What Horizontal Integration Combines
A multi-omics study may include several cohorts: groups of subjects recruited or measured in distinct study settings. If the cohorts have measurements on comparable molecular features, their subjects can be combined into one larger subject-by-feature data set. The subjects are added as rows; the feature columns are retained in common. This is horizontal integration.
The word “horizontal” describes the direction of combination in a sample-by-feature matrix. It does not mean that the same subjects were measured repeatedly, nor that additional molecular layers were added. Cohorts may contain different subjects, and their sample sizes need not be equal. What must be established is that each column in the combined matrix refers to the same measurable feature in every contributing cohort.
For cohort \(c\), let \(X^{(c)}\) be its sample-by-feature matrix, with \(n_c\) subjects and \(p\) aligned features. If the cohorts contribute distinct subjects and the features are in the same order, the combined matrix is formed by stacking the rows:
The matrix alone is not a sufficient record of the design. Each row should remain linked to its subject identifier and cohort label, along with relevant sample and assay metadata. Without cohort labels, it may be impossible to investigate whether observed differences reflect biology, study procedures, or both.
Conditions for Meaningful Combination
Horizontal integration is a data-design operation, not a guarantee that the combined measurements are directly comparable. Before stacking rows, check the meaning and provenance of the shared columns. Feature identifiers should refer to the same molecular entities or measurement targets; units, specimen types, and measurement definitions should be compatible; and any feature mapping should be documented. For example, two columns both labelled “gene expression” do not necessarily represent the same transcript-level feature or the same scale.
Cohort protocols can also differ in collection time, tissue handling, assay platform, laboratory, or processing. Some differences may be small, while others may create systematic shifts or different measurement properties. A shared feature name is not enough to rule these out. In addition, the subject groups and outcome definitions should be recorded consistently. “Case” in one cohort should not be assumed to mean the same thing as “case” in another without checking the study definitions.
The practical assumptions depend on the intended analysis. For a simple combined descriptive summary, the feature values must have a common interpretation, and the subjects included in the summary must be clearly defined. For inferential analyses that treat subjects as independent observations, the sampling and dependence structure must support that assumption. If a subject appears in more than one cohort or data source, those rows are not automatically independent subjects. Horizontal integration does not erase such design relationships.
Worked Examples
Worked Example: Stacking Two Comparable Cohorts
Suppose two hypothetical cohorts measure the same three proteins, in the same units, from the same specimen type using a documented common assay procedure. Cohort A has two subjects and cohort B has one. Their values are:
Cohort A: subject A1 has \((4, 7, 2)\), and subject A2 has \((5, 6, 3)\). Cohort B: subject B1 has \((3, 8, 4)\). The columns, in order, are proteins P1, P2, and P3, with values in hypothetical abundance units.
Each cohort matrix has three columns, and the columns have the same measurement meaning and order. Row-stacking gives:
The combined matrix therefore has three subject rows and three feature columns. The row labels must also be retained: A1 and A2 belong to cohort A, while B1 belongs to cohort B. The dimensions confirm that stacking is possible, while the stated common specimen, feature definitions, units, and assay procedure support the interpretation of columns as comparable measurements. This small example does not establish that real cohorts are exchangeable; it shows the data structure of the operation.
Worked Example: Deciding Which Columns Are Shared
Two hypothetical RNA-sequencing cohorts each report five columns. After checking the feature annotations, cohort A contains genes G1, G2, G3, G4, and G5; cohort B contains G1, G2, G3, G4, and G6. Assume the four shared genes have compatible identifiers and measurement definitions. The cohorts have 12 and 9 subjects, respectively.
The common feature set is \(\{G1,G2,G3,G4\}\), so the direct shared-feature matrix has \(12+9=21\) rows and 4 columns. The overlap is \(4/5=0.8\), or 80% of the features in either cohort. The 80% figure describes this particular feature-list overlap; it does not by itself establish that the cohorts are suitable to combine.
G5 is not measured in cohort B, and G6 is not measured in cohort A. A matrix containing all six genes would require a clear representation of the unavailable measurements, rather than silently treating them as observed zeros. For a straightforward common-feature integration, use the four verified shared genes and document that the analysis excludes G5 and G6. If feature identifiers or annotations do not establish that similarly named columns are the same targets, those columns should not be treated as shared solely to make the matrices align.
Worked Example: Checking Disease Representation Across Cohorts
Consider two hypothetical cohorts with case and control labels. Cohort A includes 10 cases and 10 controls; cohort B includes 8 cases and 12 controls. The combined data contain \(10+8=18\) cases and \(10+12=22\) controls, for 40 subjects in total. Both disease groups occur within both cohorts, so disease group is not perfectly tied to cohort membership.
