Changes in the behavioural determinants of health during the coronavirus (COVID-19) pandemic: gender, socioeconomic and ethnic inequalities in 5 British cohort studies

Background: The coronavirus (COVID-19) pandemic and consequent physical distancing measures are expected to have far-reaching consequences on population health, particularly in already disadvantaged groups. These consequences include changes in health impacting behaviours (such as exercise, sleep, diet and alcohol use) which are arguably important drivers of health inequalities. We sought to add to the rapidly developing empirical evidence base investigating the impacts of the pandemic on such behavioural outcomes. Methods: Using data from five nationally representative British cohort studies (born 2001, 1990, 1970, 1958, and 1946), we investigated sleep, physical activity (exercise), diet, and alcohol intake (N=14,297). Using measures of each behaviour reported before and during lockdown, we investigated change in each behaviour, and whether such changes differed by age/cohort, gender, ethnicity, and socioeconomic position (SEP; childhood social class, education attainment, and adult reporting of financial difficulties). Binary or ordered logistic regression models were used, adjusting for prior measures of each health behaviour and accounting for study design and non-response weights. Meta-analyses were used to pool cohort-specific estimates and formally test for heterogeneity across cohorts. Results: Changes in these outcomes occurred in in both directions ie, shifts from the middle part of the distribution to both declines and increases in sleep, exercise, and alcohol use. For all outcomes, older cohorts were less likely to report changes in behaviours compared with younger cohorts. In the youngest cohort (born 2001), the following shifts were more evident: increases in exercise, fruit and vegetable intake, sleep, and less frequent alcohol consumption. After adjustment for prior behaviour levels, during lockdown females were less likely to sleep within the typical range (6-9 hours) yet exercised more frequently; lower SEP was associated with lower odds of sleeping within the typical range (6-9 hours), lower exercise participation, and lower consumption of fruit and vegetables; and ethnic minorities were less likely than White participants to sleep within the typical range (6-9 hours), exercise less frequently, yet reported less frequent alcohol consumption. Conclusions: Our findings highlight the multiple changes to behavioural outcomes that may have occurred due to COVID-19 lockdown, and the differential impacts across generation, gender, SEP and ethnicity. Such changes require further monitoring given their possible implications to population health and the widening of health inequalities.


Introduction
The coronavirus  pandemic is expected to have far-reaching consequences on population health, particularly in already disadvantaged groups. 1 2 Aside from direct effects of COVID-19 infection, detrimental changes may include effects on physical and mental health due to associated changes to health-impacting behaviours. Change in such behaviours may be anticipated due to the effects social distancing, both mandatory and voluntary, and change in risk factors which may affect such behaviours-such as employment, financial circumstances, and mental distress. 3 4 The behaviours include physical activity, diet, alcohol, and sleep 5 -likely key contributors to existing health inequalities 6 and indirectly implicated in inequalities arising due to COVID-19 given their link with risk factors such as obesity and diabetes. 7 While empirical evidence of the impact of COVID-19 on such behaviours is emerging, [8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26] it is currently challenging to interpret for multiple reasons. First, generalising from one study location and/or period of data collection to another is challenging due to the vastly different societal responses to COVID-19 which could plausibly impact on such behaviours, such as restrictions to movement, access to restaurants/pubs, and access to support services to reduce substance use. This is compounded by many studies investigating only one health behaviour in isolation. Further, assessment of change in any given outcome is notoriously methodologically challenging. 27 Some studies have questionnaire instruments which appear to focus only on the negative consequences of COVID-19, 8 thus curtailing an assessment of both the positive and negative possible effects on health behaviours.
The consequences of COVID-19 lockdown on behavioural outcomes may differ by factors such as age, gender, socioeconomic position (SEP), and ethnicity-thus potentially widening already existing health inequalities. For instance, younger generations (e.g. age 18-30 years) are particularly affected by cessation or disruption of education, loss of employment and income; 3 and were already more likely than older persons to be in secure housing, secure employment, or stable partnerships. 28 In contrast, older generations are more appear more susceptible to severe consequences of COVID-19 infection, and have in many countries were recommended to 'shield' to prevent such infection. Within each generation, the pandemic's effects are may have had inequitable effects by gender (e.g. childcare responsibilities being borne more by women), SEP and ethnicity (e.g. more likely to be in at-risk and low paid unemployment, insecure and crowded housing).
Using data from five nationally representative British cohort studies, which each used an identical COVID-19 followup questionnaire in May 2020, we investigated change in multiple health-impacting behaviours. Multiple outcomes were used since each is likely to have independent impacts on population health, and evidence-based policy decisions are likely better informed by simultaneous consideration of multiple outcomes. 29 We considered multiple wellestablished health equity stratifiers: 30 age/cohort, gender, socioeconomic position (SEP) and ethnicity. Further, since childhood SEP may impact on adult behaviours and health outcomes independently of adult SEP, 31 we utilised prospective data in these cohorts to investigate childhood and adult SEP.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2020. . https://doi.org/10.1101/2020.07.29.20164244 doi: medRxiv preprint

