Study design
We conducted a matched cohort study based on routine health care data as applied in a previous study6. In the present analysis, we compared the rates of newly diagnosed symptoms and conditions between adult individuals with and without documented SARS-CoV-2 infection and influenza infection based on ICD-10-GM coding. Persons infected with SARS-CoV-2 during 2020 and matched contemporary controls without infection were followed until September 30, 2021, for a minimum of three and a maximum of 18 months using the date of COVID-19 onset as the index date for randomly selected match groups. Persons infected with influenza in the first half of 2018 where followed until September 30, 2019. Following the NICE guidelines on long COVID26 and the clinical case definition of post-COVID conditions proposed by the World Health Organization (WHO)27 the post-COVID phase was defined as starting three months after the initial diagnosis of COVID-19. Outpatient services are documented per quarter rather than on a daily basis in the German statutory health care billing system. A diagnosis was therefore associated with the post-COVID period if it was newly documented in the second quarter after the index date or later. This operationalization ensured a time interval of at least three months between the date of COVID-19 diagnosis and post-COVID outcome incidence.
Cohorts
The COVID-19 cohort included individuals with a birthdate before 2003 (aged 18 and older) and a polymerase chain reaction (PCR)-confirmed COVID-19 diagnosis (ICD-10 U07.1) in 2020. To calculate risk exposure time, we defined the index date by using the date of an outpatient PCR test or the date of admission to a hospital with a COVID-19 diagnosis. In rare cases where no PCR test had been billed to the insurance company and no hospital stay was recorded, other documented events, such as the start of sick leave or the first contact with the responsible physician, served to determine the index date. The contemporary control cohort included individuals who were not diagnosed with COVID-19 as ICD-10 U07.1 or ICD-10 U07.2 between January 1, 2020 and September 30th, 2021. The Influenza cohort included persons born before 2001 and an ICD-10 J10 code between January 1, 2018, and June 30, 2018 as this was the last Influenza epidemic in Germany where the dominate Yamagata variant was not covered by the available trivalent vaccine in Germany.28, 29
We excluded individuals with COVID-19 diagnosis without laboratory virus detection (ICD-10-GM: U07.2) from the COVID-19 cohort and contemporary control cohort to reduce distortions due to misclassification. We further excluded individuals who were not continuously insured with the respective health insurance company between January 1, 2019 and September 30th, 2021 (or death) and for the Influenza cohort between January 1, 2018 and September 30th 2021 (or death) respectively, to ensure that relevant outcomes and preexisting health conditions were visible in our data. For each individual, preexisting medical conditions were assessed for at least 12 months prior to the matching point of the COVID-19 and control cohorts. Starting from the index date assigned from the COVID-19 case matched individuals were jointly followed for a maximum of 18 months. Only patients from the year 2020 for the COVID-19 and contemporary control and 2018 for the Influenza cohort were selected, as this ensured that the effect was not influenced by vaccinations (Vaccination for SARS-CoV-2 in Germany started as of December 27, 2020).
Ethics and registration
The POINTED study protocol was approved by the ethics committee of the TU Dresden (approval number: BO-EK (COVID)-482102021) and adheres to all relevant administrative and legal regulations. The study was registered at ClinicalTrials.gov (NCT number: NCT05074953).
Data
The underlying data sources were set up for the “Post-COVID-19 Monitoring in Routine Health Insurance Data” (POINTED) consortium6 to study the long-lasting effects of the COVID-19 pandemic in Germany. The POINTED consortium is coordinated by the Center for Evidence-Based Healthcare (ZEGV) at the TU Dresden and consists of the German National Public Health Authority, the Robert Koch Institute, health research institutes, and statutory health insurances. It is partly funded by the German Federal Ministry of Health (BMG).
We used routine health care data from different German statutory health insurances: Techniker Krankenkasse, BARMER, DAK Gesundheit, IKK classic, AOK PLUS, and several company health insurance funds (InGef30). In total, these data cover approximately 39 million individuals, which corresponds to nearly half of the total German population. In addition to sociodemographic characteristics (age and sex) and vital status (via the date of death), we had access to comprehensive information on health care utilization in the outpatient and inpatient health care sectors. The data comprise records on diagnoses (according to the International Statistical Classification of Diseases and Related Health Problems - German Modification, ICD-10-GM), medical procedures (according to the “Operationen- und Prozedurenschluessel,” OPS; German modification of the International Classification of Procedures in Medicine, ICPM), information on outpatient medical services (according to “Einheitlicher Bewertungsmassstab,” EBM), and prescribed medications (according to the German Anatomical Therapeutic Chemical (ATC) Classification).
Matching
To minimize differences between the COVID-19 and control cohorts in terms of covariates that may confound relationships between outcomes and exposure, we applied 1:3 matching with replacement for COVID-19 to non-COVID-19 contemporary controls and 1:1 for COVID-19 to influenza patients. For each individual in the COVID-19 cohort, we selected three non-COVID-19 individuals with identical age (in years) and sex. Influenza controls were assigned using the same age, sex and disease severity (outpatient, hospital, ICU). We chose exact matching on these characteristics to facilitate stratified analysis. In addition, we accounted for the presence of covariates by propensity score matching. The estimation of the propensity score was based on logistic regression including all insured individuals.
