Worked Example: Comparing Networks Across Phenotype Groups

What Does It Mean for an Edge to Change?

A multi-omics network fitted separately in two phenotype groups can differ in its edges, edge weights, or both. A useful comparison must distinguish a visually different pair of network diagrams from evidence that a particular edge parameter differs between groups. This tutorial works through that comparison for a fixed set of candidate edges, including uncertainty, multiple testing, and a cautious summary of robust changes.

Assume that the same molecular features are measured in both groups and that the network parameters have the same interpretation in each. For a Gaussian graphical model, an edge between two features represents conditional association given the other modeled features. Its partial correlation is one possible edge weight: it is dimensionless, lies between −1 and 1, and its sign indicates the direction of conditional association. It does not by itself establish a causal effect.

The contrast must use a consistent scale. For example, subtracting partial correlations compares partial correlations; subtracting Fisher-transformed partial correlations compares transformed parameters and answers a different numerical question. Network construction, feature scaling, covariate adjustment, and missing-data handling should be comparable across groups. Otherwise a difference may reflect a change in analysis rather than a phenotype-specific network.

Assumptions and an Inference Plan

The calculations below use estimates from two independent groups and bootstrap standard errors. Their interpretation relies on several conditions. Subjects are independent within and between groups; the same variables and edge definitions are used in both groups; and the fitted models adequately represent the conditional associations being compared. The bootstrap must resample independent subjects within each group and refit the network estimation procedure. If observations are clustered or repeated within subjects, the resampling unit must instead respect that structure.

The candidate edges in the examples are specified before inspecting the group differences. This matters because choosing edges after seeing which estimates look largest can make ordinary standard errors and p-values too optimistic. If network selection is part of the analysis, uncertainty and testing should account for that selection, rather than treating a selected edge as if it had been prespecified. Covariate adjustment should also be substantively appropriate and applied consistently; it can reduce confounding by measured variables but cannot guarantee the absence of unmeasured confounding.

For a reasonably regular estimator with an approximately normal sampling distribution, estimate the standard error of \(\hat{\Delta}_e=\hat{w}_{A,e}-\hat{w}_{B,e}\) by bootstrapping. A Wald statistic compares the estimated difference with its standard error. A two-sided 95% Wald interval and two-sided p-value are then computed as follows.

The p-value expression already includes both tails: equivalently, it is \(2P(N(0,1)\ge |Z_e|)\). It is not twice \(P(|N(0,1)|\ge |Z_e|)\). A small p-value is evidence against the no-difference hypothesis under the assumed model; it is not the probability that the null hypothesis is true.

Correcting for a Family of Edge Tests

Testing many edges creates opportunities for false discoveries even when every edge truly has no difference. A declared family should include all edge hypotheses that contribute to the scientific claim, not just the edges that happen to have small p-values. One common procedure is Benjamini–Hochberg (BH), which controls the false discovery rate under independence and under certain forms of positive dependence among tests. Network-edge tests can be dependent; when the required dependence conditions are not credible, consider a procedure valid under more general dependence, such as Benjamini–Yekutieli, or a suitable resampling-based familywise approach.

For \(m\) p-values, order them as \(p_{(1)}\le\cdots\le p_{(m)}\). At target false discovery rate \(q\), BH finds the largest rank \(k\) for which \(p_{(k)}\le (k/m)q\), then rejects the hypotheses ranked 1 through \(k\). Adjusted p-values can be computed as the cumulative minimum, from largest rank to smallest, of \(\min(1,mp_{(i)}/i)\). These adjusted values allow a direct comparison with \(q\); they do not measure the probability that an individual reported edge is false.

Summarizing Robust Changes

A useful summary separates the estimated size and direction of a change from the strength and stability of evidence for it. Report the group-specific estimates, their difference, an interval, the raw p-value, and the multiplicity-adjusted result. Then examine whether conclusions are sensitive to reasonable analysis choices: for example, covariate specification, regularization settings, or the treatment of influential observations. For bootstrap-based analyses, the frequency with which an edge is selected and the frequency with which its estimated difference has the same sign can describe stability. Such frequencies are diagnostics, not automatically calibrated p-values.

“Robust” should have a stated operational meaning rather than serving as an informal synonym for “large.” One defensible summary might require an edge to pass the prespecified multiplicity procedure, have a practically meaningful estimated difference, retain its direction in sensitivity analyses, and show adequate resampling stability. The exact criteria depend on the scientific question. An edge can be statistically supported but too small to matter scientifically, or have a sizeable estimate but substantial uncertainty.

Diagnostics, Alternatives, and Limitations

Before interpreting a comparison, inspect the fitted networks and the uncertainty calculations rather than relying on a final table alone. Useful checks include confirming that both groups contain enough subjects relative to the number of features; verifying convergence and comparable sparsity; examining bootstrap distributions for skewness, heavy tails, or frequent nonconvergence; and checking whether one or two observations dominate an estimated edge. A normal Wald interval may be unreliable when the bootstrap distribution is strongly asymmetric, when estimates lie near a parameter boundary, or when the estimator is highly discontinuous. In those cases, a carefully justified bootstrap interval or a method designed for the estimator may be preferable.

Other approaches answer related but not identical questions. A global network test asks whether the networks differ somewhere, without necessarily identifying a particular edge. Edgewise tests localize differences but require multiplicity control. A permutation test can provide a reference distribution by relabeling group membership when subjects are exchangeable under the null and the full estimation procedure is repeated appropriately. Bootstrap intervals describe estimator uncertainty, while stability selection describes how frequently features or edges are selected under resampling and tuning choices. Neither should be substituted for a calibrated test without justification.

The conclusions also depend on measurement quality, group comparability, model form, and the chosen network representation. A missing or weakly measured feature can alter several estimated conditional associations. Different preprocessing or regularization choices can change both selected edges and weights. Finally, a difference in partial correlation is a difference in conditional association, not necessarily a biological interaction or a causal effect. These limitations should accompany any claim about phenotype-specific network changes.

A Practical Comparison Workflow