Study samples
We used data from four British birth cohort (c) studies, born in 1946, 32 1958, 33 1970, 34 and 2000-2002 (2001c); 35 and one English longitudinal cohort study (born 1989-90; 1990c) followed up from 13 years. 36 Each has been followed up at regular intervals from birth or adolescence; data have been collected from home interviews on heath, behavioural, and socioeconomic factors. In each study, participants gave written consent to be interviewed. Research ethics approval was obtained from relevant committees. In May 2020, during the COVID-19 pandemic, participants were invited to take part in an online questionnaire which measured demographic factors, health measures and multiple behaviours.

Outcomes
We investigated the following behaviours: sleep (number of hours each night on average; given its non-linear association with health outcomes we coded it to 0=6-9 hours and 1=≤5 hours or ≥10 hours), exercise (number of days from 0-7 per week the participants exercised for 30 mins or more; at moderate-vigorous intensity-"working hard enough to raise your heart rate and break into a sweat"), diet (number of portions of fruit & vegetable per day (from 0 to ≥6; portion guidance was provided), alcohol consumption (never to 4 or more times per week). For each, participants retrospectively reported levels in "the month before the coronavirus outbreak" and then during the fieldwork period (May 2020). Herein, we refer to these reference periods as pre and during lockdown, respectively.

Exposures
Cohort member gender was coded 0=female and 1=male for all analysis. Socioeconomic position was indicated by childhood social class (at 10-14 years old), using the Registrar General's Social Class scale-I (professional), II (managerial and technical), IIIN (skilled non-manual), IIIM (skilled manual), IV (partly-skilled), and V (unskilled) occupations. Highest educational attainment was also used, categorised into four groups as follows: degree/higher, A levels/diploma, O Levels/GCSEs, or none (for 2001 MCS we used parents' highest education as many were still undertaking education). Financial difficulties were based on whether individuals (or their parents for 2001 MCS) reported (prior to COVID-19) as managing financially comfortably, all right, just about getting by, and difficult. These were converted into cohort-specific ridit scores to aid interpretation-resulting in relative or slope indices of inequality when used in regression models. 37 Ethnicity was recorded as White and non-White-ethnicity analyses are limited to the 1990c and 2001c only owing to a lack of ethnic diversity in older cohorts.

Statistical analyses
We first conducted descriptive analyses for each outcome-calculating average levels and distributions of each outcome pre and during covid-19. We then used separate regression models to examine how gender, ethnicity and SEP were related to lockdown levels of each behaviour; models examining ethnicity and SEP were gender-adjusted. We also adjusted for reported behavioural outcomes prior to lockdown to estimate differences in outcomes attributable to COVID-19 lockdown-without this adjustment, associations were largely of the same direction but with larger effect sizes (data available upon request). To inform understanding of change in these outcomes, we also tabulated calculated change in each outcome (decline, no change, and increase) by each cohort and risk factor group. We then used logistic (sleep) or ordinal (other outcomes) regression models. To aid interpretation of possible deviation of the proportional odds assumption, we additionally used multinomial regression models with each outcome grouped into . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2020. . https://doi.org/10.1101/2020.07.29.20164244 doi: medRxiv preprint low, middle, and high levels. We conducted cohort-specific analyses and conducted meta-analyses to assess pooled associations and formally test for heterogeneity across cohort (I 2 statistic).
To account for possible bias due to missing data, we weighted our analysis using weights constructed from logistic regression models-the outcome was response during the COVID-19 survey, and predictors were demographic, socioeconomic, household, and individual-based predictors of non-response at earlier sweeps, based on previous work in these cohorts. 38 39 We also used weights to account for the stratified survey designs of the 1946c, 1990c, and 2001c.
Stata v15 (STATA corp) was used in all analyses.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Results
Cohort-specific responses were as follows: 1946c: 1241 of 1842; 1958c: 5205 of 8943, 1970c: 4247 of 10458; 1990c: 1921 of 9380; 2001c: 2677 of 9909. The following factors, measured in prior data collection periods, were associated with increased likelihood of response in this COVID-19 dataset: being female, higher education attainment, higher household income, and more favourable self-rated health. Valid outcome data was available in both pre and during measures for the following: Sleep, N=14,171; exercise, N=13,997; alcohol, N=14,297; fruit/vegetables, N=13,623.
Outcomes pre and during COVID-19 were each moderately-highly positively correlated-Spearman's R as follows: sleep=0.55, exercise=0.58, alcohol=0.82, fruit/vegetable consumption=0.81. For all outcomes, older cohorts were less likely to report change in behaviour compared with younger cohorts (Supplementary Table 1).

