit for discrete variables or a Heckman correction for continuous ones. To test the gravity of panel attrition bias we estimated a transition to unemployment model both with and without attrition bias controls. Our results are that attrition tends to increase during school vacations and that attrition is strongly correlated to geographical mobility. Our conclusion is that while attrition bias does exist, its magnitude is small and its elimination does not substantially change analysis made using the models. In addition, in some cases, since it is not possible to control simultaneously for observable and non-observable attrition bias, in some cases the best procedure is not use on controls at all.