Variance Estimation for Regression Imputed Quantiles, A first Step towards Variance Estimation for Inequality Indicators

In a sample survey only a sub-part of the selected sample has answered (total non-response, treated by re-weighting). Moreover, some respondents did not answer all questions (partial non-response, treated through imputation). One is interested in income type variables. One further supposes here that the imputation is carried out by a regression. The idea presented by Deville and Särndal in 1994 is resumed, which consists in constructing an unbiased estimator of the variance of a total based solely on the known information (on the selected sample and the subset of respondents). While these authors dealt with a conventional total of an interest variable y, a similar development is reproduced in the case where the considered total is one of the linearized variable of quantiles or of inequality indicators, and that, furthermore, it is computed from the imputed variable y. By means of simulations on real survey data, one shows that regression imputation can have an important impact on the bias and variance estimations of inequality indicators. This leads to a method capable of taking into account the variance due to imputation in addition to the one due to the sampling design in the cases of quantiles.

, COMPSTAT, Geneva