Written and maintained by CASRAI Editorial Board
Last updated
A test-negative design does not compare vaccinated people to unvaccinated people drawn from the community — it compares vaccination status between two groups who already did the same thing: showed up to care with the same symptoms and got tested. Everyone in the study, cases and controls alike, presented with a compatible clinical syndrome (commonly acute respiratory illness) and received the same diagnostic test for the pathogen the vaccine targets. Cases are the people who tested positive; controls are the people who tested negative for that pathogen but still had the qualifying symptoms, usually because a different pathogen caused them. This guide covers why that specific comparison, and not a standard case-control design pulling controls from the general population, is the design most vaccine-effectiveness (VE) studies now use — and where it still leaves bias unaddressed. For the design this one modifies, see case-control study design and case-control vs. cohort study.
The comparison the design actually makes
Enrollment starts with a case definition built on symptoms, not etiology: a patient presenting to a clinic, urgent care center, or hospital with an illness consistent with the target pathogen — influenza-like illness, COVID-19-compatible respiratory illness, acute gastroenteritis for a rotavirus study, and so on. Every enrolled patient is tested with the same assay for the same pathogen, regardless of vaccination status. That single test result then sorts patients into the two arms:
- Cases — symptomatic, tested, and positive for the target pathogen.
- Controls — symptomatic, tested, and negative for the target pathogen (the illness had some other cause: a different respiratory virus, a different enteric pathogen, or in some studies a non-infectious mimic).
Vaccination status is then compared between the two arms exactly as in a standard case-control study — as exposure history recorded before illness onset, usually from an immunization registry or verified vaccination record rather than patient recall. What differs from a standard case-control design is entirely in where the controls come from: not the general population, not a random sample of clinic patients with any complaint, but specifically the set of people who had the same presenting syndrome, sought the same care, and underwent the same test — and happened to test negative.
Why this controls healthcare-seeking-behavior confounding
Vaccination is not randomly distributed across a population, and neither is the propensity to seek medical care. People who keep up with routine vaccination tend, on average, to differ systematically from people who don’t: better healthcare access, more contact with the health system, different baseline health-consciousness. If a standard case-control study recruits controls from the general community, or from clinic patients being seen for unrelated reasons, that control group’s vaccination rate reflects the general population’s mix of high- and low-healthcare-engagement people — not the mix among people who would show up sick and get tested. Because engagement with the health system independently predicts both vaccination uptake and the likelihood of appearing in the study at all (as a case, by getting tested when sick), a mismatch between how cases and controls are recruited turns healthcare-seeking behavior into a confounder: it distorts the exposure (vaccination) prevalence in the control group relative to what the sick, tested population would actually show.
The test-negative design closes that gap by construction rather than by statistical adjustment. Cases and controls are both drawn exclusively from people who already cleared the same bar — noticing symptoms, deciding to seek care, and being tested — before etiology is even known. A control’s vaccination status is measured within the same care-seeking, testing-access population the cases came from, not the general population. Differences in health-system engagement, insurance access, and general propensity to seek care are, by design, distributed the same way across cases and controls, because both groups are conditioned on the same care-seeking-and-testing event. That is the specific bias a standard case-control design with population- or convenience-based controls does not control for, and the reason the test-negative design has become the default for observational influenza and COVID-19 vaccine-effectiveness studies rather than a niche variant.
This is a design-level fix, not a claim that healthcare-seeking behavior no longer matters at all — it matters less because it no longer differs systematically between the groups being compared, the same logic behind matching or restriction in other observational designs. See confounding variable for the general concept this is one specific, structural solution to.
Calculating vaccine effectiveness from the 2×2 table
Once cases and controls are sorted, the design produces the same 2×2 table structure as any case-control study: vaccinated vs. unvaccinated, cross-tabulated with case (test-positive) vs. control (test-negative) status. The odds ratio (OR) of prior vaccination, comparing cases to controls, converts directly to an effectiveness estimate:
VE = (1 − OR) × 100%
An OR below 1 means vaccinated people are under-represented among test-positive cases relative to test-negative controls — consistent with the vaccine reducing infection risk — and produces a positive VE. An OR at or above 1 produces a VE at or below zero: no protective effect detected, or evidence pointing the other way. In practice this OR is estimated with logistic regression so that age, calendar time (to account for the mix of circulating pathogens shifting week to week), prior-season vaccination, and other measured confounders can be adjusted for directly — the test-negative sampling scheme controls healthcare-seeking-behavior confounding by design, but it does not substitute for adjusting the other, conventional confounders any case-control analysis has to handle. See how to interpret an odds ratio and confidence interval for an odds ratio for the underlying mechanics this VE formula builds on.
A reproducible simulation: how much the confounding actually matters
The figures in this section are from a simulated dataset built to illustrate the mechanism above — not real surveillance data from any real vaccine or pathogen. The simulation is fully specified and reproducible: a seeded pseudorandom generator (mulberry32, seed 20260826) over 2,000,000 simulated individuals, with a true built-in vaccine effectiveness of 60% against a target pathogen, a second unrelated pathogen causing an identical symptom profile, and two subpopulations — “high” and “low” healthcare-seeking-behavior groups — where vaccination uptake (70% vs. 30%) and the probability of seeking care once symptomatic (90% vs. 50%) are both higher in the high-engagement group. Every number below is a direct read from that simulation’s output, not a hand-picked or adjusted figure.
