Big Stanford analysis finds fewer long COVID effects clearly tied to infection

What happened

Stanford researchers say the long list of health problems blamed on long COVID is too long. Their analysis of 14.4 billion health insurance records from more than 244 million Americans suggests that conventional methods may substantially overestimate how many effects, and how durable, are truly attributable to the virus, as first reported by Medical Xpress.

The results are said to be published in PNAS. We could not locate that paper, so the numbers below come from the Medical Xpress report and the authors' medRxiv preprint, posted in February 2026. Figures may differ in the final version.

A note on framing: the Medical Xpress item opens with a familiar symptom list (brain fog, dizziness and the like). That is background, not a result of this study.

What the analysis found

The problem the authors target is selective testing. People with other conditions that cause lasting symptoms were more likely to seek care and get tested for COVID, so their symptoms can end up blamed on the infection. The preprint says this bias drives most of the long-term effects attributed to COVID in conventional comparisons.

Their fix was to use only people who took a PCR test, then compare similar patients whose first test was positive with those whose first test was negative. They tracked 614 clinical outcomes over two years. The authors say they first reproduced earlier high-profile studies exactly, then ran negative-control checks on outcomes COVID shouldn't cause.

The preprint reports "an order of magnitude smaller" set of long-term effects, still clinically significant, persisting for up to one year. Co-senior author Jure Leskovec is quoted in the Medical Xpress item as saying 98% fewer outcomes were attributable at 4 to 12 months than with conventional approaches in the same dataset. We did not see that figure in the preprint abstract, so treat it as provisional.

The confirmed effects fall in four systems: respiratory, cardiovascular, musculoskeletal and integumentary (skin). Medical Xpress lists myopathy, non-tuberculosis pneumonia, respiratory issues and pneumothorax among the top verified effects. Some symptoms lasting beyond a year showed up at similar rates in people who tested negative, so the authors do not attribute them to COVID.

What it does not show

Leskovec himself says the finding does not mean long COVID is 98% less common. His broader point is that even 244 million patients can give precise answers to the wrong question, and that study design matters as much as data size.

• It is a preprint. We have not confirmed what changed in peer review.
• "Not attributable to COVID" does not mean "absent." Someone can have a real symptom that this method can't tie to infection.
• We could not read the full text, so we can't say how it handles some obvious questions. Claims data capture billed diagnoses, which may miss subjective symptoms like fatigue or brain fog. And people who test negative may differ from everyone else, including having had other infections. Those are questions for the authors, not criticisms we can substantiate.
• We found no response yet from WHO, NASEM or CDC.

What it means for you

Little changes. The Medical Xpress item describes the practical takeaway the same way: long COVID is real, can last months, and some people still have symptoms at a year or longer.

If you had COVID and still feel off, the useful move is the same as before: describe your symptoms and timeline to a clinician and let them check for other causes too. This study suggests that some lingering problems may stem from something other than the infection. That is a reason to look, not a reason to be dismissed.

Context: where this fits

The study lands in a field that has struggled with definitions. NIH's RECOVER initiative built a symptom-based scoring system from nearly 10,000 Americans, using 12 symptoms and finding four clusters. Of 2,231 people with a first infection on or after December 1, 2021, about 10% had long COVID after six months.

RECOVER's electronic health record analysis is an example of the conventional design, comparing infected and non-infected patients with propensity weighting across 11 million NYC-area and 16.8 million Florida patients. Conditions including dementia, hair loss, pulmonary fibrosis, dyspnea, chest pain, malaise and fatigue replicated across both cohorts. The new work challenges that kind of comparison.

A 2023 Nature Reviews Microbiology review put long COVID at no fewer than 10% of infections, with more than 200 symptoms identified. A 2026 eClinicalMedicine case-control study of multi-organ imaging and physiology is titled as reporting structural and functional deficits in association with long COVID; we saw only a snippet, so we can't speak to its findings, size or setting. We saw only a snippet of that one, so we can't speak to its size or setting.

Taken together, the picture is less a verdict on whether long COVID exists and more a fight over which problems belong under its name.