Imputation techniques in regression analysis: Looking closely at their implementation

Abstract

A problem which frequently arises in regression analysis is the presence of missing values on the explanatory variables. Imputation is a time-honoured approach to tackling it, since graphical exploration of properties of a statistical model requires a complete data matrix. This article examines the performance of five imputation techniques in two frequently used implementation procedures. Specifically, imputed values based on both the response and explanatory variables (type I) are contrasted with those based on only the explanatory variables (type II). Monte Carlo results indicate that imputed values with type I procedure may give spurious impression of high precision especially as the proportion of missing data increases. But with type II, overestimation of residual mean square error may arise. Several matrices of correlation coefficients are used and an illustrative real data example is given.