MULTIPLE IMPUTATION OF THE GUESSING PARAMETER IN THE CASE OF MISSING DATA
Grant J. Orley · Lincoln (University of Nebraska) · 2017
Missing data are a significant problem in testing. Research into strategies for dealing with it have yielded no clear consensus about the best approach to take. Accuracy of ability estimates, fairness and scoring transparency are affected by the choice of missing data handling technique. In this simulation study, we propose a technique of multiple imputation of the guessing parameter using both item difficulty and individual ability estimates. This approach was compared to several other popular strategies for imputing values, such as: treating the item as incorrect, imputing a guessing parameter of 0.5, proportion correct imputation, multiple imputation of responses, and multiple imputation of the guessing parameter value. Assessments of the accuracy of ability estimates for each technique were examined in terms of root mean-squared error (RMSE) and bias. These dependent variables were calculated both across the ability continuum and as a function of theta. The implications of these results for real-world testing are discussed. Advisor: Rafael De Ayala