Meta-Analysis for a
Systematic Review of Trials

RESEARCH QUESTION

What the Review Set Out to Answer

This meta-analysis example follows a public health research team that has completed a systematic review of randomized trials of a blood pressure self-monitoring program. They want one summary answer to a single question: by how much does self-monitoring change systolic blood pressure compared with usual care, and how consistent is that effect across trials?

The synthesis plan, registered with the protocol, was written before any results were extracted. It names the outcome, the effect measure and the model, so the pooled estimate cannot drift toward whatever the data happen to show. The team also wants to know whether the effect would be similar in a new setting, not only the average across the trials already published.

STUDY DESIGN AND DATA

Around 15 Trials, Extracted in Duplicate

A systematic review with around 15 eligible randomized trials, identified through a prespecified search under a registered protocol and screened by two reviewers. Each trial contributes group means, standard deviations and sample sizes for systolic pressure at follow-up, plus design details for the risk-of-bias appraisal. Two reviewers extracted every value independently and reconciled differences, because extraction errors are among the most common ways a pooled estimate goes wrong.

METHOD AND ASSUMPTIONS

Why Random Effects, and What Must Hold

Trials differ in populations, program intensity and follow-up, so the plan specified a random-effects model, which estimates an average effect and the variation of true effects between trials (tau-squared, estimated by REML). With only about 15 trials, confidence intervals use the Hartung-Knapp adjustment, which gives more honest coverage than the standard normal-based interval when the number of studies is small. Pooling the mean difference is appropriate because every trial measured the outcome on the same scale.

Plausibility checks during extraction show two trials reporting standard errors where standard deviations are needed, which would manufacture heterogeneity, and one trial with two intervention arms sharing a control group. The values are converted, and the arms are combined so the control participants are counted once. Heterogeneity is read from tau-squared and the prediction interval, not from the I-squared statistic alone. Risk of bias is appraised trial by trial with RoB 2 and carried into the interpretation, not left as a separate table.

OUTPUT AND INTERPRETATION

How the Results Are Read

The output has a forest plot with the pooled mean difference and its 95% confidence interval, tau-squared, and a 95% prediction interval, which shows the range in which the effect in a new trial or setting is expected to fall. The two intervals answer different questions and are explained side by side: a confidence interval that excludes zero alongside a prediction interval that includes it tells the reader that the average effect is clear but the effect in any one setting remains uncertain.

Leave-one-out analyses, a comparison with trials at high risk of bias removed, and an alternative tau-squared estimator show how far the conclusion rests on single choices. A funnel plot is included and read cautiously: with about 15 trials, just above the usual minimum for asymmetry tests, such tests have little power, and asymmetry can arise from reasons other than publication bias. The write-up states what the pooled evidence supports and where it is limited.

REPORTING

What the Researcher Received

A written report structured to PRISMA 2020, with statistical wording for the methods, matched to the output, for the team to check and incorporate. The review question, the search, the eligibility decisions and the conclusions remain with the research team.