Applications of computerized adaptive testing (CAT) to the assessment of headache impact.

Abstract

To evaluate the feasibility of computerized adaptive testing (CAT) and the reliability and validity of CAT-based estimates of headache impact scores in comparison with 'static' surveys. Responses to the 54-item Headache Impact Test (HIT) were re-analyzed for recent headache sufferers (n = 1016) who completed telephone interviews during the National Survey of Headache Impact (NSHI). Item response theory (IRT) calibrations and the computerized dynamic health assessment (DYNHA) software were used to simulate CAT assessments by selecting the most informative items for each person and estimating impact scores according to pre-set precision standards (CAT-HIT). Results were compared with IRT estimates based on all items (total-HIT), computerized 6-item dynamic estimates (CAT-HIT-6), and a developmental version of a 'static' 6-item form (HIT-6-D). Analyses focused on: respondent burden (survey length and administration time), score distributions ('ceiling' and 'floor' effects), reliability and standard errors, and clinical validity (diagnosis, level of severity). A random sample (n = 245) was re-assessed to test responsiveness. A second study (n = 1103) compared actual CAT surveys and an improved 'static' HIT-6 among current headache sufferers sampled on the Internet. Respondents completed measures from the first study and the generic SF-8 Health Survey; some (n = 540) were re-tested on the Internet after 2 weeks. In the first study, simulated CAT-HIT and total-HIT scores were highly correlated (r = 0.92) without 'ceiling' or 'floor' effects and with a substantial reduction (90.8%) in respondent burden. Six of the 54 items accounted for the great majority of item administrations (3603/5028, 77.6%). CAT-HIT reliability estimates were very high (0.975-0.992) in the range where 95% of respondents scored, and relative validity (RV) coefficients were high for diagnosis (RV = 0.87) and severity (RV = 0.89); patient-level classifications were accurate 91.3% for a diagnosis of migraine. For all three criteria of change, CAT-HIT scores were more responsive than all other measures. In the second study, estimates of respondent burden, item usage, reliability and clinical validity were replicated. The test-retest reliability of CAT-HIT was 0.79 and alternate forms coefficients ranged from 0.85 to 0.91. All correlations with the generic SF-8 were negative. CAT-based administrations of headache impact items achieved very large reductions in respondent burden without compromising validity for purposes of patient screening or monitoring changes in headache impact over time. IRT models and CAT-based dynamic health assessments warrant testing among patients with other conditions.