Meta-Analysis: Combining Evidence Across Clinical Studies

simulator intermediate ~10 min
Loading simulation...
θ̂ = 0.35 [0.18, 0.52] — significant pooled effect

With 7 studies, true effect 0.35, and moderate heterogeneity (τ²=0.05), the random-effects meta-analysis yields a pooled estimate of approximately 0.35 with a 95% CI excluding zero, confirming a consistent treatment benefit.

Formula

θ̂ = Σ w_i · θ_i / Σ w_i  where  w_i = 1 / (σ²_i + τ²)
Q = Σ w_i · (θ_i − θ̂_fixed)²
I² = max(0, (Q − (k−1)) / Q) × 100%

The Gold Standard of Evidence

Meta-analysis sits at the top of the evidence hierarchy in medicine. By systematically combining results from multiple randomized controlled trials, it provides the most reliable estimate of a treatment's effect. The Cochrane Collaboration maintains thousands of meta-analyses covering interventions across all fields of medicine, and regulatory agencies like the FDA increasingly rely on meta-analytic evidence for approval decisions.

The Forest Plot

The visualization generates a forest plot — the canonical display for meta-analysis results. Each study appears as a row with a square (effect estimate, sized by weight) and horizontal line (confidence interval). Studies with larger samples and smaller variance receive more weight. The pooled estimate appears as a diamond at the bottom, synthesizing all available evidence into a single number with its confidence interval.

Fixed-Effect vs. Random-Effects Models

The simulation uses the DerSimonian-Laird random-effects model, which assumes that each study estimates a slightly different true effect due to variation in populations, interventions, and outcome measurements. This between-study variance τ² is estimated from the data and incorporated into the weights. When τ² = 0 (no heterogeneity), the random-effects model reduces to the fixed-effect model. As heterogeneity increases, the random-effects model gives more equal weights to studies, reducing the dominance of large studies.

Interpreting Heterogeneity

The I² statistic quantifies what percentage of total variation across studies is due to real differences rather than sampling error. Cochran's Q test provides a formal significance test, but with few studies it has low power (may miss real heterogeneity) and with many studies it may flag trivially small heterogeneity. When I² is high, the pooled estimate becomes less meaningful — the real scientific question shifts from 'what is the average effect?' to 'why do effects differ across studies?' — motivating subgroup analysis or meta-regression.

FAQ

What is a meta-analysis?

A meta-analysis is a statistical method that combines the results of multiple independent studies addressing the same research question. By pooling data, meta-analysis increases statistical power, provides a more precise estimate of the treatment effect, and can reveal patterns not visible in individual studies.

How do you read a forest plot?

In a forest plot, each row represents one study. The square shows the study's effect estimate (size proportional to weight), and the horizontal line shows its 95% confidence interval. The diamond at the bottom is the pooled estimate, with its width representing the pooled CI. The vertical dashed line at zero (or 1 for ratios) represents no effect.

What is heterogeneity in meta-analysis?

Heterogeneity measures how much the effect sizes vary across studies beyond what chance alone would explain. I² quantifies this as a percentage: 0% means all variation is due to sampling error, 25% is low, 50% moderate, and 75%+ substantial heterogeneity requiring investigation.

What is publication bias?

Publication bias occurs when studies with positive or significant results are more likely to be published than negative studies. This skews meta-analyses toward overestimating treatment effects. Funnel plots and statistical tests (Egger's, trim-and-fill) can detect publication bias.

Sources

Embed

<iframe src="https://homo-deus.com/lab/biostatistics/meta-analysis/embed" width="100%" height="400" frameborder="0"></iframe>
View source on GitHub