Reporting Cohort Characteristics and Measurement Coverage

Why Cohort Characteristics and Coverage Need Separate Reporting

A cross-cohort result is difficult to judge if readers cannot tell who contributed measurements, which assays were available, or how much of each assay’s intended data was observed. A cohort table that lists only total sample size can conceal important differences: one cohort may have molecular data for nearly everyone, while another has measurements for only a subset; two cohorts may use different feature panels; or their clinical variables may use different definitions or units. Reporting these details makes the evidence interpretable. It does not, by itself, establish that cohorts are comparable.

The earlier material on validation design and cohort provenance concerns how cohorts were selected and where data came from. Here the focus is the descriptive record of cohort composition and observed measurement coverage. Such reporting should make clear which people, specimens, variables, and features are represented in each number.

Definition: A cohort characteristic is a descriptive attribute of the cohort or its participants, such as age, disease status, or sex or gender as recorded. Modality availability describes whether a participant has usable data for a modality, such as RNA sequencing or proteomics. Feature coverage describes which intended features were measured, and how completely they were measured, among the samples eligible for that assay. These are different summaries and require separately defined denominators.

Describe Cohort Characteristics With Explicit Denominators

A cohort-characteristics table typically reports the number of participants enrolled or included, followed by selected clinical and demographic summaries. For quantitative variables, the mean \(\bar{x}\) and sample standard deviation \(s\) can describe a roughly symmetric distribution; the median and interquartile range are often more informative for a skewed distribution. For categorical variables, give counts and percentages. State the units, category definitions, and the denominator for each percentage.

The denominator can differ from the cohort size. For example, age may be known for 117 of 120 participants, and disease stage may be recorded only for participants with a confirmed diagnosis. Reporting “median age, 54 years” without the number with observed age hides this distinction. A concise entry is “median age 54 years (IQR 46–63; observed for 117/120).” For a categorical entry, “stage III, 28/94 (29.8%)” identifies both the numerator and the applicable denominator.

Definitions also matter. “Baseline” should refer to a stated time point, such as the visit before treatment, and a characteristic such as disease stage should follow a stated classification. If cohorts use different definitions, show them rather than implying they are identical. Descriptive summaries do not demonstrate that two cohorts are statistically equivalent, and their values should not be interpreted as causal effects.

Worked Example: Summarizing a Small Cohort

Suppose a hypothetical validation cohort has eight participants with observed age, in years: 40, 42, 45, 47, 50, 52, 55, and 57. Disease status is recorded as case for five participants and control for three. Summarize these characteristics.

The age sum is \(40+42+45+47+50+52+55+57=388\), so the mean age is \(388/8=48.50\) years. The squared deviations from 48.50 sum to 258. The sample standard deviation is therefore:

s=\sqrt{\frac{\sum_{i=1}^{8}(x_i-\bar{x})^2}{n-1}} =\sqrt{\frac{258}{7}} =\sqrt{36.8571} \approx 6.071\text{ years}.

The cohort can be described as having mean age 48.50 years and sample standard deviation 6.071 years, with age observed for 8/8 participants. Disease status is 5/8 cases (62.5%) and 3/8 controls (37.5%). These summaries describe this hypothetical cohort; they do not imply that its age or disease-status distribution represents a wider population.

Participant-Level Modality Availability

For modality availability, first decide what counts as usable data and which participants belong in the denominator. A common report gives the number with usable data for each modality divided by the cohort’s enrolled or otherwise specified eligible sample size. If participants can have several modalities, also report the counts for their combinations. Separate modality totals do not reveal how many people have paired measurements.

Let \(n_c\) be the number of participants in cohort \(c\), and let \(a_{icm}=1\) when participant \(i\) in cohort \(c\) has usable data for modality \(m\), and \(0\) otherwise. The modality availability proportion is the number with usable data divided by the stated denominator:

\widehat{A}_{cm}= \frac{\sum_{i=1}^{n_c}a_{icm}}{n_c}.

This expression assumes the denominator represents the participants to whom the availability question applies. If eligibility for the assay is narrower than the enrolled cohort, state that denominator and also report, where useful, how many participants were excluded from assay eligibility. Do not silently use different denominators across cohorts or modalities.

