Edge Uncertainty with Bootstrap and Subsampling

Why Resample a Network?

A fitted multi-omics network is an estimate from a particular sample, preprocessing pipeline, and modeling choice. If a small change in the sampled subjects changes which edges appear, the reported graph may be fragile even when the fitted network is mathematically well defined. Bootstrap and subsampling assess this fragility by repeatedly perturbing the observed data and refitting the same network procedure.

The target here is resampling variability: how often an edge is selected, how consistently its estimated coefficient has the same sign, and how much its coefficient varies under a specified resampling scheme. These are useful descriptions of sensitivity to the observed sample. They do not, by themselves, establish that an edge exists in the population or quantify the probability that it is biologically real.

For an undirected graph, use one consistent rule for determining whether an edge is selected. For example, if fitting separate nodewise regressions, specify in advance whether an edge is retained when either endpoint selects the other or only when both do. In a multilayer model, also specify which layer or cross-layer relation the edge represents. Keep that rule fixed across resamples.

Bootstrap and Subsampling Procedures

In the ordinary nonparametric bootstrap, draw \(n\) observational units with replacement from the \(n\) observed units. Some subjects therefore appear more than once in a bootstrap sample, while others are absent. Refit the chosen network procedure to each of \(B\) bootstrap samples. The resampled rows must represent the independent sampling units: for repeated measurements or multiple omics layers collected from the same subject, resample the subject together with all of that subject’s measurements, not individual rows from each layer independently.

In subsampling, draw \(m\) distinct observational units without replacement, where \(m<n\), and refit the procedure. A common design uses a fraction of the original sample, but the choice of \(m\) is part of the analysis and should be reported. Subsamples perturb the data without duplicating subjects. Because their size differs from \(n\), their selection frequencies are not automatically comparable with bootstrap frequencies or with frequencies obtained using another \(m\).

The resampling scheme also depends on the data design. If independent units are clusters, families, or matched sets, resample the independent clusters or sets as appropriate. If observations are serially dependent, ordinary row resampling does not preserve that dependence; a design suited to the dependence structure is needed. These procedures assume that the resampled units reasonably represent the sampling variation relevant to the scientific target.

Coefficient Variability and Selection

For each edge, retain the coefficient estimate from every resample, including zero if the fitted sparse procedure sets it to zero. Report its selection frequency alongside coefficient summaries such as the resampling mean, standard deviation, or quantiles. These summaries answer different questions: selection frequency describes how often the specified rule retains the edge, while coefficient summaries describe the values returned by the fitting procedure.

A summary using all estimates incorporates both selection and magnitude variation. A summary restricted to fits where the edge is selected describes coefficient behavior conditional on selection; it can look more stable because it omits zero estimates and failed selections. State clearly which summary is reported. If the coefficient is signed, the sign frequency among selected fits can reveal whether the estimated direction is consistent, but this too is conditional on the model and resampling design.

What the Frequency Estimates

There is a precise, limited interpretation of \(\hat{\pi}_{ij}\). Imagine repeatedly generating resamples from a specified resampling mechanism conditional on the observed data, then applying the same fitting procedure to each one. Let \(\pi_{ij}(D)\) be the probability that this procedure selects edge \((i,j)\) on one such resample. Each indicator \(I_{ij}^{(b)}\) is then a zero-or-one outcome with conditional expectation \(\pi_{ij}(D)\). Averaging \(B\) independent resampling outcomes gives

If the resampling runs are independent conditional on \(D\), their indicators have conditional variance \(\pi_{ij}(D)\{1-\pi_{ij}(D)\}\), and the variance of their average is

This justifies the Monte Carlo standard error formula used above, with \(\hat{\pi}_{ij}\) substituted for the unknown \(\pi_{ij}(D)\). It concerns the accuracy of the Monte Carlo approximation to a conditional resampling probability. The latter depends on the observed data, the resampling design, the model, preprocessing, tuning choices, and edge rule. It is not \(P(\text{edge is true}\mid D)\), nor is it a confidence level or a population-level edge probability.

Diagnostics, Alternatives, and Limitations

Resampling results are most informative when the fitting procedure is clearly defined and computational failures are tracked. Check that each fit converged, that the same features and edge convention were used, and that resampled subjects retained their cross-layer alignment. Record how often edges are selected, their sign patterns, and the distribution of nonzero coefficients. A very small number of resamples gives a coarse frequency; large changes when \(B\) is increased indicate that Monte Carlo error remains material. Also inspect whether conclusions are sensitive to the subsample size, preprocessing, or tuning rule.

Penalized network procedures can be nonsmooth: a small data change may move an estimated coefficient from zero to nonzero. In high-dimensional settings, where the number of features is large relative to the number of subjects, ordinary bootstrap behavior for model selection may not provide reliable inferential intervals. Sparse estimates can also produce coefficient distributions with a point mass at zero and a broad or asymmetric nonzero part. A mean and standard deviation alone may conceal this structure. If interval estimates are reported, their construction and assumptions need justification; resampling quantiles do not automatically provide valid confidence intervals for selected edges.

The main alternatives answer different questions. Analytic standard errors or model-based intervals can be more direct when their distributional assumptions and inferential targets are appropriate, but may be difficult to justify for selected edges in high-dimensional sparse networks. Sample splitting separates fitting from assessment but uses fewer observations for each task and does not provide the same repeated-perturbation summaries. Bootstrap and subsampling are flexible diagnostics of sensitivity, not substitutes for model checking or an inferential procedure with a stated target.

In particular, do not turn a frequency cutoff into a claim about truth without a separate argument. A high frequency may reflect a stable feature of the chosen algorithm and observed data while still arising from confounding, misspecification, measurement artifacts, or a tuning choice. A low frequency may reflect weak signal or instability, but does not prove that the population edge is absent. Stability selection and formal calibration of edge selection are distinct topics.

Practical Reporting

A reproducible report should name the resampling unit and scheme, give \(n\), \(m\) when subsampling, and \(B\), and describe the estimator, preprocessing, tuning strategy, and edge-selection convention. State whether tuning was fixed or repeated in each resample. For each reported edge, distinguish its selection frequency from its coefficient summaries, indicate whether zeros were included, and label any sign frequency as conditional or unconditional. These details make clear what the numbers measure and prevent a descriptive stability summary from being mistaken for an inferential conclusion.