Main Analysis

The Main Analysis section is the heart of the module -- the unified workflow for pairwise meta-analysis. Choose an effect measure, enter study data, and receive both fixed-effect and random-effects pooled estimates, heterogeneity statistics (Q, I-squared, tau-squared), forest plot, funnel plot, publication-bias tests, and sensitivity analysis.


The workflow

Step What you do
1. Select analysis type Choose one of 15 effect measures -- determines inputs and formulas
2. Enter study data Add each study with per-study inputs (means/SDs, 2x2 cells, log-effects/SEs). Import supported
3. Calculate Engine runs full analysis and shows results

Both fixed-effect and random-effects models are always returned side by side.


Choosing the effect measure

Family Effect measures
Continuous Mean Difference (MD), Standardized Mean Difference (SMD), Paired Means, Single Mean
Dichotomous Odds Ratio (OR), Peto Odds Ratio, Risk Ratio (RR), Risk Difference (RD)
Time-to-event Hazard Ratio (HR), Incidence Rate Ratio (IRR)
Other Correlation Coefficient, Proportions/Prevalence, AUC-ROC, Diagnostic Test Accuracy (bivariate), Generic Inverse Variance

Effect measures -- Statistical methods

Mean Difference (MD)

For continuous outcomes on the same scale.

MD_i = mean1 - mean2
SE_i = sqrt(SD1^2/n1 + SD2^2/n2)
CI   = MD +/- 1.96 * SE

Standardized Mean Difference (SMD) -- Hedges' g

For continuous outcomes on different scales. Uses Hedges' g small-sample correction:

Pooled SD: sP = sqrt[((n1-1)*SD1^2 + (n2-1)*SD2^2) / (n1+n2-2)]
Cohen's d: d  = (mean1 - mean2) / sP
J          = 1 - 3/(4*df - 1)
Hedges' g  = d * J
SE(g)      = sqrt[(n1+n2)/(n1*n2) + g^2/(2*(n1+n2))]

The J factor removes small-sample upward bias of Cohen's d.

Paired Means

For within-subject (paired) designs. SE uses the correlation between paired measurements (SD of the difference).

Single Mean

Pools means from single-arm studies. Per-study SE is SD / sqrt(n).

Odds Ratio (OR)

Mantel-Haenszel (fixed-effect):

OR_i = (a*d) / (b*c)
Pooled OR_MH = sum(a*d/n) / sum(b*c/n)

Inverse-Variance (random-effects):

logOR_i  = ln((a*d)/(b*c))
Var      = 1/a + 1/b + 1/c + 1/d
SE_logOR = sqrt(Var)
weight_i = 1 / Var

Continuity correction: 0.5 added to all cells if any cell is zero.

Peto Odds Ratio

For rare-event meta-analysis. Uses the Peto one-step approximation -- no continuity correction needed.

Risk Ratio (RR)

RR_i    = (a/(a+b)) / (c/(c+d))
logRR_i = ln(RR_i)
Var     = 1/a - 1/(a+b) + 1/c - 1/(c+d)
CI      = exp(logRR +/- 1.96 * sqrt(Var))

Continuity correction (0.5) applied if any cell is zero.

Risk Difference (RD)

Computed on the proportion scale (no log transform):

RD_i = a/(a+b) - c/(c+d)
CI   = RD +/- 1.96 * SE

Hazard Ratio (HR)

Computed on the log-HR scale:

weight_i = 1 / SE^2
Pooled logHR = sum(weight * logHR) / sum(weight)
Pooled HR    = exp(pooled logHR)

SE recovery from CI: SE = (ln(upperCI) - ln(lowerCI)) / (2 * 1.96)

Incidence Rate Ratio (IRR)

For count outcomes with person-time denominators. Pools on log-IRR scale.

Correlation Coefficient

Uses Fisher z-transformation:

z_i    = 0.5 * ln((1+r)/(1-r))
Var(z) = 1/(n-3)
weight = n-3

CI computed on z-scale and back-transformed.

Proportions / Prevalence

Pools on the logit scale:

Adjusted p = (e + 0.5) / (n + 1)     [continuity adjustment]
Logit      = ln(p / (1-p))
SE(logit)  = sqrt(1/(e+0.5) + 1/(n-e+0.5))

Pooled proportion = back-transformed pooled logit.

AUC-ROC

Pools AUC values with appropriate variance weighting on the AUC scale.

Diagnostic Test Accuracy (bivariate)

Joint meta-analysis of sensitivity and specificity. Models them jointly (bivariate approach). Produces a Summary ROC (SROC) plot.

Generic Inverse Variance

When you have only effect estimate and SE per study:

weight_i = 1 / SE_i^2
Pooled effect = sum(weight * effect) / sum(weight)

Fixed-effect and random-effects models

Model Assumption Weighting
Fixed-effect All studies estimate the same true effect w = 1 / SE^2 (or Mantel-Haenszel for OR)
Random-effects True effects vary across studies (tau-squared > 0) w = 1 / (SE^2 + tau^2)

DerSimonian-Laird tau-squared estimator:

Q   = sum(w_i * (effect_i - pooled_FE)^2)
C   = sum(w) - sum(w^2)/sum(w)
df  = k - 1
tau^2 = max(0, (Q - df) / C)

If heterogeneity is negligible (low I-squared), both models give similar results. With substantial heterogeneity, random-effects is generally preferred.


Heterogeneity statistics

Statistic Formula Interpretation
Q sum(w_i * (effect_i - pooled_FE)^2) Chi-squared test for heterogeneity
df k - 1 Degrees of freedom
p-value From chi-squared distribution Significance of Q
I-squared max(0, (Q-df)/Q * 100)% Proportion of variation due to heterogeneity
tau-squared DerSimonian-Laird Between-study variance

I-squared interpretation (Cochrane Handbook): 0-40% might not be important; 30-60% moderate; 50-90% substantial; 75-100% considerable.


Publication bias

Three tests run automatically:

Egger's regression test

Weighted regression of standardised effect on precision. Non-zero intercept indicates funnel-plot asymmetry.

  • p < 0.05: Significant asymmetry -- publication bias may be present
  • p 0.05-0.10: Marginal evidence -- consider publication bias
  • p >= 0.10: No significant asymmetry -- low risk

Begg's rank correlation test

Kendall's tau between standardised effects and variances.

  • p < 0.05: Significant rank correlation -- bias may be present
  • p 0.05-0.10: Marginal evidence
  • p >= 0.10: No significant correlation -- low risk

Duval and Tweedie's Trim-and-Fill

Estimates missing studies from one side of the funnel, fills them by reflection, re-computes pooled effect. Reports adjusted effect and adjusted 95% CI alongside the original.

These tests complement each other. Converging results strengthen inference; divergent results invite caution.


Sensitivity analysis

Leave-one-out analysis: re-runs the meta-analysis omitting each study in turn. Studies whose removal substantially moves the pooled estimate are flagged as influential.


Outputs

Output Contents
Pooled effect (fixed) Point estimate, SE, 95% CI, z-statistic, p-value
Pooled effect (random) Point estimate, SE, 95% CI, z-statistic, p-value
Heterogeneity Q, df, p, I-squared, tau-squared
Per-study results Effect, SE, CI, fixed and random weights
Forest plot Per-study effects with CIs, pooled diamond, weights
Funnel plot Effect vs precision with 95% pseudo-CI funnel
Publication bias Egger, Begg, Trim-and-Fill
Sensitivity Leave-one-out, influential studies
SROC plot For diagnostic-accuracy analyses

All outputs can be exported for manuscripts.