Worked Example: Reporting Availability and Paired Measurements

Consider two hypothetical validation cohorts. Cohort A has 120 participants: 108 have usable RNA data, 90 have usable protein data, and 82 have both. The remaining counts are 26 with RNA only, 8 with protein only, and 4 with neither; these sum to 120. Cohort B has 100 participants: 95 have RNA, 60 have protein, and 57 have both. It has 38 with RNA only, 3 with protein only, and 2 with neither; these sum to 100.

In Cohort A, RNA availability is \(108/120=0.900\), or 90.0%; protein availability is \(90/120=0.750\), or 75.0%; and paired availability is \(82/120=0.6833\), or 68.3%. In Cohort B, the corresponding values are \(95/100=95.0\%\), \(60/100=60.0\%\), and \(57/100=57.0\%\). The modality totals alone would not show that the number with both modalities is 82 in A and 57 in B. Reporting the joint counts clarifies how many participants can contribute paired measurements.

These percentages describe availability in each hypothetical cohort under the stated usable-data definition. They do not explain why data are unavailable. A reader should not infer that missing measurements are random or that observed participants represent those without data.

Feature Coverage and Assay Details

Availability answers whether a participant has a modality at all. Within an available modality, some intended features may be unmeasured or unusable. For a feature \(j\), let \(b_{icj}=1\) if a valid measurement for that feature is available for eligible participant \(i\), and \(0\) otherwise. If \(N_c\) participants were eligible for that assay, feature coverage is:

\widehat{C}_{cj}= \frac{\sum_{i=1}^{N_c}b_{icj}}{N_c}.

A report should state what “valid” means for the coverage summary and how the eligible denominator was defined. It should also distinguish at least three counts where relevant: features on the intended panel, features detected at least once, and features meeting a stated sample-coverage threshold. A feature detected in one sample is not necessarily well covered across the cohort. The threshold is a reporting criterion, not proof that the feature is biologically meaningful or analytically reliable.

For cross-cohort feature comparisons, report the feature identifiers and panel definitions, not just the number of features in each cohort. The overlap between panels can be summarized by the intersection and union. For feature sets \(F_A\) and \(F_B\), the Jaccard index is the number of shared features divided by the number in either set:

J(F_A,F_B)=\frac{|F_A\cap F_B|}{|F_A\cup F_B|}.

This index describes panel overlap, not whether shared features were measured with equal quality or on comparable scales. Assay reporting should therefore also identify, as applicable, the specimen type, platform or kit and version, measurement units, detection limits, laboratory or processing batch information, and the stated quality-control criteria. Where available measurements were filtered or transformed, report enough detail for readers to understand what the reported coverage refers to; detailed analysis choices belong in the analysis report.

Worked Example: Summarizing Feature Coverage and Panel Overlap

Suppose a hypothetical proteomics assay has an intended panel of 500 protein features in a cohort of 50 participants. The assay was attempted for 48 participants; 2 of the 50 had no usable assay data. Of the 500 intended features, 460 were detected in at least one of the 48 assayed participants. Using a stated high-coverage criterion of valid measurements in at least 90% of assayed participants, 400 features meet that criterion.

The participant-level assay availability is \(48/50=0.960\), or 96.0%. The proportion of intended features detected at least once is \(460/500=0.920\), or 92.0%. The proportion of the intended panel meeting the high-coverage criterion is \(400/500=0.800\), or 80.0%. These are three different quantities: one concerns participants, while the other two concern features and use the intended panel as their denominator.

For a feature-level illustration, suppose protein P has valid measurements for 45 of the 48 assayed participants. Its coverage is \(45/48=0.9375\), or 93.75%, so it meets the 90% criterion. If protein Q is valid in 39 participants, its coverage is \(39/48=0.8125\), or 81.25%, so it does not meet that criterion. The 2 participants without usable assay data remain visible in the separate \(48/50\) availability summary; they are not silently counted as feature-level assay attempts.

Finally, suppose a second hypothetical cohort’s panel contains 450 features and the two panels share 400. Their union contains \(500+450-400=550\) features, so \(J=400/550=0.7273\). Thus 72.73% of the combined distinct feature set is shared. The report should still identify which features overlap and describe the assay platforms; the index alone does not establish measurement equivalence.

