help with meta-analysis

[Juliana Moro]

I need help, I ran a meta-analysis and the result was this:

Meta-regression Number of obs = 14
REML estimate of between-study variance tau2 = 5.777
% residual variation due to heterogeneity I-squared_res = 0.00%
Proportion of between-study variance explained Adj R-squared = 90.28%
Joint test for all covariates Model F(4,9) = 9.28
With Knapp-Hartung modification Prob > F = 0.0030
----------------------------------------------------------------------------
_meta_es | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Region1 | 3.347177 14.25074 0.23 0.820 -28.89024 35.5846
Region2 | 2.380619 13.56567 0.18 0.865 -28.30706 33.0683
Region3 | -16.88669 13.56948 -1.24 0.245 -47.58298 13.80959
Region5 | -6.193403 14.38828 -0.43 0.677 -38.74196 26.35515
_cons | 21.58 13.35803 1.62 0.141 -8.637959 51.7979


However, when I look at the individual analyses, none of significant value.
I used the metareg command. However, when I look at the individual analyses, none of the significant value.
I used the metareg command. Could someone help me explain the metaregression?

[Leonardo Guizzetti]

A meta-regression is really not different from a "plain" regression. It appears that you have run a meta-regression of your effect size onto 5 regions, and I'm guessing you have used Region4 as your base category. (If this is true, read the next paragraph. If this is not true, read it all the same.) It appears that while all your effect sizes in each region hover around a value of 22, there is quite a bit of imprecision in each of those estimates, as you can see by the large standard errors and wide confidence interval, and is likely a result of (1) having few studies in each region, and (2) a large between-study varaition, τ². Despite this, the overall model fits the data better than having represented all the studies from all regions but a single, common average. Said another way, there is evidence of region-specific estimates, so region does account for the observed heterogeneity, but those estimates are quite imprecise in themselves, and differences between them are likely to suffer from the same imprecision.

Stata allows for a compact and convenient way to express factor variables (sometimes called categorical variables). See -help fvvarlist-. I would modify your regression above to use i.region (or something similar) which will let Stata automatically create the region-specific indicator variables behind the scenes, and also allows you to modify the base level to your liking. Factor variable notation also lets Stata know that all those indicators belong together as one variable, and allows for more convenient use of commands for hypothesis testing and for margins. You can test pairwise differences between regions using -margins i.region, pwcompare- following this adjustment to your syntax.

[Juliana Moro]

Thanks for the availability. Region 4 was omitted due to collinearity. I tried to do it with the i.region command but this message appears "factor-variable and time-series operators not allowed"

[Leonardo Guizzetti]

-meta regress- certainly allows factor variable notation. If you want help troubleshooting, please post back with a reproducible data example and precisely the code you are using to set up the meta-analysis. See the FAQ for details of how to do that using -dataex-.