What Makes an Omics Design Unmatched?
An integrated dataset may contain measurements from several assays, but that does not mean the measurements came from the same people. In an unmatched design, the subject sets contributing to different assay datasets are intentionally distinct: a subject represented in one dataset is not represented in the other. The key design decision is therefore not only which assays to run, but which subjects are assigned to each assay and why.
This distinction matters because samples from different people do not provide direct within-subject evidence about how two omics layers vary together. The study may still describe each layer or compare groups, subject to appropriate sampling and analysis assumptions. But it cannot treat separate measurements as if they were paired. A clear design description makes the boundary between supported and unsupported conclusions visible before analysis begins.
The definition is about subjects, not merely files, samples, or assay names. One person may contribute several specimens, and one specimen may be divided into aliquots. If those records can be linked to the same person, the subject sets overlap even if the files have different names. Conversely, two datasets with the same feature columns are not matched merely because their measurements are comparable.
Specify the Subject Source and Assay Allocation
A useful design specification answers the same questions separately for every dataset. State who could enter the subject source, who was eligible, how participants were selected, what specimen and time point were used, and which assay was assigned. Then state whether any subjects could contribute to more than one dataset. “We collected RNA and protein data” is not enough: it does not say whether both assays were performed on each person or whether two separate groups were sampled.
| Design element | What to state | Why it matters |
|---|---|---|
| Subject source | Recruitment setting and population from which subjects were drawn | Shows which people the observations represent |
| Eligibility and sampling | Inclusion rules, exclusion rules, and how subjects entered the study | Clarifies how each dataset's subjects were selected |
| Assay allocation | Which assay was run on which subjects, specimens, and time points | Establishes whether datasets overlap at the subject level |
| Linkage | Whether a stable subject identifier connects assay records | Separates true non-overlap from incomplete record linkage |
| Departures from plan | Assay failures, unavailable specimens, withdrawals, or missing identifiers | Distinguishes the intended design from the observed data |
An unmatched design can use subjects sampled independently from the same stated source, or subjects from different sources. These are not interchangeable descriptions. If sources differ, document the difference rather than implying a common sampling frame. Whether two sources are suitable for a particular integration question requires substantive justification; the design statement alone does not make them comparable.
Worked Example: Two Assays on Deliberately Separate Subjects
Suppose a hypothetical clinic plans an exploratory study of a molecular condition. It recruits 80 eligible patients for a blood RNA assay and a different 70 eligible patients for a plasma protein assay. No patient is assigned to both assay groups. Describe the design and its implication.
Let \(S_R\) be the 80 patients assigned to the RNA dataset and \(S_P\) the 70 assigned to the protein dataset. The allocation rule gives \(S_R \cap S_P = \varnothing\), and the dataset sizes are \(n_R=80\) and \(n_P=70\). The design description should also record the clinic as the subject source, the eligibility criteria, the recruitment procedure, and the specimen and collection time for each assay.
Because the subject sets do not overlap, no row in the RNA dataset is paired with a protein measurement from that same patient. The study can summarize the RNA measurements and protein measurements separately. It cannot estimate a within-patient RNA–protein relationship by matching rows across these datasets. That limitation follows from the allocation, not from the number of features measured.
Intentional Non-Overlap Is Not Lost Correspondence
A design is not unmatched merely because the exported data contain no usable pairs. It is important to distinguish what was planned from what can currently be linked. If the protocol assigned different subjects to each assay, non-overlap is intentional. If the same subjects were intended to provide multiple assays but identifiers were omitted, corrupted, or separated from the data, correspondence has been lost in the records. If a subject was assigned both assays but one assay failed, the design still planned matched measurements, while the observed data contain an incomplete assay record for that subject.
These situations have different implications. Identifier problems may be resolvable using a verified linkage key or source records; they should not be “solved” by guessing matches from row order or similar measurements. Assay failures and unavailable specimens are forms of incomplete measurement, not evidence that separate sampling was intended. The earlier material on missingness gives context for recording and examining within-assay missingness; here the central concern is accurately naming the design and its correspondence.
Worked Example: Planned Pairs With Incomplete Linkage
In a hypothetical protocol, all 60 enrolled subjects are assigned both an RNA assay and a protein assay. Both assays yield a measurement for every subject, but an export retains subject identifiers for only 56 RNA records and 57 protein records. The overlap among records with identifiers is 53 subjects.
The protocol defines the intended design: each of the 60 subjects was assigned both assays. Therefore, this is a planned matched design, not an intentionally unmatched design. In the export, 3 of the 56 identified RNA records and 4 of the 57 identified protein records lack a counterpart among the identified records, since \(56-53=3\) and \(57-53=4\). The other records may be linkable if identifiers can be recovered, but the counts alone do not reveal which records correspond.
The correct action is to preserve the unmatched records as unresolved and check the source files or linkage documentation. It would be incorrect to pair rows by their positions or to describe the protocol as separate sampling. If the missing identifiers cannot be recovered, report the loss of linkage and distinguish the intended design from the analysis-ready data.
Why Separate Samples Cannot Supply a Within-Subject Association
The limitation can be seen directly from the definition of covariance. For paired measurements \(X_i\) and \(Y_i\) on the same subject \(i\), sample covariance uses products of deviations belonging to the same person:
With deliberately separate samples, there is no observed \(Y_i\) for the subject who supplied \(X_i\). The marginal measurements provide the separate values and means, but not the subject-specific products in this formula. The following numerical illustration shows why pairing the rows arbitrarily does not fix the problem.
