1
$\begingroup$

I am conducting a Meta-Analysis of a clinical intervention in randomised trials. I am using the RevMan Software of the Cochrane Foundation. The problem regards the extraction process of relevant outcome results that are typically needed for Meta-Analysis.

  1. I have one primary outcome, which is "symptom level of disease X" but different scales to measure symptom level of "disease X". Lets say "Scale A, Scale B, Scale C".
  2. The different study papers report the outcomes in different ways
    • one study only gives as result a "mean change score" with a Confidence Interval 95%, but the program requires a "Mean" as well as "Standard Deviations" (SD) for the "Intervention" as well as for the "Control". Is it possible to get the Mean and the SD from the single given "change score"? (The baseline values for each group are given.)
    • another study only gives the results in a comparison of "percentage of people having a cut point of XYZ"
    • while other studies give the Mean Difference and Standard Deviations of different outcome scales
  3. The "properties" of the analysis in Review-Manager are as follows:
    • "Continuous Data Type" (since the scales have different measurements/pointage system from 1-20 or 1-10)
    • Statistical Method: "Inverse Variance" (As far as I understand this is useful for different data type entry possibilities)
    • Analysis Model "Random Effects" (It seems to offset the heterogeneity)
    • Effect Measure: "Mean Difference"

Will it be sensible to conduct the Meta-Analysis in this way?

I already tried finding answers in the Cochrane Handbook, RevMan Tutorials, as well as a lot of articles and parts of books about meta-analysis.

Improve this question
$\endgroup$
6
  • $\begingroup$ I need to take a deeper look at your question to answer it (and I don't know the software). I would suggest that you read a bit about meta-analysis (there's a book by Koricheva for ecologist) and that you try to compute by hand your effect sizes. $\endgroup$
    – Emilie
    Apr 7, 2015 at 13:10
  • $\begingroup$ Ok, thank you. Yes, I have read several published articles as well as short parts from different statistic books and am still continuing the reading process. I will try get informed about how to compute the effect sizes. The problem is that I am required to use the software to produce the corresponding graphs. $\endgroup$
    – Sebastian
    Apr 7, 2015 at 13:20
  • $\begingroup$ Is it not so that the given "change scores" e.g. "-4,76" are already the effect size? My problem is how to arrive at the Mean Difference and Standard Deviation from that. Maybe it will be better to post this as a question by itselt. $\endgroup$
    – Sebastian
    Apr 7, 2015 at 13:43
  • $\begingroup$ Hi, yes it is possible but a lot of these calculations will have to be done outside of RevMan. I'll write a more detailed explanation, but here are some quick points: 1) back calculate the SE from the CI and then calculate the SD from that; 2) use Standardized Mean Difference instead of Mean Difference for multiple scales; 3) you can combine Cohen's d from the binary and continuous meta-analyses (meta-analysis.com/downloads/…) $\endgroup$
    – abousetta
    Apr 7, 2015 at 13:58
  • $\begingroup$ I opened a new, more detailed question about the "change score" problem. $\endgroup$
    – Sebastian
    Apr 8, 2015 at 12:12

1 Answer 1

Reset to default
2
$\begingroup$

Part of this answered has has already been given by @abousetta in a comment but repeated here as comments may disappear.

Assuming that you are constrained by the software to do this you can back calculate standard errors from confidence intervals by reversing the usual formula.

using the standardised mean difference is a typical way of dealing with primary studies using different tools to measure the same concept.

Inverse variance is indeed the standard way of weighting studies. It is not totally correct though that the random effects model offsets heterogeneity. It is a different model which assumes there is no true single underlying summary effect but rather a distribution of them whose mean and variance it estimates.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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