Exploring predictors of post-COVID-19 condition among 810 851 individuals in Sweden

Categories: Disease & Virus

October 30, 2025

Study design and data sources

This population-based cohort study is part of the SCIFI-PEARL project (Swedish Covid-19 Investigation for Future Insights – a Population Epidemiology Approach using Register Linkage), which is a nationwide multi-register, regularly updated, observational study of the COVID-19 pandemic in Sweden²⁰. The SCIFI-PEARL project links a broad range of national and regional healthcare registers using the unique Swedish personal identification number²¹ and forms a pseudonymized dataset. In the present study, data on positive SARS-CoV-2 polymerase chain reaction (PCR) test results were obtained from the National Register of Notifiable Diseases (SmiNet). Diagnoses of COVID-19, PCC, and comorbidities were retrieved from the National Patient Register (NPR), including both specialist inpatient care and outpatient visits, and from two regional databases of all public and most private primary healthcare (VAL and VEGA, in Region Stockholm and Region Västra Götaland, respectively) using the ICD-10 Swedish version (ICD-10-SE) diagnosis codes. Prescriptions of medication were retrieved from the National Prescribed Drug Register. Additional data on COVID-19-related intensive care unit (ICU) stays were obtained from the Swedish Intensive Care Register. Data on COVID-19 vaccination were retrieved from the National Vaccination Register (NVR), and data on death, emigration, demographic, and socioeconomic data from the Total Population Register and the Longitudinal Integrated Database for Health Insurance and Labour Market Studies (LISA). Information on kinship was retrieved from the Multigeneration Register and information on cohabitation was retrieved from the Apartment register. The study was approved by the Swedish Ethical Review Authority (Dnr: 2020-01800 with several amendments) who waived the requirement of informed consent. Data were obtained from the register holders after approved applications.

Study population, cohort definition, and follow-up

As PCC mainly is diagnosed in primary healthcare in Sweden²² and we have access to data from the primary healthcare databases VAL and VEGA, this study included all residents from these two regions (about 40% of the Swedish population) who were registered with their first COVID-19 (based on either a positive PCR test [83%] or ICD-10-SE diagnosis codes U07.1 or U07.2 as main or secondary diagnosis) during the extensive PCR testing period in Sweden (i.e., 1 August 2020 to 9 February 2022) and were ≥18 years of age at time of registered infection. All individuals with any symptom of ongoing COVID-19 were recommended to take a PCR test during this period, which was easy to access and free of charge. Date of registered COVID-19 was considered the index date, and every individual was followed with start from 28 days after index date until a valid PCC diagnosis (see Outcome section), death, emigration, move out from the two regions, or end of the study (30 November 2023), whichever came first. Individuals who died, emigrated, or moved out from the two regions within 28 days after index date were not at risk of PCC as captured in the study and were thus not included. Finally, 810 851 individuals were included in the study cohort (Fig. 1). In a sensitivity analysis, 90 days interval between index date and a valid PCC diagnosis was used instead of 28 days, resulting in 807 517 individuals in the study cohort.

**Fig. 1: Flowchart of the study population.**

Outcome – PCC

A valid clinical diagnosis for PCC was defined as ICD-10-SE code U09.9 in NPR, VEGA, or VAL as the main or secondary diagnosis ≥28 days after index date. A minimum of 28 days between index date and PCC diagnosis was required, as a PCC diagnosis within 28 days was interpreted as a likely misclassification relating to health effects of the acute infection rather than PCC. In the sensitivity analysis, we used a 90-day interval between index date and a valid PCC diagnosis³.

Evaluated risk factors

A predetermined list of potential risk factors of interest was selected by literature review and clinical knowledge and included factors reflecting personal sociodemographic data, health conditions, healthcare contact behaviors, COVID-19-related factors, area-level socioeconomic status (SES), as well as PCC in family and cohabitants (as proxies for genetic and shared environmental factors).