Sleep
The average (mean) amount of sleep (hours per night) were either similar or slightly higher during compared with before lockdown. In each cohort, the variance was higher during (Table 1) Table 1).
In logistic regression models, the following were associated with either low or high sleep levels during lockdown (in models adjusting for prior levels): female gender, lower SEP (lower childhood class, education attainment, and financial difficulties), and ethnic minority status ( Figure 2). Females and ethnic minorities reported more change in sleep levels during lockdown than males or White participants (Supplementary Table 1). Analyses using multinomial logistic regression suggested that in 2001-1970c males had lower risks of both lower and higher levels of sleep (compared with 6-9 hours), and that lower education and ethnic minority status were associated with higher risk of lower sleep levels during lockdown. However, confidence intervals around these estimates were wide (Supplementary Table 2).

Exercise
Mean exercise frequency levels were similar during and before lockdown ( Table 1). As with sleep levels, the variance was higher during, reflecting both reduced and increased amounts of exercise during lockdown ( Figure 1). The following were associated with lower exercise levels during lockdown (in models adjusting for prior levels): male gender (in 2001-1970c only; I 2 =60.4%), lower SEP (all cohorts; I 2 for education=2.0%), and ethnic minority status (I 2 =0%; Figure 2). These findings were largely similar in both ordered and multinomial logistic regression (Supplementary Table 2), and reflected a greater proportion of female, higher SEP, and White participants reporting increases in exercise levels during lockdown (Supplementary Table 1).

Alcohol intake
In 2001c, a larger fraction of participants reported transitions to never drinking during lockdown than in older cohorts (Table 1, Figure 1, and Supplementary Table 1). The 2001c also reported a greater transition to lower numbers of drinks consumed in a given day of drinking (data available upon request). In older cohorts, more participants reported never and the highest frequencies of alcohol (≥4 times/week) in lockdown compared with prior (Table 1 and Figure 1).
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted July 31, 2020. . https://doi.org/10.1101/2020.07.29.20164244 doi: medRxiv preprint The following were associated with lower frequency of alcohol consumption during lockdown (in models adjusting for prior levels): female gender (in 2001c only; I 2 =78.6%) and ethnic minority status (I 2 =0%; Figure 2). This reflected consistent lower levels of drinking in ethnic minority groups, and increased consumption in White participants (Supplementary Table 1). Associations between SEP and alcohol consumption differed by SEP indicator and cohort (I 2 for education=67.3%); lower education attainment was associated with more frequent drinking in 2001c yet less frequent drinking in 1958c and 1970c ( Figure 2). These findings were largely similar in both ordered and multinomial logistic regression (Supplementary Table 2).

