Methodology defaults, effect measures, and changeable assumptions

Why defaults must be visible

A meta-analysis application should never hide the statistical defaults that shape the result. The default effect measure, heterogeneity estimator, sparse-data rule, interval method, and display scale can all change the reported estimate or its uncertainty.

Coventra therefore treats defaults as starting points, not as methodological endorsements. The review team should confirm that each default matches the protocol, outcome type, clinical direction, study design, and journal reporting requirement.

Default effect measures

• Pairwise binary outcomes default to odds ratio (OR). Users can select OR, risk ratio (RR), or risk difference (RD) for binary pairwise outcomes.
• Pairwise continuous outcomes default to mean difference (MD). Users can select MD or standardized mean difference (SMD) in the normal project setup. Ratio of means (ROM) and weighted mean difference are also available in advanced configurations.
• Time-to-event outcomes default to hazard ratio (HR). The analysis is performed on the natural log HR scale using generic inverse-variance inputs when a reported HR and uncertainty are available.
• Single-arm binary outcomes use a pooled proportion. Single-arm continuous outcomes can use a raw mean or an incidence rate depending on whether the extraction records mean/SD/N or events/person-time.
• Network meta-analysis inherits the selected outcome scale. Binary network workflows default to OR unless another supported summary measure is selected; contrast-based network workflows accept TE and seTE on the chosen effect scale.

Required explicit scales for generic estimates

Generic inverse-variance and contrast-based network analyses require an explicit effect measure. A missing TE/seTE scale should fail validation rather than silently becoming MD, because the same numeric TE column can represent a mean difference, standardized mean difference, log hazard ratio, log odds ratio, or another estimand.

This means the extraction sheet and the analysis settings must agree on the selected effect measure before a result is publication-ready.

Recommended, package-native, and manual method sources

The analysis workspace separates the scientific estimand from package-controlled computational defaults. Effect measure, analysis family, confidence level, and model reporting remain explicit; package-default mode does not silently choose OR instead of RR, MD instead of SMD, or another estimand.

Recommended settings are the default method source. They send REML where supported, the Hedges SMD correction, explicit sparse-binary settings, and GLMM/logit settings for single-arm proportions. Hartung-Knapp is applied only on supported aggregate random-effects paths when k > 2 and tau2 > 0 after the initial fit. With exactly three studies, Coventra records a sensitivity note comparing Hartung-Knapp and classic random-effects intervals.

Native package-default mode remains available as an advanced reproducibility option. In pairwise and single-arm workflows this is shown as meta 8.5-0 native defaults; in contrast NMA it is shown as netmeta 3.5-0 native defaults. The actual package and version returned by R are saved with each result. Manual mode sends every exposed setting explicitly and rejects incomplete configurations.

Each fitted result records the method-source profile, package name and version, and effective settings returned by R.

• In the validated meta 8.5-0 image, package defaults include REML tau estimation and classic random-effects confidence intervals. Hartung-Knapp is not automatically enabled merely because the study count is small.
• In the validated netmeta 3.5-0 image, omitting method.tau for contrast NMA resolved to DL; recommended settings explicitly send REML.
• Package-default mode is rejected for one-stage IPD mixed models and sparse arm-level binary NMA where an explicit model choice is required.
• Manual single-arm proportion and rate controls remain blocked until the complete compatible GLMM parameter set is exposed.

Default model settings

• Pairwise binary method: Mantel-Haenszel (MH) by default, with Inverse, Peto, and GLMM available where compatible with the selected summary measure.
• Between-study variance estimator: REML by default. Changeable estimators include DL, PM, HM, ML, HS, SJ, HE, and EB where supported by the underlying R package and analysis type.
• SMD correction: Hedges by default. Cohen and Glass are also available.
• Hartung-Knapp style random-effects confidence intervals: native package defaults use the installed package behavior; recommended settings apply HK only for supported aggregate random-effects paths when k > 2 and tau2 > 0; manual mode lets users send HK/classic explicitly.
• Confidence level: 95 percent by default. The API accepts a level setting, normally 0.95.
• Primary display: both common-effect and random-effects summaries are shown by default. Plot settings can hide either diamond without changing the underlying data.
• Continuity correction: 0.5 by default for sparse binary cells where a correction is needed.
• Double-zero binary studies: not included by default in pairwise relative-effect synthesis. Users can opt to include all studies where the selected method supports that decision.
• Prediction intervals: off by default in forest plot display. Users can enable them for random-effects analyses when the study count and model make interpretation defensible.

What is changeable by the user

• Analysis family: pairwise, network meta-analysis, or single-arm analysis is selected before extraction and can be changed later from the extraction workspace.
• Outcomes and effect measures: users define outcome type and effect measure during extraction setup or project settings.
• Pairwise model options: users can choose native meta defaults, recommended settings, or manual settings. Manual mode exposes binary method, tau-squared estimator, SMD method, HKSJ interval behavior, confidence level, continuity correction, double-zero handling, multi-arm strategy, and common/random display.
• Plot options: forest style, columns, labels, sort order, axis handling, back-transformation, log-spaced ticks, prediction interval display, output format, colors, and study exclusions can be changed without editing extracted data.
• NMA options: reference treatment, effect measure, tau-squared estimator, ranking direction, requested analysis modules, and common/random display can be changed where supported by the NMA workflow.
• Bayesian pairwise options: prior mean and SD for the overall effect, tau prior family and scale/range/rate, and prior-sensitivity execution are changeable.
• Advanced extraction modules: generic/time-to-event estimates, subgroup/meta-regression fields, cluster design fields, paired/crossover fields, matched-binary fields, and IPD/covariance metadata are opt-in so basic users do not see unnecessary columns.

Assumptions that need human verification

• The app cannot prove that OR, RR, RD, MD, SMD, HR, proportion, rate, or mean is the correct estimand for the clinical question.
• The app cannot prove that all effects are oriented so that benefit and harm point in the intended direction.
• The app cannot prove transitivity in an NMA, proportional hazards for HRs, correct handling of missing outcome data, or absence of selective reporting.
• The app cannot infer a defensible ICC, within-person correlation, or matched-pair covariance when the source paper does not report it.
• The app cannot turn two-stage study-level IPD-derived estimates into one-stage participant-level mixed-effects modeling.
• The app cannot make funnel plot asymmetry diagnostic of publication bias; asymmetry remains a signal requiring methodological interpretation.

Reference checkpoints for verification

• Cochrane Handbook Chapter 6 for effect measures, log scales, unit-of-analysis issues, direct extraction of adjusted estimates, and time-to-event outcomes.
• Cochrane Handbook Chapter 10 for pairwise meta-analysis, inverse-variance weighting, binary and continuous models, heterogeneity, subgroup analysis, meta-regression, Bayesian approaches, and sensitivity analysis.
• Cochrane Handbook Chapter 11 for network meta-analysis concepts, transitivity, coherence, ranking, and confidence in network evidence.
• Cochrane Handbook Chapter 13 for missing evidence, funnel plots, small-study effects, and asymmetry test caveats.
• Cochrane Handbook Chapter 23 for cluster-randomized, crossover, and multi-arm randomized trial variants.
• Cochrane Handbook Chapter 26 for individual participant data reviews and the distinction between aggregate data synthesis and IPD analysis.