Comparison of data sampling methods on IRT parameter estimation

Abstract

Data sampling methods are promising for analysis of large-scale data sets to reduce computing time and resources. These methods include uniform (random), and leverage-based sampling methods with a recent one called shrinkage leverage-based method. In this study, we compared data sampling methods for accuracy of item parameter estimates in IRT models. In addition, we introduced a new method of sampling, adjusted shrinkage leverage-based (Adj-SLEV) method. We analyzed two samples from PISA 2012 mathematics data set that were normally and non-normally distributed. Random sampling provided the most accurate Rasch item parameter estimates. The method with the highest accuracy varied depending on the type of item parameter for 2-pl and 3-pl models, if each parameter was evaluated individually. Adj-SLEV did not necessarily provide the highest accuracy for each type of item parameter individually, however, consistently provided a good trade-off when all parameters in a model were evaluated together.