Comorbidity Timing Changed Long COVID Risk Estimates Across a 6.1 Million-Patient Cohort
In a national analysis of Long COVID risk, the timing used to define preexisting illness changed the size of the signal in a meaningful way. Using electronic health records from 6,130,413 adults with documented COVID-19 across 49 U.S. states, investigators compared comorbidity exposure windows anchored either before SARS-CoV-2 infection or before Long COVID diagnosis. The result was not a minor bookkeeping difference. Defining comorbidities before infection rather than before Long COVID diagnosis produced 8%–314% higher adjusted attributable risks and 3%–34% higher adjusted relative risks across 16 major comorbidities.
One sentence captures the methodological hazard: exposure timing can change the apparent biology of post-acute disease.
The study was built around a common but often underappreciated problem in observational medicine. In chronic disease research, the date chosen to classify a baseline condition can shift who is counted as exposed, especially when a disorder is recognized late, coded inconsistently, or underascertained in routine care. Here, that issue mattered because Long COVID is itself a time-dependent diagnosis, and comorbidity status can look different depending on whether the clock starts at infection or at later clinical recognition of persistent symptoms. The data showed that window choice can inflate or dampen estimated risk in ways that are large enough to alter clinical inference and public health prioritization.
Delayed Recognition Can Make Risk Look Smaller Than It Is
The direction of bias was especially important for comorbidities prone to delayed clinical recognition or incomplete capture in the record. If a condition is only documented after the acute infection, then anchoring exposure to the later Long COVID diagnosis can misclassify some patients as unexposed during the period that matters most. In practical terms, this can pull risk estimates toward the null and make a preexisting vulnerability appear less important than it really is. The investigators directly framed exposure-window specification as a major source of bias in studies of Long COVID risk.
The analytic strategy was correspondingly ambitious. The team used ensemble cross-fitted double/debiased machine learning to adjust for individual- and county-level confounders, a choice designed to reduce distortion from the many clinical and social variables that travel with both infection outcomes and chronic disease burden. That approach is relevant because Long COVID does not occur in a vacuum; age, geography, neighborhood context, prior disease burden, and health system access all influence who gets diagnosed, who gets followed, and who gets labeled with persistent post-viral illness.
Across 16 major comorbidities, the pattern was consistent enough to matter. Measuring comorbidity before infection yielded larger adjusted attributable risk estimates and larger adjusted relative risks than measuring it before Long COVID diagnosis. The spread across conditions was wide, with some risk estimates rising only modestly and others increasing by more than threefold. For epidemiology, that kind of range is a warning that classification rules are not neutral technical details. They are part of the causal machinery that shapes the final estimate.
Risk Declined Over Time, but Disparities Stayed Visible
The cohort spanned 2020 through 2024, and comorbidity-associated risks generally declined over that period. The decline may reflect changing infection waves, improved acute management, vaccination effects, altered health-seeking behavior, evolving diagnostic practices, or a changing composition of infected patients. The preprint does not isolate a single explanation, and it should not be forced into one. What it does show is that the relation between comorbidity and Long COVID was not fixed across the pandemic years, which means temporal context belongs in every interpretation of the data.
Disparities did not disappear as the risk estimates fell. Substantial demographic, socioeconomic, geographic, and multimorbidity-related differences persisted. That finding matters because Long COVID risk stratification often depends on variables that are unevenly distributed across populations already carrying heavier burdens of chronic disease and lower access to care. If a patient lives in a county with thinner diagnostic coverage, or belongs to a group less likely to have early or complete comorbidity documentation, the measured risk can diverge from the clinical reality.
The study’s emphasis on county-level confounding is important here. Long COVID is partly a disease of access, not only of infection. Diagnosis depends on the availability of clinicians, the willingness of patients to seek care, and the local infrastructure needed to record persistent symptoms in a way that can be found in data systems. Those conditions vary widely across the United States, and the analysis suggests that the resulting disparities remain visible even after advanced adjustment.
Why Exposure Windows Matter in Long COVID Science
The pathophysiologic rationale is straightforward even if the data architecture is complicated. Comorbidities such as cardiometabolic disease, immune dysregulation, and other chronic conditions may shape the host response to SARS-CoV-2 infection, influence inflammatory persistence, and alter the probability that lingering symptoms become clinically apparent. If the exposure is misdated, the biologic sequence is blurred. The patient still has the disease, but the study may fail to count it at the point when it most plausibly contributes to downstream risk.
This has direct implications for future Long COVID research. Risk models that rely on routine electronic records should specify the exposure window clearly and test whether their estimates hold when the comorbidity definition is moved across clinically reasonable time anchors. Without that step, comparisons across studies may be misleading even when the underlying populations look similar. Two papers can appear to disagree on the importance of the same comorbidity while actually using different clocks.
The preprint also included additional numerical analyses by introducing two new cohorts to compare comorbidity exposure-window definitions more comprehensively. That design choice reinforces the central message: measurement strategy is not a side issue in longitudinal EHR research. It can change the magnitude of risk estimates enough to affect how patients are grouped for surveillance, prevention, and follow-up.
What Clinicians and Public Health Teams Should Take from the Data
For clinicians, the immediate use is cautious interpretation. A patient’s documented comorbidity burden should be understood as dynamic, especially in the months surrounding infection and follow-up. For public health teams, the findings argue for stricter harmonization of exposure definitions in Long COVID surveillance and risk stratification tools. If exposure windows are inconsistent, high-risk groups may be overcalled in one analysis and undercounted in another, making targeted prevention strategies harder to defend.
This was a method-focused study, not an interventional trial, and it does not establish causality for any single comorbidity. It also cannot fully solve the underlying problem of incomplete EHR capture. Even so, the message is strong: temporal misclassification can materially distort Long COVID risk attribution. In a syndrome already defined by diagnostic uncertainty, the calendar date used to define comorbidity status can change the story the data tell.