Social determinants of recovery from ongoing symptoms following COVID-19 in two UK longitudinal studies: a prospective cohort study

Abstract

Introduction: Social gradients in COVID-19 exposure and severity have been observed internationally. Whether combinations of pre-existing social factors, particularly those that confer cumulative advantage and disadvantage, affect recovery from ongoing symptoms following COVID-19 and long COVID is less well understood.

Methods: We analysed data on self-perceived recovery following self-reported COVID-19 illness in two UK community-based cohorts, COVID Symptom Study Biobank (CSSB) (N=2548) and TwinsUK (N=1334). Causal effects of sociodemographic variables reflecting status prior to the COVID-19 pandemic on recovery were estimated with multivariable Poisson regression models, weighted for inverse probability of questionnaire participation and COVID-19 infection and adjusted for potential confounders. Associations between recovery and social strata comprising combinations of sex, education level and local area deprivation were estimated using the intersectional multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA) approach. Further analyses estimated associations with variables reflecting experiences during the pandemic.

Results: Gradients in recovery from COVID-19 along the lines of social advantage were observed in intersectional MAIHDA models, with predicted probability of recovery lowest in female strata with lowest education and highest deprivation levels (CSSB: 55.1% (95% CI 44.0% to 65.1%); TwinsUK: 73.9% (95% CI 61.1% to 83.0%)) and highest in male strata with highest education and lowest deprivation levels (CSSB: 79.1% (95% CI 71.8% to 85.1%); TwinsUK: 89.7% (95% CI 82.5% to 94.1%)). Associations were not explained by differences in prepandemic health. Adverse employment, financial, healthcare access and personal experiences during the pandemic were also negatively associated with recovery.

Conclusions: Inequalities in likelihood of recovery from COVID-19 were observed, with ongoing symptoms several months after coronavirus infection more likely for individuals with greater social disadvantage prior to the pandemic.

Keywords: COVID-19; SARS-CoV-2; Social Medicine; Sociodemographic Factors.

Figure 1. Directed acyclic graph describing hypothesised causal pathways. Proposed directed acyclic graph (DAG) used to generate minimal adjustment variable sets for models estimating the total causal effect of exposure variables on the outcome of self-perceived COVID-19 recovery. Data only available in CSSB or TwinsUK are coloured in blue and purple, respectively, while key unobserved potential confounders are coloured in gold. The DAG is structured approximately in order of data generation/crystallisation from left to right, and variables with similar time of data generation/crystallisation are grouped for clarity into ‘super nodes’. The proposed DAG is ‘saturated’, in that each variable is hypothesised to be caused by all earlier variables. BMI, body mass index; CSSB, COVID Symptom Study Biobank.

Figure 2. Association between prepandemic sociodemographics and recovery from COVID-19 in CSS Biobank and TwinsUK cohorts. Relative risk ratios (RRs) and 95% CIs from Poisson regression models testing association between recovery from COVID-19 and various prepandemic sociodemographic exposure variables, among individuals with self-reported COVID-19 infection. Unadjusted recovery rates and group sizes from unweighted samples are also shown. Results for each exposure variable originate from models with distinct adjustment variable sets as follows: age: (sex); sex: (age); ethnic group: (age, sex); education: (age, sex, ethnic group, first language (CSSB only)); region: (age, sex, ethnic group, first language (CSSB only), education); rural–urban classification (RUC): (age, sex, ethnic group, first language (CSSB only), education, region); prepandemic employment status: (age, sex, ethnic group, first language (CSSB only), education, region, RUC); local area deprivation: (age, sex, ethnic group, first language (CSSB only), education, region, RUC, prepandemic employment status). Models included participant weighting for inverse probability of questionnaire response and selection into the analysis sample. CSSB, COVID Symptom Study Biobank; IMD, Index of Multiple Deprivation; GCSE, General Certificate of Secondary Education; GNVQ, General National Vocational Qualification.

Figure 3. Predicted probability of COVID-19 recovery for social strata of sex, education level and local area deprivation from MAIHDA models in CSS Biobank and TwinsUK cohorts. Predicted probabilities and approximate 95% CIs from MAIHDA mixed-effects logistic regression models testing association between recovery from COVID-19 and social strata for: (a) all CSSB cohort participants with self-reported COVID-19, (b) CSSB participants with diagnosed or self-reported long COVID, (c) all TwinsUK participants with COVID-19. Fixed effects included in models were age group, ethnic group, sex, education level and local area deprivation. Strata labels: F=female, M=male; NoDegree=less than degree level education (including not stated/prefer not to say), Degree=undergraduate degree level or higher; IMDQ1-IMDQ5=Index of Multiple Deprivation Quintile 1–5, where 1 is most deprived 20% of areas, and 5 is least deprived 20%. Models included participant weighting for inverse probability of questionnaire response and selection into analysis samples. CSS Biobank, COVID Symptom Study Biobank; MAIHDA, multilevel analysis of individual heterogeneity and discriminatory accuracy.

Figure 4. Associations between sociodemographics collected during the COVID-19 pandemic and recovery from COVID-19 in CSS Biobank and TwinsUK cohorts. Relative risk ratio and 95% CIs from Poisson regression models testing association between various sociodemographics collected during the COVID-19 pandemic and recovery from COVID-19, among individuals with self-reported COVID-19 infection. Results for each exposure variable originate from models with distinct adjustment variable sets, including prepandemic sociodemographic factors, prepandemic health characteristics, COVID-19 acute illness factors and sociodemographic factors collected during the pandemic as potential confounding factors as appropriate according to the proposed directed acyclic graph. Models included participant weighting for inverse probability of questionnaire response and selection into the analysis sample. CSS Biobank, COVID Symptom Study Biobank.