A post hoc audit analyzing 377 two-group comparisons across 140 pediatric pneumonia studies has demonstrated that reported within-group dispersion varies substantially by reporting format, altering pooled effect estimates. Researchers Feng Li and Ling Zou evaluated dispersion plausibility and conversion provenance to develop an exploratory screening framework for biomarker meta-analyses. The investigation focused on standardized mean differences, which rely directly on reported within-group standard deviations to serve as denominators in statistical synthesis.
The audit compared coefficients of variation, Hedges g, and ratios of means between directly reported mean and standard deviation data and quantile-derived values. Because reporting groups were not exchangeable, researchers treated the contrast between reporting types as purely descriptive. A screening trigger of a coefficient of variation below 0.15 emerged during the audit, prompting the authors to examine dispersion thresholds ranging from 0.10 to 0.25 alongside mechanism checks, source-label inspections, and denominator-sensitivity syntheses.
Directly reported comparisons exhibited a markedly lower median coefficient of variation of 0.214 compared to 0.719 for quantile-derived comparisons. Across the evaluated thresholds, direct reporting accounted for 97.5 percent to 99.4 percent of all flagged instances, representing 98 of the 99 flagged comparisons identified. Furthermore, directly reported comparisons yielded a median signed Hedges g of 1.58 versus 0.29 for quantile-derived metrics, accompanied by median ratios of means of 1.41 compared to 1.27.
To investigate potential reporting errors, researchers conducted source-page inspections across all 66 original publications that contributed flagged direct comparisons. Although standard deviation or conventional mean plus-or-minus standard deviation notation was confirmed in all 66 source documents, this inspection could not validate underlying numerical dispersion. Statistical mechanism evaluations, including square root of sample size rescaling checks and slope tests, failed to identify a definitive operational error mechanism responsible for the unusually low dispersion values.
The methodology was further evaluated against an external illustrative COVID-19 review and scale-structure comparators, where flags occurred in 22 of 30 median and interquartile range comparisons and four of five mean and standard deviation comparisons. Across 52 of 57 evaluable interquartile range arms, reported review standard deviations were arithmetically compatible with Hozo range-based conversion formulas using interquartile range endpoints and review sample sizes under an absolute tolerance of 0.011.
Despite this numerical alignment, arithmetic compatibility did not confirm the exact conversion algorithm used, leaving source-to-review mappings unresolved. The sensitivity syntheses revealed that denominator assumptions strongly dictate effect sizes. When flagged comparisons were excluded, conventional pooled standardized mean differences for four evaluated cytokines dropped from an initial range of 1.94 to 2.47 down to 0.92 to 1.80. Meanwhile, pooled ratios of means ranged between 1.48 and 1.74, and prediction intervals remained wide under both synthesis strategies.
The authors concluded that reporting format correlates with relative dispersion and that pooled standardized mean differences remain highly sensitive to denominator choices. Identifying a low coefficient of variation should prompt thorough source verification and sensitivity testing rather than automatic data exclusion, mathematical rescaling, or correction. Li and Zou stressed that their workflow functions strictly as an exploratory screening heuristic with unknown diagnostic performance, warning that thresholds derived from pediatric pneumonia biomarkers cannot be generalized across other clinical fields without independent calibration.
Subject of Research: Statistical dispersion and denominator auditing in biomarker meta-analyses
Article Title: Auditing denominators in biomarker meta-analysis: a screening framework for dispersion plausibility and conversion provenance
Article References: Li, F., & Zou, L. (2026). Auditing denominators in biomarker meta-analysis: a screening framework for dispersion plausibility and conversion provenance. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03089-4
Image Credits: AI Generated
DOI: 10.1186/s12874-026-03089-4
Keywords: biomarker meta-analysis, standardized mean differences, coefficients of variation, pediatric pneumonia, within-group standard deviations, Hedges g, ratios of means, dispersion plausibility, conversion provenance, quantile-derived data, denominator sensitivity, Hozo formula, interquartile range, cytokines, screening heuristic, source-label inspection, prediction intervals
News Source: Ophelia Keating. (November 18, 2026). Audit Reveals Directly Reported Biomarker Data Drive Most Low Dispersion Flags. Scienmag.



