Did you know that aspirin following MI doesn’t work for those with Gemini and Libra star signs?

No, it’s true*. The ISIS-2 trial, which demonstrated the mortality benefits for anti-platelet agents after myocardial infarction with a p<0.00001 only showed benefit for people born in ten of the twelve signs of the zodiac.So if you believe statistics, and randomised trials, then you could save 1/6th of the antiplatelet bill by not giving it to this lot.

The problem of subgroup analysis is fraught with statistical, philosophical, and I would argue emotional problems. We all want to find out not just if something works, but in whom does it work best, or worst. In oncology, we want to reduce the toxic effects of our therapies. (Even if at times this doesn’t seem to be the case.) With expensive biological agents in rheumatology, we want to keep the heathcare budget in balance yet maximising the benefit to patients. And to do this we look at subgroups.

The ‘obvious’ problem is that if you take 20 subgroups, you’d expect 1:20 ( which is 5 in 100, which is 0.05 …) to be ‘significant’. The less obvious thing is that chance may throw up three spurious association in 20 subgroups, around 1 in 7.
A defence is that ‘it’s biologically plausible’. Perhaps. But set yourself a challenge. Pick a treatment, and split the patients. For example, acute asthma and magnesium sulphate infusions. Now, imagine it works better in the more poorly patients. Come up with an explanation. Flip it – it works better in less poorly patients. Explain that. When I’ve tried it on wards, I’ve found most of us can do this in under a minute.

Subgroups are exploratory. They suggest. They hint. If repeated and repeated, they may be true. However, you should hold tight when you see the next subgroup analysis and remember that the true* explanation may actually be astrological after all.

* OK – so ‘true’ may be a lie.