(© arloo - stock.adobe.com)
Flawed Comparisons May Have Inflated Long COVID Symptom Counts
In A Nutshell
- A study of 244.7 million US patients tested how study design shapes long COVID findings.
- An older method linked COVID-19 to unrelated outcomes such as firearm injuries, signaling testing bias.
- Comparing only tested patients cut linked outcomes from 262 to 43 early, and 369 to none after a year.
- Lingering symptoms are not ruled out, but their cause is uncertain.
Telling a patient that COVID-19 made them more likely to be shot afterward would sound absurd. Yet an older style of long COVID research produced results like that when put to the test on a nationwide set of insurance records, tying the virus to firearm injuries.
Researchers from Stanford and other institutions dug through 14.4 billion insurance claims from 244.7 million US patients to find out how many health problems after COVID-19 are caused by the virus. A more careful comparison method found 83.5% to 100% fewer health outcomes that could be confidently linked to COVID-19 than an older design used in earlier studies, depending on how much time had passed since infection.
That does not make long COVID imaginary. But hundreds of health problems blamed on the virus in earlier research, especially ones lasting beyond a year, may have been wrongly attributed to COVID-19. Many could trace back to whatever sent a person to the doctor in the first place, according to the study in the Proceedings of the National Academy of Sciences (PNAS).
Testing Habits Distorted Earlier Long COVID Comparisons
Most big long COVID studies follow one recipe. Scientists gather people who tested positive and compare their later health with people never diagnosed. Any extra illness in the first group gets credited to COVID-19.
That recipe hides a trap. Many people get tested because they are already at a clinic or hospital, sometimes for an unrelated problem such as diabetes. Much of the comparison group had not been to a clinic when enrolled and had less medical care on record. Waiting rooms tend to hold sicker people than living rooms. A gap in later health can therefore reflect who walked into a doctor’s office rather than what the virus did. Researchers call this setup an “unrestricted control” design.
Even people who tested negative stood out. Compared with untested people with similar medical histories, they had higher risk for 82.5% of the health outcomes studied in the first four months, a sign that getting tested at all flagged people with other health issues.
To fix this, the team enrolled patients at their first COVID-19 lab test between March 2020 and March 2021, then compared those testing positive with those testing negative, a test-based design. Both groups had sought testing, making them more alike, and researchers also matched them on four years of prior medical history and basic demographics. The sample was balanced across age, race, income, and education. Follow-up covered 614 health outcomes over two years.
Impossible Results Exposed the Old Method’s Flaw
As a reliability check, the team examined 49 outcomes COVID-19 could not plausibly cause, such as firearm injuries, drowning, and cerebral palsy in adults. Scientists call these negative controls, a trap for a flawed method.
Under the older design, 53.1% of them appeared linked to COVID-19, including a 76% higher likelihood of injury from lightning, drowning, electrocution, and similar causes in the first four months. The test-based design found no such link. Only 4.1% of the unrelated outcomes showed a clear pattern, and those pointed toward lower risk, a hint that the new method may slightly understate some real effects.
Only Five Long COVID Outcomes Held Up After Four Months
At one to four months after infection, the older design flagged 262 health outcomes as linked to COVID-19, compared with 43 for the test-based design. Between five and twelve months, only five outcomes remained. After the first year, none did, versus 369 under the older approach. To count, an outcome needed a clear statistical signal and at least a 10% higher risk than the comparison group.
Early on, confirmed outcomes touched many body systems. Muscle disease showed roughly triple the risk and hair loss nearly double, while heart palpitations and poor nutrition made the list. By five to twelve months, muscle disease, hair loss, blood clots in the lungs, and breathing failure were still elevated. Pneumonia, shortness of breath, chest pain, and fatigue added only a sliver of risk, yet they are common enough that the extra burden still mattered.
