Function rocTest

Tests the difference of the areas under two ROC curves, by DeLong's method or the bootstrap. Equivalent to pROC::roc.test() in R.

Curves from the same response are paired, as pROC pairs them (identical responses; where a predictor is missing, both curves are rebuilt on the observations where neither is). Paired DeLong: Z = (θ₁ − θ₂)/σ with σ² from the placements' covariances (DeLong, DeLong and Clarke-Pearson 1988), Z = 0 when both are 0, and the interval θ₁ − θ₂ ∓ zσ. Unpaired DeLong: D = (θ₁ − θ₂)/√(S₁ + S₂) on pROC's Welch-Satterthwaite degrees of freedom (when both variances are 0, D = 0 with p-value 1 if the AUCs are equal, and the degrees of freedom are NaN). Bootstrap: D = (AUC₁ − AUC₂)/sd of the resampled differences, with R's random stream. DeLong's method needs at least two controls and two cases in each curve, as does the stratified bootstrap. The method is DeLong's unless given (pROC switches to the bootstrap silently for curves of different directions; here DeLong's test gives pROC's warning).

Tests the difference of the areas under the ROC curves of two predictors of the same response (paired); observations with a missing response or predictor are dropped from both.