A group of researchers at Harvard's T.H. Chan School of Public Health has published a series of papers claiming that living near nuclear power plants is associated with higher rates of cancer. The earliest of the series, led by Yazan Alwadi, appeared in December 2025 in Environmental Health and examined cancer incidence in Massachusetts. An expanded nationwide version followed in March, in Nature Communications, this time looking at cancer mortality rather than incidence. Additional papers appeared in the Journal of Exposure Science & Environmental Epidemiology, covering lung, breast, and colon cancers, and most recently in the European journal Environmental Epidemiology, which looked at total mortality instead of cancer alone.
The method behind all of them is straightforward to describe. The authors build a "proximity score" for every ZIP code or county based on distance from nuclear plants — within 120 kilometers in the Massachusetts study and 200 kilometers in the national ones, roughly 75 and 125 miles. Every location inside those circles counts as exposed, with closer places weighted more heavily, and exposure is cumulative where circles overlap. They then run regressions of cancer outcomes on that proximity score plus a set of control variables, convert the resulting coefficients into a Relative Risk for each age and sex group, compute an Attributable Fraction, and finally estimate how many deaths from 2000 to 2018 can be attributed to the plants. The authors state that their methodology and results "provide the scientific basis and public health justification for an expanded research program."
Deric Tilson and Adam Stein, who had previously called the earlier studies nonsense, spent months replicating the analysis. The authors supplied eight lines of code and answered a few questions about their control variables — which, they say, did not reproduce the results. Most of the replication was done from first principles combined with trial and error. Once they matched the first national cancer-mortality study closely enough, they applied the same methodology to other landmarks.
The results are difficult to take seriously on their face. Applied to Costco warehouses, the method attributes more than 2.2 million cancer deaths — over 20 percent of all cancer deaths in the country between 2000 and 2018. Private four-year colleges, including Harvard, account for 1.59 million; public colleges for 626,212; Superfund sites for 895,258; and state capitals for 132,389. Combined, those landmarks alone explain more than half of every cancer death in the period.
The replication applied the Harvard team's method to ordinary landmarks and found it attributed enormous cancer deaths to Costco warehouses, private and public colleges, Superfund sites, and state capitals.
Professional sports venues produce equally strange numbers. Living near an NFL stadium is associated with 418,520 deaths, near an NBA arena with 471,794, near a Major League Baseball field with 944,769, near an NHL rink with 126,751, and near an MLS pitch with 50,889. Baseball fields alone therefore cause nearly 19 times as many attributed deaths as soccer pitches. NASCAR tracks, which one might expect to carry more emissions and smoke, account for only 79,278. With sports included, the method attributes over 7.85 million deaths — three-quarters of the total. And Harvard, examined on its own, accounts for 111,003 deaths, compared with 115,586 for nuclear power plants.
The core problem is that proximity is not exposure. The authors' own coauthors have described, in a webinar presentation, a mechanism in which radioactive dust escapes during refueling, passing through containment and shielding undetected by the facility's sensors. But measured exposure tells a different story. Most closely monitored U.S. nuclear workers receive no measurable annual dose at all, and when workers are exposed, the average dose is only 2 percent of the occupational limit. If the people actually on-site receive, at most, one-fiftieth of the legal occupational limit, there is no plausible pathway by which residents 5, 10, 25, 50, 120, or 200 kilometers away would receive a measurable dose from the same plant.
The authors acknowledge the limitation themselves, at least in part: in talks, they describe their papers as merely ecological studies that show association and never prove causality. That is a legitimate category of research, but it raises a question the papers do not answer — what the mechanism of exposure would be. The replication suggests the answer may be that there is none, and that the method is capturing something else entirely.
That something else appears to be the methodology itself. Tilson and Stein report that no matter what landmark they tested, they never obtained a negative result. Even when they fed the method randomized outcomes rather than real cancer data, it still produced positive results. The variation, they conclude, is random noise sloping consistently in one direction. Most distance-based exposure studies look at much smaller areas — under 10 kilometers, about 113 square miles — whereas the Harvard studies claim exposure across circles of 48,519 square miles, roughly the size of Mississippi. The method, in their account, is blind to radiation exposure and dose; it understands only coordinates, proximity, covariates, and death counts, and from those it can produce a number and a positive result without being able to explain why.
The practical consequence is that the papers do not demonstrate a novel mechanism by which nuclear plants contribute to cancer. They demonstrate an association of data points across enormous circles. Sophisticated statistics, the authors argue, can produce extremely precise nonsense when the study design fails to identify a causal mechanism — and that nonsense is especially dangerous when applied to something culturally plausible that yields a positive result. The attributable number of deaths from this methodology is probably zero, they conclude, while the attributable number of bad papers is at least four.
What remains uncertain is whether the original authors' failure to test their method against placebo landmarks was an oversight or a deliberate choice, and whether peer reviewers at the journals examined the sensitivity of the results to the choice of landmark. Tilson and Stein's replication, while thorough, was partly trial and error, and they note that the authors' supplied code was insufficient to reproduce the published results. Whether the same critique would apply to other proximity-based studies in the literature is an open question, since this replication covers only one family of methods.
For readers who want the bottom line: the studies do not establish that nuclear power plants cause cancer. They establish that a particular statistical approach, applied to cancer mortality data, produces alarming associations for nuclear plants — and for Costco, baseball fields, and Harvard. Absent an exposure pathway and a method that can distinguish signal from noise, those associations carry no evidentiary weight.