Choosing effect measures

Binary outcomes

• Risk ratio. Useful when event risks are interpretable and studies report events and totals.
• Odds ratio. Common for case-control designs and logistic models; interpret carefully when events are common.
• Risk difference. Useful when absolute effects are important and baseline risks are comparable. Computed and pooled directly from reported events and totals, like risk ratios — including studies with zero events in an arm. Results stay on the absolute scale, where a difference of 0 means no effect.

Continuous outcomes

• Mean difference. Use when all studies measure the outcome on the same scale.
• Standardised mean difference. Use when studies measure the same construct on different scales.
• Change scores. Document whether you analyse final values, change from baseline, or adjusted differences.

Time-to-event outcomes

Hazard ratios are appropriate when follow-up time and censoring matter. Do not mix hazard ratios with simple binary events unless the protocol has a defensible transformation or hierarchy for doing so.

Prevalence and single-group outcomes

Prevalence, incidence, count, and single-arm mean outcomes often need different transformations and interpretation from comparative effects. Use random-effects models when heterogeneity across populations is expected, and report prediction intervals when they are meaningful.

Network meta-analysis measure scope

NMA requires compatible comparative effects across a connected set of treatments. Keep the NMA measure set narrower than the full pairwise catalogue unless your data, assumptions, and protocol justify the broader choice. Axelium's NMA workflow is centred on clinical comparative outcomes with reviewer-approved treatment definitions.

A network outcome is analysed as an odds ratio, risk ratio or risk difference from each arm's events and total, or as a mean difference or standardised mean difference from each arm's mean, standard deviation and total. A hazard ratio can only be used where studies report the comparison itself. The measure is chosen per outcome in the review's settings, and every result is shown on that measure's natural scale: no difference is 1 for a ratio and 0 for a difference, and the direction of benefit you set for the outcome decides which way the ranking reads.

Ecology and evolution measures

Ecology/evolution reviews can use the log response ratio (lnRR), the standardised mean difference (Hedges' g), the Fisher z-transformed correlation, the log coefficient of variation ratio (lnCVR), the log variability ratio (lnVR), or a generic inverse-variance estimate for a protocol-defined effect. These are computed from arm-level means, standard deviations, and sample sizes (or from correlations), or supplied pre-computed. Analysis runs on the transformed scale and results are back-transformed for reporting: response and variability ratios return to the ratio scale (null of 1), and correlations return via the inverse Fisher transform (null of 0). These measures often depend on clear moderator, dependency, and covariance assumptions, especially when one study contributes multiple effects.

Reported confidence intervals

When a study reports an estimate with its confidence interval rather than arm-level data, the width of the interval gives its standard error: width ÷ 2·z(1 − α/2), which is 3.29 for a 90% interval, 3.92 for 95% and 5.15 for 99% (Cochrane Handbook §6.3.1). For ratio measures — risk ratio, odds ratio, hazard ratio — the width is taken on the log scale. A standard deviation calculated from the interval of a mean uses the same multiplier.

Not every interval is 95%. Trials often report 97.5% or 99% intervals for interim-adjusted or multiplicity-adjusted results, and some report 90%. Extraction records each interval's level from the paper's own wording and reads it as 95% only when the paper names none. Read as 95%, a 97.5% interval would give the study about 30% too much variance, and so too little weight; a 99% interval about 73%; a 90% interval about 30% too little. The extraction page names each interval's level, and flags a value whose quoted sentence gives a different level from the one recorded.

A study that reports only its estimate — no events and totals, no means and standard deviations — is pooled from that estimate by inverse variance, when the paper says it is the measure being pooled. An estimate that names no measure or a different one (a hazard ratio under an odds-ratio outcome, a raw mean difference under a standardised one) is left out rather than pooled on the wrong scale, and so is an estimate outside its own interval. Hazard ratios pool from the reported ratio and its interval, as before.

RevMan export notes

RevMan export supports the common Cochrane handoff for included and excluded studies plus supported quantitative outcomes. Single-arm, qualitative, ecology-specific, and unsupported measures may be listed as notes rather than modelled as RevMan comparisons. Risk-of-bias assessment exists in Axelium, but the current RevMan quality export is limited; check RevMan-specific quality fields after import.

Practical checklist

• Pre-specify the measure in the protocol when possible.
• Keep direction of benefit consistent across studies.
• Do not combine adjusted and unadjusted estimates without a rationale.
• Check the level of each reported interval — 97.5% and 99% intervals are common in interim analyses.
• Check that transformed estimates remain interpretable to the target audience.
• Document any override from the suggested measure.

www.axelium.ai · Analysis