Nonrandomly Missing Data in Multiple Regression Analysis: An Empirical Comparison of Ten Missing Data Treatments

Abstract

This study investigated the effectiveness of ten missing data treatments within the context of a two-predictor multiple regression analysis with nonrandomly missing data. Five distinct types of missing data treatments were examined: deletion (both listwise and pairwise methods), deterministic imputation (with imputations based on the sample mean, simple regression and multiple regression), stochastic imputation (mean, simple regression and multiple regression), maximum likelihood estimation (ML) and multiple imputation (MI). Design factors included in the study were sample size, total proportion of missing data, and the proportion of missing data occurring in the upper stratum of each predictor. The success of each method was evaluated based on the sample estimate of R^2 and each standardized regression coefficient. Results suggest that the stochastic multiple regression imputation
procedure evidenced the best performance in providing unbiased estimates of the parameters of interest. Deterministic imputation approaches and the stochastic mean imputation approach resulted in large amounts of bias in the estimates.