Holding the same simulated case group (people who sought care and tested positive for the target pathogen) constant, two different control groups were compared:
| Design | Cases: vaccinated / unvaccinated | Controls: vaccinated / unvaccinated | OR | Estimated VE (95% CI) |
|---|---|---|---|---|
| Test-negative (controls = test-negative, same care-seeking population) | 31,622 / 62,017 | 113,130 / 83,474 | 0.376 | 62.4% (61.8–63.0%) |
| Standard case-control (controls = random population sample, not conditioned on care-seeking) | 31,622 / 62,017 | 101,589 / 95,015 | 0.477 | 52.3% (51.5–53.1%) |
The true effectiveness built into the simulation is 60%. The test-negative estimate (62.4%) lands close to that value, with the small remaining gap attributable to ordinary sampling variation. The standard case-control estimate, using the identical case group but population-based controls, is biased downward by roughly 8 percentage points — because the control group’s vaccination rate (proportionally more low-engagement, lower-vaccination-rate people, since it wasn’t conditioned on having sought care) understates how common vaccination actually is among the population that would show up sick and get tested. That gap is entirely attributable to healthcare-seeking-behavior confounding as modeled here; the simulation holds every other factor equal specifically to isolate that one mechanism.
What the test-negative design does not fix
Solving healthcare-seeking-behavior confounding does not make the design bias-free. Four issues recur in the methodological literature on test-negative VE studies and need to be handled on top of the core design, not assumed away by it:
- Differential test sensitivity by vaccination status. If vaccinated people who do get infected tend to have lower viral loads or shorter shedding windows, a test with imperfect sensitivity can miss a higher proportion of true vaccinated cases than unvaccinated ones — misclassifying them as test-negative controls and biasing the VE estimate upward. This is a real, specifically-discussed threat in the literature, not a hypothetical.
- Non-specific vaccine effects on the control-defining pathogens. The design assumes the vaccine has no effect, positive or negative, on the “other” pathogens that put people in the control group. If the vaccine happens to reduce (or increase) susceptibility to those other pathogens too — a documented possibility for some vaccines — the control group’s composition shifts in a way unrelated to the target pathogen, and the VE estimate is biased.
- Confounding by indication and frailty. Vaccination is not randomly assigned even within the care-seeking population — clinicians may preferentially vaccinate (or fail to vaccinate) people who are frailer or have more comorbidities, and those same factors can independently affect susceptibility to either pathogen. The test-negative design controls for healthcare-seeking behavior specifically; it does not control for this on its own, which is why VE analyses still adjust for age, comorbidity, and calendar time in the regression model.
- Depth-of-testing differences. If clinicians test more thoroughly, or with a different threshold, for vaccinated vs. unvaccinated patients — for instance, testing an unvaccinated patient more readily “just in case” — that reintroduces a selection difference between the groups the design is otherwise built to avoid.
None of these make the design worse than a standard case-control alternative — they are additional threats layered on top of a design that has already removed one large, structural source of bias. They are the reason VE studies using this design report adjusted odds ratios with explicit covariate lists, rather than treating the crude test-positive-vs-test-negative comparison as sufficient on its own.
Where the design is used
The test-negative design is now the standard observational method behind most published seasonal influenza vaccine-effectiveness estimates and has been used extensively for COVID-19 vaccine-effectiveness surveillance, typically run through networks of participating clinics and hospitals that enroll patients meeting a standard respiratory-illness case definition and test them with a validated assay. Its main practical advantage over a standard case-control or cohort design, beyond the confounding control described above, is operational: it can be run entirely on patients who were already going to be tested for clinical reasons, without needing to separately recruit or track an unvaccinated comparison cohort.
Frequently asked questions
Is the test-negative design the same as a standard case-control study?
It is a specific variant of case-control design, not a separate design type. The case-control logic — comparing exposure (vaccination) odds between an outcome-positive and an outcome-negative group — is identical. What’s specific to the test-negative design is how controls are selected: from among people who also sought care and were tested, but tested negative, rather than from the general population.
Does the test-negative design require a randomized trial to validate it?
No — it’s an observational design used specifically because a randomized trial (withholding vaccination from part of a population) is usually neither ethical nor practical once a vaccine is in routine use. Its value is that it approximates some of a trial’s internal validity for the specific confounder (healthcare-seeking behavior) that would otherwise be hardest to control observationally.
Can the test-negative design be used for anything besides vaccines?
The same case/test-negative-control logic has been applied outside vaccine-effectiveness research — for example, to estimate the effectiveness of other preventive interventions where the exposure and the diagnostic test can be similarly structured — but vaccine effectiveness against acute infections is by far its dominant use, and the confounding-control argument above is specific to that context (healthcare-seeking behavior driving both the exposure and study entry).
Why not just adjust for healthcare-seeking behavior statistically instead of changing the control group?
Healthcare-seeking behavior is difficult to measure directly and completely — there is no single clean variable that captures “propensity to seek care and get tested.” Structural solutions (built into who becomes eligible for the study) are generally more robust than adjustment for a confounder that’s hard to fully specify and measure. The test-negative design’s advantage is removing this specific confounder from the comparison by construction, rather than depending on having measured it well enough to adjust for it.
How is this different from a nested case-control study?
They solve different problems. Nested case-control design is about efficiently sampling controls from within an existing cohort’s risk sets to avoid measuring an expensive exposure on everyone. The test-negative design is about which population controls are drawn from at all — people who sought care and tested negative — specifically to remove healthcare-seeking-behavior confounding in vaccine-effectiveness estimation. A study could, in principle, combine elements of both, but they answer different design questions.