Lingering Symptoms Remain Real, but Their Cause Is Uncertain
Authors stressed that the study “does not indicate that symptoms beyond one year are not genuine but that their attributability to COVID-19 is uncertain.” A lingering problem could come from a condition that predated testing, from COVID-19 at levels this method could not detect, or from something else.
Some people who tested negative may have been seeking care for serious illnesses, and 20.2% of the comparison group later tested positive within two years, both of which would make COVID-19’s effects look smaller. Because the work compared two methods on a single database, the reductions show how much the choice of comparison group changed results there, not a correction to every earlier paper.
Fair comparisons decide whether long COVID counts can be trusted. A method that links COVID-19 to gunshot injuries deserves little confidence when sorting out fatigue or chest pain, and future studies need comparison groups who sought the same kind of care.
Disclaimer: This article summarizes published research for general information only and is not medical advice. Anyone with health concerns should consult a qualified health care professional.
Paper Notes
Limitations
Researchers relied on deidentified insurance claims data from the Komodo Health database, which is built for billing rather than clinical research and may lack clinical detail. The authors note that the design still detected well-established acute COVID-19 symptoms, which supports using claims for this purpose. Patients were enrolled between March 2020 and March 2021, and the observation period ended April 4, 2023. People who took only at-home COVID-19 tests were not captured unless they later presented to the health system, and because the analysis required a lab-based PCR test, results apply only to people who sought health care. Patients were tracked using a hashed identifier based on first name, last name, gender, and date of birth, which carries some risk of misidentification, though results held after filtering out overlapping patients. The authors report that the sample was geographically well-balanced and balanced by age, race, income, and education.
Two of the 49 negative control outcomes were significantly lower in test-positive patients, which the authors attribute to limiting the study to people who were tested, a possible source of mild bias that could understate real effects. Other factors could also push estimates slightly downward: all patients were required to survive 30 days after testing, and 1.2% of test-positive patients died in that period versus 0.62% of test-negative patients. Between 0.6% and 1.1% of controls tested positive in any 30-day period, and cumulative rates reached 4.2% within 119 days, 10.1% within 359 days, and 20.2% within two years. The researchers also required a risk ratio of at least 1.1 to count an effect as meaningful, so small real effects may have gone undetected (a risk ratio of 1.1 means a risk 10% higher than the comparison group’s). As an observational analysis, the study cannot establish what caused symptoms that were not attributed to COVID-19.
Funding and Disclosures
Authors declare no competing interest. The paper acknowledges support for first author Hamed Nilforoshan from a Stanford Knight-Hennessy Scholarship and a National Science Foundation Graduate Research Fellowship. Additional support came from several National Science Foundation grants, the Stanford Data Applications Initiative, the Wu Tsai Neurosciences Institute, the Stanford Institute for Human-Centered AI, the Chan Zuckerberg Initiative, Amazon, Genentech, GSK, Hitachi, SAP, and UCB. The claims data were obtained from Komodo Health under a data use agreement with Stanford University and cannot be shared publicly. The analysis code is available on GitHub at snap-stanford/longcovid.
Publication Details
Hamed Nilforoshan, Julia Reisler, Erfan Jahanparast, Michael Moor, Steven N. Goodman, Stefan Wager, and Jure Leskovec authored “Revised estimates of types and duration of long COVID symptoms based on claims records from 245 million US patients.” Nilforoshan, Reisler, and Jahanparast contributed equally. Authors are affiliated with Stanford University, the University of California, Berkeley, and ETH Zurich. The study was published as an open access research article (Medical Sciences) in the Proceedings of the National Academy of Sciences (PNAS), Vol. 123, No. 44, e2604414123, on October 27, 2026, under a Creative Commons Attribution License 4.0. The paper was edited by Elizabeth A. Stuart of Johns Hopkins University and received February 11, 2026, and accepted September 9, 2026. Correspondence: Steven N. Goodman, [email protected]. DOI: 10.1073/pnas.2604414123.
