Clustered data occur frequently in ophthalmologic research. One always has two eyes per individual (one level of nesting) and, in some instances, one has multiple measurements obtained on an eye (e.g. assessments of visual field in different regions of an eye) (more than one level of nesting). Thus, the appropriate unit of analysis is an omnipresent issue. The ideal solution in the one level of nesting case is to use the eye as the unit of analysis, and account for the correlation between eyes while performing the analysis. Several newer methods have been developed in the last 5-10 years to i accomplish this goal, including the general linear model approach (Rosner, 1984) for normally distributed outcome variables; the polychotomous logistic regression approach (Rosner, 1984) and estimating equation approaches (Liang and Zeger, 1986) for binary outcome variables. In the previous grant proposal, these methods were compared with simpler, more commonly used methods, including analysis of the left and right eye separately or an analysis based on the better or worse eye. These methods were compared on two real ophthalmologic datasets and revealed important differences between the newer vs. standard methods. An important goal for this renewal application is to extend these investigations using simulation methods where the appropriate model is fixed by design and one is interested in whether important differences emerge among methods. In addition, there is an important need to extend the methodology for clustered data to the areas of: (i) ordinal clustered data, where an eye is scored on an ordinal rather than on a cardinal or binary scale, (ii) survival data in a clustered data setting, where time to failure is the key outcome variable, (iii) clustered data with a more general correlation structure such as in the analysis of visual field data with region-specific rather than eye-specific outcome information. A key issue in the development of new methodology is to provide software that is available to a maximum number of statisticians, epidemiologists and ophthalmologists. We have made important strides in converting some of our software to a PC environment and, in some cases, to be SAS-compatible and will continue to be mindful of these issues in the development of new methods. The research has important implications for ophthalmology and in many other clinical specialties (e.g. otolaryngology, dentistry, coronary artery disease) where clustered data are the rule rather than the exception.
Showing the most recent 10 out of 19 publications
