[R-meta] WG: Results interpretation publication bias

[Wilma Charlott Theilig]

________________________________


Dear Guido,

thank you very much for the recommendations. I have done the Thompson-Sharp test and also a Limit meta-analysis.

The Thompson-Sharp test is also significant:

Test result: t = -2.02, df = 77, p-value = 0.0468

Sample estimates:

bias se.bias intercept se.intercept

-1.3745 0.6801 0.4728 0.0543


For the limit meta-analysis, I obtained the following:


Random effects model COR 95%-CI z pval

Adjusted estimate 0.3872 [0.3267; 0.4445] 11.55 < 0.0001

Unadjusted estimate 0.3605 [0.3111; 0.4080] 13.49 < 0.0001


Quantifying heterogeneity:

tau^2 = 0.0549; I^2 = 98.4% [98.2%; 98.5%]; G^2 = 23.5%


Test of heterogeneity:

 Q d.f. p-value

 4751.13 78 0


Test of small-study effects:

Q-Q' d.f. p-value

 1267.74 1 0


Test of residual heterogeneity beyond small-study effects:

 Q' d.f. p-value

 3483.39 77 0


However, my conclusions according to this would not really change compared to the results I got before, apart from the fact that the trim and fill method does not give good results for some reason. Do you notice anything unusual about the results perhaps, or any possibility as to what that might be? It doesn't seem to be due to the large heterogeneity, maybe it really is due to the one study?

I would appreciate a short feedback!


Best regards

Wilma

________________________________
Von: Dr. Guido Schwarzer <guido.schwarzer using uniklinik-freiburg.de>
Gesendet: Montag, 3. April 2023 18:44:43
An: R Special Interest Group for Meta-Analysis; Theilig, Wilma Charlott
Cc: Michael Dewey
Betreff: Re: [R-meta] Results interpretation publication bias

@Michael: the funnel plots show contours for significance levels p < 0.01 and p < 0.05. These contour-enhanced funnel plots (see doi: 10.1016/j.jclinepi.2007.11.010 ) can be used to evaluate whether funnel plot asymmetry is due to publication bias (i.e., if small unprecise studies are only "published" if they are statistically significant which is not the case here).

@Wilma:
 I agree with Michael that the large funnel plot looks odd. I guess that the single very precise study at the top right side is triggering the statistical significance of tests for funnel plot asymmetry. This said, the result of the trim-and-fill method does not make any sense here as all added studies have correlations above 1!
Did you have a look at the result of the Thompson-Sharp test for funnel plot asymmetry (e.g., meta::metabias(..., method = "Thompson")? This test works better than the Egger test in heterogeneous meta-analyses.
Furthermore, you could have a look at the results of some sensitivity analyses such as limit meta-analysis (metasens::limitmeta), Copas selection model (metasens::copas) or selection models (metafor::selmodel).

Best,
Guido


	[[alternative HTML version deleted]]