Prediction Intervals, Subgroup Analysis, and Meta-Regression

Prediction intervals, subgroup analysis, and meta-regression are important meta-analysis methods for understanding variation in true effects and exploring potential sources of heterogeneity. This session begins by distinguishing a confidence interval for the average pooled effect from a prediction interval, which estimates the range within which the true effect of a new study may plausibly fall, assuming that the new study is drawn from a population comparable to those represented in the meta-analysis. Students will examine how between-study variance (tau^2), the number of included studies, and uncertainty in the estimated pooled effect influence the prediction interval. They will also learn how to interpret prediction intervals that cross the null or neutral effect value, recognizing that a statistically significant average effect may coexist with the possibility of different effects, including no effect, in individual settings.

The second component focuses on subgroup analysis to investigate whether effect estimates differ across predefined categories of study or participant characteristics. Students will learn how categorical moderators can be used to compare pooled effects across subgroups and how statistical tests can evaluate differences between subgroup estimates. The session emphasizes that differences in statistical significance within separate subgroups do not, by themselves, establish a statistically significant difference between those subgroups. Instead, an appropriate test of the between-subgroup difference is needed. Students will also consider the importance of defining subgroup hypotheses in advance and interpreting subgroup findings cautiously, particularly when the number of studies within each subgroup is small or when multiple comparisons are conducted.

The final component introduces meta-regression, which examines the association between study-level characteristics and effect sizes using continuous or categorical moderators. Students will learn how to interpret meta-regression coefficients and their confidence intervals, assess whether moderators help explain heterogeneity, and understand the concept of residual heterogeneity that remains after moderators are included in the model. The session will also discuss important limitations, including ecological bias, overfitting, and the limited statistical power associated with a small number of studies. Finally, students will distinguish analyses planned before examining the results from exploratory analyses conducted to generate new hypotheses. By the end of this session, students should be able to interpret prediction intervals, explain the purpose and limitations of subgroup analysis and meta-regression, and evaluate potential sources of heterogeneity without overstating the strength of the evidence or making unsupported causal conclusions.