Multilevel Meta analysis in Stata 14

[Katherine Picho]

Good Morning!

I currently use Stata 14.


I'm conducting a meta-analysis. I have 125 effect sizes from about 70 studies. To account for dependence in effect sizes (ES) due to multiple ES per study I posited a 3 level multi level model.



I set it up such that:
Level 1 = Variance of effect size

Level 2 = Effect Sizes
Level 3 = Study characteristics.


1. I have a study level id, and an id for effect sizes within studies.
2. I have sorted my data such that the ids are nested-- ie level2 ids are nested in level 3 ids. ( sorted data by level3id and level2id): My data is set up like this:


level3id level2id
1 1
1 2
2 3
2 4 etc


I am solid (expert) at doing regular (ie non- meta-analytic) HLM in the HLM software and have had no previous problems using it; but for some reason my file is being created with errors for the multilevel meta. I am currently trouble shooting but given how many hours and days I've spent spinning my wheels on this issue in HLM, I figure I should perhaps give this a shot in Stata. My stata skills are generally strong (although I've never done multilevel modeling in Stata).

Can anyone point me to resources on conducting multilevel meta analysis in Stata for continuous outcomes (ie Cohen's D)? Sample syntax code et al would be helpful.



Thank You,

Katherine Picho

[Marcos Almeida]

Hello Katherine,

I wonder if this hand out wouldn't apply to your needs:http://www.ats.ucla.edu/stat/stata/e...x/chapter8.htm

Best,

Marcos

[Katherine Picho]

Hello Marcos

Thank You for this. I saw this earlier but...the reason I wasn't so sure about it is that I was expecting to see the ids used at different levels to denote nesting specified in the syntax (see example below) but didn't.

  mixed mathgain || schoolid: || classid:, variance reml

The syntax at ATS reminds me of a syntax I used for regular meta-analysis and not multilevel. This is why I came to the forum....I will let you know what I find....Thanks again for your help.

Katherine

[Joseph Coveney]

I'm not familiar with meta-analysis of effect sizes. I'm curious why you've modeled the effect size variance as nested under the effect size. It would make more sense to my naive intuition if the (inverse of the) variance were included in a two-level model as an analytical weight, [aweight] or equivalent.

[Katherine Picho]

Hi Joseph

I'm sorry that was an error. Basically in setting up the HLM, effect size and its variance are at level 1;-- and the variance would also be the case. I think in classical meta analysis, the weighted variance is usually applied (I've done that before and actually do have those values (i.e. aweight) in my current dataset) -- but in the hierarchical/ multilevel Models, from what I've read-- they suggest having variances (not inverse of var) at level 1; although I think Joop Hox mentions something about inverse of var being used in some instances (don't recall at the top of my head).

At any rate-- I apologize for that mistake up there... I'm still looking for resources to carry out MLM meta so if you know of any,please let me know...

thanks!!

[AnnaMarie Jones]

Hi Katherine

I have just come to the same problem as you. Did you find a solution?

Thanks
Anna-Marie

[Katherine Picho]

Anna-Marie. No I did not (as far as setting this up was concerned).

However, I did use robumeta command which takes into account dependencies in ES. I can send you some articles on this if you're still interested!