How to mitigate selection bias in COVID-19 surveys: evidence from five national cohorts (preprint)

Background: Non-response is a common problem, and even more so during the COVID-19 pandemic where social distancing measures challenged data collections. As non-response is often systematic, meaning that respondents are usually healthier and from a better socioeconomic background, this potentially introduces serious bias in research findings based on COVID-19 survey data. The goal of the current study was to see if we can reduce bias and restore sample representativeness despite systematic non-response in the COVID-19 surveys embedded within five UK cohort studies using the rich data available from previous time points. Methods: A series of three surveys was conducted during the pandemic across five UK cohorts: National Survey of Health and Development (NSHD, born 1946), 1958 National Child Development Study (NCDS), 1970 British Cohort Study (BCS70), Next Steps (born 1989-90) and Millennium Cohort Study (MCS, born 2000-02). We applied non-response weights and utilised multiple imputation, making use of covariates from previous waves which have been commonly identified as predictors of non-response, to attempt to reduce bias and restore sample representativeness. Results: Response rates in the COVID-19 surveys were lower compared to previous cohort waves, especially in the younger cohorts. We identified bias due to systematic non-response in the distributions of variables including parental social class and childhood cognitive ability. In each cohort, respondents of the COVID-19 survey had a higher percentage of parents in the most advantaged social class, and a higher mean of childhood cognitive ability, compared to the original (full) cohort sample. The application of non-response weights and multiple imputation was successful in reducing bias in parental social class and childhood cognitive ability, nearly eliminating it for the former. Conclusions: The current paper demonstrates that it is possible to reduce bias from non-response and to a large degree restore sample representativeness in multiple waves of a COVID-19 survey embedded within long running longitudinal cohort studies through application of non-response weights or multiple imputation. Such embedded COVID-19 surveys therefore have an advantage over cross-sectional COVID-19 surveys, where non-response bias cannot be handled by leveraging previously observed information on non-respondents. Our findings suggest that, if non-response is appropriately handled, analyses based on the COVID-19 surveys within these five cohorts can contribute significantly to COVID-19 research, including studying the medium and long-term effects of the pandemic.

Risk factors for long COVID: analyses of 10 longitudinal studies and electronic health records in the UK (preprint)

The impact of long COVID is increasingly recognised, but risk factors are poorly characterised. We analysed questionnaire data on symptom duration from 10 longitudinal study (LS) samples and electronic healthcare records (EHR) to investigate sociodemographic and health risk factors associated with long COVID, as part of the UK National Core Study for Longitudinal Health and Wellbeing. Methods Analysis was conducted on 6,899 adults self-reporting COVID-19 from 45,096 participants of the UK LS, and on 3,327 cases assigned a long COVID code in primary care EHR out of 1,199,812 adults diagnosed with acute COVID-19. In LS, we derived two outcomes: symptoms lasting 4+ weeks and symptoms lasting 12+ weeks. Associations of potential risk factors (age, sex, ethnicity, socioeconomic factors, smoking, general and mental health, overweight/obesity, diabetes, hypertension, hypercholesterolaemia, and asthma) with these two outcomes were assessed, using logistic regression, with meta-analyses of findings presented alongside equivalent results from EHR analyses. Results Functionally limiting long COVID for 12+ weeks affected between 1.2% (age 20), and 4.8% (age 63) of people reporting COVID-19 in LS. The proportion reporting symptoms overall for 12+ weeks ranged from 7.8 (mean age 28) to 17% (mean age 58) and for 4+ weeks 4.2% (age 20) to 33.1% (age 56). Age was associated with a linear increase in long COVID between age 20-70. Being female (LS: OR=1.49; 95%CI:1.24-1.79; EHR: OR=1.51 [1.41-1.61]), poor pre-pandemic mental health (LS: OR=1.46 [1.17-1.83]; EHR: OR=1.57 [1.47-1.68]) and poor general health (LS: OR=1.62 [1.25-2.09]; EHR: OR=1.26; [1.18-1.35]) were associated with higher risk of long COVID. Individuals with asthma also had higher risk (LS: OR=1.32 [1.07-1.62]; EHR: OR=1.56 [1.46-1.67]), as did those categorised as overweight or obese (LS: OR=1.25 [1.01-1.55]; EHR: OR=1.31 [1.21-1.42]) though associations for symptoms lasting 12+ weeks were less pronounced. Non-white ethnic minority groups had lower 4+ week symptom risk (LS: OR=0.32 [0.22-0.47]), a finding consistent in EHR. Associations were not observed for other risk factors. Few participants in the studies had been admitted to hospital (0.8-5.2%). Conclusions Long COVID is clearly distributed differentially according to several sociodemographic and pre-existing health factors. Establishing which of these risk factors are causal and predisposing is necessary to further inform strategies for preventing and treating long COVID.