I would like to accumulate the association between BMI and investigated feature expressed with OR. The problem is that different studies adopted different variable intervals in their Cox regression models, e.g.:
study 1: BMI (per 1 unit) OR = 1.056 (1.002; 1.099)
study 2: BMI (per 3 units) OR = 1.19 (1.15-1.23)
Is there a way to meta-analyze above data in order to conclude about the estimate per 1 unit of BMI?
Can you please clarify? Cox regression models usually produce hazard ratios not odd ratios.
Stephen
Of course you are right. There should be HR instead of OR. But I have the same problem with results from logistic regression that produce ORs.
OK. Thanks. Here is what to do.
1. Transform the HR, OR or whatever to the log-scale
2. Do the same to the confidence limits
3. Calculate the standard error on the log-scale as (approximately) (upper CL-lower CL)/4
4. Express them as values per BMI unit by dividing by the number of BMI units to which they apply
5 Carry out standard meta-analysis
See attached solution in MathCad
Sorry there was a typo in entering one of the confidence limits, I hope that this is right now.
This answer may be late for this purpose, but a suggest these websites as a source for further reading about prognosis systematic review and meta-analysis.
http://prognosismethods.cochrane.org/our-publications
http://www.york.ac.uk/inst/crd/index_guidance.htm
Take a look at the attached JAMA article that describes how to pool regression coefficients. Hope you find this helpful.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis