Abstract:
|
The researcher organizing a randomized clinical trial (RCT) seldom has the advantage of a truly random sample of patients. Most often, those patients are available cases, sometimes referred to as a sample of convenience. As a result, it is the treatment randomization that drives statistical inference, not the properties of random samples drawn from (hypothetically) infinitely large populations. The patients randomized in a RCT constitute a local population, the population that is the target of statistical inference. Although statisticians are aware generally of this limitation to the scope of inference, there is a continuing reliance on infinite population methodology. This results in inferences that are less exact than they could be. In this paper, I review the prejudices against the wider use of available randomization-based methodology, permutation, or randomization tests, and provide an introduction to a wider range of options grounded in randomization-based bootstrap sampling distributions.
|