Partial non-response to financial variables of the 2006 Insee Housing survey#New correction procedures and imputation methods comparisons.

Sophie O’PREY

The Insee Housing survey contains financial information in which appear a number of missing values - "Refusal" or "Do not know" - and outliers. It is necessary to correct these variables before processing them because non-response will produce a bias in estimation and reduce the accuracy of estimators. The procedures described in this document concern two types of variables possessing certain particularities. First of all, figures for financial investments and rent had been under-estimated and so had to be corrected. Secondly, with the question about income it proved necessary to resort to imputation methods taking interval-censored data into account, since it was possible to answer in two ways, by giving an exact figure or else by indicating the interval in which it was comprised. Further to this study, some of the imputation procedures were compared. As regards the distribution of the corrected variables, there proved to be little difference in cases where areas of interest were large enough because of the generally low non-response rates and the sufficiently large sample size. However estimations for small sub-populations were of poorer quality, in particular as regards ratios which featured imputed variables both in the numerator and denominator, such as housing budget share. In this case it is preferable to use small area estimation or at any rate to take into account the confidence intervals of estimators, especially if one wishes to make comparisons in time or between sub-populations.

Documents de travail
No F0901
Paru le : 01/02/2009