|
Abstract:
|
Much of the debate on effectiveness vs. efficacy conflates three issues. First, there has been discussion of whether the pharmacologic effect of a drug should or should not be diluted by the lack of effect in people who don't take it. Second, there is a problem of defining the pharmacologic effect when some of the effect is to make people take it who would not have taken the control, or to make people not take it who would have taken the control. Third, it is often assumed that an undiluted effect must be estimated by some version of per-protocol analysis in spite of the well-known bias. I will give precise definitions of the undiluted pharmacologic effect. I will discuss methods of estimating it different from per-protocol analysis. In some cases these methods can be based on randomization with plausible additional assumptions. In other cases, the additional assumptions can at least be made transparent, and sometimes quantitative bounds can be given for how wrong they could be.
|
