Dear Murthy, variance estimation methods for singly imputed data have been around for quite a while now and there is a comprehensive review of these in chapters 21 and 22 of the J. Wiley book Survey nonresponse produced after the 1999 Portland conference on nonresponse. There are the two-phase approach (with a nonresponse model), the model approach (with a data model), jackknife and bootstrap to name a few.
In these approaches, the imputation classes need not match exactly the strata. What is more important is the quality of the modelling that leads to a given imputation method and this includes the construction of imputation classes. Here are practical details on two of the approaches: Bootstrap: When you do your bootstrap imputation, you have to use the same imputation model and the same imputation class definition as for the full sample. You want to make sure that you are not using the strata for imputation classes but really the imputation classes. Model approach: When you compute the residuals, it has to be within the imputation classes and not the strata. In the formulas, the strata come into play only through the final weight. This is because in deriving the formulas with the model approach, the sampling weights - and therefore strata - are constants. Eric Rancourt Statistics Canada -----Message d'origine----- De : nmi13 [mailto:[EMAIL PROTECTED] Envoyé : 11 février, 2004 16:55 À : impute Objet : IMPUTE: Re: variance estimation. Dear any, I have a doubt on variance estimation in the presence of singly imputed data? I would be thankful if any one can please help me. Q: Let me say that I have imputed the data using K=5 imputation calsses, these K imputation classes are different from the strata of the original survey? Now if I find the variance using stratified design, rather than the original survey design which is Multistage stratified design. Is it correct or worng to do it in this way? Thanks in advance for any help, comments, suggestions and corrections on this question. Regards Murthy.N.Mittinty