FOREWORD LIMBURG COMPLIANCE SYMPOSIUM
There could not be worse experimental animals on earth than human beings; they complain, they go on vacations, they take things they are not supposed to take, they lead incredibly complicated lives, and, sometimes, they do not take their medicine. This last fact led directly to the symposium organized by Els Goetghebeur at Limburg University in March 1995, on the analysis of non-compliance in clinical trials.
To the lay person it may seem blindingly obvious that if we know that many of the people in a clinical trial took only half of the intended dosage of the drug being tested, then this fact should influence our conclusions about the drug's true efficacy. Perhaps it then seems strange that the inclusion of compliance measurements in the analysis of clinical trials has been fiercely resisted by much of the statistics community, and if not resisted simply ignored. However, the reason is not a frivolous one. Statistics deals with the analysis of complicated noisy phenomena, never more so than in its applications to biomedical research, and in this noisy world the intent-to-treat analysis of a randomized double-blinded clinical trial stands as a flagpole of certainty amidst the chaos. This technology can be impressively effective, yielding honest conclusions from just dozens of subjects, while less disciplined surveys involving millions prove nothing, but intent-to-treat analysis means ignoring everything except each subject's randomly assigned treatment group, and in particular ignoring his or her compliance with the intended drug dosage.
However, technology is destiny, in statistics as in the rest of the scientific world, and the technology of compliance measurement has leapt forward in the past decade. Urquhart and de Klerk discuss the two most noteworthy advances, electronic pill bottle monitors and the use of chemical blood markers such as phenobarbital and digoxin. The two methods are complementary, either one offering an almost voyeuristic insight into each subject's pill-taking habits. This information is simply too good to ignore. The question underlying the Limburg conference is how to take compliance information into account without fatally compromising the conclusions of a randomized clinical trial.
There is a lot at stake here. First of all, compliance data can sometimes cast a disturbing light on even the most scrupulously obtained intent-to-treat conclusions. Pocock and Abdalla offer an example from their own analysis of a three-arm study of cardiology patients, testing a betablocker and a diuretic versus placebo. As expected, the diuretic group showed an unpleasant tendency toward increased serum cholesterol levels, but unexpectedly the same effect showed up in the beta-blocker group. However, it was discovered that 30 per cent of the beta-blocker group was also taking diuretics, and the beta-blocker cholesterol effect disappeared when this fact was incorporated in the analysis. Besides broadening the definition of 'compliance', and proving again the dictum that humans are the worst experimental units, Pocock and Abdalla's story emphasizes a crucial fact about intent-to-treat analyses: they measure the efficacy of regimens, not necessarily of the drugs involved in the regimens.
Compliance analysis offers a solution to the regimen/drug puzzle. Pursued more aggressively it may also offer more. It seems possible to use variable compliance, as observed in a randomized