Negative Controls

A negative control is a repetition of an experiment or analysis under conditions where the hypothesized mechanism cannot operate but the suspected biases still can. If the procedure returns a positive result anyway, the result is an artifact of the method, not evidence for the mechanism. Borrowed from bench biology (leave out an essential ingredient; inactivate the active ingredient; test an impossible target), the idea was formalized for observational epidemiology by Lipsitch, Tchetgen Tchetgen & Cohen (Epidemiology 2010), who distinguished negative-control exposures (a variable that shares the exposure’s confounding structure but has no causal path to the outcome) from negative-control outcomes (an outcome the exposure cannot cause but that shares its confounders). ¹

Why observational studies need them

Non-causal associations arise from mismeasurement, confounding, and selection bias. Adjustment methods (stratification, matching, regression, weighting) handle measured confounders; negative controls are one of the few tools that can detect unmeasured confounding and analytic error. The canonical examples from Lipsitch et al.: influenza-vaccine mortality benefits in the elderly that shrink when a pre-season “control period” is used (the vaccine can’t reduce deaths before flu circulates), and the phantom MS–infectious-mononucleosis association that collapsed under a recall-bias control. The selection principle: a good negative control reproduces a condition that cannot involve the hypothesized mechanism but is very likely to involve the same sources of bias. Choosing one requires subject-matter knowledge, just like choosing adjustment variables. ²

Case study: the Harvard nuclear-proximity series

A 2025–2026 series of papers by Alwadi and colleagues at the Harvard T.H. Chan School of Public Health reported that residential proximity to US nuclear power plants is associated with elevated cancer incidence and mortality, with risk declining by distance. The Massachusetts study (Environmental Health, Dec 2025) built an inverse-distance-weighted “proximity score” summing over all plants within 120 km of each ZIP code’s population center, regressed 2000–2018 cancer-registry incidence on that score with covariates, and attributed ~10,800 female and ~9,800 male cancer cases (3.5–4.1% attributable fraction) to plant proximity; national follow-ups extended the design to mortality at the county level with a 200 km radius. ³

Deric Tilson and Adam Stein (Breakthrough Journal, Oct 2026) ran the obvious negative control: they re-implemented the same proximity-score methodology and pointed it at landmarks with no plausible carcinogenic emission pathway. The method “attributed” over 2.2 million cancer deaths (2000–2018) to Costco warehouses — more than 20% of all US cancer deaths in the period — 1.59 million to private four-year colleges, 895,258 to Superfund sites, 944,769 to MLB fields versus 50,889 to MLS pitches, and 111,003 to Harvard itself. Every landmark tested produced a positive result; no landmark produced a negative one. Their diagnosis: the score is confounded by whatever clusters near populous or affluent areas (demographic and healthcare-access gradients), and because it is cumulative over everything within a 120–200 km radius, it functions as a smooth spatial-density proxy that can be story-matched to any landmark chosen in advance. The papers’ authors had never placebo-tested the method against other landmarks. ⁴

The episode also illustrates the dose side of the argument: monitored US nuclear workers — maximally exposed by construction — typically receive zero to ~2% of the occupational annual dose limit, so a mechanism delivering measurable dose to residents 50–200 km away would contradict the dosimetry record. A negative control and a physical implausibility check are different tools, but here they converge: the association is produced by the methodology, not by radiation. ⁵

Failure modes negative controls catch

• Unmeasured spatial/demographic confounding — the Alwadi case: proximity scores inherit every gradient that covaries with geography.
• Recall bias — the MS–mononucleosis case.
• Selection and immortal-time bias — nasal-corticosteroid and perioperative-medication examples in Lipsitch et al.
• Analytic bugs and pipeline artifacts — a control that should return null but doesn’t localizes the error to the pipeline rather than the world.

Limitations

A negative control that returns null does not prove the original result is causal — it only removes the specific bias the control was built to share. Controls are also only as good as their bias-matching: a control exposed to a different confounding structure than the original association detects little. Lipsitch et al. define precise conditions under which a non-null control implies the existence and direction of uncontrolled-confounding bias, and note that controls complement, rather than replace, design-stage rigor. ⁶

Cross-domain connections

• adversarial-collaboration — Tilson & Stein’s replication is the informal version: hostile re-implementation with the original authors’ partial cooperation (eight lines of code that did not reproduce the results). A formal adversarial collaboration would have required the placebo tests before publication.
• cold-fusion — the same epistemology at the bench: Langmuir’s pathological-science symptoms include results that never converge and effects at the edge of detectability; the proximity-score series shows the statistical-science analog — a method whose output is always positive and never converges toward zero for impossible exposures.
• llm-replication-economics — both are audit stories: one audits economics results with LLM re-implementation, the other audits an epidemiology method with placebo landmarks; the negative control is the general technique underlying both.
• assortative-mating — the marriage-market IV literature solves the same confounding problem from the other direction: instead of placebo tests, Goldman–Gracie–Porter exploit Bernoulli sex-ratio variation whose randomness is directly checkable against theory. Negative controls and credible instruments are complementary answers to unmeasured confounding.
• schelling-commitment-and-focal-points — pre-registered placebo tests are a commitment device: they burn the researcher’s option to story-match positive results after the fact.

Sources

• Marc Lipsitch, Eric Tchetgen Tchetgen, and Ted Cohen 2010 — Negative Controls: A Tool for Detecting Confounding and Bias in Observational Studies — the canonical methods paper (Epidemiology 2010)
• Deric Tilson and Adam Stein 2026 — Things that Apparently Cause Cancer — the Costco/landmark placebo replication (Breakthrough Journal, Oct 2026)
• Yazan Alwadi, John S. Evans, Joel Schwartz, Carolina L. Zilli Vieira, David C. Christiani, Brent A. Coull, and Petros Koutrakis 2025 — Residential proximity to nuclear power plants and cancer incidence in Massachusetts, USA (2000–2018) — the criticized proximity-score study (Environmental Health, Dec 2025)

Footnotes

1. Marc Lipsitch, Eric Tchetgen Tchetgen, and Ted Cohen 2010 — Negative Controls: A Tool for Detecting Confounding and Bias in Observational Studies ↩

2. Marc Lipsitch, Eric Tchetgen Tchetgen, and Ted Cohen 2010 — Negative Controls: A Tool for Detecting Confounding and Bias in Observational Studies ↩

3. Yazan Alwadi, John S. Evans, Joel Schwartz, Carolina L. Zilli Vieira, David C. Christiani, Brent A. Coull, and Petros Koutrakis 2025 — Residential proximity to nuclear power plants and cancer incidence in Massachusetts, USA (2000–2018) ↩

4. Deric Tilson and Adam Stein 2026 — Things that Apparently Cause Cancer ↩

5. Deric Tilson and Adam Stein 2026 — Things that Apparently Cause Cancer ↩

6. Marc Lipsitch, Eric Tchetgen Tchetgen, and Ted Cohen 2010 — Negative Controls: A Tool for Detecting Confounding and Bias in Observational Studies ↩