The clinical question--"Which treatment(s) for which patients with what set of subgrouping characteristics working by what mechanism(s)?"--rests at the heart of differential therapeutics. Experimentally, this question reduces to a test of how well we can predict the outcome of treatment using the treatment conditions plus other moderating and mediating variables. Reflecting the discussions held at a recent National Institute of Mental Health (NIMH) conference on psychosocial treatments, and using pediatric anxiety disorders as a case in point, we discuss the problem of prediction in treatment outcome studies from the standpoint of definition of terms, using the general linear model of prediction. We also outline types of studies that may be useful in testing potential predictors, and put forward a possible matrix of predictor variables as currently implemented in an NIMH-funded treatment outcome study of pediatric obsessive-compulsive disorder (OCD). We conclude by making specific suggestions for implementing a broader approach to the study of predictors.