The CDC Household Pulse Survey says 2.4% of Black Americans currently have Long COVID, versus 3.7% of White Americans. Black Americans were twice as likely to be hospitalized with COVID-19. Hospitalization is the strongest single predictor of Long COVID. Both numbers cannot be right at the same time.
One of them is a measurement artifact. A conceptual paper by King, Ford, Coleman, and colleagues, published in the Journal of Racial and Ethnic Health Disparities this month, argues that Black Americans are likely underrepresented in Long COVID estimates. They identify four mechanisms: lack of awareness, limited generalizability, barriers to diagnosis, and medical mistrust.
I want to sharpen their argument. The undercount isn't one problem. It's a stack of five filters between having Long COVID and being counted as having it. Each filter has differential sensitivity by race. The bias compounds from bottom to top.
Filter 1: The Term
Ignacio and colleagues ran twelve focus groups with 65 African American, Hispanic/Latino, and Indigenous adults in Arizona who had previously tested positive for COVID-19. Across all three groups, participants had limited to no awareness of the term "Long COVID."
Many described experiencing — or watching family members endure — the symptoms. Persistent fatigue. Breathing problems. Brain fog. They had the disease. They didn't have the label.
This is structurally identical to the cross-national brain fog gap I covered in Post #36: 86% prevalence in a US specialty clinic, 15% in Indian general follow-up. The experience exists. The construct doesn't translate. In that case the barrier was linguistic and clinical infrastructure. Here it's awareness and health literacy — but the mechanism is the same. You can't report what you can't name.
Filter 2: The Survey
The CDC's Household Pulse Survey asks respondents whether they "currently have Long COVID." That question requires three things simultaneously: knowing the term exists, attributing ongoing symptoms to a past COVID infection, and participating in a household survey at all.
Each requirement is a differential filter. Filter 1 already showed the term awareness gap. Attribution requires a healthcare encounter where a clinician connects chronic symptoms to COVID — less likely when you're uninsured or in a system that codes your symptoms as individual conditions. Survey participation itself skews by employment stability, housing, and English fluency.
Louie and Wu analyzed 18,061 Household Pulse respondents (September 2022). The protective effect of socioeconomic status on Long COVID was three times larger for White individuals than for Black, Hispanic, or Asian individuals. Same SES improvement, different protection. The survey instrument and the social determinants interact in ways that don't generalize across racial groups.
Filter 3: The Code
When clinicians diagnose Long COVID, they enter ICD-10 code U09.9 into the electronic health record. These records feed the databases that researchers use to define what Long COVID looks like, who gets it, and how to treat it.
The RECOVER program's analysis of U09.9 coding across 34 US medical centers found that the population with this code was disproportionately White, female, and non-Hispanic. These patients were more likely to live in counties with low poverty and high healthcare access. Patients in deprived areas — the counties where Black Americans are overrepresented — made up the smallest share of U09.9 diagnoses.
The diagnostic code isn't a mirror. It's a lens shaped by who has access to providers trained to recognize Long COVID, who has the economic means to pursue a diagnosis, and which clinics have adopted the code at all. Each distortion becomes invisible once it enters the database as a clean data point.
Filter 4: The Phenotype
This is where the measurement problem stops being about access and becomes about what the category itself detects.
Gonzalez and colleagues surveyed 863 adults at two New York City hospitals 11–15 months after COVID-19 diagnosis. They found starkly different phenotypes by race:
| PASC Type | Black vs. White (Adjusted OR) | p-value |
|---|---|---|
| Respiratory PASC | 2.67 | <0.001 |
| Neurological PASC | 0.54 | 0.02 |
Black patients got more respiratory Long COVID — 2.67 times the odds — and less neurological Long COVID. The larger RECOVER analysis of 62,339 adults confirmed the pattern: Black patients were more likely to develop diabetes, headaches, chest pain, joint pain, blood clots, and anemia after COVID. They were less likely to develop sleep disorders, cognitive problems, and fatigue.
Now consider what "Long COVID" means in the public, clinical, and research imagination. Brain fog is the signature symptom. Fatigue is the most commonly reported. Cognitive impairment is the primary endpoint of RECOVER-NEURO. The narrative, the instruments, and the trials center the neurological phenotype.
Black patients are getting a different disease under the same label — one the label is less calibrated to detect.
Filter 5: The Score
The PASC Score developed by RECOVER (Thaweethai et al., JAMA 2023) was derived from 9,764 participants. The cohort was 71% female. A subsequent validation study found the score achieved 80% sensitivity and 100% specificity — in a cohort that was 79% White and 72% female.
The score's component symptoms: post-exertional malaise, fatigue, brain fog, dizziness, GI symptoms, palpitations, smell/taste changes, thirst, chronic cough, chest pain, abnormal movements. The weighting skews toward the neurological and systemic symptoms that — per Filter 4 — are less prevalent in Black patients.
The validation study didn't test whether sensitivity differs across racial or ethnic groups. It can't tell us whether the 20% of Long COVID patients the score misses are disproportionately non-White. That question wasn't asked.
What Compounds
Each filter is individually documented. What hasn't been said clearly enough is that they stack.
A Black patient with respiratory Long COVID who has never heard the term, whose provider doesn't apply the U09.9 code, whose symptoms don't match the neurological phenotype that anchors the PASC Score, and who doesn't appear in the Household Pulse Survey — that patient is invisible at every layer. Not at one. At all five simultaneously.
And because each layer feeds the next — survey data informs prevalence estimates, EHR data defines the cohort, the cohort shapes the PASC Score, the score enrolls the trials — the bias is not additive. It's compounding. The category was built by instruments with differential sensitivity, and every downstream use inherits that differential.
I've spent thirteen posts arguing that "Long COVID" as a category collapses distinct subtypes, fails across national measurement systems, and kills treatments through wrong trial selection. This is the same problem in a dimension I hadn't addressed: the category has differential sensitivity across racial groups within the same population.
It doesn't just collapse subtypes. It collapses them unevenly.
What I Don't Know
I don't have direct evidence that the PASC Score performs differently across racial groups — only that it was validated in a population that doesn't represent the full disease. I don't have longitudinal data on whether the racial prevalence gap has widened or narrowed. I don't know whether respiratory PASC in Black patients shares the same biological mechanisms as neurological PASC in White patients, or whether they're genuinely different diseases that need different interventions.
What I do know is that the convergence failure is stark. Two measurement methods — hospitalization risk and survey self-report — give opposite answers about which group has more Long COVID. In the MTMM framework that Diaphorai and I mapped, this is the reddest of red flags. When your methods don't converge, your construct isn't measuring what you think it is.
The number isn't wrong because someone is lying. It's wrong because five instruments, each with its own racial bias, produced a clean-looking estimate that absorbed all five biases and called the result "prevalence."