Comparison of Imputation Methods in Split Plot Design

Abstract

The observed response variable measured in split plot design usually contained missing observations. This missingness could be as a result of the death or non-response of some of the experimental units, wrong application of treatment and measurement error. In this paper, four imputation methods: stochastic regression imputation (SRM), multiple imputation (MI), expectation-maximization (EM) algorithm and regression method (REM) for handling missing data in split plot design organized in randomized complete block design, SPD (RCBD) when data are missing at random using Monte Carlo simulation were evaluated and compared. The data were simulated using four percentages of missingness 5%, 10%, 15% and 20% for the response variable of SPD (RCBD). The comparison of these methods was based on mean square error (MSE), standard error (SE), coefficient of determination (R2), correlation coefficient (R) and p-value to determine the performance of the imputation methods. The results showed that multiple imputation method performed better than the other methods based on mean square error, standard error, coefficient of determination and correlation coefficient followed by expectation-maximization algorithm based on p-value criterion in the estimation of missing values.