In this paper, we present a case study of the imputation in a complex household survey - the first wave of the German Panel on Household Finances (PHF). A household wealth survey has to be built on a questionnaire with rather complex logical structure mainly because the probes of many wealth items have to be proceeded on both intensive and extensive margins. Hence the number of potential predictors for each imputation model grows and more non-compliance can confront standard modelling due to, e.g., irregular missing patterns, interdependent logical constraints, data anomalies etc. Our model selection procedure borrows the techniques for the out-of-sample prediction to handle the overfitting often associated with the introduction of a large number of predictors. We also take the measures to produce ex ante evaluation for modelling which can be more efficient than the common diagnosis done after imputation in practice. Solutions for the difficulties in the real data and questionnaire structures are also presented. On the other hand, we incorporate the rich flagging information in developing various measures of item-nonresponse to access this complication from logical structure. We find that information loss due to the contagion of item-nonresponse between variables is not serious in our imputed data.
Multiple imputation Model selection Panel on household finance item-nonresponse evaluation