Greenhalgh recently described ‘Ten ways of cheating with statistics’ [1]. Here are a few tips on how to ‘cheat’ with the design of a clinical trial. ‘Cheat’ is actually the wrong word – what is really meant is the clever design of a study to increase its chances of yielding ‘positive’ results (i.e. results that apparently demonstrate the effectiveness of the treat- ment under scrutiny). The simplest approach is to conduct a study without a control group. Most conditions improve over time and regression towards the mean will also help to normalize parameters that were abnormal at an initial reading. Thus, with repeated measurements of clinical endpoints, one will almost invariably find an apparent overall improvement. This apparent improvement can be entirely unrelated to the therapeutic intervention applied. The trick is to ignore this well-known fact and conclude that the treatment was effective. Controlled clinical trials lacking design features that minimize bias are more prone to generate a positive result than studies that incorporate design features such as placebo controls, blinding and randomization [2]. Failure to blind subjects, therapists, those assessing out- come measures and those analysing the data can all lead to biased results and interpretation. Randomization is used to prevent the experimenter from allocating subjects to treatment groups in a biased way, and to achieve groups that are balanced for important prognostic factors. The success of randomization in terms of balance is, how- ever, a judgement call. One way to obscure differences between groups that favor the experimental group is to apply a statistical test of significant difference. Such tests are conservative because they are designed to err towards finding no difference. They can yield no statistically significant difference, even if there are clinically relevant differences in important prognostic factors. The inappro- priate use of tests of significant difference to establish baseline comparability between groups is widespread and often remains unrecognised. There are other subtle methods that can be used to demonstrate that ineffective therapies work, including equivalence or non-inferiority trials. For example, one could conduct an under-powered equivalence trial com- paring an experimental therapy with a ‘gold standard’ therapy; because of the small sample size the trial would fail to show a difference. Consequently, one could con- clude (falsely) that the experimental therapy was as effec- tive as the gold standard. Alternatively, one could carry out an adequately powered equivalence trial and use an in- effective comparator treatment. This would actually show that both treatments are ineffective, but the trick is to convince the reader that this evidence demonstrates the effectiveness of the treatments. Perhaps the most reliable way to fool people with clinical trials is to use a comparator therapy that causes a deteriora- tion of your primary clinical outcome measure. In a typical parallel group design this will create an inter-group difference favoring the experimental, ineffective treat- ment. Consequently, you need only to convince the reader that this was due to the effectiveness of your therapy, and at the same time omit the fact that the comparator inter- vention led to a deterioration of the control group. Experienced, critical professionals will find these tech- niques far too obvious, therefore we need to consider even more refined ways of producing false-positive results. A report recently published by Paterson et al. [3] provides a subtle example of this concept. Consider a group of patients who, at entry to a trial, are asked whether they prefer acupuncture (treatment A) or homeopathy (treatment B) for their condition. Those who prefer treatment A are allo- cated to treatment arm A and those who prefer treatment B go to arm B. Both groups are then randomized to receive either the preferred therapy or standard GP care. Patients 99 DDT Vol. 9, No. 3 February 2004 editorial 1359-6446/04/$ – see front matter ©2004 Elsevier Ltd. All rights reserved. PII: S1359-6446(03)02912-X How to show that an ineffective therapy works ‘ …the aim of clinical trials is not to prove that therapy X works, but to test whether or not it works.’ Edzard Ernst and Peter H. Canter, Complementary Medicine, Universities of Exeter and Plymouth www.drugdiscoverytoday.com