Abstract
Generalizability theory is a theory enabling comprehensive reliability analyses by dealing with several variability source like item, task, rater, time, etc. However, missing data causes an inability in making analyses and predictions. In literature, there is limited number of studies comparing missing data imputation methods in generalizability theory. In this direction, the aim of the study is to compare G and Phi coefficients obtained from generalizability theory analyses according to full data matrix and missing data imputation methods with different sample sizes and missing data rates. In this research, a Monte Carlo simulation study was conducted in order to compare performances of missing data imputation methods for missing data for two-facet crossed designs differently from limited studies in literature. In the simulation study, 36 conditions as “number of examine x missing data ratio x method (3 x 4 x 3)” were investigated. In addition, G and Phi coefficients were calculated for full data matrices for each sample size. Bias between G and Phi coefficients was compared by using imputation methods with these calculated values. As a result of this comparison, it is found that reliability predictions closest to the full data come from EM method instead of missing data. If EM is used instead of missing data, reliability predictions quite similar to reliability coefficients predicted from full data matrix is found in most of the conditions. The lowest level of relative error and bias for reliability coefficients is found with EM and LR methods respectively. PM is the method with the highest error and bias considering all conditions. In addition, the current study found that as the missing data rate increase, difference, error and bias values increase for all methods according to full data matrix.
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call