Methodology

Every statistic on this site is recomputed from the underlying study data at build time, so the numbers are auditable and consistent across pages and the data-entry tool. Where possible we store the raw 2×2 contingency table (events and totals in the treatment and control arms) rather than only a reported effect size.

Effect size

For a 2×2 table with treatment risk p₁ = a/n₁ and control risk p₂ = c/n₂, the risk ratio is RR = p₁ / p₂ and the percent improvement is (1 − RR) × 100. The 95% confidence interval uses the Katz log-RR variance Var(ln RR) = 1/a − 1/n₁ + 1/c − 1/n₂, with limits exp(ln RR ± 1.96·SE). When any cell is zero, a Haldane–Anscombe continuity correction (adding 0.5 to each cell) is applied for the variance.

Studies that report only an adjusted RR, odds ratio, or hazard ratio with a confidence interval are entered as a precomputed effect; the standard error is recovered from the CI width on the log scale. OR/HR values are treated as RR-approximations and flagged with a “~”.

Number needed to treat / harm

The absolute risk reduction is ARR = p₂ − p₁. When treatment reduces risk, NNT = ⌈1/ARR⌉; when it increases risk, NNH = ⌈1/ARI⌉ where ARI = p₁ − p₂. For precomputed effects, NNT is only shown when a baseline (control) risk is reported, via ARR = control risk × (1 − RR).

Pooling

Studies for a given outcome are combined with a DerSimonian–Laird random-effects model on the log scale. We report the pooled RR with its 95% CI, the heterogeneity statistics (Cochran's Q, between-study variance τ², and I²), and each study's weight. Reviews and other sources flagged as excluded from pooling are displayed but not combined, to avoid double-counting their constituent trials.

Evidence status

A topic is labeled favorable when the pooled CI excludes 1 in the beneficial direction, harmful when it excludes 1 in the adverse direction, limited data when there are fewer than 3 studies or too few events, and inconclusive otherwise.

Limitations

Pooled estimates are only as good as the extracted data and the comparability of the included studies. Observational data are subject to confounding; rare outcomes produce wide intervals. This site does not provide medical advice—see the disclaimer in the footer.