Estimação de variância em pesquisas por amostragem com dados imputados pelo método Hot-Deck: Uma comparação empírica

Abstract

In many sample surveys, it is common the presence of missing data, a problem also known as non-response. This problem occurs due to possible failures in obtaining some of the information that should be collected in the sample. Ignoring non-response, by analyzing the observed data only, is a simple way to approach the problem, but it could produce bias in the survey results. One of the techniques for treating non-response is by imputation, that is by replacing the missing data with appropriately chosen values, seeking to eliminate or reduce the non-response bias. Using imputation, the data set is completed so that valid estimates of population parameters and their variances could be obtained more easily. However, the validity of such estimates needs to be carefully evaluated, as imputation may also be subject to estimation bias. In this work, a Monte Carlo simulation study is used to evaluate empirically statistical properties of estimators of a finite population mean under a missing completely at random and missing-at-random response mechanisms. The population mean estimators used in the study correspond to the average of the data observed among the sample respondents and the average of the data after hot-deck imputation, which is a popular imputation method. The study also compares variances estimators of methods based only on the respondent data, treating of imputed data as true values, application of the usual Jackknife, application of the Rao and Shao Jackknife, Multiple Imputation, and the Approximate Bayesian Bootstrap. The results of the study demonstrate, for each non-response mechanism, which methods estimate appropriately the population parameter of interest and the respective variances.