Looking at the funnel plot in fig. 3 in the linked meta-analysis [1], the high precision studies (standard error < 0.268) are clustered around an effect size of 0. The overall positive effect derives exclusively from the medium and low-precision studies, with several visible outliers well to the right influencing the mean significantly. The medium and low-precision studies had problems with one or more of the following:

"randomisation; allocation concealment; blinding of therapists (intervention supervisors); blinding of participants; blinding of outcome assessors; handling of incomplete data (use of intention-to-treat analysis); selective reporting and any other risk of bias." [1]

Also, the result was only statistically significant for poor quality control groups that didn't engage in any shared activity:

"When the control group involved either no contact (eg, waiting list, usual care; p<0.01) or education (eg, computer course, health lectures; p=0.01) the estimate was statistically significant. Where the control condition was exposed to an active control (eg, stretching; p=0.17) or social group (p=0.62), the effect size was still positive but no longer statistically significant." [1]

In summary, a few poor quality studies with large bias account for most if not all of the effect reported in this meta-analysis.

[1] http://bjsm.bmj.com/content/early/2017/03/30/bjsports-2016-0...