Identifying individual characteristics that mediate an effect of treatment, though often vitally important, can be extremely challenging. The usual approach in clinical trials requires the analysis of possible interactions between the treatment and one or more covariates. This approach suffers from serious shortcomings related to issues of model specification, statistical power, and multiple comparisons. A rich literature in clinical trials has shown that these conventional methods of identifying subgroups of patients who should or should not be treated with a specific intervention often falsely discover subgroups that appear to respond to therapy differently from the average patient, or fail to find subgroups that do in fact respond unusually positively or unusually negatively. Thus, a common stance is to dismiss almost all subgroup findings as inherently suspect. The panelists in this session worry lest throwing the baby out with the bathwater in this way might ultimately inefficient or even dangerous. This session discusses three novel, but rigorous, approaches to subgroup analysis, The three panelists come to the problem with different philosophical and statistical approaches. Dr. Scott Solomon is a cardiologist at Brigham and Women’s Hospital in Boston. His interest lies in finding drugs that are particularly useful for subsets of the population at risk of major cardiovascular events. He will present an example of a biologically deduced subgroup successfully validated by an independent analysis of data from a large randomized clinical trial. Dr. Richard Simon, a statistician from the National Cancer Institute, will present methods of identifying subgroups of patients who are likely to respond to, or who are likely not to respond to, specific agents. He is involved in trials of cancer therapeutic agents, many of which are highly toxic, so that treating unnecessarily can cause serious harm. He will describe how to develop and internally validate a predictive classifier for each endpoint of interest, translating the problem to one of predictive classification, rather than a series of unvalidated subset analyses. The third panelist, Dr. Herbert Weisberg, a statistician from Causalytics, LLC, will introduce a radically different approach that does not involve interaction effects. The basic method utilizes a special outcome variable, the cadit, that has a known statistical relationship to the causal effect (rate difference or mean difference). The cadit is a simple function of the individual’s exposure status (e.g., drug vs. placebo) and outcome value. A statistical model, such as ordinary least squares or logistic regression, can then be derived using the cadit as the dependent variable. This model can suggest which of the covariates are related to the causal effect and the strength of the relationships. The results can be used to identify individuals or subgroups that are most likely to benefit, or least likely to be harmed. The panelists will discuss their own approaches as well as the approaches of the other members of the panel.