Exchangeable and comparison-specific heterogeneity priors are blocked by the covariance constraint
Statement
Letting between-study heterogeneity variances differ across comparisons in a network requires the induced random-effects covariance matrix to remain positive semidefinite, the condition Lu and Ades call second-order consistency, because the variance of a contrast not involving the reference treatment is a function of two diagonal entries and their covariance. Four structures have been built under that constraint: one common variance; proportional variances whose prior means differ by comparison type; separate priors on the variances together with a prior on the correlations, whose hyperparameters are found by numerically solving a nonlinear moment-matching system; and an informative Wishart prior on the precision matrix. Two structures that a sparse network most needs are still unavailable. An exchangeable model that borrows heterogeneity information across comparisons has no known specification that also meets second-order consistency, and a model in which each comparison receives its own externally derived informative prior cannot be delivered by the Wishart route, whose moments cannot be matched comparison by comparison, and can be delivered by the separate-prior route only by solving a fresh moment-matching system for every new network. The setting where this matters is the ordinary one: in the reference review of network characteristics the median network holds 21 studies with a median of 2 studies per comparison, so per-comparison heterogeneity is essentially uninformed by the data and has to come from outside.
Why it is open
The constraint is a joint condition on the whole covariance matrix while both missing structures are specified marginally. An exchangeable hierarchical prior on the comparison-specific variances, and a set of separate external priors on them, both place distributions on quantities that are constrained functions of the underlying matrix, and no construction is known that induces prescribed marginals for those quantities while guaranteeing the matrix stays positive semidefinite. The routes that do guarantee validity either fix the borrowing strength through a single degrees-of-freedom parameter, as the Wishart family does, or make the induced marginals an awkward function of the parameterization, so matching them to published targets requires numerical solution network by network, which is the impracticality that keeps the approach out of applied use. Because the parameter is close to unidentified in the networks concerned, a wrong variance structure does not show up in the fit, which removes the empirical pressure that would otherwise force the construction to be found. Any solution also has to be implementable by applied analysts in general-purpose samplers rather than merely derivable, which is a stiffer requirement than the derivation itself.
What has been tried
| Work | What it contributes |
|---|---|
| Turner, Dominguez-Islas, Jackson, Rhodes and White 2019, Statistics in Medicine | Builds and compares the four approaches to informative priors for multiple heterogeneity variances under second-order consistency, and states in its discussion both that it is not clear how to assume exchangeability while meeting the requirement and that an approach allowing unequal variances with different informative priors would be desirable but is not available. |
| Lu and Ades 2009, Biostatistics | Models between-trial variance structure in mixed treatment comparisons and establishes the second-order consistency condition that any unequal-variance structure must satisfy; the constraint that blocks both missing structures. |
| Lu and Ades 2004, Statistics in Medicine | Defines the contrast-based mixed treatment comparison model in which heterogeneity variances attach to comparisons rather than arms, fixing the parameters whose priors are at issue. |
| Thorlund, Thabane and Mills 2013, BMC Medical Research Methodology | Builds and compares four replacements for the homogeneous variance assumption, including exchangeable and consistency-constrained variance structures and two informative variance-prior schemes, evaluated on two illustrative datasets; the earlier attempt at the same problem, without external predictive distributions attached comparison by comparison. |
| Turner, Jackson, Wei, Thompson and Higgins 2015, Statistics in Medicine | Supplies the 80 predictive distributions by outcome and comparison type that a comparison-specific prior structure would attach edge by edge; the external evidence that currently cannot be used at that granularity. |
| Rhodes, Turner and Higgins 2015, Journal of Clinical Epidemiology | The corresponding predictive distributions for continuous outcomes, extending the set of external priors that the missing structure would need to accommodate. |
| Rhodes, Turner, White, Jackson, Spiegelhalter and Higgins 2016, Statistics in Medicine | A pseudo-data device for implementing informative lognormal heterogeneity priors in standard software, which addresses the implementation half of the problem for a single variance but not the joint validity constraint for many. |
| Lewandowski, Kurowicka and Joe 2009, Journal of Multivariate Analysis | Vine and onion constructions for random correlation matrices, the separation-strategy machinery that would let a valid covariance be built from standard deviations and a correlation matrix with a tunable borrowing structure. |
| Nikolakopoulou, Chaimani, Veroniki, Vasiliadis, Schmid and Salanti 2014, PLoS ONE | The description of 186 published networks that supplies the median of 21 studies per network and 2 studies per comparison, which is why per-comparison heterogeneity cannot be estimated from within the network. |
Probable solution or research direction
Attack the construction rather than the constraint: parameterize the covariance by the separation strategy, a vector of standard deviations for the reference contrasts and a correlation matrix under a vine or onion prior, then derive by simulation-based calibration the map from that parameterization to the marginals of the comparison-specific variances, and choose hyperparameters so those marginals match the published predictive distributions. Tabulate the solution once per network size and comparison-type pattern so an analyst does not solve a nonlinear system for every new network, which is the step that currently makes the approach impractical. For exchangeability, place the hierarchy on the parameters of the valid construction rather than on the constrained variances themselves, and measure by simulation how much borrowing that actually achieves relative to a common variance. Evaluate all of it on the criteria a decision cares about, interval coverage for treatment contrasts and how often the ranking changes, at network sizes drawn from the empirical distribution, and ship the survivor in multinma or netmeta, since a structure that has to be hand-coded will not be used.
Verification
Verdict. Confirmed open. One reader proposed this and the reviewer accepted it because the cited source expressly identifies both exchangeability under second-order consistency and comparison-specific informative priors as unresolved. The evidence is a single paper, but it is the state of the art for the problem and the two gaps are named in its own discussion rather than inferred.
Revised by the full-text reading. Applied 2026-07-26 full-text reading of the consolidated library (687 papers).
- Citation corrected
- normalized a citation year: ‘Turner, Dominguez-Islas, Jackson, Rhodes and White 2018, Statistics in Medicine’ -> ‘Turner, Dominguez-Islas, Jackson, Rhodes and White 2019, Statistics in Medicine’ 10.1002/sim.8044 is cited with two different years across the registry. CrossRef gives 2018 online and 2019 in print; the issue year is used throughout.
Source
Derived from reading:nma-32 of the reviewed corpus.