Why Cohort and Platform Need to Be Diagnosed Together
A molecular feature can differ between clinical groups because of biology, but it can also differ because samples came from different cohorts or were measured on different assay platforms. These explanations may coexist. A pooled case-control comparison can therefore be misleading if cases and controls are distributed differently across cohorts and platforms. Before interpreting a group difference, inspect which combinations of cohort, platform, and group are actually represented.
A cohort is the study or recruitment set from which a subject came; a platform is the assay technology or measurement system used for an omics layer. A cohort-by-platform table summarizes sample availability across both dimensions. Adding group counts to each cell shows whether cases and controls can be compared under similar measurement conditions. The table is a diagnostic of the design, not a correction for confounding by itself.
Let \(Y_i\) denote a quantitative molecular feature or clinical outcome for subject \(i\), \(G_i\) its group indicator (for example, case \(=1\), control \(=0\)), \(C_i\) its cohort, and \(P_i\) its platform. A useful first model for diagnosing an adjusted group difference is \(Y_i=\beta_0+\beta_GG_i+\gamma_{C_i}+\delta_{P_i}+\varepsilon_i\). Here \(\gamma\) and \(\delta\) represent cohort and platform effects, with a reference level or an equivalent constraint used for each set of categorical effects. The model assumes that the conditional mean is additive, that the specified effects capture systematic mean differences, and that residuals have mean zero. For conventional exact small-sample inference, independent normally distributed residuals with a common variance are additionally assumed; descriptive comparisons do not require that full distributional assumption.
Start With the Design Table
Make one row for each cohort-platform combination and report the number of cases and controls separately. Include zero-count combinations rather than omitting them. A group with no representation in a cell has no direct within-cell comparison there. Also check whether an assay platform occurs in more than one cohort and whether each cohort uses more than one platform. Those patterns reveal which effects are being contrasted by the data.
Worked Example: Reading a Cohort-by-Platform Table
Suppose a hypothetical study has 40 subjects, two cohorts, and two platforms. The table shows the number of cases and controls in each cell.
| Cohort | Platform | Cases | Controls |
|---|---|---|---|
| Cohort 1 | A | 10 | 0 |
| Cohort 1 | B | 0 | 10 |
| Cohort 2 | A | 0 | 10 |
| Cohort 2 | B | 10 | 0 |
There are no within-cell case-control comparisons: every cohort-platform cell contains only one group. It would be incorrect, however, to conclude from that fact alone that the group effect is not identifiable in an additive model. The pattern of group labels across the four cells matters. Here the case pattern is 1, 0, 0, 1, in the order shown.
For two cohorts and two platforms, an additive combination of cohort and platform effects must have zero difference-in-differences. If its cell values are \(a_{11},a_{12},a_{21},a_{22}\), then \(a_{11}+a_{22}-a_{12}-a_{21}=0\). For the group indicator in this table, the contrast is \(1+1-0-0=2\), not zero. Thus the group indicator is not an additive combination of the cohort and platform indicators, and its coefficient can be identified in the additive model. But that identification is model-dependent: the model assumes no cohort-by-platform interaction that could explain the same contrast. Moreover, there are no direct within-cell group comparisons, so the result should not be described as supported by matched cases and controls in each cell.
The general linear-model check is precise. Let \(Z\) be the design matrix containing the intercept and nuisance indicators for cohort and platform, and let \(g\) be the vector of group indicators. The group coefficient is separately identifiable from those nuisance effects when \(g\) is not in the column space of \(Z\), subject to the rest of the design being estimable. If \(g\) is in that column space, adding a group coefficient does not add a distinct direction to the model; different combinations of group, cohort, and platform coefficients yield the same fitted values. Lack of shared cohort-platform cells is not, by itself, this criterion.
Feature Summaries as Assay Diagnostics
For each informative cohort-platform cell, summarize each feature using its sample size, mean, standard deviation, and, where useful, a plot of its distribution. Look for broad platform shifts across many features, unusual variance, outliers, and patterns that occur only in one cohort. These summaries can flag measurement differences, but a platform-associated shift is not automatically a technical artifact: biological composition may also differ across cells. Group-stratified summaries, reference or quality-control samples, and knowledge of how samples were assigned to platforms help distinguish these explanations.
