Meta analysis(corrected for sample size)

[suganthiny kumarasamy]

Dear stata users.
I am doing meta analysis.
i have some 1 study have very narrow interval (larger sample size and negative effect)
and 4 studies have wider confidences interval (small sample size and +ve effect).
In stata does the meta analysis random effect model corrected for sample size?
if not is there is way i can correct for sample size?
and any suggestion deal with this situation



Thanks
sugan

[Weiwen Ng]

I don't know which command you're using. But in the user-written metan command, all the studies are basically weighted by the inverse of their variance (which is generally related to sample size), unless you manually specify otherwise.

[suganthiny kumarasamy]

thanks you weiwen , i am using "metan" command.

[suganthiny kumarasamy]

Hi I have another question
I am using metan command
Below is the two study which is significant, when combine result become not significant. What is the reason combine result not significant?

OR(95%(CI))
Study1 80.25(5.57, 1156.15) significant
Study2 3.11(1.80, 5.37) significant
Combine 12.02(0.52, 277.85) not significant

Thanks

[Weiwen Ng]

That is interesting. I'm not able to comment much further, because I don't know the individual studies. I am not as familiar with the theory of the random effects model in meta analysis. But, I suspect you probably ran a random effects analysis. The model attempted to characterize the heterogeneity between studies, and perhaps it found that the heterogeneity was very large, thus rendering the main effect not significant.

I'd caution you that an OR of 80, with a 95% CI of 5.57 to 1156.15 is enormous. It's so large that I have to wonder about the quality of the study. It would have to be tiny, for one. The magnitude of the treatment effect is also implausible for just about any healthcare intervention (this is my field) I can think of; it probably is equally implausible whatever your field (because, if the intervention were so effective, why isn't everybody doing it by now, and why aren't there numerous other studies?).

An OR of 3.11 for the other study is also quite large, but more plausible.

Earlier, you said there were 5 studies in total. The larger study had a negative effect. The smaller ones had positive treatment effects. In meta analysis, in my opinion (going with what I've been taught), most of the work is qualitative. You have to decide if the individual studies are of sufficient quality to be able to analyze them in the first place. You have to decide if the interventions are similar enough to be pooled. There's a lot of other stuff going in. After the meta analysis, you want to look for any evidence of publication bias via funnel plots, which are a pretty inexact method, but they are what we have.

All of this is to say that while you only showed 2 of the 5 total studies, I have to wonder if one of those studies is even plausible to begin with.

[Carlo Lazzaro]

Suganthiny:
my (less than) 5 cents after Weiwen's superb advice: have you ruled out any mishap in data entry before running -metan-?

[suganthiny kumarasamy]

Thanks so much for you clear explanation Weiwen, I have data that each predictors have different studies , predictor 1 has 5 studies and predictor 2 has 2 studies , i did
the metan randon effect model by predictors that's why the second question i ask for only predictor 2 result.
i did not enter the data entry for the studies Carlo.

Thanks

Kind Regards
sugan