After matching individuals with COVID-19 and controls, we excluded individuals from the match groups if they died before the beginning of the post-COVID phase, i.e., within the quarter of the COVID-19 diagnosis or the following quarter. We also excluded individuals with COVID-19 who lacked a matching partner. When analyzing specific health outcomes, we further excluded individuals from the analysis if the considered outcome was documented in two of the four quarters preceding in the outpatient setting or once in the inpatient setting. To maintain cohort balance on covariates, complete match groups of COVID-19 and control cases were excluded if the outcome was preexisting in the individual with COVID-19 or all of their matched non-COVID-19 contemporary control cases. For estimation, data from individuals in the contemporary control cohort were weighted with the inverse number of individuals remaining in the respective match group (i.e., weights between 1/3 and 1) to ensure that total weights in the control cohort added up to the number of individuals in the COVID-19 cohort. Due to a smaller pool of potential Influenza patients, only a subgroup of COVID-19 patients could be used and influenza patients generally had to be included multiple times in the matching process.
Outcomes
Although the ICD-10 catalog lists codes for post-viral disease (B94.8) and as of 2021 also for post-COVID (U09.9), these codes may largely underestimate the proportion of affected patients.21 For this reason, we follow the widely used strategy to define post-COVID by symptoms and conditions associated with it. Based on published literature, previous work developing a core outcome set31, and the clinical expertise of the author team, we selected a large set of 96 outcomes covering multiple organ systems and diagnosis/symptom complexes (Supplemental Table 1). These outcomes constitute new-onset morbidity documented in ICD-10-GM codes by a physician or psychotherapist in the inpatient or outpatient sector within the statutory healthcare system. Of these 96 symptoms and conditions, seven were selected to represent the WHO post-COVID clinical case definition (malaise/exhaustion, chronic fatigue syndrome, dyspnea, respiratory insufficiency, chest pain, cognitive impairment, memory disorder). These symptoms and conditions cover the three main clusters of persistent fatigue as well as respiratory and cognitive problems. Furthermore, four conditions were selected for the less common but potentially more severe tissue damages (pulmonary embolism, lung damage, pericarditis, and myocarditis). Lastly, two negative control endpoints were defined as melanoma and tinea-pedis. Both endpoints are assumed not to be caused by a SARS-CoV-2 or Influenza infection, but subject to the same unmeasured exposures such as health seeking behavior (detection bias) after an infection or in case of contemporary controls also to lockdown effects. 32
Covariates
We used information on preexisting chronic conditions as available health records from 2019 and 2017, respectively to adjust for potential confounders in the relationship of exposure (COVID-19) and endpoints. The approach is the same as in a previous study6. For each individual, we used information on preexisting health conditions in the four quarters preceding the index date. The 34 prevalent morbidities were based on published evidence and clinical expertise. In addition, we included age, sex, and the number of recorded inpatient and outpatient contacts as covariates. In line with previous studies33, we included the severity of COVID-19 as a stratification feature and differentiated between (1) individuals with outpatient diagnoses of COVID-19, (2) individuals with a hospitalization with COVID-19, and (3) individuals requiring intensive care and/or mechanical ventilation with COVID-19 or influenza.
Statistical analyses
The incidence rates (IRs) of the endpoints per 1,000 person-years were estimated. Differences in IRs between COVID-19 and non-COVID-19/influenza patients were estimated using Poisson regression models to estimate incidence rate ratios (IRRs). As a prerequisite, we derived aggregated information on each health outcome by counting incident cases of the respective endpoint within the COVID-19 and control cohorts. Since the number of incident cases for each outcome varied across the match groups of contemporary controls, we assigned weights to the remaining cases that added up to 1. The pooling of individual-level data was not possible due to data protection restrictions. Each authorized institute calculated the required aggregate statistics and provided them to ZEGV, where regressions based on combined aggregate data were performed.
To synthesize evidence across datasets, point estimates from aggregate matched data were found to be the same compared with Poisson regression based on individual level data34. The characteristics of Poisson regression applied to aggregate count data allowed for consistent estimation of incidence rates regardless of the distribution of the outcome on the individual level when the conditional mean function is correctly specified.35 While the variance estimates for a 1:1 matching are the same model estimates on individual as well as aggregate data level, the variance estimates from aggregates for 1: M matching tend to be larger, meaning that the statistical significance of the presented effects may be underestimated. However, in the case of a 1:1 matching with replacement from a comparatively small pool of Influenza controls the variance is larger in the models based on aggregate data. To address this issue, simulations were conducted and it was determined that weighting by natural persons per control case provided appropriate variance estimates. Utilizing a main advantage of Poisson regression, we adjusted for differences in times at risk (time between the index date and the end of the observation period or death) due to inclusion of these times as offset in the model. Stratified aggregation enabled us to deploy separate estimators for age, sex, and severity of the infection.
To investigate the persistence of the endpoints we employed the Kaplan–Meier estimator.36 This estimator allowed to approximate the course of the symptoms and conditions under each censoring due to death or end of observation time for the 6 quarters after the index date. The absence of the diagnosis was interpreted as a loss of the symptom or condition. All analyses used the statistical programming language R version 3.6.337.