As we know when analysing the odds ratios in meta analysis we use the log transform of the OR to get the effect size. I think most papers/analyses do this but the log transformation never seems to be discussed in the stats methods. Is is that I've not looked at enough papers ? Or is it assumed obvious that we take logs ?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
I think your last assumption is the correct one. The problem with the odds ratio itself is that its null value is 1, its lower bound is 0 and its upper bound is positive infinity. This lack of symmetry about the null makes it hard to work with. Contrast this with the log odds ratio whose null is 0 and whose bounds are negative and positive infinity. The standard error of the log odds ratio is also easier to work with and confidence and other intervals also have symmetry. The log odds ratio is also what is worked with in logistic regression.
Of course for presentation of the final results you may want to exponentiate the estimates and their confidence interval endpoints to obtain odds ratios as you may feel they are more easily interpretable.
-
$\begingroup$ I think you're right, mdewey, thank you for your comment. I suppose very often we say we have a binary endpoint and analysis with logistic regression - but we don't say "our endpoint isn't binary - its the logit" - its understood. Maybe its the same with ORs and random effects meta analysis $\endgroup$ Commented May 23, 2018 at 19:22