Worked Example: Spotting a Platform-Associated Feature Shift
Suppose a hypothetical set of pooled quality-control samples has been measured on both platforms. A molecular feature is recorded on a log-abundance scale. The cell means and sample standard deviations are:
| Cohort | Platform | QC sample size | Mean log abundance | Standard deviation |
|---|---|---|---|---|
| Cohort 1 | A | 4 | 5.0 | 0.20 |
| Cohort 1 | B | 4 | 7.0 | 0.24 |
| Cohort 2 | A | 4 | 5.2 | 0.22 |
| Cohort 2 | B | 4 | 7.1 | 0.25 |
Within Cohort 1, the platform difference is \(7.0-5.0=2.0\) log-abundance units. Within Cohort 2, it is \(7.1-5.2=1.9\) units. The similar differences, together with the small cohort shifts within each platform (0.2 units on A and 0.1 on B), are a warning that raw values from A and B may not be directly comparable for this feature. The standard deviations are also similar in these summaries, so there is no obvious large variance increase on one platform in this small QC sample.
This comparison supports a platform-associated shift in the QC measurements; it does not establish why the shift occurred or prove that subject-level group differences are technical. A useful next check is whether the same shift appears in appropriate reference samples and across many features. QC replicates are not substitutes for cases and controls, and these summaries alone do not justify a biological group conclusion.
Crude and Adjusted Outcome Comparisons
Compare the pooled group summaries with within-cell or model-adjusted comparisons. A crude difference averages over the cohort-platform composition actually observed in each group. If those compositions differ, the crude difference may reflect composition as well as group. The additive model estimates a common group contrast after accounting for cohort and platform, but its interpretation depends on the additivity assumption and on adequate information in the design.
Worked Example: A Pooled Difference Reverses Within Cells
Consider a hypothetical feature measured in two cohorts on platforms A and B. The listed means are in feature units. Within every cohort-platform cell, the case mean is exactly 2 units above the control mean, but the group counts are unbalanced.
| Cohort-platform cell | Case count | Case mean | Control count | Control mean |
|---|---|---|---|---|
| Cohort 1, A | 8 | 12 | 1 | 10 |
| Cohort 1, B | 2 | 14 | 1 | 12 |
| Cohort 2, A | 1 | 22 | 8 | 20 |
| Cohort 2, B | 1 | 24 | 8 | 22 |
The pooled case mean is
The pooled control mean is
The crude difference, case minus control, is \(14.17-19.89=-5.72\) feature units. Yet the four within-cell differences are \(12-10=2\), \(14-12=2\), \(22-20=2\), and \(24-22=2\) units. The reversal occurs because most cases are in the lower-baseline cells, while most controls are in the higher-baseline cells. Under an additive model with a common group effect, the adjusted group contrast is 2 units, not the crude pooled difference. The table provides direct group overlap in every cell; the key diagnostic is the mismatch in cell composition.
The computation uses the stated cell means and counts. It does not supply a standard error or establish statistical significance, because the individual-level variability is not given. The appropriate conclusion is that the pooled mean difference is misleading as a group comparison in this hypothetical design, while the consistent within-cell contrasts suggest a positive adjusted difference under the common-effect model.
When the Group Effect Is Not Separately Identified
Worked Example: Perfect Group-Platform Confounding
In a second hypothetical study, cases were measured only on Platform A and controls only on Platform B. Both cohorts used both platforms, and each cohort contains 10 cases on A and 10 controls on B. Suppose the observed feature means are 12 for cases and 8 for controls in each cohort.
The platform indicator is identical to the group indicator: \(G_i=1\) exactly when the platform is A. In the model \(Y_i=\beta_0+\beta_GG_i+\delta I(P_i=A)+\gamma_{C_i}+\varepsilon_i\), the fitted mean depends on the sum \(\beta_G+\delta\) for Platform A and the baseline plus cohort effect for Platform B. The observed 4-unit difference identifies that combined contrast, not separate values of \(\beta_G\) and \(\delta\). For example, \((\beta_G,\delta)=(4,0)\) and \((0,4)\) give the same fitted group/platform means. The group effect is therefore not separately identifiable from platform in this design.
The fact that both platforms appear in both cohorts does not solve the problem, because group and platform remain perfectly aligned. New measurements that put both groups on comparable platforms, suitable bridge samples, or a design with genuine group-platform overlap would be needed to distinguish these explanations. A model cannot recover a separate group effect merely by including both terms when the design does not distinguish them.
