Purchase this article with an account.
Patrick Johnston, John Rodriguez, Keith Lane, Richard Abelson; A Simple Approach to the Analysis of Correlated Dry Eye Data. Invest. Ophthalmol. Vis. Sci. 2012;53(14):544.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Group comparison studies in dry eye research typically include subjects in independent groups (treatment or placebo, normal or dry eye, etc.) with measurements (staining scores, blink rate, etc.) taken on both eyes. Since these pairs of observations are usually positively correlated, analysis is properly conducted by correlated data models, such as those based on generalized estimating equation (GEE) methods. This study showed that in the highly balanced context of correlated eye data, average eye scores analyzed by simpler and more familiar independence models (t-tests and their generalized linear model (GLM) generalizations) provide identical results. This independence model alternative is proposed as a straightforward option for statisticians and ophthalmic researchers alike.
Two methods were compared: (a)correlated data pairs analyzed by a GEE correlated data model, and (b)average scores analyzed by a GLM independence model. Robust standard errors were used for both GEE and GLM models, and a variety of working correlation structures were considered for GEE models.Several common data types were considered: normal linear models for continuous data (staining scores), gamma multiplicative models for continuous data (tear film breakup time), Poisson multiplicative models for count data (blink rate), and binomial logistic models for binary data (proportion of subjects with no tear film breakup).Additional generalizations and limitations were investigated, including adjustment for within-subject and between-subject covariates, missing data, and other combinations of link functions and probability distributions.
For all of the data types outlined above, and for any other combination of link function and probability distribution, identical results (estimates of group means and their standard errors) were obtained from both methods: correlated data models applied to correlated data pairs, and independence models applied to average eye scores.
Independence GLM models are simpler and more familiar than correlated data GEE models. This study showed that, in the context of correlated eye data, and subject to certain limitations such as missing data, both methods gave the same results, thus providing the researcher with a simple approach to the analysis of correlated eye data.
This PDF is available to Subscribers Only