Anchor Samples for Bridging Modality-Specific Cohorts

Why Anchor Samples Matter

Suppose one cohort has measurements on modality \(X\) and a second cohort has measurements on modality \(Y\), but few or no subjects have both. Separate summaries of the two modalities do not, by themselves, reveal which values of \(X\) tend to occur with which values of \(Y\). A group of subjects measured on both modalities can provide that missing link. The usefulness of this link depends not only on how many such subjects are measured, but also on which subjects and cohorts they represent.

An anchor sample is not a mathematical device for creating subject-level pairings between unrelated people. Its paired records supply direct information about cross-block dependence. That information can help connect modality-specific cohorts, provided the relationship measured among anchors is relevant to the target population. The earlier material on Estimating Cross-Omics Associations From Partial Overlap addresses association estimation with partial overlap; here the emphasis is on how the number and placement of jointly measured subjects govern the bridge.

What the Anchor Sample Estimates

For scalar features \(X\) and \(Y\), the target cross-block covariance is \(\operatorname{Cov}(X,Y)=E[(X-\mu_X)(Y-\mu_Y)]\), where \(\mu_X=E(X)\) and \(\mu_Y=E(Y)\) refer to the target population. For \(n\) anchor subjects, the sample covariance and correlation are

These quantities use the within-subject pairing. Reordering the \(Y_i\) values relative to the \(X_i\) values generally changes both estimates, even though the separate modality-specific distributions stay exactly the same. For feature vectors, the same idea yields a cross-covariance matrix: each entry pairs one feature from one modality with one feature from another modality on the same anchor subjects.

Using an anchor estimate for a broader cohort requires assumptions beyond having paired observations. The anchors must be measured comparably to the target data, their sampling and missingness must not select subjects in a way that distorts the relationship of interest (or that selection must be accounted for), and the cross-modality relationship must transport to the target population. If the target spans different cohorts or subgroups, anchor coverage should reflect those differences. A large anchor sample from one narrow subgroup cannot establish that the same relationship holds everywhere else.

Anchor Count and Precision

More independent anchors generally provide more information about a cross-block relationship, but the gain is not a guarantee of low uncertainty if the sample is unrepresentative or the measurements are noisy. Under independent, identically distributed bivariate normal observations, an approximate interval for a population correlation \(\rho\) can be constructed on the Fisher-transformed scale. Define \(z_r=\frac{1}{2}\log\{(1+r)/(1-r)\}\). For moderate or large \(n\), its approximate standard error is \(1/\sqrt{n-3}\). A 95% interval on that scale is \(z_r\pm1.96/\sqrt{n-3}\); applying the inverse transformation \(\tanh(\cdot)\) gives an approximate interval for \(\rho\). This is an approximation, not an exact small-sample result, and relies on the stated distributional and sampling conditions.

The formula makes the role of anchor count explicit: increasing \(n\) decreases the approximate standard error. But count alone is not enough. Dependence among subjects, influential observations, restricted ranges, and subgroup imbalance can make the nominal precision misleading or make an estimate irrelevant to the target.

Placement: Which Subjects Should Be Anchors?

Placement has at least two meanings. First, anchors should cover the cohorts, strata, or covariate ranges to which the cross-block estimate will be applied. Second, when estimating a relationship such as a regression slope, anchors need useful variation in the predictor. A sample concentrated near one predictor value may estimate a relationship poorly even if its total size appears adequate.

For a simple linear model \(Y_i=\beta_0+\beta_1X_i+\varepsilon_i\), assume the conditional mean is linear, errors are independent with mean zero and constant variance \(\sigma^2\), and the observed \(X_i\) values are treated as fixed (or are independent of the errors). The least-squares slope is \(\hat{\beta}_1=\sum_i(X_i-\bar{X})(Y_i-\bar{Y})/\sum_i(X_i-\bar{X})^2\). Given the \(X_i\) values, its variance is

This follows by writing \(\hat{\beta}_1-\beta_1=\sum_i(X_i-\bar{X})\varepsilon_i/S_{XX}\), where \(S_{XX}=\sum_i(X_i-\bar{X})^2\). Independence and equal error variance give \(\operatorname{Var}(\hat{\beta}_1\mid X)=\sum_i(X_i-\bar{X})^2\sigma^2/S_{XX}^2=\sigma^2/S_{XX}\). Thus, greater spread in anchor \(X\) values reduces slope variance under this model. It does not justify extrapolating a linear relationship beyond the range or populations represented by the anchors.

Bridging Cohorts and Choosing an Anchor Allocation

When modality-specific cohorts differ, the target relationship must be specified before deciding where to place anchors. Let \(G\) denote cohort membership. The overall covariance decomposes as

The first term summarizes within-cohort cross-modality covariance, averaged over the target cohort distribution. The second captures association between cohort-specific means. Consequently, a pooled anchor covariance can reflect both within-cohort relationships and between-cohort differences. If anchors come only from one cohort, they can estimate that cohort's relationship but cannot, without a transport assumption, reveal whether the within-cohort relationship is the same elsewhere or determine the target's between-cohort contribution.

If the goal is a target-population estimate, allocate anchors to the cohorts and important subgroups that define that target. If the goal is a within-cohort relationship, ensure enough anchors within the cohort for that estimate. When a cohort is rare in the target but scientifically important, disproportionate anchor sampling may be useful, but the resulting estimate must be combined using target-population weights rather than treated as a simple unweighted average. If the relationship is believed invariant across cohorts, anchors can be pooled to improve precision, but that invariance is an assumption to scrutinize, not a consequence of having many paired subjects.

Alternatives, Diagnostics, and Limits

Complete overlap, in which every subject has every modality, gives direct paired information throughout the study but may be costly or infeasible. A deliberately chosen anchor subset is more economical, though it trades measurement cost for dependence on how representative and well placed the subset is. With no anchors, approaches such as model-based imputation or other partial-overlap methods may still estimate cross-block quantities under additional modeling and identification assumptions; separate marginal data alone do not identify an arbitrary cross-block covariance. Anchors contribute empirical paired evidence, but do not remove assumptions about selection, measurement comparability, or transport.

Before using an anchor estimate, check the pairing and data alignment, compare anchor and target cohort composition, inspect the observed range of each feature, and look for influential measurements. Examine whether the estimated relationship differs meaningfully across cohorts or key strata when enough anchors exist to make that comparison informative. Sparse coverage in a subgroup is not evidence that the relationship is identical there. A linear model should also be checked for systematic departures from linearity and changing residual spread; wider predictor coverage cannot repair a misspecified relationship.

A common mistake is to claim that a large anchor count guarantees a valid bridge. The reasoning is wrong because sample size mainly controls sampling variability under the assumed model; it does not fix selection bias, poor cohort coverage, assay incompatibility, or an incorrect transport assumption. The correction is to assess both precision (how much information the anchors provide) and relevance (whether those anchors represent the target relationship). Anchor estimates can be precise for the wrong population.