I am conducting meta-analysis by using R package (metafor), it seems i have a significant heterogeneity, so I need to do sub-group analysis guided by gender (categorical) and age (continuous) variables. My question is what I can add to the syntax to achieve the sub-group analysis. If you suggest another package its okay.
Dear Hamdan,
There are several ways to achieve that in metafor package.
1. You can run a separate random-effects model, using "subset" argument (to select studies by specific variable e.g., gender). Then you combine results from subgroups analysis with fixed effect model (specifying variable characterizing subgroup as a moderator "mods="). P-value
for a moderator coefficient will then represent a p-value for the subgroup difference.
2. Directly run a random-effects model with moderator ("mods=")
3. Run a mixed-effect model with rma.mv function (subgroup as moderator and author/study as random effect)
Everything depends on your needs. Approaches #1 and #3 are equivalent - we use fixed moderator and we allow for a different amount of heterogeneity in each sugroup (in #1 by running independent models; #3 by mixed-effect approach). In #2 we assume that amount of heterogeneity is equal in both subgroups, which is not directly what we want.
In general, the idea in approach #1 is the same as subgroup analysis in other meta-analytic software (e.g., RevMen)
More details on coding: http://www.metafor-project.org/doku.php/tips:comp_two_independent_estimates
To add to what Jakub Morze has said:
If you want to perform a subgroup analysis for Gender:
add to rma.uni or rma.mv the code 'subset=(Gender=="Male")' --or however you have coded this part. Note that the == means exactly equal to. != can be used for not equal to. This will give you an exact estimate of the subgroup meta-analytical effect seperately for Male and Female.
There is no clear equivalent for a subgroupanalysis on age. Usually, subgrouping age as a continuous variable into categorical subgroups is meaningless (or reduces information) so approach 2 and 3 above might work better.
You can achieve this using the code
'mods = c(Age,Gender)' or using 'yi~Age+Gender' .
Note that this will give you coefficients of whether Age (and Gender) influence the relationship between X and Y. Hence, this analysis does not produce the estimate of the subgroup effect as above. This can be computed from the combination of the intercept and the moderator coefficient, but approach 1 is more straightforward for Gender.
Good luck.
Dear Jakub Morze and Robert Suurmond many thanks for your comprehensive explanation.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis