Cross-Sectional Analogue Designs: Population-Analogue, Cohort-Analogue & Case-Control-Analogue Sampling for Rare Exposure vs Rare Outcome
On this page
Reading the simulator in this article's language: gate A · Single-Gate is the population-analogue design; gate B · Two-Gate is the case-control-analogue design (sampled by outcome); gate C · Test-Based is the cohort-analogue design (sampled by exposure — here the "exposure" is the index-test result). The older "case–cohort" wording on that third gate is clarified near the end.
Cross-Sectional Analogue Designs: How You Draw the Sample Decides the Design
A diagnostic-accuracy study is cross-sectional by nature: the candidate predictor (the index test — our X) and the outcome (disease status on the reference standard — our Y) are captured in the same snapshot, on the same patients, at the same time.
But that single label — "cross-sectional" — hides three very different recruitment strategies. To guide how you actually collect the data, it helps to categorise a cross-sectional study by how the sample is drawn. Drawn one way it behaves like a population survey; drawn another, like a cohort; drawn a third, like a case–control study. That is why we call them analogue designs: each cross-sectional sampling scheme is an analogue of another primary study design.
The interactive above shows the mechanism — where you place the sampling gate decides which cells of the 2×2 table fill at nature's ratio and which are fixed by design.
Start here: the only question that picks the design is — what is rare?
Before arguing about metrics, ask two blunt questions about your setting:
- Is the predictor / exposure (X) rare? Would consecutive recruitment leave you with too few exposed patients — too few index-test-positives, or too few carriers of the risk factor?
- Is the outcome / disease (Y) rare? Would consecutive recruitment leave you with too few diseased cases?
Those two answers — and nothing more — pick the analogue design. This is the piece that is easy to lose: the design is chosen by rarity, not by taste.
| Analogue design | How the sample is drawn | Appropriate when… | Rarity condition |
|---|---|---|---|
| Population-analogue design consecutive or random recruitment |
Recruit everyone eligible — consecutively, or a random sample of them | Both the predictor and the outcome accrue naturally in adequate numbers | X not rare AND Y not rare |
| Cohort-analogue design sampled by exposure |
Draw the sample on the exposure — enrich on the predictor X | The predictor / exposure would otherwise be too scarce to study | X is rare |
| Case-control-analogue design sampled by outcome |
Draw the sample on the outcome — enrich on the disease Y | The disease / outcome would otherwise be too scarce to study | Y is rare |
The rarity map (the piece most versions leave out). Read it as: choose the analogue whose sampling axis fixes whatever is rare. If both X and Y are rare, enrich on both (a two-way enriched sample) and read every prevalence-dependent number with extra caution.
Translating to diagnostic-accuracy research: the exposure / predictor (X) is the index-test result (or any candidate predictor), and the outcome (Y) is disease status on the reference standard. So a rarely-positive index test is a "rare X" problem (reach for the cohort-analogue), while a rare disease is a "rare Y" problem (reach for the case-control-analogue).
1. Population-analogue design (consecutive or random recruitment)
- How you sample: take everyone who meets the eligibility criteria — consecutively, or as a random sample of them — and give all of them both the index test and the reference standard.
- When it is appropriate: when X is not rare and Y is not rare. Natural accrual already yields enough exposed patients and enough diseased patients.
- What you get: the observed 2×2 fills at the population's true ratio, so every quantity is directly interpretable — sensitivity, specificity, PPV, NPV and the observed prevalence all estimate the target population.
- The catch: in a low-prevalence setting this design quietly becomes an imbalanced sample (too few cases) with unstable estimates — which is precisely the signal to switch designs.
Example (illustrative): a walk-in fever clinic enrols every consecutive adult with undifferentiated fever and runs both a rapid multiplex antigen panel (index test) and blood-culture / PCR (reference standard) on all of them. Fever is common and a positive panel is common, so the natural sample is informative as it stands — a population-analogue design.
2. Cohort-analogue design (sampled by exposure)
- How you sample: draw the sample on the exposure / predictor. Keep (almost) all of the exposed — the index-test-positives, or carriers of the rare risk factor — and take a defined subsample of the unexposed. Both arms then receive the reference standard.
- When it is appropriate: when X is rare — the predictor is so uncommon that consecutive recruitment would leave almost no exposed patients to learn from.
- Why "cohort": a classic cohort study is defined by exposure and looks forward to the outcome. Sampling a cross-section on the exposure is the cross-sectional analogue of that logic — so the forward, exposure-anchored quantities (the predictive values, PPV/NPV — the column metrics) stay interpretable, while sensitivity and specificity are now distorted by the enriched design and must be read with care.
Example (illustrative): a wearable ECG patch flags "possible atrial fibrillation" in only about 4% of wearers. To study how well that flag predicts confirmed AF, you keep every flagged (exposed) wearer plus a random 1-in-20 sample of the un-flagged, and send both groups for a cardiologist-adjudicated reference read. You sampled on the exposure because the exposure was rare — a cohort-analogue design.
3. Case-control-analogue design (sampled by outcome)
- How you sample: draw the sample on the outcome. Deliberately assemble a group with the disease (cases) and a comparison group without it (controls) from the same source population, then measure the index test in both — all in one cross-sectional snapshot, not a follow-up.
- When it is appropriate: when Y is rare — the disease is so uncommon that consecutive recruitment would yield too few cases for a stable estimate.
- Why "case–control": a case–control study is defined by outcome and looks back at exposure. Sampling a cross-section on the disease is its analogue — so the outcome-anchored discrimination (sensitivity and specificity, the row metrics) is preserved, while PPV, NPV and the observed prevalence are now fixed by your case:control ratio and must not be reported as clinical values.
- Watch for spectrum bias: hand-picking florid cases against super-healthy controls (the "two-gate" trap) inflates apparent sensitivity and specificity. Draw both arms from the same clinical stream.
Example (illustrative): a new serum biomarker for pancreatic cancer — a disease far too rare in a general clinic for consecutive sampling to capture. You assemble a group with histologically confirmed pancreatic cancer and a matched comparison group without it, and measure the biomarker in both. You sampled on the outcome because the outcome was rare — a case-control-analogue design.
A note on naming: it is the "cohort-analogue", not the "case–cohort"
Earlier phrasings of this material called the exposure-sampled design a "case–cohort analogue." That wording is misleading and is corrected here. The design that samples by exposure to handle a rare X is the cohort-analogue design — because it mirrors a cohort study, which is defined by exposure. Reserve case-control-analogue for the design that samples by outcome to handle a rare Y. Two axes, two names:
- Sample by exposure (rare X) → cohort-analogue design.
- Sample by outcome (rare Y) → case-control-analogue design.
(“Case–cohort” is a real and distinct sampling scheme within cohort studies — a random subcohort plus all incident cases — but it is not what the third analogue design refers to, which is why the clearer “cohort-analogue” label is used throughout.)
Putting it together — the 30-second decision
- Neither X nor Y is rare → population-analogue design (consecutive / random recruitment).
- The predictor / exposure X is rare → cohort-analogue design (sample by exposure); trust the predictive values, treat Se/Sp cautiously.
- The disease / outcome Y is rare → case-control-analogue design (sample by outcome); trust Se/Sp, do not report design-driven PPV/NPV or prevalence.
The deeper point: the analogue design does not change the test — it changes which numbers survive the sampling. Choose the analogue that preserves the numbers you most need to trust.
Appendix: when the recruited 2×2 table itself is imbalanced
Rarity picks the analogue; a finer, second question is whether the recruited table is balanced on the index test (columns) and on the reference standard (rows), and at what prevalence — because that governs which metrics stay stable. The map below is a quick reference for the eight combinations.
- Balanced Index – Balanced Reference – Low Prevalence — Problem: sensitivity unstable. Fix: enrich diseased cases.
- Balanced Index – Balanced Reference – High Prevalence — Problem: none (the ideal). Fix: report all metrics.
- Balanced Index – Imbalanced Reference – Low Prevalence — Problem: PPV low, NPV inflated. Fix: case-enrichment.
- Balanced Index – Imbalanced Reference – High Prevalence — Problem: specificity unstable. Fix: add non-diseased.
- Imbalanced Index – Balanced Reference – Low Prevalence — Problem: accuracy misleading, sensitivity poor. Fix: ROC / likelihood ratios.
- Imbalanced Index – Balanced Reference – High Prevalence — Problem: specificity poor. Fix: emphasise AUROC.
- Imbalanced Index – Imbalanced Reference – Low Prevalence — Problem: double bias, apparent accuracy misleading. Fix: enrichment + robust metrics.
- Imbalanced Index – Imbalanced Reference – High Prevalence — Problem: accuracy unreliable (specificity collapse). Fix: enrichment + AUROC / robust metrics.
Related reading
- Balanced vs. imbalanced diagnostic accuracy — the companion piece on how index / reference imbalance distorts each metric.
- Cohort design in etiologic research — the parent design the cohort-analogue mirrors.
- Case–control design in etiologic research — the parent design the case-control-analogue mirrors.
- Diagnostic accuracy basics — sensitivity, specificity and the predictive values referenced above.