1. Generalise your findings from an unrepresentative group.Example: The RALES trial showed spironolactone helped in heart failure - but practice showed that this wasn't the case for anyone with renal failure or mild LV dysfucntion [not included in the trial].
2. Find a dodgy comparator. Example: Compare high dose Lipitor with less potent doses of Pravastatin, as in the recent TNT study.
3. Use a surrogate end point, not a clinically important one.Example: If you have an expensive anti-Alzheimer's drug, show it makes some differences to cognitive function and then claim this will result in less need for institutionalisation, reduced disability, fewer deaths or adverse events, lower carer burden and decreased health-care costs. A recent trial of donepezil shows it doesn't.
4. Always emphasise the relative rather than absolute benefits.Example: Treating patients with moderate to severe hypertension will prevent more strokes (ARR = 8%; NNT = 12) than treating mild hypertension (ARR = 0.6%; NNT = 166), even though the relative risk reduction for antihypertensives is identical (40%) for both groups.
5. Emphasise statistical significance and play down effect size.
Example: An Australian trial in 6000 patients found that ACE inhibitors were beter than diuretics in elderly hypertensive patients. The much more powerful ALLHAT trial didn't.
6. Dig deep - there's always good news in subgroup analyses.Example: Pfizer's PRAISE trial of amlodipine found a highly significant survival benefit in a non-ischaemic paient subgroup. Not seen in subsequent studies.
7. De-emphasise harmful effects - or even better, don't measure them at all.Example: Vioxx and cardiovascular risk - why did it take four years to show this? So much for post marketing surveillance.
8. Composite end points can show anything if you try.Example: The UKPDS trial of intensive glycaemic control found a significant benefit on "first diabetes-related events" but this was made up of 21 end points. Most of this effect comprised reduction in retinal photocoagulation, with no changes in diabetes-related deaths and all-cause mortality.
9. Clinician-initiated end points can mean anything.Example: Endpoints like revascularisation procedures and initiation of dialysis are arbitrary, proxy endpoints that may vary with the environment and may not reflect the natural history of the disease.
10. Secondary endpoints may save the day.Example: The ELITE I trial of elderly patients with heart failure using either losartan or captopril found no difference in renal function as the primary end-point. An unexpected decrease was seen in the secondary end-point of all-cause mortality favouring losartan, not confirmed by subsequent trials.
11. Conflated trials: aggregate the data, confuse the punters.Example: the PROGRESS study was in effect two trials, with patients in one arm randomised, according to clinician preference, to perindopril plus indapamide or perindopril alone. The separate results for each trial showed perindopril alone had no outcome effect, a result de-emphasised in several interpretations of PROGRESS results recommending perindopril be initiated post-stroke.
12. It's a class effect!Example: Class effcts of ACE inhibitors in patients with stable cardiovascular disease and preserved left ventricular function? Not according to mixed results from HOPE, EUROPA, and PEACE studies.
13. Do an equivalence trial with fuzzy marginsExample: The INJECT trial of thrombolytics.
14. Sponsored trials have sunny summaries.Example: "The inconsistencies in data analysis and reporting suggests to us a biased attempt to present ESSENCE in a positive light. Four of 7 authors and 4 of 7 members of the trial executive committee were, or had previously been, drug company employees; the trial executive chairman and the lead author both received company research grants; and the company's research and development centre undertook data co-ordination."
15. Negative trials never see daylight.Examples: Glaxosmithkline's latest Serevent data on paradoxical bronchoconstriction. A prospective follow up of 126 trials submitted to the ethics committee of a major Sydney tertiary hospital, those with significantly positive results were more likely to be published (85 vs 65% over 10 years), and be published earlier (median time to publication 4.8 years vs 8.0 years) than trials showing nil effect.
Sources:
http://pharmawatch.blogspot.comInternal Medicine Journal Vol. 35 Issue 10 Page 611 October 2005 Cautionary tales in the clinical interpretation of therapeutic trial reports I. A. Scott, P. B. Greenberg
