Data across studies can be combined at the level of either p-values (probability pooler methods) or effect sizes (effect size meta-analysis). [36, 55, 56] When p-values are combined, at a minimum one should take into account also the direction of effects, but the magnitude of the effects is not taken into account. When effect sizes are used, there are several models that can be used, depending on whether between-study heterogeneity is taken into account or not, and if the former, how this is done. In general, fixed effects approaches that ignore between-study heterogeneity are better powered than random effects approaches and thus more efficient for discovery purposes. However, there is a trade-off for increased chances of false-positives. For effect estimation and predicting what effects might be expected in future similar populations, random effects are intuitively superior in capturing better the extent of the uncertainty. Commonly, random effects are estimated with a 95% CI that captures the uncertainty about the mean effect, but ideally one should also examine the uncertainty of the distribution of effects across populations. This is provided by
