Answers are grounded in this paper and cite their sources.
Test-negative studies recruit ‘cases’ who test positive for a particular disease; ‘controls’ are patients undergoing the same tests for the same medical reasons and who test negative. The design is often used for vaccine effectiveness studies but has also, with certain assumptions, been used to assess risk factors for disease. However, although the test-negative design (TND) has been widely used, and also explored theoretically, there has been little agreement on what type of bias the TND is intended to address, or what the corresponding Directed Acyclic Graph (DAG) is. In this paper, we consider the situation where the only available disease outcome is that available from the natural process of people being tested in a particular population (i.e. no additional testing is done for the study), and those who test positive (the cases) are compared to the rest of the general population (or a sample of general population controls), not all of whom have been tested. In this situation, we show that the bias that the TND is intended to address arises from misclassification of the outcome due to incomplete testing. This bias occurs irrespective of whether the entire source population is studied, or a standard case-control study (with general population controls) is conducted. This bias can be eliminated or minimized by restricting the source population to the subpopulation who get tested. Once again, this restriction works whether all of this tested subpopulation is studied, or if controls are sampled from the non-cases. Using either design within the tested subpopulation controls (fully or in part) for determinants of who gets tested, including health-seeking behaviour. We then briefly consider other biases that may occur in a TND study (e.g. imperfect testing, residual confounding). However, we emphasize that these additional biases generally can occur in any observational study, and are not unique to the TND. The unique bias which the TND is intended to address is misclassification of the outcome due to incomplete testing.
No takes yet. Share an insight, caveat, or question.
Pearce et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: