The concept of meta-analysis was addressed previously, essentially pulling together data from a range of different studies and assuming that they are only (fundamentally) different by chance, or differ by real things too as well as chance, and you’re seeking an average effect across the average of these differences. The maths under this takes each study as an item, and comes up with a weighted average of the effect sizes.

There’s another way of looking at this:Now if instead of thinking that the studies are themselves a ‘thing’ and you look for the average of that, you instead think ‘each of these studies gives us information about patients and their outcomes, and we should look at those as the unit of analysis’ then you’re getting conceptually to “A general framework for the use of logistic regression models in meta-analysis.”

This approach essentially uses a regression approach (like linking shoe size and height?) so you can predict the ‘true’ effect of an intervention by drawing a line through the study data that you have to predict the truth. It gives a simple (well, singular) approach that can be used to explore all sorts of things; like diagnostic test accuracy, normal meta-analysis and mixed treatment comparision meta-analysis.It works around some rather difficult things when we have tiny numbers of events and studies without events. And it lets us undertake examination of subgroups or explanatory variables a lot more easily.

Why is this important? Well, it’s probably the case that quite a few of our meta-analysis paper give results which aren’t right using the approximations we currently use. We also have trouble with many approaches leading to confusion, and one approach might be easier, and lets us use the same methods for IPD meta-analysis too.

(And it’s just much more intellectually satisfying!)

– Archi