A influência da variabilidade dos dados na qualidade de imputação de dados faltantes

Abstract

Imputation methods were developed with the purpose of defining estimates for missing data in a database and, in this way, solving possible problems generated by the loss of such information. In this study the objective is to evaluate if the variability of the data influences the results obtained after applying an imputation method. From complete real databases, from experiments conducted in the Randomized Block Design, some with larger and others with less variability, incomplete databases were generated with the withdrawal of different amounts of data. Subsequently, the Free Distribution Multiple Imputation method was applied, generating complete databases from the imputation. The results of the research confirm the importance of evaluating the variability of data before joining the application of an imputation method to obtain complete databases. For the data of this study, it was verified that the variability of the same influenced in a negative way when high and in cases in which the variability was low the imputed values are closer to the real ones. This confirms the importance of evaluating the variability of the data before choosing to apply the imputation method.