A Comparison of the Prediction Accuracy for Latent Traits of Replenished Items in Multidimensional Adaptive Testing

The purpose of this study was to predict latent traits of replenished items in multidimensional adaptive testing. To do this, a simulation study was conducted using LASSO method to predict the latent traits of items measured. Simulation data were generated by combining test length, item discrimination, covariance matrices and item selection methods. This study evaluated correct specification rate, item parameter estimation accuracy, test overlap ratio and item exposure bias to determine optimal testing condition. The main results are as follows. First, correct specification accuracy exceeded 80% across most conditions. Second, item selection methods optimized for multidimensional adaptive testing are better than traditional method in every conditions. Third, correct specification rate of item selection method that considered prior distribution showed high even with short test lengths. Fourth, longer test length, better item parameter estimation accuracy, test overlap ratio and item exposure. Based on these findings, implications for predicting the latent traits of new items and suggestions for future research are discussed.