Article Text

Download PDFPDF

The objectives of medical research
  1. M J R HEALY
  1. 23 Coleridge Court Milton Road
  2. Harpenden AL5 5LD

    Statistics from

    Request Permissions

    If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

    Editor—Dr Marlow’s interesting annotation on high frequency ventilation,1 with its pithy epigraph, adopts a common attitude to the methodology of medical research. This holds that, to be “scientific,” a study must specify a null hypothesis and then attempt, using data, to disprove it. This agrees with the writings of Karl Popper, but goes back further to the founding father of modern statistics, R A Fisher, who2 wrote: “Every experiment may be said to exist only to give the facts a chance of disproving the null hypothesis.”

    In recent years many statisticians have come to take a much broader view of research methodology. In applied fields such as medicine, engineering, and agriculture, null hypotheses—that two treatments are equal in their effects—are often neither plausible nor interesting, and it must be remembered that the null hypothesis specifies exact equality, not merely negligible difference. Instead, the existence of some difference between the treatments is taken for granted and the study aims at establishing its size, whether it is large enough to be important, or perhaps small enough to be ignored. The calculation of confidence limits aims precisely at demarcatirig the range of true differences which are consistent with the data; whether or not these include zero is of secondary importance. I have referred to this type of study as “technological.”3

    Schwartz et al 4 make a similar distinction between explanatory and pragmatic clinical trials. They point out that in the latter case the exact equivalence of two treatments is seldom plausible and it is necessary to arrive at a definite recommendation that one treatment is to be preferred to another. The type I error rate is then 100%—the null hypothesis that the treatments are equivalent is always rejected. For the same reason the type II error rate is zero. As the treatments are unlikely to be exactly equivalent (and if they are, it does not matter which is recommended), neither of these errors is of any interest; the important error is that of recommending an inferior treatment, one which they refer to as type III. This formulation has not been so widely accepted, but is well worth consideration at the design stage of a technological study.