Increasingly in biomedical studies, health status is inferred through a series of questionnaire item responses. Challenges for analyzing associations between such responses and risk factors include multiplicity - many indicators must be combined to derive summary statements about health status, and measurement error persons' self-report fluctuates due to causes other than substantive health changes. In order to deal with these challenges, the authors propose a strategy which comprises three methods: 1) score the item responses, then regress the score on predictors; 2) regress each item response on predictors, accounting for within-person associations; and 3) summarize and analyze the item responses jointly, using a latent variable model. The authors develop modeling and diagnostic procedures for method 3. They then show how the three-method analytic strategy can be used to solve the problem of determining which aspects of vision are associated with self- reported functioning in activities that require seeing at a distance. They demonstrate that methods 2 and 3 illuminate basic findings from method 1 by adding specificity, describing patterns as well as severities of health impairments, and identifying isolated items that relate to risk factors differentially than others. They conclude that the three-method strategy specifies how risk factors determine questionnaire-based health outcomes substantially better than any of the methods in isolation.
|Number of pages||14|
|Journal||American Journal of Epidemiology|
|State||Published - 1 Dec 1999|
- Discrete data
- Latent class
- Latent variable
- Multivariate regression