Stabilizing Maximum Likelihood Estimation with a Damping Factor in the Initial Phase of Computerized Adaptive Testing

Maximum Likelihood Estimation (MLE) is a widely used ability estimation method in Item Response Theory (IRT)-based CAT applications. However, traditional MLE is highly sensitive to initial responses, often leading to substantial fluctuations and estimation instability, particularly in short tests or small item pools. This study investigates the effects of incorporating a damping factor into MLE at the early stages of CAT to mitigate undue ability estimate fluctuations. Using Monte Carlo simulations based on a 3-Parameter Logistic (3PL) model in R, we examine the performance of the adjusted MLE compared to standard MLE, Maximum A Posteriori (MAP), and Expected a Posteriori (EAP) estimation methods. Results indicate that damping improves MLE stability, reducing extreme ability fluctuations and enhancing estimation accuracy, particularly in short tests and small sample conditions.