Worked Example: Identical Marginals, Different Pairings
Consider hypothetical measurements with \(X=(1,2,3,4)\) and \(Y=(1,2,3,4)\), each recorded on a different set of four subjects. Both sample means are \(\bar{x}=\bar{y}=2.5\). If one were to pair the values in the displayed order, the covariance calculation would be:
If the same \(Y\) values were instead paired in reverse order, the numerator would be \((-2.25)+(-0.25)+(-0.25)+(-2.25)=-5\), so the covariance would be \(-5/3=-1.667\). The two arrangements have exactly the same marginal values and means, yet give opposite covariance signs. Since the samples are separate, neither arrangement is a real subject-level pairing. Thus the within-subject covariance is not identified by these marginal datasets alone.
This argument does not say that no joint relationship can ever be studied with separate samples. It says that the observed separate measurements do not themselves reveal which value of one layer belongs with which value of the other in an individual. A population-level relationship may be addressed only with an appropriate design and assumptions; it must not be presented as a directly observed within-subject association.
Audit the Allocation Before Integration
A practical audit starts from the protocol and subject-level records, not from a visual impression of the matrices. For each dataset, record the source and assay allocation, then compare subject identifiers across datasets. Classify every apparent non-overlap as intentional separate sampling, an assay or specimen failure, an identifier or linkage problem, or an unresolved case. Do not collapse these categories into a single label such as “unmatched.”
For each assay, specify the subject source, eligibility, specimen, time point, and which subjects were assigned to it.
Use verified identifiers to determine whether the contributing subject sets intersect. Do not infer identity from row order, similar values, or matching feature profiles.
Separate planned non-overlap from assay failure, absent specimens, missing identifiers, and other deviations from the plan.
Describe available subject-level and dataset-level comparisons without claiming a within-subject link that was not observed or reliably recovered.
The audit relies on several conditions for a defensible description: the protocol or recruitment records must accurately describe assignment; identifiers used for linkage must be unique and reliable; and the unit called a “subject” must be consistent across datasets. If one table has one row per aliquot and another one row per person, counts of rows cannot be treated as counts of subjects without resolving that difference. When any of these conditions is uncertain, report the uncertainty instead of forcing a design classification.
Design Choices, Alternatives, and Failure Modes
The main alternative is to assign more than one assay to the same subjects, producing matched observations for those assay layers. That allocation makes subject-level cross-layer comparisons possible for subjects with both measurements, though it does not by itself guarantee adequate data or remove other sources of bias. A separate-sample design may be chosen for practical or scientific reasons, but it trades away direct subject-level correspondence. These are design distinctions, not a ranking: the appropriate choice depends on the research question and what specimens and resources are available.
| Situation | How to describe it | Design implication |
|---|---|---|
| Different subjects deliberately assigned to each assay | Intentionally unmatched datasets with stated sources and allocation | No observed within-subject cross-layer pairing |
| Same subjects intended for both assays; identifiers unavailable | Planned matched design with incomplete record linkage | Check records; do not invent pairs |
| Same subjects intended for both assays; one assay fails for some | Planned matched design with incomplete measurements | Report which measurements are absent and why, where known |
| Subject sources differ across assays | Separate source descriptions for each dataset | Do not imply a common sampling frame without evidence |
Common warning signs include a protocol that says “paired samples” while the analysis files have no shared identifiers, a report that calls samples unmatched solely because some assays failed, and a matrix join based on row number. Another failure mode is treating sample labels as subject identifiers when one person supplied repeated specimens. Each can misstate how the data were collected and create false correspondence.
An unmatched design also has limits beyond the absence of pairs. Differences between assay datasets may reflect differences in subject source, recruitment, specimen handling, collection time, or measurement process rather than a biological cross-layer pattern. The design statement should expose these features so that later analytic choices can account for them. It should not imply that separate samples are interchangeable or that a shared clinic name alone establishes comparability.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A hypothetical study recruits 40 people for RNA sequencing and a different 35 people for metabolomics, with no person in both groups. What is the design, and what within-subject claim is unavailable?
- A protocol assigns both assays to 25 subjects, but a data export has no identifiers in the metabolomics file. Is the intended design unmatched? What should be checked?
- Four RNA values and four protein values come from separate subjects. Explain why placing the values in matching row order does not provide a valid sample covariance.
- Of 50 subjects assigned two assays, 46 have both measurements, 3 have only the first assay, and 1 has only the second. How many subjects have at least one measurement, and is the intended allocation unmatched?
Solutions. (1) The design is intentionally unmatched because the two subject sets are disjoint. It does not provide observed within-person RNA–metabolite pairs or a direct within-subject association. (2) No. The intended design is matched; check source records and linkage documentation for the missing identifiers, and do not infer pairs from row order. (3) Covariance requires products of deviations for measurements from the same subject. Separate samples provide no such pairing, and arbitrary row ordering can produce different covariances from the same marginal values. (4) There are \(46+3+1=50\) subjects with at least one measurement. The intended allocation is matched because both assays were assigned to each subject; the observed data have incomplete measurements for 4 subjects.