Study design and study settings

We conducted a population-based cohort study of adults who received commercially available serologic tests for SARS-CoV-2 antibodies to assess how mutually exclusive combinations of serologic and molecular results, supplemented with survey data on prior COVID-19, were associated with symptoms and health domains consistent with long COVID (S1 Fig in S1 File). The survey was conducted in Kaiser Permanente Southern California (KPSC) and Colorado (KPCO) integrated health care delivery systems, which comprise employer-based and individual insurance, and Medicare and Medicaid plans. KPSC serves approximately 4.8 million members and KPCO serves approximately 0.5 million members. For context in these study settings, the first California and Colorado COVID-19 cases were identified in January and March of 2020, respectively. California experienced a peak of Alpha variant mid-May 2021, Delta from June to mid-December 2021, and Omicron starting in late December 2021. Alpha peaked in Colorado prior to May, Delta from May to December, and Omicron after December 2021 [28]. Vaccines (Pfizer-BioNTech BNT162b2, Moderna mRNA-1273, and Janssen JNJ-7846725) became available in December 2020. This study was conducted as part of a multisite collaboration to assess the clinical roles of serology across our health systems [29,30].

For varied indications (e.g., patient request, physician order for symptoms, research studies, community events), clinical staff collected nasopharyngeal, anterior nares, saliva, or oropharyngeal specimens for SARS-CoV-2 molecular testing, conducted phlebotomy for serology, and conducted laboratory assays in each health system. The health systems employed the Abbott anti-nucleocapsid IgG test, Roche anti-nucleocapsid total IgG, and three anti-spike receptor binding domain assays (Siemens ADVIA Centaur SARS-CoV-2 Spike Total IgG/IgM, Spike IgG COV2G, and sCOVG Spike IgG) [31,32]. Anti-spike antibodies can be induced by natural infection or vaccination, while anti-nucleocapsid antibodies are induced only by infection. SARS-CoV-2 molecular (RNA) testing results were derived from real-time reverse transcription polymerase chain reaction assays made by Roche (Rotkreuz, Switzerland) and ThermoFisher (Waltham, MA), and transcription mediated amplification assay for the qualitative detection of SARS-CoV-2 by Aptima on the Panther (Hologic, Inc., San Diego, CA). A positive test was defined in accordance with the manufacturer’s instructions for each test.

Both health systems maintained comprehensive electronic health record (EHR) databases, with details from health care encounters, including diagnoses, procedures, laboratory tests and results, pharmacy records, and vaccination data. Information from state immunization registries were also available. Data were coded according to a common data model to facilitate extraction using distributed code.

Participant recruitment

Eligible participants were KPSC and KPCO members ≥18 years with ≥1 SARS-CoV-2 serology result before the study start date (August 4, 2021) who spoke English or Spanish, had not opted out of research, and had an available contact method. If members had >1 serology test, the index test was the positive serology result most proximal to August 4, 2021; if no positive result was identified, the index test was the most proximal negative result. Serologic results on or after the first COVID-19 vaccine dose were excluded. Eligible members had at least 12 months of enrollment in a KP health plan prior to the index serology test to capture covariates. Members with a negative serology test and no RNA tests were excluded as COVID-19 could not be confirmed. For members with multiple RNA results before the index serology test date, the most proximal to the index serology was designated as the “index” result, selecting positive RNA test results over negative, and negative over missing results. For positive index serologies, the RNA test prior to, or on the day of, the positive serology test, was selected. To classify serology as negative, we required ≥14 days between the index serology and the RNA test to ensure adequate time elapsed to seroconvert. To be classified as RNA negative, all RNA tests prior to the index serology had to be negative.

A priori, we created four exposure groups based on the index serology test and RNA tests drawn before the index serology test date (Table 1). Among serology-positive participants, the Ab+/RNA+ group (1) was considered to have had SARS-CoV-2 infection with an expected immune response whereas the Ab+/RNA- or RNA missing group (2) was considered to have had a false-positive serology or not have been tested by SARS-CoV-2 RNA due to less severe infection. Among serology-negative participants, the Ab-/RNA+ group (3) was presumed to have an impaired or waning immune response whereas the Ab-/RNA- group (4) was planned as a control group. However, the majority of participants in the Ab-/RNA- group reported prior COVID-19 in the survey (described below), so this group was subdivided into an Ab-/RNA-/no prior COVID-19 group consistent with a “true control” group (4a) and an Ab-/RNA-/prior COVID-19 group consistent with prior self-reported COVID-19 (4b).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. COVID-19 test groups.

https://doi.org/10.1371/journal.pone.0304262.t001

Eligible members in each health system were selected using stratified random sampling technique by the study site, four exposure groups, and language. All eligible Spanish-speaking members were sampled from all groups. All eligible English-speaking members in the Ab+/RNA+ and Ab-/RNA+ groups at KPCO and Ab-/RNA+ group at KPSC were sampled to achieve target sample sizes.

Statistical power was calculated for a priori comparisons in any dyspnea between the Ab+/RNA+ vs. Ab-/RNA+ groups and Ab+/RNA+ vs. Ab-/RNA- groups. The assumed power was 80%, with an alpha of 0.05, design effect of 2, and prevalence of dyspnea of 25% and 15% for groups 3 and 4 respectively. The sample sizes needed for a difference in proportions of 0.103 between the Ab+/RNA+ vs. Ab-/RNA+ groups and 0.071 between Ab+/RNA+ vs. Ab-/RNA- groups were 1500 for Ab+/RNA+ group, 400 for Ab-/RNA+ group, and 700 for the Ab-/RNA- group.