This arrangement permits comparisons that take cohort membership into account, although it does not guarantee that the groups are otherwise comparable or that there is no cohort effect. Now consider a different design: cohort A contains 20 cases and no controls, while cohort B contains no cases and 20 controls. Here the case indicator is exactly determined by cohort: every case is from A and every control is from B. A difference between the combined case and control averages could reflect disease status, cohort, or both.
The arithmetic identifies the design issue: in the second arrangement, the two groups have zero overlap in cohort membership. Merely stacking the 40 rows does not resolve this confounding. The subject and cohort labels must be examined before interpreting a combined group comparison.
Why Cohort Labels Matter: A Short Justification
Even when all cohorts use the same feature columns, a pooled feature average can combine within-cohort values with differences between cohort averages. For one feature, let \(\bar{x}_c\) be the mean in cohort \(c\), and let \(n_c\) be its subject count. The mean after stacking is:
To see why, write the sum of all observed values as the sum of the cohort-specific sums. In cohort \(c\), that sum is \(n_c\bar{x}_c\). Dividing the total sum by the total number of subjects gives the displayed expression. Thus the pooled mean is a sample-size-weighted average of cohort means, not generally an equally weighted average of cohorts.
For example, if a feature has cohort means 10 and 14 units in cohorts with 10 and 30 subjects, respectively, then the pooled mean is \((10\times10+30\times14)/(10+30)=520/40=13\) units. It is closer to 14 because the second cohort contributes more subjects. If cohort membership is also associated with case status, a pooled case-control difference can mix cohort composition with the biological contrast of interest. Keeping cohort labels makes such structure visible and allows the analysis to reflect the design.
Diagnostics, Alternatives, and Limitations
Before or after constructing a combined matrix, inspect whether the assumptions are plausible. Useful checks include:
- Compare feature identifiers, annotation versions, units, and column ordering across cohorts.
- Check that specimen type, assay definitions, and outcome labels have compatible meanings.
- Tabulate subjects by cohort and key study groups to identify sparse or absent combinations.
- Summarize and visualize feature distributions by cohort to identify large shifts, unusual ranges, or cohort-specific patterns.
- Confirm that each row represents the intended subject or sample and that repeated observations are not being counted as unrelated subjects.
A distributional difference across cohorts is a warning to investigate, not automatic proof of a technical artifact: cohorts may differ biologically as well as procedurally. Conversely, similar distributions do not prove that every feature is measured comparably. When measurement definitions or specimen types are incompatible, forcing the columns together can make the resulting matrix misleading.
Row-stacking is one choice. An alternative is to analyze cohorts separately and report cohort-specific estimates, or to combine appropriately defined estimates using meta-analysis. Separate analyses preserve cohort-specific results but may be less convenient when a joint subject-level analysis is needed. Meta-analysis combines estimates rather than simply treating every row as though it came from one undifferentiated source; it requires estimates that address a sufficiently comparable question. Neither alternative fixes incompatible feature definitions or an unrecorded design problem.
The main limitation of horizontal integration is that larger row count does not automatically mean greater effective information. Cohort-specific measurement shifts, differing populations, imbalanced group representation, and dependence between observations can affect interpretation. Nor does a common feature set imply that all subjects have the same measured features in every practical sense: missingness and assay failures may still occur within cohorts. Those issues should remain distinguishable from features that a cohort never measured.
Common Mistake
Wrong reasoning: “Both files have columns named G1 through G500, so I can stack the rows, discard the cohort column, and analyze all subjects as one homogeneous sample.”
Correction: Matching labels are only a first check. Verify that the columns refer to the same features and compatible measurements, retain cohort membership, and examine how cohort relates to the subject groups of interest. A combined matrix is a representation of the data, not evidence that cohort effects or confounding are absent.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Cohort A has 24 subjects and cohort B has 31. Both have measurements on 120 verified shared features. What are the dimensions of the row-stacked matrix?
- Two cohorts use the same column label for a protein, but one assay reports concentration per millilitre and the other reports a normalized intensity score. Is the label alone enough to establish a shared comparable feature?
- In a combined data set, every case comes from cohort A and every control comes from cohort B. What design concern arises?
- Why should cohort labels be retained after row-stacking, even when the feature columns are aligned?
- Give one reason to analyze cohorts separately rather than immediately row-stack them.
Solutions. (1) The matrix has \(24+31=55\) rows and 120 columns. (2) No. The measurement definitions and units differ, so the columns are not directly comparable on the basis of the shared label alone. (3) Disease group and cohort are perfectly confounded in this design, so a group difference cannot be separated from a cohort difference using these data alone. (4) Cohort labels allow analysts to inspect cohort-specific patterns and account for the study structure; removing them hides possible shifts or imbalances. (5) For example, cohorts may have incompatible measurement procedures or substantially different populations, making a single pooled subject-level analysis inappropriate without further harmonization or justification.