Methods & limitations
Eligible data
Version 2.0.35 pools independent, two-group arithmetic means, standard deviations and sample sizes. Reported medians with Q1, Q3 and n can first be converted to approximate means and SDs using Wan et al. (2014), equations 14 and 16. The pooled outcome is then the mean difference or Hedges’ g, not a difference in medians. Paired measurements, repeated observations, survival outcomes and event proportions are not supported. A percentage of cells measured in individual participants is treated as a continuous outcome; it is not treated as a binomial event count.
Each analysis uses one published outcome category, clinical group, comparator and material. A category defines the cell population, unit and denominator/gate. The effect direction is always clinical group minus comparator. Cross-disease and cross-tissue pooling is deliberately blocked.
Effect measures
Mean difference: y = mean₁ − mean₀; sampling variance v = SD₁²/n₁ + SD₀²/n₀. Hedges’ g uses the pooled within-group SD, correction J = 1 − 3/(4(n₁+n₀−2)−1), and large-sample variance v = (n₁+n₀)/(n₁n₀) + g²/[2(n₁+n₀)]. J is the usual approximation, not the exact gamma-function correction.
Median / IQR conversion
Choose Mean + SD or Median + IQR separately for each group. Enter median, Q1, Q3 and n; a scalar IQR width alone is insufficient. Estimated mean = (Q1 + median + Q3)/3. Estimated SD = (Q3 − Q1) / [2 Φ⁻¹((0.75n − 0.125)/(n + 0.25))]. This implementation uses the finite-sample quantile approximation in equation 16, not the tabulated expectation in equation 15. Approximately normal data are assumed; strong skewness, bounded distributions, or very small groups may make conversion unreliable. Extra uncertainty from estimating the summaries is not included in confidence intervals. Original summaries are retained, estimated rows are labelled, and the reported-only filter excludes any row with an estimated group.
Pooling and uncertainty
Inverse-variance weights are 1/(v + τ²). Fixed effect sets τ² to zero. Random effects use either DerSimonian–Laird (method of moments) or Paule–Mandel (iterative solution of generalized Q = k−1). All intervals use a normal critical value of 1.959963984540054 and 95% confidence. I² = max(0, (Q−(k−1))/Q) × 100%, where Q is computed using fixed-effect weights.
At least two independent studies are required. With few studies, uncertainty and heterogeneity estimates are unstable and normal intervals can be too narrow. This version does not implement Hartung–Knapp intervals, prediction intervals, REML, meta-regression or formal funnel asymmetry tests. Model selection requires scientific judgment.
Dependence and source quality
Only one series per publication may enter a pooled analysis. An optional shared cohort label links known overlapping participants across publications: at most one row per nonempty shared label may enter an analysis. Leave it blank when no overlap is known; blank labels do not establish independence. Curators must identify overlap across papers. The software cannot discover unrecorded shared participants. Exact versus digitized extraction is recorded; a sensitivity filter can exclude digitized series. Digitization error is not incorporated in sampling variance.
Reproducibility and figures
CSV and JSON exports contain selected data, identifiers, extraction provenance, effect estimates and model settings. SVG forest exports include the original references and software citation. Leave-one-out analyses re-estimate the entire model after omitting each study. Funnel plots are exploratory; asymmetry does not establish publication bias. Software results require verification against a prespecified review protocol before publication.