Meta-Regression

There is no absolute minimum, but the widely cited rule of thumb is at least ten studies per moderator variable. With fewer studies the model is prone to overfitting, standard errors are underestimated, and results are unlikely to replicate. When the number of available studies is small, subgroup analysis or simple heterogeneity exploration is more appropriate than a full regression model.

Fixed-effect meta-regression assumes that after accounting for moderators all residual heterogeneity is zero — that is, only sampling error remains. Mixed-effects meta-regression adds a random-effects term (tau-squared) to allow for unexplained between-study variability beyond what the moderators capture. Mixed-effects models are generally preferred because the assumption of zero residual heterogeneity is rarely tenable in practice.

No. Meta-regression is observational at the study level and cannot establish causation. Moderator variables are often correlated with each other — for example, more recent studies may also use larger samples — making it difficult to isolate the unique contribution of any single characteristic. Findings should be interpreted as exploratory or hypothesis-generating unless supported by strong prior theory and pre-registration.