Obtaining the power for a ROC analyses

The below is from Thom Baguley (delivered to psych-postgrads jiscmail list):

Power calculations for ROC analyses can be obtained in R and are the same for apriori and post-hoc scenarios:

install.packages('pROC')
library(pROC)
power.roc.test(ncases=41, ncontrols=72, auc=0.73, sig.level=0.05)

     One ROC curve power calculation 

         ncases = 41
      ncontrols = 72
            auc = 0.73
      sig.level = 0.05
          power = 0.9897453

The problem is where to get the true population AUC from and what it means. If you plug in an observed AUC then power is simply a monotonic function of the p value (or something close to it) and hence uninformative. (Power will be high for significant effects and low for non-significant ones regardless of the true power of the test). If you plug in the true population AUC you get the actual power, but the true value is unknown (or worse still may not be fixed but is a random variable depending on many other variables). The best practice here is either not to do it or be cautious and realise that it means if the true auc were .73 (which it almost certainly isn't) then an identical replication (which is a near-impossible ideal) would have 99% power.