An important step in providing treatment is prediction of which treatments will be effective for which patients. Common methods for prediction of efficacy, however, are inconsistent with the assumptions of the standard placebo-control paradigm for establishment of efficacy. This is because in prediction of the observed response, one ignores the distinction between response attributable to non-specific factors and that attributable to active-treatment factors. This paper presents a paradigm for the use of log-linear analysis that allows for the development of predictive methods that take into account this distinction. An example with simulated data demonstrates how, if this distinction is ignored, one can reach misleading conclusions and make non-optimal treatment decisions as a result of an inaccurate cost-benefit analysis of the treatment.