Why design-specific handling matters
Ordinary two-arm formulas assume independent observations. Cluster-randomized, crossover, paired, matched, split-body, and IPD-derived analyses violate that assumption in different ways.
Coventra exposes design-specific extraction modules so advanced users can collect the fields needed for valid variance handling without forcing those columns on basic users.
Cluster randomized trials
For clustered data, observations within clusters are correlated. A common adjustment uses a design effect: DE = 1 + (m - 1)*ICC, where m is average cluster size and ICC is the intracluster correlation coefficient.
An effective sample size can be approximated as n_eff = n / DE. If the paper reports a cluster-adjusted effect and SE, generic inverse-variance synthesis is usually preferable to reconstructing from raw counts.
- Opt-in fields include study design, average cluster size, ICC, and design effect.
- Missing ICC should not be guessed silently.
- Imputed ICC values require sensitivity analysis.
Crossover and paired continuous outcomes
Paired designs depend on within-person or within-pair correlation. The most direct paired mean-difference input is mean_diff with sd_diff and paired_n, giving Var(MD_paired) = sd_diff^2 / paired_n.
If only period-specific SDs are available, sd_diff can be reconstructed only with an assumed correlation r: sd_diff = sqrt(sd_e^2 + sd_c^2 - 2*r*sd_e*sd_c). That r must be justified and sensitivity-tested.
- Opt-in fields include paired N, mean difference, SD difference, within-person correlation, and paired SMD denominator.
- Carry-over and period effects remain study-design judgments, not automatic calculations.
Matched binary outcomes
Matched binary synthesis should use discordant pairs. If b is treatment-event/control-non-event and c is treatment-non-event/control-event, the matched odds ratio is OR_matched = b / c and Var(log OR_matched) = 1/b + 1/c.
Concordant pairs help describe event frequency but do not identify the matched odds ratio in the same way. Entering only independent arm totals can ignore the matching.
IPD-derived estimates
IPD can support participant-level mixed models, time-to-event models, and covariate interactions. Coventra's extraction sheet records IPD model metadata and study-level adjusted estimates; it does not perform full one-stage IPD mixed-effects modeling from raw participant rows.
Two-stage IPD-derived synthesis uses a study-level TE and seTE from each study's participant-level model. The analysis record should document covariates, covariance structure, link function, and whether estimates are comparable.
Reporting checklist
- Identify non-standard designs before extraction.
- Enable the relevant extraction module before collecting data.
- Prefer published adjusted TE/seTE when the design-specific covariance is not recoverable.
- Report ICCs, design effects, assumed correlations, discordant-pair counts, or IPD model covariates.
- State clearly when Coventra performed two-stage study-level synthesis rather than one-stage IPD modeling.