Clinical Trials

A typical setting for testing a hypothesis in clinical trials would be a comparison of two clot-busting drugs on mortality following cardiac surgery or whether a drug is more effective than a placebo (aspirin, say, in preventing heart attacks within a five-year span). In each case two groups of individuals are recruited having similar medical profiles and then randomly assigned to one protocol or other. We want to test whether any observed difference between the outcomes can be considered statistically significant. Typically the subjects are not informed as to which group they belong. One starts with is a fervent hope that there is a meaningful difference between the two outcomes. To do this one begins by invoking a null hypothesis H0: any difference one observes is due solely to chance. The null hypothesis is the straw man of statistical testing in which one plays the devil’s advocate in the hope that it can be refuted. This is analogous to proving a theorem in mathematics by assuming the result is false and then showing that this leads to a contradiction implying that the result must in turn have been true all along. If H0 is true there will still be detectable differences due to sampling errors but can we say, with some degree of certainty, that the difference goes beyond experimental error and can be attributed to some real effect?