Testing Cross-Layer Edge Differences

Why Test an Edge Difference Directly?

A network estimated in one phenotype group may show an edge that is absent from the estimated network in another group. That visual contrast is a reason to investigate, but it is not itself evidence that the underlying conditional associations differ. An edge may pass a significance threshold in one group and fail it in another simply because the groups have different sample sizes or precision. The relevant question is a direct comparison of the group-specific edge parameters.

Here, a cross-layer edge connects a feature in one omics layer to a feature in another. The test concerns their association conditional on the other measured features in the model. It is an association comparison, not a test that one feature causes changes in another.

The partial correlation is dimensionless and ranges from -1 to 1. Its sign indicates the direction of the conditional association, and its magnitude indicates its strength on the partial-correlation scale. In a Gaussian model, \(\rho_{ij\mid -ij,g}=0\) is equivalent to conditional independence of features \(i\) and \(j\) given the remaining features. A test of equality is broader: two groups can have nonzero associations in both groups and still differ in their values.

Likelihood-Ratio Test and Its Assumptions

Suppose there are \(G\) independent groups, with \(n_g\) independent observations in group \(g\). Assume each group follows a multivariate Gaussian distribution, with its own mean vector \(\mu_g\) and positive-definite covariance matrix \(\Sigma_g\). The feature set and its meaning must be the same across groups. Group means are nuisance parameters and may differ; the test compares conditional associations, not means.

For one prespecified edge, the alternative model estimates the group-specific covariance matrices without requiring the partial correlations to be equal. The null model imposes the constraint that the edge partial correlation is equal across all \(G\) groups, while allowing other parameters, including means and other covariance features, to vary as specified by the model. Let \(\hat{\ell}_a\) and \(\hat{\ell}_0\) be the maximized log likelihoods under the alternative and null models. The likelihood-ratio statistic is

Because the alternative contains the null as a special case, \(D\) cannot be negative, apart from numerical optimization error. Under the null, and under regularity conditions such as an interior parameter value and sufficiently informative samples, \(D\) is approximately chi-square distributed. Equality of one scalar edge parameter across \(G\) groups imposes \(G-1\) independent restrictions, so the reference distribution is \(\chi^2_{G-1}\). For two groups, it is \(\chi^2_1\). This is an asymptotic calibration, not an exact finite-sample guarantee.

Why the Gaussian Determinant Formula Works

For a multivariate Gaussian sample, the log likelihood contains a term involving the log determinant of the fitted covariance matrix and a trace term involving the residual covariance. Under the unrestricted alternative, the maximum-likelihood covariance estimate is the residual covariance, using divisor \(n_g\), and the fitted trace equals \(p\). Under a null constraint on partial correlations, each group retains a free overall covariance scale: multiplying its covariance matrix by a positive constant does not change its partial correlations. At the constrained maximum, differentiating the likelihood with respect to that scale also gives a fitted trace of \(p\) in each group. Thus the trace terms cancel in the difference of maximized log likelihoods, leaving

where \(\widehat{\Sigma}_{0,g}\) and \(\widehat{\Sigma}_{a,g}\) are the fitted covariance matrices for group \(g\) under the null and alternative, respectively. This determinant expression is appropriate for the stated Gaussian likelihood setup with scale-invariant partial-correlation constraints. In other likelihood models or with other constraints, calculate \(2(\hat{\ell}_a-\hat{\ell}_0)\) from the actual fitted likelihoods rather than assuming a determinant-only formula.

More Than Two Groups and Multiple Edges

For \(G\) groups, the equality hypothesis for one edge is \(H_0:\rho_{ij\mid -ij,1}=\cdots=\rho_{ij\mid -ij,G}\). It imposes \(G-1\) restrictions. A significant omnibus test says that not all group-specific values are equal; it does not identify which groups differ. Follow-up contrasts require their own hypotheses and appropriate multiplicity control.

Network analysis usually considers many candidate edges. Testing each at level \(\alpha\) without adjustment can make the probability of at least one false rejection much larger than \(\alpha\). Define the testing family before interpreting results—for example, all cross-layer edges between two specified layers, or all edges in a prespecified network. Two common adjustments are:

• Bonferroni: for \(m\) tests, compare each p-value with \(\alpha/m\). This controls the family-wise error rate, the probability of one or more false rejections, without requiring independent tests.
• Holm: sort the p-values from smallest to largest and compare the \(k\)th smallest with \(\alpha/(m-k+1)\), stopping at the first failure. This also controls family-wise error under arbitrary dependence and is never less powerful than the basic Bonferroni procedure.
• Benjamini–Hochberg: sort the p-values and find the largest rank \(k\) for which \(p_{(k)}\le kq/m\); reject through that rank. Under independence or certain forms of positive dependence, this controls the false discovery rate, the expected proportion of false rejections among rejections. It does not control the probability of any false rejection.

Diagnostics, Alternatives, and Limits

Before trusting a likelihood-ratio result, check that the same features, preprocessing rules, and conditioning set were used across groups. Inspect sample sizes and covariance estimates for near-singularity, and assess whether strong outliers or marked departures from Gaussian behavior make the model implausible. A non-positive determinant, failed optimization, or a statistic that is negative beyond rounding indicates a computational or model-fitting problem, not a valid test result.

The ordinary chi-square reference can be unreliable when the number of features is large relative to the group sample sizes. In particular, when a group has too few observations to estimate a full covariance matrix, the ordinary Gaussian maximum-likelihood fit may be singular. Penalized estimators such as graphical lasso are useful for estimating sparse networks, but comparing their selected edges does not by itself give an ordinary likelihood-ratio test with the degrees of freedom above. Selection and penalty tuning change the sampling distribution; a standard test applied after selecting edges on the same data may also be anti-conservative. The target edge or testing strategy should therefore be justified, and inferential calibration must match the estimation and selection procedure.

A Wald test or an approximate test based on transformed partial correlations can be less computationally demanding than fitting the constrained likelihood, but relies on its own standard-error approximations. A direct likelihood-ratio test has the advantage of comparing nested models, but may require challenging constrained optimization. Neither approach makes a poorly specified model reliable. Finally, separate tests of whether an edge is zero in each group answer a different question: “present in one group but not the other” is not equivalent to “the edge parameters differ.”

Common Mistake

• Wrong reasoning: “The edge is significant in group 1 but not significant in group 2, so it differs between groups.” The two significance decisions do not test the difference; they also depend on group-specific uncertainty.
• Correction: Test equality of the group-specific edge parameters directly, using the constrained-versus-unconstrained likelihood comparison or another valid direct contrast. Then account for the full family of edges when interpreting discoveries.
• Wrong calculation: Halving the likelihood-ratio statistic when substituting determinant ratios. In the Gaussian formula, each log determinant ratio is multiplied by its group sample size \(n_g\); the factor 2 from the definition \(D=2(\hat{\ell}_a-\hat{\ell}_0)\) is already accounted for.