Abstract
Purpose:
To apply alternative statistical approaches to evaluate the parameters most closely associated with a dichotomous indicator of patient status (affected (AFF) or unaffected (UNAFF)).
Methods:
In 57 patients treated with plaquenil, multifocal ERG testing was used as the “gold standard” to determine AFF (n=19) or UNAFF (n=38) status. Multiple other clinical parameters were then assessed for their association with affected status: retinal thickness (RT), determined by optical coherence tomography (OCT) for nine ETDRS subfields; mean deviation (MD) from Humphrey 10-2 visual field (VF) testing; and color fundus and autofluorescence imaging grades. To identify the parameters most strongly associated with affected/unaffected status, we first used a cross-validation approach, creating two subsamples via a random number generator. For each parameter, we fit a logistic regression model on the first subsample to compute area under the ROC curve (AUROC) and identify the optimal cutpoint, using Youden’s J. We then determined the sensitivity and specificity of that optimal cutpoint in the second subsample. Our second analytic approach used polynomial models to identify the best-fitting curve to the data for each parameter, followed by a series of hierarchical models to find the overall best-fitting model, while adjusting for age and other pertinent covariates.
Results:
Approach 1 (cross-validation): For inner inferior RT, the optimal cutpoint for the first half-sample was 278.5 um, with AUROC of 0.98, sensitivity of 100%, and specificity of 88.2%. When this cutpoint was applied in the second half-sample, results were not as strong: sensitivity=75%, specificity=90.5%. When the second half-sample was analyzed independently, the optimal cutpoint was 245.6 um and AUROC was 0.94. Approach 2 (polynomial modeling): The linear model had the best fit for all parameters (based on -2 log likelihood). Inner inferior RT (odds ratio, 0.94; 95% confidence interval (CI), 0.89 - 1.00, p=.045) and VF MD (odds ratio, 0.55; 95% CI, 0.31 - 0.99, p=.047) were the variables most closely associated with AFF/UNAFF status.
Conclusions:
The difference between optimal cutpoints for the two half-samples shows a shortcoming of the cross-validation approach to identify a generalizable optimal cutpoint. Results of the polynomial modeling approach were consistent with results based on computations of AUROC.
Keywords: 464 clinical (human) or epidemiologic studies: risk factor assessment •
503 drug toxicity/drug effects •
688 retina