Analyzing Data with Missing Continuous Covariates by Multiple Imputation Using Proper Imputation

Missing covariate data occur inevitably in various scientific researches. the response variable of interest in these studies may be continuous or categorical and the covariates may have a continuous or discrete nature. multiple imputation (mi) procedures may be used to properly or improperly impute the missing data several times and to find parameter estimates by combining the pseudo-complete-case analyses of the imputed data-sets. although many efforts in the literature have been placed on analyzing continuous response data with missing covariates using mi, models for ordinal response data with missing covariates have received less attention. in this paper four different models for imputation of a missing continuous covariate, of which three are proper and one improper, are compared in models for ordinal responses. all models can be easily implemented in existing software. data from a steatosis study is used to illustrate the use of these models. the importance of using a fuller model for imputation compared to that of the analysis model is finally underlined