Balancing Two-Way Unbalanced Data with Replication

Mean Substitution, Regression Estimation, Expectation Maximization and Multiple Imputation techniques were considered for imputing missing data in two-way design under the experimental assumption that the missing data follow monotone distribution with varying missing data densities. The performance, relative efficiency and suitability of the techniques were evaluated and compared using the standard error, the root mean square error and the relative efficiency index as statistical tools. The relative efficiency index is a better measure of efficiency and more powerful than the standard error and the root mean square error. Moreover, the Euclidean Distance is a sufficient measure of relative efficiency and appropriate for evaluation of trend of performance of techniques across levels of missing
data. All techniques evaluated performed relatively more efficient at lower percent of missing data; therefore, the efficiency of the data imputation techniques decreases with increasing proportion of missing data. The Regression Estimation technique is the most stable and efficient method for intermediate and high density missing data; while the Mean Substitution and Multiple Imputation technique are most efficient for low density missing data.