Methodologies

Subgroup analysis, meta-regression, and moderator interpretation

How subgroup fields and study-level covariates connect to moderator analyses, including formulas, defaults, limitations, and reporting requirements.

17 min readUpdated June 15, 2026

Subgroup analysis

Subgroup analysis partitions studies by a categorical label, such as dose group, design, region, risk-of-bias stratum, or population category. Coventra's optional extraction module adds subgroup fields so rows can be stratified without overloading the basic extraction sheet.

A subgroup model estimates pooled effects within categories and may compare between-subgroup differences. The comparison is observational at the study level unless subgroup membership was randomized or prespecified in a design that supports causal interpretation.

Meta-regression formula

A simple random-effects meta-regression can be written as y_i = beta_0 + beta_1*x_i + u_i + e_i, where y_i is the study effect, x_i is a study-level covariate, u_i is the between-study random effect, and e_i is within-study sampling error.

Coventra's meta-regression uses a selected study-level covariate and the fitted meta-analysis. When no covariate is provided, the default is study year, but publication workflows should explicitly select a meaningful covariate.

The residual heterogeneity tau^2_residual describes remaining between-study variance after the moderator. It does not prove that the moderator is causal.

Changeable settings

  • The optional extraction module can be enabled before extraction or later from the extraction analysis plan.
  • The covariate name is user-selectable in the meta-regression request.
  • The same effect measure, binary method, tau estimator, HKSJ setting, confidence level, SMD method, and plot options can be carried into moderator analyses.
  • Bubble plot rendering uses plot options for colors and display.

Main limitations

  • Study-level meta-regression cannot estimate participant-level effect modification.
  • Moderator analysis is usually underpowered with few studies.
  • Multiple exploratory moderators inflate false-positive risk.
  • Covariates extracted inconsistently across studies can create artificial signals.
  • A single influential study can drive a moderator coefficient.

Reporting checklist

  1. State whether each subgroup or moderator was prespecified.
  2. Report the number of studies per subgroup or across the covariate range.
  3. Describe how missing moderator values were handled.
  4. Report the model, covariate scale, coefficient, confidence interval, p-value if used, and residual heterogeneity.
  5. Label exploratory moderator findings as exploratory.