Which Type of Meta-Analysis Do You Need? Pairwise, Network, IPD, Diagnostic Accuracy, Proportion or Dose-Response: A Decision Tree

The types of meta-analysis differ in the question they answer, the data they need and the PRISMA extension that governs reporting (Table 1). In every case the meta-analysis sits inside a systematic review, and the model must match the research question.

Work through the questions in order (Table 2). The first “yes” gives the design; a pairwise meta-analysis is the default when every answer is “no”.

IPD is a data source as much as a design: an IPD network meta-analysis or IPD diagnostic review is possible when raw data exist for several comparisons or tests [6].

Pairwise meta-analysis

Select the effect measure before extraction. Risk ratios and odds ratios suit binary outcomes, with risk differences for absolute effects; mean differences suit outcomes on the same scale, and standardised mean differences (Hedges’ g) pool different instruments measuring one construct. Pool time-to-event outcomes as hazard ratios using the log HR and its standard error, not as dichotomised counts at an arbitrary time point.

Quantify heterogeneity with τ² and I², and report a prediction interval, because I² is the proportion of variability due to between-study differences, not its magnitude [9]. In a random-effects model, prefer REML or Paule–Mandel estimators over DerSimonian–Laird [7], and consider the Hartung–Knapp adjustment for the confidence interval, especially when studies are few [8].

Network meta-analysis

Assess transitivity by comparing the distribution of effect modifiers across comparisons, then test consistency between direct and indirect evidence with node-splitting or design-by-treatment interaction models [10]. Present the network geometry and report rankings (SUCRA or P-scores) with their uncertainty, alongside effect estimates rather than in place of them [2].

Individual participant data (IPD)

A one-stage model analyses all participants at once, stratified by study; a two-stage approach estimates effects within each study before pooling [6]. Separate within-study treatment–covariate interactions from across-study associations to avoid ecological bias [11]. Report the data obtained, data checking and studies without IPD, as PRISMA-IPD requires [3].

Diagnostic test accuracy

Sensitivity and specificity are correlated through threshold variation, so pooling them separately is incorrect. The bivariate model estimates both jointly, with their correlation, and gives a summary point and confidence region [12]. The HSROC model [13] is equivalent without covariates and is used when interest lies in the summary curve across thresholds.

Proportion meta-analysis

Untransformed proportions near 0 or 1 have unstable variances. The Freeman–Tukey double arcsine transformation has been advocated for prevalence [14], but its back-transformation can give seriously misleading results, and generalised linear mixed models with a logit link are now preferred [15].

Dose-response meta-analysis

Category-specific estimates within a study share a reference group and are correlated. The Greenland–Longnecker method estimates this covariance from reported cases and totals [16]; restricted cubic splines then test non-linearity [17]. Assign doses using category midpoints and document rules for open-ended categories during data extraction.

Confirm feasibility during protocol development, before registration. Check each type’s minimum data (Table 3) against a scoping search and pilot data extraction.

Funnel plots and Egger’s test assume a comparative effect size and should not be used with fewer than ten studies [18]. For diagnostic test accuracy reviews, use Deeks’ funnel plot, which plots the log diagnostic odds ratio against 1/√(effective sample size) [19]. For proportion meta-analysis, funnel asymmetry is hard to interpret because there is no null effect against which small studies would be selectively published [20].

Grade the certainty of evidence for each main pooled estimate with GRADE, using the network-specific approach for NMA estimates [21] and the GRADE approach for diagnostic tests in accuracy reviews [22]. PRISMA 2020 asks you to report how certainty was assessed [1].

• Running a network meta-analysis on a disconnected network, or without assessing transitivity.
• Pooling sensitivity and specificity in two univariate models instead of a bivariate model.
• Choosing a random-effects model only because I² exceeds 50%, rather than on the expected heterogeneity of the evidence synthesis.
• Treating dose categories as independent studies in a dose-response meta-analysis.
• Omitting GRADE certainty of evidence for the primary pooled estimate.

For reporting detail, see our guide to meta-analysis reporting standards, and for reading the output, how to read a forest plot.