Assumptions, Diagnostics, and Alternatives
A diagnosis should make its assumptions explicit. The additive model presumes that cohort and platform shifts can be represented by separate effects and that the group contrast is sufficiently common across the included cells. If the group effect varies by platform or cohort, or if platform performance differs by cohort, an interaction may be needed. But adding interactions also requires enough observations across the relevant combinations; sparse or empty cells can make a more flexible model unestimable. Compare cell-level plots and contrasts with model residuals rather than relying on one fitted coefficient.
- Check the design matrix. Confirm that the group indicator adds information beyond the cohort and platform indicators. Inspect empty cells, zero group counts, and redundant columns.
- Compare like with like. Where both groups occur in the same cohort-platform cell, calculate or plot within-cell differences. Report the cell counts alongside each contrast.
- Inspect feature distributions. Look for shifts, spreads, outliers, or irregular patterns that are concentrated on one platform or in one cohort. A mean alone can hide these problems.
- Check model adequacy. Large systematic residual patterns or inconsistent cell-specific group differences challenge an additive common-effect summary.
- Trace sample processing. Review whether group, cohort, and platform assignments arose together, and whether reference samples or technical replicates provide cross-platform comparisons.
Stratified within-cell comparisons make fewer assumptions about how cohort and platform effects combine, but they require overlap and may be imprecise in small cells. An additive regression can use the design more efficiently when its assumptions are credible, but it can be misleading if interactions are important. A random-effects model may describe variation among cohorts when its assumptions are appropriate; it does not repair perfect group-platform confounding. Re-measuring samples or adding bridge samples can provide information that the original design lacks. No statistical adjustment can guarantee separation when the group indicator is exactly explained by the nuisance indicators.
Common Mistakes and Limits
- Wrong: “There are no cells containing both cases and controls, so the group effect is necessarily unidentifiable.” Correction: Check whether the group indicator is in the span of the cohort and platform indicators. Some crossed patterns, such as the 1, 0, 0, 1 pattern above, are identifiable under an additive model even without within-cell comparisons.
- Wrong: “Cases and controls have different pooled means, so the feature is a disease signal.” Correction: Compare group composition across cohort-platform cells and inspect adjusted or within-cell contrasts before interpreting a pooled difference.
- Wrong: “Including cohort and platform in a regression always removes their influence.” Correction: Adjustment requires an estimable design and a credible model. It cannot separate effects that the observed design aliases, and omitted interactions can still distort interpretation.
- Wrong: “A platform-associated feature shift proves that all apparent biological differences are artifacts.” Correction: It is a warning that needs context, reference measurements, and group-aware comparisons; platform association alone does not identify the cause.
These checks diagnose the credibility of comparisons; they do not establish that a feature causes an outcome, nor do they guarantee transportability to another population or assay setting. Report the cohort-platform counts, the comparisons actually supported by the design, the model assumptions, and any unresolvable aliasing. When the design only supports a combined cohort/platform contrast, say so rather than assigning it uniquely to biology.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A study has cases only on Platform A and controls only on Platform B, with both platforms represented in each cohort. What additional fact determines whether group and platform effects can be separated?
- In a two-cohort, two-platform design, the group indicator across cells is 1, 0, 0, 1. Does the lack of both groups in each cell alone show that a group coefficient is not identifiable under an additive model?
- In a hypothetical cell, cases have mean 15 units and controls have mean 12 units. State the within-cell difference and one important limitation of interpreting it.
- Why can pooled case and control means have the opposite ordering from all within-cell comparisons?
Solutions. (1) Check whether the group indicator is in the column space of the cohort and platform indicators. If it is, the effects are not separately identifiable; if not, the group coefficient can be identified under the model. (2) No. The additive difference-in-differences of the group pattern is \(1+1-0-0=2\), not zero, so that pattern is not an additive combination of cohort and platform indicators. The coefficient is model-dependent and lacks direct within-cell comparisons. (3) The within-cell case-minus-control difference is \(15-12=3\) feature units. It is a cell-specific association, not proof of causation, and its precision and consistency elsewhere are not known from these means alone. (4) The groups can have different proportions in high- and low-baseline cells. Their pooled means then reflect both within-cell group differences and cell composition, allowing a reversal such as the one calculated above.