A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Meta-analysis statistically combines the results of multiple independent studies addressing the same research question to produce a single, more precise summary estimate. It weights each study inversely proportional to its variance (larger, more precise studies get more weight) and produces a pooled effect size with a narrower confidence interval than any individual study. The critical choice is between a fixed-effect model (assumes all studies estimate the same true effect) and a random-effects model (assumes study-specific true effects drawn from a distribution, with between-study heterogeneity). The I-squared statistic quantifies the proportion of total variability due to heterogeneity rather than sampling error. Meta-analyses are threatened by publication bias (studies with positive results are more likely to be published) and require systematic review methodology to identify all relevant studies, not just the convenient ones.
Individual studies are often too small to detect clinically important effects with adequate power. Meta-analysis addresses this by combining results across studies, effectively increasing the sample size and producing more precise estimates. But it is not simply adding up patients — each study is treated as the unit of analysis, and its result is weighted by its precision (inversely proportional to the variance of its effect estimate). A study of 10,000 patients gets more weight than a study of 50 patients because its estimate is more precise.
The fundamental distinction is between fixed-effect and random-effects models. A fixed-effect model assumes every study estimates the same true underlying effect — differences between observed effect sizes are due entirely to sampling variation. A random-effects model assumes that each study has its own true effect size, drawn from a distribution of effects, and the meta-analytic goal is to estimate the mean of that distribution. The choice matters: when heterogeneity is present, the fixed-effect confidence interval is too narrow (it ignores between-study variability), and the random-effects model is more appropriate. The I-squared statistic quantifies heterogeneity — the proportion of total variability attributable to true differences between studies rather than chance. Values above 50% are conventionally considered substantial heterogeneity.
Forest plots are the visual workhorse of meta-analysis. Each horizontal line represents one study: the point estimate and its confidence interval. The diamond at the bottom represents the pooled estimate, with its width showing the pooled confidence interval. Studies with wider confidence intervals (less precise) have less influence on the diamond. When the lines are scattered widely and the diamond's confidence interval is narrow, you have high precision but high heterogeneity — and the single pooled number may be misleading as a summary.
The most serious threat to meta-analysis validity is publication bias. If studies with non-significant results are less likely to be published, the meta-analysis systematically overestimates the effect. Funnel plots (plotting study precision against effect size) provide a visual diagnostic: an asymmetric funnel, with small studies missing on the null side, suggests bias. Statistical corrections (trim-and-fill, selection models) can partially adjust for this, but the best protection is a comprehensive systematic review that searches for unpublished data. The distinction between meta-analysis (the statistical method) and systematic review (the comprehensive literature search methodology) is important — meta-analysis without systematic review is a quantitative synthesis of a biased sample.