Meta-analysis: known sampling variance, and the two correlations

Status — Stable (univariate meta_V), Stable (bivariate V =)

Univariate meta-analysis via meta_V(v) has been in DRM.jl since the Gaussian core. The bivariate known-sampling-covariance path — one known 2×2 block per study — arrived with meta_vcov_bivariate and the V = keyword, and is parity-verified against drmTMB 0.7.0 (tools/parity_biv_meta.R).

1. What makes meta-analysis different

In an ordinary regression every observation contributes the same unknown residual variance, and the model estimates it. In meta-analysis each row is a study, and each study arrives with its own sampling variance already known — computed from that study's sample size and design, not estimated from the data in front of you.

So the total variance of study splits in two:

The quantity you want is almost always — the between-study heterogeneity, how much the true effects genuinely differ once you have accounted for the fact that small studies are noisy. Getting this split wrong in either direction is the classic meta-analysis error: attribute sampling noise to heterogeneity and you overstate genuine variation; absorb heterogeneity into the sampling term and you understate it.

Univariate

fit = drm(bf(@formula(yi ~ x + meta_V(vi)), @formula(sigma ~ 1)),
          Gaussian(); data = dat)
sigma(fit)      # the HETEROGENEITY SD -- not sqrt(v + sigma^2)

meta_V(vi) marks the data column holding the known sampling variances. The fitted sigma is heterogeneity alone; the known is added per row inside the likelihood and never absorbed into it.

2. Two responses, and now two correlations

The moment you have two outcomes per study — say a mean difference in growth and a mean difference in survival, measured on the same animals — a second correlation appears, and the two are routinely conflated.

	symbol	known or fitted?	what it is
Sampling correlation	(inside )	known	the two effects were measured on the same animals, so their sampling errors covary
Residual / heterogeneity correlation		fitted	do studies with a larger true growth effect also have a larger true survival effect?

Only the second is a scientific finding. The first is an artefact of the study design — real, but not news. Per study:

V = meta_vcov_bivariate(v1, v2; cor12 = 0.6)   # known sampling correlation

fit = drm(bf(mu1    = @formula(y1 ~ x),
             mu2    = @formula(y2 ~ x),
             sigma1 = @formula(sigma1 ~ 1),
             sigma2 = @formula(sigma2 ~ 1),
             rho12  = @formula(rho12 ~ 1)),
          Gaussian(); data = dat, V = V)

fit.scales[:rho12][1]     # the HETEROGENEITY correlation

If you omit V, rho12 absorbs the sampling correlation and you will report a between-study correlation that is partly a measurement artefact. That failure is silent — the fit converges and the number looks plausible. It is worth constructing V even when you believe is small.

Supplying the sampling covariance

meta_vcov_bivariate(v1, v2)                      # independent sampling errors
meta_vcov_bivariate(v1, v2; cor12 = 0.6)         # a shared correlation
meta_vcov_bivariate(v1, v2; cor12 = r_per_study) # one per study
meta_vcov_bivariate(v1, v2; cov12 = c_per_study) # covariances directly

At most one of cor12 / cov12; the default is independence. Each 2×2 block must be a valid covariance matrix ( ) and is checked at construction, so an impossible input fails immediately rather than as a strange optimisation failure later.

3. Coming from R

drmTMB writes the bivariate case as

drmTMB(bf(mu1 = y1 ~ x + meta_V(V = V), mu2 = y2 ~ x,
          sigma1 = ~1, sigma2 = ~1, rho12 = ~1),
       family = c(gaussian(), gaussian()), data = dat)

DRM.jl takes the object as a fit-call keyword instead:

drm(bf(mu1 = @formula(y1 ~ x), mu2 = @formula(y2 ~ x), …),
    Gaussian(); data = dat, V = V)

This is a deliberate divergence, and the reason is mechanical rather than a preference: meta_V(V = …) puts a keyword argument inside a formula, which Julia's @formula macro cannot represent at all. So V sits with tree, K and A — the other "structure supplied alongside the data" arguments. Writing meta_V(...) inside a bivariate formula errors with a pointer to the keyword.

Matrix(V) produces drmTMB's dense   block-diagonal form (stacking order ), and that matrix is accepted back directly, so a V built in R can be handed over unchanged.

How wrong can your cor12 guess be?

Often you know v1 and v2 exactly but must estimate the sampling correlation. Measured (true sampling correlation 0.6, true heterogeneity ρ = −0.35, 40 replicates):

assumed cor12	bias in rho12
0.0	+0.062
0.3	+0.030
0.6 (correct)	−0.003
0.9	−0.035

The bias is linear in the error of the guess, at roughly

bias in rho12  ≈  -0.10 × (assumed cor12 − true cor12)

— a 10:1 attenuation. Being wrong by 0.3 costs about 0.03. So a plausible-but-imperfect cor12 is far better than none, and it is worth reporting which value you assumed.

And if you have no idea at all, still pass V with cor12 = 0 rather than omitting V: supplying the sampling variances matters separately from the correlation, and omitting V entirely was measurably worse (+0.091) than assuming independence (+0.062). Evidence: docs/dev-log/evidence/2026-08-16-a12-biv-meta-recovery.md.

4. What is not covered

• Known covariance with structured effects. V combined with phylo/relmat/animal/spatial markers on the bivariate route is refused; phylogenetic meta-analysis with known sampling covariance is a later slice.

• REML with V — ML only.

• Cross-study sampling covariance. Only row-paired 2×2 blocks are supported. A dense matrix carrying off-block entries is refused, not silently truncated, because dropping them would fit a different model than the matrix describes.

• Univariate meta_V remains diagonal. The bivariate path is the dense one.

5. Verifying your own fit

The heterogeneity split is exactly where a meta-analysis goes wrong quietly, so it is worth one check: fit with and without V and confirm rho12 moves.

with    = drm(bfbiv, Gaussian(); data = dat, V = V)
without = drm(bfbiv, Gaussian(); data = dat)
with.scales[:rho12][1], without.scales[:rho12][1]

If they are identical, V is not reaching the likelihood — check that its length matches the number of studies.

See also

• meta_vcov_bivariate — the constructor and its validation.

• The R↔Julia phrasebook: rosetta.md.

• tools/parity_biv_meta.R — the native-vs-Julia comparator behind the status claim at the top of this page.