Equating Health Measures: Testing the Practical Implications Final rept. (Sept 30 2000-Sept 29 2003)

摘要

The purpose of the project was to develop evidence-based guidelines for the equating health outcomes measurement. Equating relates scores from different measures upon a common mathematical metric. The trustworthiness of these studies may depend upon choices made in designing the equating study. We evaluated the effects of IRT model choice, sample size, and characteristics of linking items on equating results. 31,500 calibrations were completed and results compared to reference equatings and external health variables. For this report, we rank-ordered the calibrations based on R(sup 2) values. We then selected the '100 worst' calibrations and identified the percentage of time each value of each parameter was represented in this subset of calibrations. Results were compared both across and within models. Sample sizes below 400-500 were overly represented among the 100 worst. The random selection of items for linking worked as well as most, and better than some, item selection procedures. A clear advantage was evident for using 12 or more linking items. In both the within and across model comparisons, the GRM performed the best and the GPCM performed the poorest. The GPCM was the model least likely to converge. The PCM, though it out-performed the GPCM and was the most likely to converge, did not perform as well as the GRM. The results suggest other directions for analyses including identifying 'thresholds' for values of each parameter that produce reliable equating results.