Personal sociodemographic factors included: age at first infection (categorized as 18-39, 40-79, and ≥80 years of age due to a non-linear association with the outcome [Supplementary Fig. 1]), sex (women, men), country of birth (Sweden, non-Sweden), parents’ country of birth (both from Sweden, one from Sweden, both non-Sweden, and unknown), education levels (primary [≤9 years], upper secondary [10-12 years], tertiary [≥13 years], and unknown), income levels (four quartile groups based on the distribution of disposable income per consumption unit in the study cohort), occupation (healthcare workers, other essential workers [including teachers, service sector workers, police and security services, postal and delivery workers, cleaners, and taxi, bus, and tram drivers], non-essential workers, and unemployed), marital status (married, not married), and region of residence (Region Stockholm, Region Västra Götaland). All personal sociodemographic factors, except age, were obtained before the pandemic, i.e., at the end of 2019.

Personal health conditions were defined based on specific comorbidities during 2015-2019 and medication use during 2018-2019. Comorbidities included cardiovascular diseases (heart failure, ischemic heart diseases, stroke, peripheral vascular disease, thromboembolic disease, arrythmias, and other cardiac diseases); respiratory diseases (COPD, asthma, and other respiratory diseases), metabolic syndrome components (hypertension, type 2 diabetes), chronic kidney disease, immune disorders and immune suppression, autoimmune diseases, fibromyalgia, and psychiatric disorders (dementia, bipolar disorder and schizophrenia, depression and anxiety, stress-related disorders, and other mental disorders). The ICD-10-SE and Anatomical Therapeutic Chemical Classification System (ATC) codes defined these conditions (Supplementary Table 1). The total number of these specific comorbidities for each individual observed in 2015-2019 was additionally calculated and grouped into three categories (0, 1-2, and ≥3). Number of healthcare (primary and specialist care) contacts in 2019 (categorized as 0, 2]) was used as a surrogate of personal healthcare seeking behaviors.

COVID-19-related factors included: variants of concern (VOC) period when particular virus variants were dominating as the first infection occurred (preAlpha, Alpha, Delta, and Omicron [Supplementary Table 2]), severity of first acute infection (non-hospitalized, hospitalization without ICU, and ICU admission), and any vaccination ≥14 days before first infection (no, yes).

Area-level SES was calculated as the proportion of inhabitants in each specific geographical area (Demographic Statistical Areas, DeSO)²³ with an annual income lower than the first quartile of the national average income level in 2019 and further categorized into low (>0.4; i.e. 40% of inhabitants with low annual income) or high (≤0.4) area-level SES according to a non-linear association with the outcome (Supplementary Fig. 3). Thus, study individuals living in the same specific geographical area will have the same area-level SES.

We also considered PCC diagnosis in family and cohabitants as potential risk factors, defined as: PCC cases at any time in the core family (yes, no or uncertain [family members not living in the included regions]), and PCC cases at any time in cohabitants (yes, no or uncertain). The core family was defined as biological parents and children.

Statistics and reproducibility

Descriptive analysis (count and proportion) for each risk factor is presented according to PCC status at the end of follow-up. Penalized smoothing splines was used to illustrate and determine the appropriate categories for age, number of healthcare contacts in 2019, and area-level SES, using R package “pspline”. Crude incidence rates of PCC (per 1000 person-years) are presented in each variable category. Cox proportional hazards models, with time-on-study as the underlying time scale, were used to estimate hazard ratio (HR) with 95% confidence interval (95%CI). First, separate crude univariable models were run for each risk factor, then a full model including all risk factors simultaneously was run. To identify the most important risk factors, backwards stepwise selection was applied, starting with the full model. Factors were removed if their significance level was p ≥ 0.1 and reintroduced if p ≤ 0.05. To access the robustness of the variable selection, LASSO regression using Bayesian Information Criterion (BIC) and adaptive methods was conducted, and the results were compared with those from the backwards stepwise selection. The backwards stepwise selection was further performed separately in subgroups for each stratum of VOC period and severity of the acute infection.

In the sensitivity analysis using a 90-day interval between index date and PCC diagnosis, a full Cox regression model with all risk factors was run and compared to the main analysis (with 28-day interval), and backwards stepwise selection was performed as described. For all Cox regression models, the proportional hazards (PH) assumption was checked using graphical diagnostics and it was not violated. All analyses were conducted with STATA 18 statistical software.