Ophthalmic data is of necessity bivariate. Also, many scales used in ophthalmology are ordinal in nature. For example, the ETDRS diabetic retinopathy scale is an ordinal scale ranging from 10 = no retinopathy to ? 60 = proliferative diabetic retinopathy (PDR). Most commonly, the ordinal grades for the eyes of a subject are combined into a 15 category ordinal level scale and progression is determined by a 2 or 3 level change in the person-level scale. However, much information is lost by collapsing an eye level scale to a person-level scale since there can be differences in retinopathy grades and progression in retinopathy for individual eyes. Furthermore, information is also lost by categorizing change as ? 2 level or 3 level change rather than treating the scale as an ordinal scale. Similar issues arise in AMD where drusen size is used to characterize subjects with early AMD. Drusen size is usually coded as an ordinal variable and one wants to look at risk factors that are associated with change in drusen sizes. The goal of specific aim 1 is to propose new analyses of clustered ordinal longitudinal data that would be applicable to diabetic retinopathy scale and drusen size. The goal of specific aim 2 is to assess risk prediction for ophthalmic endpoints such as AMD using the eye as the unit of analysis. A common method used to assess discrimination of risk prediction rules is the area under the ROC curve (or C statistic). However, most literature concerning estimating the C statistic is based on the person as the unit of analysis where each person is considered an independent unit of analysis. In the previous cycle of this grant we developed methods for assessing AUC for ophthalmic endpoints using the eye as the unit of analysis, based on logistic regression models where each person is followed for the same length of time.
In specific aim 2, we now propose to extend this method to the case of familial data such as in the Seddon longitudinal cohort where there can be several people in the same family and one has two levels of nesting for person within family and eye within person. Another issue in risk prediction is that in some studies subjects are followed for variable lengths of time (such as the AREDS study). There are extensions proposed of the C statistic that are applicable to survival analysis, but they use the person as the unit of analysis.
In specific aim 3, we propose to extend these methods based on the eye as the unit of analysis. Finally, in specific aim 4 we propose to continue our work on translating statistical methods for clustered data to the ophthalmologic community. This would be achieved by a combination of (a) papers (b) posters and/or presentations at important ophthalmologic meetings (c) giving short courses on Intermediate Statistical Methods at ARVO. Although there has been some movement towards using the eye as the unit of analysis in ophthalmology, the issue is frequently ignored and people are often presenting results based on the person and/or a single eye with a consequent loss of power. Hence, specific aim 4 is of high priority in virtually every specialty of ophthalmology.
This proposal will consider (a) more efficient use of longitudinal data in ophthalmic studies based on ordinal scales, with the eye as the unit of analysis (b) assessment of measures of discrimination for risk prediction rules in ophthalmic studies collected in families where the eye is the unit of analysis (c) assessment of measures of discrimination for risk prediction rules with variable follow-up time using the eye as the unit of analysis (d) innovative approaches for translating work on clustered data methods to the ophthalmic community.
