Use of Correlated Data Methods in Ophthalmology

点击翻译按钮获取中文摘要

Project Summary / Abstract Some exposures in ophthalmology may not have an immediate effect, but instead a lag is necessary. For example, there is a literature on possible cataractogenic effects of corticosteroid eyedrops (CS) among uveitis patients. However, the precise impact of dose and/or duration of use are unknown. Also, the lag between CS administration and development of cataract is unknown. Another possible application is to study the effects of dietary and/or supplement use on development of AMD, where a lag effect is also likely to occur. The goal of specific aim 1 is to use latency analysis methods for ophthalmological endpoints. Latency methods have been used in pharmacoepidemiology, but to our knowledge, have never been used for ophthalmologic endpoints. The AREDS study was a landmark study in the epidemiology of AMD. A byproduct of this study was the development of the AREDS severity scale which is an ordinal scale ranging from 1 for no AMD to 9+ for advanced AMD (AAMD). The usual analysis for risk factors is a time-to-event analysis based on the Cox Proportional Hazards Model, where the event is reaching grade 9+. This is a valid, but inefficient analysis. There are many eyes (about 40%) which show changes (either an increase or decrease), but which don't develop AAMD. There are risk factors which are associated with these changes, but all such changes are treated as censored data and are considered “non-events”. In Aim 2, we propose to use an ordinal regression model for changes between successive visits which would provide a more efficient use of the data. There have been previous multi-state ordinal models proposed, but separate models are fit for each possible transition and are not integrated into an overall assessment of risk for specific covariates. This has application not only for AMD, but also for other ordinal scales used for other ophthalmologic conditions, such as diabetic retinopathy. For Aim 3, we propose to continue our work on applying correlated data methods to risk prediction for endpoints such as AUC. We will specifically compare methods for estimating AUC for small samples, extensive numbers of tied prediction scores and presence of both bilateral and unilateral subjects. In addition, we will incorporate clustered data methods for estimation of NRI, which to our knowledge, has never been done before. In Aim 4, we will continue our work on translation of clustered data methods for the eye research community including (a) correlated data methods in survival analysis, (b) analysis of longitudinal binary ocular data, and (c) sample size/power calculations based on the eye as the unit of analysis.