societal and individual acceptance of a health care intervention should depend on the balance of its benefits, harms, and costs. The measurement of those benefits and harms, the balance of harms and benefits, and how that balance compares with the balance of harms and benefits of an established treatment are at the heart of most clinical research. This section discusses some core principles in comparisons and evaluations of health care interventions.

Many possible solutions or interventions exist for every human complaint and every affliction. Some interventions gain general acceptance and become standard treatment. Some of these have strong scientific evidence to support their use. That is, investigators have tested formal hypotheses about the interventions according to established principles and have found that they are clearly superior. Others have little or no scientific evidence to support their use but have become accepted as effective because of their long-term use. For practicing clinicians, clinical research often addresses the question asked by many patients, “This treatment has worked for me in the past, so why should I switch to another, less established treatment?” The way to answer this question is a head-to-head comparison of the two treatments. In designing a study to this end and analyzing the results, a number of considerations can be important. This section describes some of these considerations, with particular emphasis on how they apply to the problem of predicting the response to an intervention.

The data set from a randomized study of two treatments will contain many different variables. Among them is a measure of outcome (e.g., blood pressure at the end of the study); this variable would be the dependent variable in a multivariable model (the goal of the model would be to predict the end-of-study blood pressure). Another variable would be treatment assignment (Drug A or Drug B); treatment assignment would be a predictor variable (or an independent variable). Other predictor variables might include age, sex, ethnicity, pretreatment blood pressure, and dietary salt intake. One form of a multivariable model would include all of the predictor variables and might show that several predictor variables were significant predictors of the end-of-study blood pressure. One of them would be Drug A and another might be salt intake. In a multivariable regression model, these two would be independent predictors of outcome. This result would not prove that salt intake was a mediator of the response to Drug A, in the sense that Drug A had a greater effect in the presence of a low-salt diet versus high-salt diet. However, the model could be set up in a different way, with so-called interaction terms reflecting the extent to which the effects of Drug A and salt intake vary as the levels of the other variables vary. If the