BMA of IRT models from simulated data

Bayesian Model Averaging is a known tool for model selection, estimation, and prediction, which is known to be of lower risk than the common practice of choosing a single model (Fragoso and Neto 2015). Simulated data of 2000 participants were used for IRT models which were then compared using BMA. Predicted Person Scores from each model were then compared as predictors to the True Scores that were simulated and thus known. Person scores and SEs were computed using R code from Rights et al., (2018), with BMA done via R package as suggested by Starkweather (2011). IRT summaries are provided and IIC and ICC plots are supplied in appendix for reference. BMA allows us to consider all models simultaneously and is a reasonable choice since we have multiple viable models present (Hinne et al., 2020). Ultimately, RASCH and 1pl IRT models seem more appropriate for prediction of Person Scores than 2pl or 3pl models, at least for our simulated sample with 5 items.