Abstract
PURPOSE: Application of computerized adaptive testing (CAT) can improve the assessment of patient-reported health outcomes by reducing patient burden. We aimed to reduce patient burden of CATs further by optimizing a standard error reduction stopping rule (SER; minimum change in SE(θ) after each CAT step). METHODS: We extracted PROMIS Anxiety and Depressive Symptoms CAT responses (mean age = 13.7, male = 50.3%) from the Dutch-Flemish PROMIS Assessment Center and estimated theta levels (θ) and standard errors (SE(θ)) for each step. The default stopping rules were a minimum/maximum of 4/12 items administered, respectively, or a minimum precision of SE(θ) < 0.32. We imposed increasing SER thresholds (0.01-0.20) and compared the following outcome criteria: mean efficiency of the CAT (M(efficiency); 1 - SE(θ)(2)/n(items)), mean number of items administered (Mn(items)), the mean SE(θ) of all respondents (M(SE(θ))), and mean T-score difference compared to default stopping rules (M(∆T)). RESULTS: Default stopping rules showed a mean efficiency of 0.88 and1.27, Mn(items) = 9.98 and8.13, and M(SE(θ)) = 0.36 and0.38 for respectively the Anxiety and Depressive Symptoms item banks. We optimized the SER value with a differential efficiency function, resulting in shorter, more efficient CATs (Anxiety: mean efficiency = 1.08, Mn(items) = 5.58, M(SE(θ)) = 4.24, M(∆T) = 0.04; Depressive Symptoms: mean efficiency = 1.45, Mn(items) = 4.79, M(SE(θ)) = 4.15, M(∆T) = 0.58). For participants reporting no problems, this results in fewer items administered, but a decrease in measurement accuracy and biased T-scores, which may be relevant depending on the goal of assessment. CONCLUSIONS: We conclude that the current approach allows us to determine an optimal SER threshold that improves measurement efficiency, especially when floor/ceiling effects are present in the target population. The threshold values will vary depending on the θ distribution of the target population and the IRT model parameters.