The UK National Core Studies Longitudinal Health and Wellbeing initiative draws together data from multiple UK population-based longitudinal studies and analyses these data to answer priority pandemic-related questions. By conducting similar analyses within each study and pooling results in a meta-analysis, we can provide robust evidence to understand how the pandemic has impacted population health, and support efforts to mitigate effects going forward. Data were obtained from eight long-running UK population-based longitudinal studies, each of which had conducted surveys during the pandemic (which we refer to as COVID surveys). Details of the design, sample frames, current age range, timing of the most recent pre-pandemic and COVID surveys, response rates, and sample size are in Additional file 1: Table S1 [12,13,14,15,16,17,18,19,20,21,22,23,24,25].
Five studies were age homogenous birth cohorts (where all individuals within each study were of similar age): the Millennium Cohort Study (MCS), the index children from the Avon Longitudinal Study of Parents and Children (ALSPAC-G1), Next Steps (NS, formerly the Longitudinal Study of Young People in England), the 1970 British Cohort Study (BCS70), and the 1958 National Child Development Study (NCDS). Three age heterogeneous studies (each covering a range of age groups) were included: Understanding Society (USOC), the English Longitudinal Study of Ageing (ELSA), and Generation Scotland: the Scottish Family Health Study (GS). Finally, the parents of the ALSPAC-G1 cohort were treated as a fourth age heterogeneous study population (ALSPAC-G0).
Analytical samples were restricted to participants of working age, defined as those aged 16–66 years based on the current state pension age in the UK , and to those who had recorded at least one health behaviour outcome in a COVID-19 survey and had valid data on all covariates. Most studies were weighted to restore representativeness to their target populations, accounting for sampling design where appropriate, and differential non-response to pre-pandemic and COVID surveys . Weights were not available for GS. Details of the weighting applied within each study are in Additional file 1: Table S1.
In this section, we describe all variables in the analysis. Full details of the questions and coding used for each cohort are in Additional file 2.
Exposure: employment status change
Employment status change (or stability) was coded in six categories based on the status both prior to the pandemic and at their first COVID-19 survey: stable employed (reference group), furloughed (i.e. from work to furlough), no longer employed (i.e. from employed to non-employed), became employed (i.e. from non-employed to employed), stable unemployed (i.e. unemployed at both points), and stable non-employed (i.e. not available for employment at either point, including in education, early retirement, caring responsibilities, sick, or disabled).
Outcomes: health behaviours
We examined diet, physical activity, and sleep. Participants self-reported fruit and vegetable consumption (≤2 portions per day vs more portions ), time spent exercising (<3 days a week for 30 min or more vs more frequent exercise within recommended levels ), and hours of sleep (outside the typical range of 6–9 h vs within that range ) both during and pre-pandemic. However, this information, used for our main analyses, was only available in some studies (MCS, NS, BCS, NCDS, USOC), whereas others (ALSPAC, GS, ELSA) only had information on change since the start of the pandemic (see Additional file 2). Based on these levels or on the information on changes in health behaviours since the start of the pandemic, we additionally created dichotomous outcomes indicating change from before to during the pandemic (in comparison to no change or change in the other direction): more portions of fruit/vegetables, fewer portions of fruit/vegetables, more time spent exercising, less time spent exercising, more hours of sleep, fewer hours of sleep, a shift from outside to within the typical sleep range of 6–9 h, and a shift from within to outside the typical sleep range of 6–9 h. All information on behaviours during the pandemic was from surveys conducted between April and July 2020 (inclusive).
Confounders and moderators
Potential confounders included sex, ethnicity (non-white ethnic minority vs white, including white ethnic minorities), age, education (degree vs no degree), UK nation (i.e. England, Wales, Scotland, Northern Ireland or other), household composition (based on presence of a spouse/partner and presence of children), pre-pandemic psychological distress, pre-pandemic self-rated health (excellent-good vs fair-poor), and pre-pandemic health behaviour measures, where available.
We examined modification of the associations by sex, education (degree vs no degree holders), and age in three categories: 16–29, 30–49, and 50 years or more (with age-homogeneous cohorts included in the relevant band).
Within each study, each outcome was regressed on employment status change, using a modified Poisson model with robust standard errors that returns risk ratios for ease of interpretation and avoids issues related to non-collapsibility of odds ratios [31, 32]. After estimating unadjusted associations, confounder adjustment was performed in two steps. First, a “basic” adjustment including socio-demographic characteristics: age (only in age heterogeneous studies), sex, ethnicity (except the BCS70 and NCDS cohorts who were almost entirely white), education, UK nation (except ALSPAC, GS, and ELSA which only had participants from a single country), and household composition. Second, a “full” adjustment additionally including pre-pandemic measures of psychological distress, self-rated health, and health behaviours. Moderation by sex, age, and education was assessed with stratified regressions using “full” adjustment.
Both stages of adjustment are relevant because our exposure, employment change, incorporates pre-pandemic employment status, which may have influenced other pre-pandemic characteristics such as mental health, self-rated health, and health behaviours (see Additional file 1: Fig. S8). By not controlling for these pre-pandemic characteristics, the basic adjusted risk ratios may represent both newly acquired behaviour and/or continuation of established (pre-pandemic) behaviour. In contrast, the full adjustment risk ratios block effects via these pre-pandemic characteristics and can therefore be interpreted as representing the differential change in health behaviour between exposure groups which is independent of these pre-pandemic characteristics. For the outcomes that directly capture changes in health behaviour, the full adjustment did not include pre-pandemic levels of the behaviour in question, as pre-pandemic levels of that behaviour are incorporated within the change outcome. This means that even full adjustment risk ratios estimated for these outcomes may partially reflect associations with pre-pandemic behaviour.
The overall and stratified results from each study were pooled using a random effects meta-analysis with restricted maximum likelihood in Stata. To explore the role of potential moderators, a subgroup analysis was performed which meta-analysed findings separately within each category of each moderator and performed a test of between-group differences. For stratified results, a test of group differences was performed using the subgroup meta-analysis command. Some studies could not contribute estimates for every comparison due to differences in the ages sampled, measures used, and sparsity of data. For a small number of exposure-outcome comparisons, reliable estimates could not be computed because the outcome prevalence was low (≤2). While such selective exclusion could potentially lead to bias, the low numbers of events mean that the corresponding within-study estimates would be so imprecise that their exclusion is unlikely to lead to considerable bias (see Additional files 3 and 4 for more details and sensitivity analyses showing that results were robust to different low cell count exclusion thresholds). We report heterogeneity using the I
2 statistic: 0% indicates estimates were similar across studies, while values closer to 100% represent greater heterogeneity. While we could have undertaken a multivariate meta-analysis of all exposure categories simultaneously, for ease of interpretation, we instead conducted a series of univariate meta-analyses, bearing in mind the consistency of results from these approaches generally observed elsewhere [33, 34]. We performed a multivariate meta-analysis with one outcome in a subset of the studies as a sensitivity analysis, and differences from the individual univariate meta-analyses were negligible (results not shown).