COVID-19 has been reported to cause multisystem effects lasting years [Z. Al-Aly, E. Topol, Science 383, 830-832 (2024)]. These findings are primarily based on cohort studies comparing test-positive COVID-19 patients to controls without known COVID-19 infection, tested or untested [M. Cai, Y. Xie, E. J. Topol, Z. Al-Aly, Nat. Med. 30, 1564-1573 (2024)]. However, such comparisons can be biased by the reasons people were tested. Using 14.4 billion insurance claims from 244.7 million US patients, we find that most long-term effects attributed to COVID-19 by unrestricted control designs likely reflect bias due to selective testing of individuals with non-COVID conditions, causing long-term effects of those conditions to be misattributed to COVID-19. We removed this bias by analyzing only tested individuals, comparing test-positive to test-negative patients. Across 614 clinical outcomes followed over 2 y, the test-based design resulted in 83.5% fewer symptoms being attributable to COVID-19 between 30 to 119 d after infection (43 attributable symptoms vs. 262 for the unrestricted control design), a 98.2% reduction at 120 to 359 d (5 vs. 286), and a 100% reduction at 360 to 719 d (0 vs. 369). The unrestricted control design also attributed unrelated outcomes (e.g., firearm injuries) to COVID-19 at a high rate, whereas the testing-based design did not. The test-based control design may modestly underestimate some effects. This study does not indicate that symptoms beyond one year are not genuine but that their attributability to COVID-19 is uncertain.
the test-based design resulted in 83.5% fewer symptoms being attributable to COVID-19 between 30 to 119 d after infection (43 attributable symptoms vs. 262 for the unrestricted control design)
observational comparator differences and threshold-qualified outcome categories do not identify an equivalent reduction in symptom burden
The unrestricted control design also attributed unrelated outcomes (e.g., firearm injuries) to COVID-19 at a high rate, whereas the testing-based design did not.
negative-control outcomes directly demonstrate markedly different rates of implausible positive associations within this dataset
the frequency of signs and symptoms in the test positive group decreased to that of the test-negative group within approximately 1 y
thresholded non-detection cannot establish equality in an observational contrast vulnerable to collider bias and subsequent infection among controls
it may have fewer symptoms—and ones of shorter duration and lower overall severity—than previously thought
severity was not measured and null-directed biases limit duration inference from claims-based comparator differences
Derived from the full evaluation — not a separate score.
Strengths
Applying both designs to the same claims source makes differences in estimated outcomes directly visible rather than relying on comparisons between unrelated studies.
↳ Results, Number and Duration of Long COVID Signs and Symptoms; Fig. 1
The analysis combines 49 negative-control outcomes, a negative-test placebo exposure, and detection of known acute symptoms to examine distinct aspects of design performance.
↳ Results, Comparing False Attribution Rates, Biases Due to Selective COVID-19 Testing, and Study Design Validation; Fig. 3
The paper explains how collider bias, subsequent control infections, and survivor selection can attenuate estimates, and limits its target population to tested healthcare-seeking patients.
↳ Discussion, limitations paragraphs
Limitations
The 43-versus-262 comparison counts categories crossing an RR and multiplicity-adjusted uncertainty threshold. It therefore combines comparator effects with threshold choice, precision, and differences in cohort composition.
↳ Methods M2–M3; Results, Number and Duration of Long COVID Signs and Symptoms
Conditioning on testing and later infection among controls can attenuate contrasts. Finding no qualifying late-window outcomes does not establish equal symptom rates or individual recovery within one year.
↳ Results, Characterizing Long COVID Symptoms and Recovery Trajectory; Fig. 4; Discussion
The Abstract describes bias as removed and outcome-category counts as symptoms, while the Discussion infers lower severity without measuring severity. These formulations exceed the narrower comparator-design evidence.
↳ Abstract; Discussion, first paragraph
The same-source comparison and negative-control analyses support the conclusion that comparator choice can generate substantial spurious attribution in claims-based long-COVID research. Four-year history weighting and aligned enrollment strengthen the design without resolving unmeasured differences or conditioning-related bias. The count of threshold-qualified outcomes is not a direct measure of symptom burden, and late-window non-detection does not establish recovery. These distinctions constrain the interpretation of the abstract and Figure 4 while preserving the paper's methodological contribution.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Strong3.8
Confidence highThe same-data comparator contrast, negative-control outcomes, and placebo exposure meaningfully extend the methodological debate over long-COVID attribution. The advance concerns design sensitivity rather than definitive symptom duration or severity.
Temporal alignment, four-year history weighting, and negative and acute-outcome checks provide a credible observational strategy. The threshold-dependent count comparison and incompletely quantified collider, contamination, and measurement biases prevent treating the test-based estimates as unbiased effects.
The sequence from comparator construction to validation and trajectories is followable. Central wording about removed bias and fewer symptoms materially exceeds what a contrast of threshold-qualified outcome categories establishes.
The Discussion distinguishes attribution from symptom burden and engages unrestricted-control studies, prospective RECOVER evidence, and relevant bias mechanisms. Comparisons between claims outcome counts and preselected self-reported symptoms, and extrapolation across prior cohorts, are less secure.
Lower confidence on Methodological Rigour — domain match limited.
Caveats4 of 4 checks
The supplied text lacks an ethics-review determination or exemption declaration, and Reference 59 has an access date after publication. Neither note establishes an error in the empirical results.
Conflict of interest: declared · Ethics approval: not stated · Data availability: declared · Code availability: declared
56 of 56 checkable references verified
59 references in manuscript 3 have no canonical index record — counted, but not index-checkable 4 references confirmed by manual review
References not verified automatically are resolved manually within 24 hours.
No retraction or expression of concern found in Crossref (includes Retraction Watch).
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
The method has been executed at national scale and is accompanied by analysis code and validation checks. Its immediate application is research design, with triangulation still needed before changing clinical guidance or burden estimates.
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
This systematic review and meta-analysis based on large MRI studies demonstrated that SVD is a common health problem affecting about one-fifth of the adult population.
TentativeOverall, ML approaches show promise for scalable diagnosis and risk stratification in iRBD, but progress is constrained by methodological bias, inconsistent endpoints, data imbalance and a lack of explainable, externally validated models.
TentativeBeaumont, Singh, Soares, et al. · 2026 · Basic medicine