I'm about to make a meta analysis of a particular topic. I'm using Comprehensive Meta-Analysis (CMA) ver. 3.0 for that reason. However, after I input all the necessary data, the software refused to create the forest plot because the studies included have asymmetrical confidence interval.

I checked out that the default for 'allowed asymmetric CI ratio' is 1.10. I've changed the ratio to 2.0, which is the maximum input available. Even after that, there is one study that has asymmetrical CI ratio > 2.0.

What I'm asking is: 1. Is there any solution to minimize the asymmetry of the CI of the studies? 2. Is there any solution to alter the 'allowed asymmetric CI ratio' to more than 2.0 so no study will be excluded in forest plot in CMA?

If it turns out that a transformation can minimize the asymmetry of the CI, is the data still valid and presentable?


1 Answer 1


Firstly, asymmetric confidence intervals such as for odds ratios, hazard ratios, log transformed analyses and many other situations are perfectly appropriate. Usually there's some transformation that makes them almost symmetrical (up to rounding, not always when the number of events and/or sample size is very small).

Secondly, I suspect the warning indicates that the program is presumably (I don't know for sure, but the warning does not make sense otherwise) indicates that the program is trying to treat the data from each study as following a N(estimate, SE) distribution. For that purpose it is probably trying to get the SE from the confidence interval width assuming that a 95% CI is given by estimate $\pm$ 1.96 $\times$ SE. In your case that would seem to be very inappropriate. So, do not ignore the warning.

Thirdly, more appropriate things to do depend on the type of data. E.g. if your estimates are all hazard, odds or rate ratios from large trials with many events, then working on the log-scale would likely solve the issue (one can back transform after there analysis). However, other situations may require very different solutions and those may not be covered by your software. The book by Borenstein et al. that they recommend on their webpage is a good introduction into the easier use cases.

  • $\begingroup$ Hi Björn! Thanks for the answer! It turns out that I'm about to make a single-armed meta analysis l. And the effect size I'll be using is Objective Response Rate (ORR), Disease Control Rate (DCR), Progression-Free Survival (PFS) for cancer, which of these results are in percentages. Hence I believe it will look like a meta analysis of proportion. Is data transformation still eligible? Furthermore, other than percentages, other parameter available is median and lower and upper bound, can that parameter be used for meta analysis? $\endgroup$ Commented Feb 1, 2019 at 9:27
  • $\begingroup$ A logical transformation for proportions is the logit log(p)-log(1-p). E.g. for 30% (=0.3) you get -0.847 (if you use logarithms with base e). Also keep in mind the limitations of single arm information (a great read on that is "ICH E10 Choice of Control Group and Related Issues in Clinical Trials"), in particular that you would not expect such rates to be the same, at all, ever (other than by chance), if trials differ in duration, different assessment criteria are used, populations differ etc. so that it's hard to judge if results are better than if patients had not received treatment. $\endgroup$
    – Björn
    Commented Feb 1, 2019 at 9:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.