Technical reasons are presented as to why therapist should be included as a random design factor in the nested analysis of (co)variance (AN[C]OVA) design commonly used in psychotherapy research. Incorrect specification of the ANOVA design can, under some circumstances, result in incorrect estimation of the error term, overly liberal F ratios, and an unacceptably high risk of Type I errors. Review of studies indicates that the great majority of investigators continue to ignore this issue. Computer simulation studies revealed that considerable bias can be introduced by not specifying therapist as a random term. Finally, a reanalysis is presented of data from 10 psychotherapy outcome studies that indicated that therapist effects vary considerably and at times can be large. More recent studies that implement better quality controls appear to demonstrate less variance due to therapist. The implications of these results for the design of future studies are discussed.In most psychotherapy research studies, each participating psychotherapist sees more than one patient. Over 10 years ago, Martindale (1978) presented evidence that this apparently minor methodological detail was an important statistical consideration. On the basisof reanalysesofseveral examplesdrawn from the published literature, he concluded that the failure to include the therapist as a blocking or stratification factor in the statistical design could seriously distort testsof statistical significance.The implications of Martindale's paper were sobering. His methodological opinion was that, strictly speaking, almost all of the published findings in the psychotherapy research literature applied only to the specific samples of therapists used in each study. If researchers want to generalize their results to a larger population of therapists, he argued, then they must treat therapists as a random factor in the statistical analyses. His own review of the literature, however, indicated that this practice was very much the exception rather than the rule. Unhappily, adoption of his methodological counsel would result in serious reductions in statistical power in most studies. Furthermore, Martindale's statistical reanalyses suggested strongly Preparation of this article was supported in part by National Institute of Mental Health (NIMH) Grant MH-40472 and NIMH Career Development Award MH-00756 to Paul Crits-Christoph.We wish to acknowledge the generosity of the investigators who lent us their data, including Aaron T. Beck, Tom Borkovec, Robert DeRubeis, Dolores Gallagher, Steve Hollon, Stan Imber, Bernard Liberman, Lester Luborsky, Tom McLellan, Paul Pilkonis. William Piper, Larry Thompson, George Woody, and Charlotte Zitrin. We also thank David Kenny for his helpful suggestions and comments on a draft of this article.Correspondence concerning this article should be addressed to Paul Crits-Christoph, Hospital of the University of Pennsylvania, 308 Piersol Building. 3400 Spruce Street, Philadelphia, Pennsylvania 19104-4283 that the erroneous statistical desig...