Individual results in forest plot meta-analysis

[Maike_protein]

Dear all,

In this meta-analysis (https://www.ncbi.nlm.nih.gov/pubmed/30475963) it seems that for instance in figure 2 for lean body mass improvements individual studies did not find significant better results with protein supplementation compared to no protein supplementation (since the interval lines crosses the 0 line).
For example: Mitchell et al https://academic.oup.com/ajcn/article/106/6/1375/4823157 crosses the line while you can find a significant interaction effect in the article of Mitchell et al, see table 3. The time effect was also significant but the diet effect not. Is the absent diet effect the reason why the study of Mitchell et al. crosses the 0 line in the meta-analysis? Or is it because the authors used SMD (standardized mean difference) which is more relative and uses the pooled standard deviation? (since SMD = (new treatment improvement - placebo improvement) / pooled standard deviation. Or has the small sample size of the study influence on this result?
I got confused because normally you see the significant results from the individual studies reflected in the meta-analyses right? But maybe that is only the case for absolute effect sizes and not SMD? I tried to find information about this on the internet but I could not find it. I hope you can help.

FYI: I added the pdf's of the mentioned articles in the appendix.

Thank you in advance!

 

[hlsmith]

Post the table and breakdown what your are writing about. I see the forest plot in the MA, but I don't have time to read the whole entire Mitchell paper. Table 2 seems to show values then just pvalues - yes they list a significant interaction. How about Fig 2A, shows a significant Lean Mass change. It likely gets weird trying to pool these studies together, especially given the interaction term addressing the study design. I don't work with a lot of 2factor analyses. Is your issue that there seems to be an effect but it isn't listed in the MA? Perhaps try to plot a disordinal interaction plot. The crude lean mass values don't seem impressive, RDA not change, 2RDA had slight change but lots of variability.

 

[Maike_protein]

Thank you for your quick response. I added the figures and table. I think you're right the significant difference does not have enough power when it is pooled, but I still think it is hard to understand why this happens. Especially since I thought you normally see the significant individual results back in the forest plots.

 

[hlsmith]

In general, I am sorry I don't have an examples in front of me, but given publication bias didn't prevent studies from getting published, you can easily have a bunch of null studies, just because they didn't cross the 0.05 threshold but were close. Then you pool them and you then get an overall significant effect, since now you have more data (power). Though, if you have to control for random effects in the MA, that may make your SEs larger since you are now controlling for between and within study variability. So things may still not be "significant". I am not saying you have this in your scenario, but just brining it up.