Survey development and content

The investigators developed a survey instrument to measure 31 potential long COVID symptoms and health domains, based on conditions and symptoms reported to be associated with long COVID, with existing, validated scales for quality of life, gastrointestinal symptoms, anxiety/depression, cognition, pain, dyspnea, sleep, and fatigue from the Patient-Reported Outcomes Measurement System (PROMIS); chest pain from the Seattle Angina Questionnaire; palpitations (including tachycardia, heart flutters, and extra or missed heart beats) from the Health and Quality of Life Outcomes for Arrythmia; headaches limiting ability to work from the ID-Migraine Questionnaire; fainting/lightheadedness from the Scales for Outcomes in Parkinson’s Disease Autonomic Dysfunction; sense of smell and taste from the University College of London Patient Led Research Questionnaire on Long COVID; hair loss from the Hamilton-Norwood and Ludwig scales; and newly developed items for menstrual symptoms (S1 Table in S1 File) [33–39]. To maximize recall, items referred to the last 7 days, except hair loss, lightheadedness, headaches, and chest pain, which referred to the last 4 weeks, and menstrual symptoms, which referred to the last 3 months. For participants who reported prior COVID-19, a checklist was developed for 31 symptoms (e.g., fatigue, brain fog, body aches) that were persistent (≥ 3 months), recurrent (went away for ≥ 2 weeks and then returned) or new (occurred ≥ 2 months after having had COVID-19). The survey also assessed race, ethnicity, gender, whether participants had or thought they had prior COVID-19, hazardous drinking using the Alcohol Use Identification Test (AUDIT-C) [40], current tobacco use or vaping [41,42], and vaccination status. The survey was professional translated into Spanish and pre-tested by a multidisciplinary clinical and research group, survey researchers, and a physician who treats long COVID followed by a pilot test with 100 members. Surveys were administered electronically using Research Electronic Data Capture (REDCap) [43] and on paper.

Recruitment and survey procedures

Initial recruitment occurred ≥ 12 weeks after serology testing by email and text messaging with a link to a REDCap face page containing information about the study and a link to the survey. After electronic recruitment attempts, one recruitment attempt was made by US mail. Spanish surveys were provided if the EHR indicated Spanish as a preferred language. Recruitment occurred between November 18, 2021, and April 20, 2022, and data collection closed at the end of June 2022. Participants were renumerated $10. Participants were contacted to verify if key information about outcomes were missing; participants with non-response or missing data to prior COVID-19, symptom and health domain measures, or key covariates were excluded from the final study population.

Ethics approval and study reporting

The KPSC Institutional Review Board approved the study and waived the requirement for written informed consent due to minimal risk to study participants. Potential participants received information about the study and agreed to the use of survey answers and electronic health record information. Study reporting follows the Strengthening the Reporting of Observational Studies in Epidemiology checklist.

Data analysis

Variables extracted from the EHR included age, sex, language preference, insurance enrollment, RNA and serology test dates, test results, and vaccination dates. We compared characteristics of participants and non-respondents using EHR data. Data quality was assessed by examining missing data, congruence between reported prior COVID-19 and RNA results, and timing of reported COVID-19 relative to COVID-19 surges in each region.

After accounting for the stratified sampling design and applying weights to account for sampling and survey response probabilities, we described self-reported clinical and demographic participant characteristics using weighted percentages, means (SD) and standard errors, such as the proportions with prior COVID-19 and persistent, recurring, and new symptoms. For PROMIS measures, raw scores were summed, reverse scored when appropriate, and converted to standardized T scores in which a score of 50 represents the mean score of a reference population (i.e., the US general population), with an SD of 10. For PROMIS measures with previously defined general population cut-offs, those were applied to identify higher levels of impairment in the outcomes [34]. When published cut-offs were only derived from clinical samples (e.g., patients with gastrointestinal conditions), the 75^th percentile was the cut-off. For measures with no published cut-offs, any reported symptom was the cut-off. For measures asked as a single question, the dichotomized responses were used as the outcome. Multivariable logistic regression was applied, accounting for the sampling strategy and including the health system, age, sex, race, ethnicity, and time between serology test date and survey date as covariates. Separate models were run for each measure as the outcome. To assess our hypothesis that antibody presence was associated with long COVID among people with a history of COVID-19 infection, we compared group 1 (Ab+/RNA+) to group 3 (Ab-/RNA+). To check the validity of our data, we assessed whether people who had COVID-19 with an expected immune response would report more impairment than people who had not had COVID-19 by comparing group 1 (Ab+/RNA+) to group 4a (true controls, Ab-/RNA-/no prior reported COVID-19).

Secondary and sensitivity analyses

In secondary analyses, group 1 (Ab+/RNA+) was compared to group 2 (Ab+/RNA-) and participants who were hospitalized with their initial infection were compared to those who were not. In the overall sample, we also examined the physical and mental health by vaccine status. In sensitivity analyses, we applied linear regression to examine T scores as continuous variables for measures with general population estimates [34], comparing group 1 (Ab+/RNA+), group 3 (Ab-/RNA+), and group 4a (true control group). Due to waning antibody levels over time [44,45], we examined the association between antibody status and fair/poor physical health and fair/poor mental health for individuals in which the time between the RNA test and the antibody test was 3 or fewer months. Finally, we assessed the association between antibody status and symptoms adjusting for the modified Charlson co-morbidity score assessed in the year prior to the index serology test date [46]. Otherwise, secondary and sensitivity analyses adjusted the same covariates as the primary analyses.