Use of receiver-operator characteristic (ROC) curve analysis to evaluate predictors of response to clomipramine therapy

Abstract

We used receiver-operator characteristic (ROC) curve analysis to evaluate predictors of response to clomipramine in obsessive-compulsive disorder (OCD). Previously, we identified response predictors among 230 OCD patients who received clomipramine in a placebo-controlled, multicenter clinical trail. We found that at baseline a later age of OCD onset, low scores on the Hamilton Depression scale, and high scores on items 3 and 8 of the Yale-Brown Obsessive Compulsive Scale predicted good response. Certain early side effects also predicted outcome. We fitted a logistic regression model containing baseline information and then calculated each patient's estimated response probability by substituting individuals' values in the regression equation. Next we compared the estimated response risks with each patient's known outcome. Finally, we produced a ROC curve by plotting the true positive and false positive rates for various cutoff points of the risk scores. The same steps were followed for Weeks 1 through 4, adding information about early side effects and weekly response. We found that baseline information predicted outcome better than chance, and predictive ability increased with data on side effects and early response.