Because trials to compare digital and analog mammography rely on complex designs, cost large amounts of money and involve the health of many women, accurate power and sample size calculations are essential. In these trials, radiologists read many films using different sorts of digital machines, and with varying forms of image processes. Measures of the accuracy of detection are the hit rate, and the false alarm rate. Each doctor generates several accuracy scores, one for each condition. These measures are correlated. If one uses an accuracy measure for analog machines as a predictor of accuracy on digital machines, another layer of complexity is added. This experimental design leads to analyzing a General Linear Multivariate Model for repeated measures, both fixed and random predictors, and a full model in every cell. For this design, several sorts of power calculations are proposed. The conditional power can be calculated for a specific set of predictor values. Unconditional power is the expected value of conditional power over all possible stochastic realizations of the random predictors. Quantile power is the conditional power with the noncentrality that corresponds to a realization of the random predictors that occurs with specified probability. The theoretical derivations will lead to easily used software and descriptive examples. These will give radiologists tools for rapid calculation and an interactive, graphical approach to sample size determination.
Showing the most recent 10 out of 13 publications
