Experiments with Variance Estimation from Survey Data with Imputed Values
Hyunshik Lee, Eric Rancourt, and Carl E. Särndal
Missing data occur in almost all surveys and frequently some form of imputation is used to obtain a completed data set. It is well known that the ordinary variance formula applied to the data with imputed values generally underestimates the variance. There have been some proposals to remedy this problem. One is the well known multiple imputation. This method, however, requires generating two or more completed data sets, which may be seen as a disadvantage in some applications. Recently, Sa¨rndal proposed a variance estimation method for single imputation using a model-assisted approach to the problem. Rao also proposed a method based on the two-phase sampling approach for single imputation. In this article, these methods are studied by the Monte Carlo technique together with other methods including multiple imputation methods, under 12 artificially generated populations representing a variety of forms and three response mechanisms of which two are confounded, that is, they depend on the values of the variable of interest.
Model-assisted approach; ratio imputation; multiple imputation; Monte Carlo study.