Model Selection for Monotonic Polynomial Item Response Models

Abstract

One flexible approach for item response modeling involves use of a monotonic polynomial in place of the linear predictor for commonly used parametric item response models. Since polynomial order may vary across items, model selection can be difficult. For polynomial orders greater than one, the number of possible order combinations increases exponentially with test length. I reframe this issue as a combinatorial optimization problem and apply an algorithm known as simulated annealing to aid in finding a suitable model. Simulated annealing resembles Metropolis-Hastings: A random perturbation of polynomial order for some item is generated and acceptance depends on the change in model fit and the current algorithm state. Simulations suggest that this approach is often a feasible way to select a better fitting model.