Comparison of Methods of Estimating Missing Values in Randomized Complete Block Designs (RCBD) for Various Number of Missing Values

Abstract:

In virtually all types of research, missing value(s) are common occurrences. Several methods and techniques are available for handling this issue. This study aims at comparing the result of using some of the techniques (Do nothing, List-wise Deletion, Overall Mean Imputation, Group Mean Imputation and Least squares Imputation techniques) in handling missing values in Randomized Complete Block Designs (RCBDs). Research data on outcome of an experiment conducted in July 2010 in the department of Animal Science, University of Nigeria Nsukka was employed. The data is made up of a random sample of size 36, on the effect of stocking Densities on the weight of birds at varying ages. Weight gain was used as parameter for measurement. Missing value(s) were introduced into the original complete data set randomly in three different levels of n = 1, 2 and 3. A two-way analysis was carried out, when the data is complete and when there are missing value(s) using the different methods of handling missing values considered. Results showed that the model assumptions of the RCBD was the same, both when the data is complete and when using the different methods of handling missing value(s) employed at different levels of n = 1,2 and 3 missing value(s). There is significant effect of the stocking densities when the data is complete and when using the different method of handling missing value(s) at different levels of n = 1, 2 and 3 missing values. On the contrary, there are significant differences between the MSE of the analysis with the complete dataset and when the different methods of handling missing value(s) at different levels (n=1, 2 and 3) of missingness are created. The complete data showed an MSE of 0.538. On the other hand, for n=1, 2 and 3 missing values respectively, the “Do Nothing” technique generated MSEs of 0.559, 0.573 and 0.577 (average 0.570). Listwise Deletion showed MSEs of 0.564, 0.534 and 0.714, (average 0.604). Random Mean Imputation showed MSEs of 0.537, 0.629 and 0.776, (average 0.657). Group Mean Imputation generated MSEs of 0.553, 0.704 and 0.785, (average 0.681). Least Squares Imputation produced MSEs of 0.553, 0.527 and 0.527, (average 0.530). Hence, the Least Squares Imputation, with consistently small MSEs and the closest average MSE to the true MSE, is recommended among the methods studied.