Combining Coverage Rates and Checking Comparability

When a summary across cohorts is useful, calculate it from the underlying counts and denominators. An unweighted mean of cohort percentages gives each cohort equal influence, regardless of its size. The overall rate across eligible observations instead weights each cohort by its denominator. If cohort \(c\) has \(y_c\) valid measurements among \(N_c\) eligible measurements, the combined rate is:

\widehat{C}_{\mathrm{combined}} =\frac{\sum_c y_c}{\sum_c N_c} =\frac{\sum_c N_c\widehat{C}_c}{\sum_c N_c}.

The second equality follows because \(y_c=N_c\widehat{C}_c\), so the combined numerator is \(\sum_c N_c\widehat{C}_c\). This pooled descriptive rate is appropriate only when the counts refer to the same kind of measurement and the denominator has the same meaning across cohorts. If cohorts used different panels or eligibility rules, report cohort-specific rates and clarify their definitions rather than presenting a misleading pooled percentage.

Reporting choice	What it helps readers judge	What it cannot establish
Cohort-specific counts and percentages	Who was represented and how much data were available within each cohort	That cohorts are statistically equivalent
Joint modality availability	How many participants have paired data across modalities	Why participants lack a modality
Feature-level coverage and panel overlap	Which intended measurements are shared and how completely features were observed	That shared features have comparable assay performance
A pooled coverage rate	Overall coverage across eligible observations when definitions match	That coverage is equal across cohorts or that missingness is ignorable

Useful checks include verifying that modality-combination counts sum to the stated cohort denominator, that numerators do not exceed denominators, and that feature counts reconcile with the intended panel. Large changes in denominators across adjacent table entries should prompt a check of eligibility and missing data. If the same clinical characteristic uses different time points, categories, or units across cohorts, flag the difference explicitly. These checks identify reporting inconsistencies; they do not substitute for investigating the reasons for missingness or assay differences.

Common Mistakes and Practical Limits

Using only the enrolled sample size as every denominator. This can make percentages wrong when a variable or assay applies to only a subset. State the denominator beside each summary.
Equating modality totals with paired availability. Knowing the number with RNA and the number with protein does not determine how many participants have both. Report the joint count or the relevant combinations.
Calling any detected feature “well covered.” Detection in one participant can coexist with extensive missingness. Give per-feature coverage or a transparent threshold-based summary.
Averaging cohort percentages to describe all observations. An unweighted average is not generally the pooled rate. Use the ratio of summed counts to summed denominators when a common pooled quantity is meaningful.
Treating coverage as evidence of representativeness or comparability. A high coverage percentage does not show that missing data are unrelated to participant characteristics, and overlapping panels do not guarantee equivalent assays. Report the descriptive facts without making those unsupported claims.

Every summary also has limits: a table cannot resolve measurement error, remove differences in eligibility, or make unlike definitions identical. Its purpose is to expose the information readers need to evaluate those issues, not to replace that evaluation.

Check Your Understanding

Answer each question using the stated denominators; concise calculations are sufficient.

A cohort has 80 participants, of whom 64 have usable RNA data and 50 have usable protein data. If 40 have both, how many have RNA only, protein only, or neither, assuming everyone is classified?
For a categorical characteristic, 18 of 60 participants with observed status are in category A, but status is missing for 5 of the 65 cohort members. What percentage should be reported, and what denominator should accompany it?
A cohort has 200 eligible feature measurements, of which 170 are valid. A second cohort has 20 eligible measurements, of which 10 are valid. Find the combined rate and the unweighted mean of the two cohort rates. Why do they differ?
Two assay panels each contain 300 features, with 240 features shared. Calculate their Jaccard index and state what it does not show.
A report states that 98% of intended features were detected at least once. Give one additional feature-coverage summary and one assay detail that would improve interpretation.

Solutions. (1) RNA only is \(64-40=24\); protein only is \(50-40=10\); neither is \(80-(24+10+40)=6\). (2) Report \(18/60=30.0\%\) among participants with observed status, and state that status was observed for 60/65; do not use 65 as the denominator for that percentage without specifying how missing status is handled. (3) The combined rate is \((170+10)/(200+20)=180/220=81.82\%\). The cohort rates are 85.0% and 50.0%, whose unweighted mean is 67.5%; they differ because the cohorts have different denominators. (4) The union has \(300+300-240=360\) features, so \(J=240/360=0.6667\). This describes panel overlap, not assay quality or equivalent measurement scales. (5) Report, for example, the number and proportion meeting a stated sample-coverage threshold, with its denominator; also report the assay platform or kit version, specimen type, measurement units, or applicable detection limits.