Hello!
I would really appreciate some wider knowledge on my issue.
I am running a meta-analysis to look at the risk ratios associated with the risks of incidence/mortality of breast or ovarian cancers in BRCA2+ and BRCA1/2+ carriers. Due to how I must split the data into groups (BRCA2+ and OC risk for example), I get lots of small groups.
My issue is, some of the sets only contain 2/3 studies. I tried random effects using Revman but because I have a low number of studies I am unsure whether it can be accurate. When I used fixed effects I couldn't produce the charts the same way which was what initially prompted me to use random effects.
Really at a loss just now as to whether I can keep this as random effects but explain the different weights of my results...how would you approach this?
A fixed-effect model is only appropriate if the research designs for all studies in the meta-analysis are very similar. In the twelve systematic reviews/meta-analyses my colleagues and I have published, we have never used a fixed-effect model. The random-effects model is the more conservative (and safer) model.
FYI, note that "effect" is singular in "fixed-effect" model but plural in "random effects" model. The distinction is important. In the fixed-effect model, you are assuming that there is only one kind of variance (i.e., within-study variance) whereas in the random-effects model, you are willing to assume there is also between-study variance. If there is no between-study variance in a meta-analysis, the random-effects model gives you the same result as the fixed-effect model.
Hello!
It seems that you are familiar with the concepts of the fixed- and random-effects models. Then, you probably have considered the random-effects model to be more appropriate, because having one true effect size is rarely a logical assumption. However, in your situation, it seems that the random-effects model may not be appropriate. With three effect sizes, your estimate of the true variance (tau-squared) will have very poor precision. But can you use the fixed-effect model instead? Yes. However, you should interpret the results as descriptive statistics (i.e., the weighted average of the effect sizes you have). Although you are willing to estimate the population effect size (in the random-effects model, the distribution of the population's true effect sizes), your data do not allow for such an inference.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis