I am conducting a meta-analysis with differing amounts of heterogeneity in subgroups. Since the data structure is nested (effect sizes nested in studies) I took this into account by adding random effects for studies and effect sizes within studies. Profile likelihood plots peaked at their estimates. A likelihood ratio test comparing the model with different amounts of heterogeneity and a model with a common estimate of heterogeneity was significant. However, after obtaining confidence intervals (CIs) for $ \tau^2_\text{between} $ and $ \tau^2_\text{within} $ except one all the lower bounds of the CIs for $ \tau^2_\text{within} $ were zero. The one significant estimate is still moderate.
Would it be appropriate to simplify the model to a two-level model with random effects only at the study level, thereby only accounting for between study heterogeneity, in this situation?
