ABSTRACT
The topic addressed in this chapter is, how many readers and cases, usually abbreviated to sample size, should one employ to conduct a ROC study? As noted in Chapter 08, a typical design sets alpha equal to 5% and beta equal to 20%; the latter corresponds to 80% power. Sample size estimation involves: (a) performing a pilot study in order to determine the magnitudes of the variance components and (b) making an educated guess regarding the true performance difference between the two modalities, termed the anticipated effect size. Since the non-centrality parameter was defined in DBMH Chapter 09, in this chapter sample size estimation is illustrated using the DBMH method. Hillis has derived the conversions between DBMH and ORH parameters. An online appendix describes the corresponding implementation using the ORH method. Two examples, widely used in the methodology literature, are given using functions in RJafroc. A cautionary note by Kraemer and colleagues, regarding the use of the observed effect size as the anticipated effect size, is summarized. Prediction accuracy of sample size estimation, investigated in a recent study, is summarized. The approach in this chapter is to utilize the confidence interval for the observed effect size as a guide in choosing the anticipated effect size. It is shown that the common practice of specifying effect size as the difference of two AUCs can be misleading. Suggested is an alternative that uses the relative change in the separation parameter corresponding to a given AUC as a measure of effect size.
