Test-Negative Design

Test-negative design (abbreviated TND) is a modified observational case-control study architecture widely deployed in infectious disease epidemiology and public health surveillance to estimate real-world vaccine effectiveness (VE) against acute respiratory, enteric, and vector-borne pathogens. Formalized for seasonal influenza surveillance by Gaston De Serres, Danuta M. Skowronski, and colleagues in the mid-2000s and analyzed using causal-inference graphs by Michael L. Jackson, Lisa A. Jackson, and Marc Lipsitch (2013–2016), the test-negative design restricts enrollment to individuals presenting to healthcare facilities with a standardized clinical syndrome—such as influenza-like illness (ILI), severe acute respiratory infection (SARI), or acute gastroenteritis—who undergo laboratory diagnostic testing for the pathogen of interest. Participants whose specimens test positive for the target pathogen are classified as cases (D = 1), whereas participants from the same clinical care setting whose specimens test negative are classified as controls (D = 0). By sampling both cases and controls from the subpopulation of patients who actively seek medical evaluation for identical symptoms, the test-negative design cancels unmeasured confounding driven by differential healthcare-seeking propensity and differential access to clinical testing.

Overcoming Healthcare-Seeking Bias in Post-Licensure Surveillance

Once a prophylactic vaccine receives regulatory licensure and enters population-wide rollout, ethical standards and operational realities generally preclude placebos in randomized controlled trials. Public health agencies must therefore rely on observational biostatistics to monitor post-licensure vaccine effectiveness across evolving viral variants, multi-year waning immunity trajectories, and clinically vulnerable subpopulations. In traditional population-based cohort studies or community case-control studies, observational estimates of vaccine effectiveness are vulnerable to healthy user bias and confounding by indication. Specifically, an individual's unobserved propensity to seek medical care (H) strongly influences both the probability of receiving preventive immunization (V = 1) and the probability of presenting to a clinic and receiving a laboratory diagnosis when mild or moderate respiratory symptoms occur (T = 1).

In a conventional community case-control design where cases are laboratory-confirmed patients at a clinic and controls are randomly sampled from the general community Registry, conscientious healthcare seekers are overrepresented among diagnosed cases because untreated community residents with mild infections never undergo diagnostic swabbing. This differential ascertainment biases the exposure odds ratio upward, underestimating true vaccine protection. Conversely, in hospitalized cohorts where frail patients with chronic cardiopulmonary comorbidities are preferentially vaccinated, unadjusted comparisons can make an effective vaccine appear ineffective or harmful.

Emergence from Seasonal Influenza Sentinel Networks

To eliminate the logistical cost of community control sampling while neutralizing healthcare-seeking variation, sentinel physician networks in Canada, Europe, the United States, and Australia adopted the test-negative sampling frame for annual influenza vaccine evaluation. Clinicians systematically collect nasopharyngeal swabs from all consenting patients presenting within a specified window (typically 7 days or fewer from symptom onset) with fever and cough or sore throat. All specimens are assayed via reverse-transcription quantitative polymerase chain reaction (RT-qPCR) before vaccination history is linked. Because neither the patient nor the ordering clinician knows at the moment of care-seeking whether the respiratory syndrome is caused by influenza virus (I₁) or by a non-influenza respiratory pathogen (I₂, such as rhinovirus, respiratory syncytial virus, parainfluenza, or seasonal human coronaviruses), both groups share the same underlying healthcare-seeking behavior.

Odds Ratio Estimation of Vaccine Effectiveness

Let a source population experience two mutually exclusive classes of acute syndromic illness over a defined surveillance period: infection with the vaccine-targeted pathogen (I₁) and infection with non-target pathogens causing clinically indistinguishable symptoms (I₂). In a 2 × 2 contingency table of tested patients stratified by vaccination status (V ∈ {0, 1}) and diagnostic test result (D ∈ {0, 1}):

• a = Vaccinated patients testing positive for the target pathogen (V = 1, D = 1)
• b = Vaccinated patients testing negative for the target pathogen (V = 1, D = 0)
• c = Unvaccinated patients testing positive for the target pathogen (V = 0, D = 1)
• d = Unvaccinated patients testing negative for the target pathogen (V = 0, D = 0)

The test-negative exposure odds ratio (OR_TND) and the corresponding vaccine effectiveness (VE) are estimated via logistic regression (adjusting for calendar week, age group, and clinical comorbidity) as:

OR_TND = (a × d) / (b × c), and VE = (1 − OR_TND) × 100%

Core Identification Assumptions and Incidence Density Proportionality

As demonstrated by Jackson and Nelson (2013) and Sullivan, Tchetgen Tchetgen, and Cowling (2016), OR_TND provides an unbiased estimator of the true relative risk of medically attended target-pathogen illness (RR_MA) without requiring the rare-disease assumption, provided four structural conditions hold:

1. No Direct Vaccine Effect on Non-Target Pathogens (The Negative Control Condition): Vaccination against the target pathogen (V) neither increases nor decreases the biological incidence of non-target syndromic illnesses (I₂). Under this assumption, test-negative controls function as an embedded negative control outcome that mirrors the exposure distribution of the underlying source population at risk of seeking care.
2. Equal Healthcare-Seeking Propensity Across Etiologies: Among individuals who develop symptoms, the probability of seeking clinical care (S = 1) given healthcare-seeking propensity H is identical for target (I₁) and non-target (I₂) infections once conditioned on clinical severity tier (outpatient clinic vs. inpatient hospitalization).
3. Perfect Diagnostic Test Specificity and High Sensitivity: The laboratory assay must not misclassify true target infections as negative controls or vice versa.
4. Conditioning on Healthcare-Seeking (S = 1) Blocks the V ← H → S Path: When propensity H acts multiplicatively on care-seeking probabilities, the ratio of vaccinated to unvaccinated test-negative controls (b / d) cancels the healthcare-seeking ratio in the test-positive numerator (a / c).

