Feature-Specific Platform Offsets Versus Cohort-Wide Shifts

Why the Pattern Across Features Matters

When two cohorts are measured on different platforms, their recorded values may differ for more than one reason. A platform may read a particular feature consistently high or low, while a broad shift may move many features in roughly the same direction. Those patterns matter: a feature-specific offset suggests that feature-level calibration may be relevant, whereas a broad shift suggests a common component worth investigating across the measured panel.

The pattern is evidence, not a diagnosis. A shift shared across many features could arise from a technical change, a real cohort-wide biological difference, or both. Distinguishing these possibilities requires attention to the study design and to measurements that connect platforms, such as shared specimens or reference materials. The earlier material on bridge samples and reference materials supplies useful design context; here the focus is how to interpret the pattern across features.

A Model for Common and Feature-Specific Differences

Suppose specimen \(i\) is measured for feature \(f\) on platform \(c\), and \(Y_{ifc}\) is the resulting log-scale abundance. Consider the additive model

Here, \(\theta_{if}\) is the specimen's underlying level for feature \(f\); \(g_c\) is a platform-wide shift shared across features; \(a_{fc}\) is a feature-specific platform offset; and \(\varepsilon_{ifc}\) is measurement error. The model is a working description, not a guarantee that real data obey it.

The split between \(g_c\) and \(a_{fc}\) is not identifiable without a convention: adding a constant to \(g_c\) and subtracting it from every \(a_{fc}\) gives the same fitted values. One common convention is to require the feature-specific offsets to sum to zero across the features under consideration, so the common component represents their average. The chosen feature set and convention therefore affect how the components are described. A robust summary such as a median can also be useful when a few features have unusually large offsets, but it does not remove the need to define the target set of features.

The model assumes that an additive description on the selected log scale is reasonable, that feature definitions are comparable across platforms, and that relevant measurements are not distorted by unmodelled specimen-by-platform interactions. If the same specimen is measured on both platforms, the comparison also requires that its underlying biological state has not changed between measurements. These assumptions should be checked against how samples were collected and processed.

Shared Specimens Separate Measurement Contrasts from Specimen Levels

A specimen measured on both platforms provides a direct paired contrast. For platforms A and B, define \(D_{if}=Y_{ifA}-Y_{ifB}\). Under the additive model, the underlying specimen level cancels:

This cancellation is the key justification for using shared specimens. The contrast no longer includes \(\theta_{if}\), so it can reveal how the platform difference varies across features without comparing different specimens' biology. It still includes measurement error and any platform-by-specimen effects not represented in the model. A single paired contrast is not enough to establish that a pattern is stable; repeated shared specimens help assess its consistency.

Summarizing Common and Feature-Specific Components

For each feature, first average its paired contrasts over the \(n\) shared specimens:

Then summarize the set of \(\bar{D}_f\) values with a common center, such as their mean or median. If the mean is used, the estimated common contrast is \(\hat{g}=\frac{1}{F}\sum_{f=1}^{F}\bar{D}_f\), where \(F\) is the number of included features. The deviation \(\bar{D}_f-\hat{g}\) describes how far feature \(f\)'s estimated contrast lies from that common center. A median center can reduce the influence of a few extreme features, but must be interpreted as a robust summary of the included feature panel, not as a universal platform correction.

These calculations assume that the selected features form a meaningful comparison set. For example, if a large fraction of features truly change together between specimen groups, the center of their contrasts may include biological change. A broad shift estimate is therefore conditional on the features and design used to calculate it.

Why Unpaired Cohorts Do Not Identify the Source

If platform A measures one cohort and platform B a different cohort, the observed cohort difference includes both specimen-level differences and measurement differences. Writing \(C\) for cohort, a difference in observed feature means can be expressed schematically as

Without shared specimens, suitable controls, or a design that otherwise separates cohort and platform, the terms cannot generally be disentangled from that difference alone. Even if every feature shows the same observed change, that pattern is compatible with a common technical shift, a real broad biological change, or a combination. The pattern can motivate a hypothesis; it cannot resolve the confounding.

Diagnostics, Alternatives, and Limits

Useful diagnostics examine both the overall pattern and departures from it. Plot \(\bar{D}_f\) across features, or display its distribution, to see whether most contrasts cluster around a common value and whether a small number depart sharply. Also inspect the contrasts across shared specimens: a feature whose contrast changes substantially from specimen to specimen may violate the simple additive description. Check whether the pattern persists when problematic measurements or a small set of extreme features are examined separately. Such checks identify instability; they do not prove a cause.

The main alternatives provide different kinds of information. Shared specimens estimate paired platform contrasts while controlling for each specimen's underlying feature level. Reference materials can provide a common benchmark across platforms, while negative-control features can help assess whether a shift appears in features not expected to reflect the biology of interest. Unpaired cohort comparisons can describe observed differences, but cannot by themselves separate platform from cohort when those factors align. These tools are complementary; none makes an unidentifiable design identifiable merely by summarizing more features.

A broad additive shift is also scale-dependent. A nearly constant difference on a log scale represents a roughly constant multiplicative ratio on the original scale; it is not a constant additive difference there. If platform differences vary with abundance or dispersion, a location-only description may be inadequate. Missing measurements, feature selection, and different feature panels can further distort apparent cross-feature patterns. Finally, even a highly consistent pattern across features and shared specimens identifies a measurement contrast under the model—not its physical mechanism or its relevance to a biological question.

Common Mistake

• Wrong reasoning: “Nearly every feature is shifted by the same amount, so the shift must be a platform artifact.”
• Correction: A common pattern is consistent with a broad platform effect, but also with a cohort-wide biological difference or a mixture. Shared specimens or reference materials strengthen the measurement comparison; design and external knowledge are still needed to assess the source.