Confronti tra misure di accuratezza per la stima di mancate risposte in indagini longitudinali

Abstract

In this paper, dealing with the problem of estimating missing values concerning a certain Y variable, we have first tried to define 8 different estimators, derived from a general imputation model. Then 7 different accuracy measures (or quality indicators of the imputation process) have been proposed and commented. Finally, starting from a concrete and strongly skewed population an ignorable non response mechanism has been simulated, in order to compare estimators and quality indicators. Results showed that it could be useful to compare estimators using more than one quality indicator only. In particular, both the difference among true and estimated values and the discrepancy existing between the true and the partially estimated distribution should be ever evaluated, in order to point out the best estimator whatever could be the population density form.