Collider-Stratification Bias from Clinical Selection

Despite its strengths, the test-negative design explicitly conditions on a post-exposure variable: presenting to a clinic and receiving a diagnostic test (T = 1). In the language of Directed Acyclic Graphs, T is a common effect (a collider) of vaccination status (V → T, if vaccinated individuals are more or less inclined to seek testing when mildly ill), true infection status (I₁, I₂ → T), and unmeasured health determinants or high-risk occupations (U → T). Conditioning on T = 1 can induce collider bias along the pathway V → T ← U → I₁ unless the investigator adjusts for baseline clinical severity, occupational testing mandates, and demographic covariates, or restricts the sample to severe syndromes (such as hypoxic hospitalization) where care-seeking probability approaches 100% regardless of H.

Violation of the Viral Interference Independence Assumption

If immunization against pathogen I₁ biologically alters susceptibility to non-target pathogens I₂, the control group's vaccination odds are distorted. For instance, if live-attenuated viral immunization transiently stimulates mucosal innate interferon signaling that cross-protects against non-target respiratory viruses (nonspecific viral interference), vaccinated individuals will be depleted from the test-negative control group (b decreases), biasing OR_TND upward and underestimating true VE. Conversely, if natural infection with influenza temporarily blocks rhinovirus infection through epidemiological interference, unvaccinated individuals (who experience more influenza) will experience fewer rhinovirus control illnesses, inflating VE.

Diagnostic Sensitivity, Specificity, and Exposure Misclassification

Both outcome misclassification (imperfect RT-qPCR or rapid antigen sensitivity Se and specificity Sp) and exposure misclassification (inaccurate self-reported vaccination records) attenuate test-negative estimates toward the null (VE = 0%). As shown by Jackson and Rothman (2015), imperfect diagnostic specificity (Sp < 1.0) is far more damaging in a TND than imperfect sensitivity (Se < 1.0), particularly at the tails of an epidemic wave when true pathogen prevalence among symptomatic patients is low and false-positive cases contaminate the case group. Similarly, delayed swabbing (> 5–7 days post-symptom onset, when upper-respiratory viral shedding declines) causes true cases to test negative and migrate directly into the control denominator (b and d), diluting the contrast between groups. Epidemiologists apply quantitative bias analysis and correct for imperfect gold standard bias to bound these distortions.

Beyond seasonal influenza and SARS-CoV-2 mRNA booster surveillance, the test-negative architecture has been adapted to evaluate oral rotavirus vaccines among infants presenting with acute dehydrating diarrhea, cholera vaccines during enteric outbreaks, and cluster-randomized trials of Wolbachia-introgressing Aedes aegypti mosquitoes for dengue fever control (Anders et al., 2018), where febrile patients testing positive for dengue virus by NS1 antigen or RT-qPCR are compared against febrile patients testing negative across treated and untreated neighborhood clusters.

Why must test-negative studies adjust for calendar time?

In almost all vaccine rollouts and seasonal epidemics, both population vaccine coverage (V) and the ratio of target pathogen circulation (I₁) to non-target pathogen circulation (I₂) vary substantially across calendar weeks. Consequently, calendar time functions as a strong confounder (V ← Time → D). Failing to stratify or adjust for calendar week or two-week intervals via conditional logistic regression or cubic splines will confound the odds ratio even if vaccine effectiveness is constant.

Does the test-negative design measure protection against asymptomatic infection?

No. Because enrollment into a standard test-negative study requires presenting with a defined clinical syndrome (such as fever and cough), the design estimates vaccine effectiveness against medically attended symptomatic illness, which is a composite of vaccine protection against acquiring infection and protection against progressing from infection to clinical care-seeking. Estimating protection against asymptomatic infection requires a prospective cohort with scheduled periodic swabbing or a nested case-control study.

1. Jackson, M. L., & Nelson, J. C. (2013). The test-negative design for estimating influenza vaccine effectiveness. Vaccine, 31(17), 2165–2168. https://doi.org/10.1016/j.vaccine.2013.02.053
2. Sullivan, S. G., Tchetgen Tchetgen, E. J., & Cowling, B. J. (2016). Theoretical basis of the test-negative study design for assessment of influenza vaccine effectiveness. American Journal of Epidemiology, 184(5), 345–353. https://doi.org/10.1093/aje/kww064
3. Fukushima, W., & Hirota, Y. (2017). Basic principles of test-negative design in evaluating influenza vaccine effectiveness. Vaccine, 35(36), 4796–4800. https://doi.org/10.1016/j.vaccine.2017.07.003
4. Lewnard, J. A., Patel, M. M., Jewell, N. P., Verani, J. R., Kobayashi, M., Tenforde, M. W., & Dean, N. E. (2021). Theoretical framework for retrospective studies of the effectiveness of SARS-CoV-2 vaccines. Epidemiology, 32(4), 508–517. https://doi.org/10.1097/EDE.0000000000001366