Article Text

PDF

Towards evidence based medicine for paediatricians
  1. Bob Phillips
  1. Correspondence to Dr Bob Phillips, Centre for Reviews and Dissemination, University of York, York YO10 5DD, UK; bob.phillips{at}doctors.org.uk

Statistics from Altmetric.com

I-squared

No, not −1, the self-multiplication of that fancy imaginary number that helps aircraft designers make wings work properly, but a (semi) quantitative assessment of how much heterogeneity there is in a meta-analysis: I2.

You will recall that the idea of heterogeneity (mixed-up-ness) comes in both statistical and clinical flavours.1 This measure—(I2)—assesses the statistical aspect. It is often to be found at the bottom of a forest plot, near some other numbers (τ2 and χ2).

The principle of I2 is straightforward. It gives you an idea of the ‘percentage of variation which is beyond that you'd expect by chance alone’. It can be interpreted, approximately, like this:

  • 0%–25%: just chance, really,

  • 25%–50%: bit more than just chance, but unless clinically odd, probably acceptable. With a grain of salt.

  • 51% plus: Woah! Really, quite different. Should you be doing a summary estimate here at all?

(Flashback again: the underlying assumption of fixed-effect meta-analysis is that the only variation between the studies is chance. Random effects gives you an average effect over an average population in an average trial with the average outcome measure.2)

There are problems with I2; if there are very few studies (<5), it probably underestimates heterogeneity, and if there are lots of studies (>40 or so) then it certainly overestimates it. And it is actually a point estimate in itself (and should be given with its own CI… but that is just too complicated to try…). And it is not really giving you a value you can assess the absolute amount of ‘extra’, to try to assess how clinically meaningful that ‘beyond chance'ness would be. But frankly, if those sorts of things really bother you then you will probably be able to interpret τ2 directly and calculate your own prediction intervals.3

References

View Abstract

Footnotes

  • Competing interests None declared.

  • Provenance and peer review Commissioned; internally peer reviewed.

Request permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.