Fruit and vegetable intake
Fruit and vegetable intake was broadly similar pre and during lockdown, although increases in consumption were most frequent in 2001c compared with older cohorts (Supplementary Table 1). The following were associated with lower fruit and vegetable consumption (in models adjusting for prior levels): male gender (in 2001c only; I 2 =75.8%), and lower SEP (more consistently in younger cohorts for childhood social class and education; I 2 for education=52.9%); there was no strong evidence for association with ethnic minority status ( Figure 2 and Supplementary Table 1). These findings were largely similar in both ordered and multinomial logistic regression (Supplementary Table 2).
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Main findings
Using data from 5 national British cohort studies, we estimated change in multiple health behaviours pre and during COVID-19 lockdown in the UK (May 2020). Where change in these outcomes was identified, it occurred in both directions-ie, shifts from the middle part of the distribution to both declines and increases in sleep, exercise, and alcohol use. In the youngest cohort (2001c), the following shifts were more evident: increases in exercise, fruit and vegetable intake, sleep, yet reduced alcohol consumption frequency.
Across all outcomes, older cohorts were less likely to report changes in behaviour. After adjustment for prior behaviour levels, during lockdown females were less likely to sleep within the typical/recommended range (6-9 hours) yet exercised more frequently. Lower SEP was associated with lower odds of sleeping within the typical range (6-9 hours), lower exercise participation, and lower consumption of fruit and vegetables. Ethnic minorities were less likely during lockdown to sleep within the typical range (6-9 hours), exercise less frequently, yet had lower regular alcohol consumption. Our findings suggest-for most outcomes measured-to a potential widening of inequalities in healthimpacting behavioural outcomes which may have been caused by the COVID-19 lockdown.
Comparison with other studies In our study the youngest cohort reported increases in sleep during lockdown-similar findings of increased sleep have been reported in many, 13 17 18 24 but not all 8 previous studies. Both too much and too little sleep may reflect, and be predictive of, worse mental and physical health. 40 41 In this sense, the increasing dispersion in sleep we observed may reflect the negative consequences of COVID-19 and lockdown. Females, those of lower SEP, and ethnic minorities were all at higher risk of both lower and higher sleep levels. It is possible that lockdown restrictions and subsequent increases in stress-related to health, job, and family concerns-have affected sleep across multiple generations and potentially exacerbated such inequalities. Indeed, recent work using household panel data in the UK has observed marked increases in anxiety and depression in the UK during lockdown that were largest amongst younger adults. 4 Our finding on exercise add to existing somewhat mixed state of the evidence. Some studies have reported declines in both self-reported 12 23 and accelerometery-assessed physical activity, 19 yet this is in contrast to others which report an increase, 22 and there is corroborating evidence for increases in some forms of physical activity since online searches for exercise and physical activity appear to have increased. 21 As in our study, another also reported that males had lower exercise levels during lockdown. 20 While we cannot be certain that our findings reflect all changes to physical activity levels-lower intensities exercises were not assessed nor activity in other domains such as in work or travelthe inequalities observed with respect to gender, SEP, and ethnicity may be cause of public health concern.
As for the impact of the lockdown on alcohol consumption, concern was initially raised over the observed rises in alcohol sales in stores at the beginning of the pandemic in the UK 42 and elsewhere. As with other outcomes our finding suggest movement in actual consumption (frequency) in both directions, with the younger cohort tending to decrease consumption. Existing studies appear largely mixed, suggesting increases consumption, 9 16 26 with others reporting decreases; 11 12 23 25 others also report increases, yet use instruments which appear to particularly focus on . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted July 31, 2020. . https://doi.org/10.1101/2020.07.29.20164244 doi: medRxiv preprint capturing increases and not declines. 8 10 Different methodological approaches and measures used may account for inconsistent finding across studies, along with differences in the country of origin and characteristics of the sample.
The closing of pubs and bars and associated reductions in social drinking likely underlies our finding of declines in consumption amongst the youngest cohort. Increases in fruit and vegetable consumption observed in this cohort may have also reflected the considerable social changes attributable to lockdown, including more regular food consumption at home. However, in our study only positive aspects of diet (fruit and veg consumption) were captured-we did not capture information on volume of food, snacking and consumption of unhealthy foods. Indeed, one study reported simultaneous increases in consumption of fruit and vegetables and snacks. 11 Further research using additional waves of data collection is required to empirically investigate if the changes and inequalities observed in the current study persist in future. If the changes persist and/or widen, given the relevance of these behaviours to a range of health outcomes including chronic conditions, COVID-19 infection consequences and years of healthy life lost, the public health implications of these changes may be long-lasting. vegetable intake is only one component of diet. As in other studies investigating changes in such outcomes, we are unable to separate out change attributable to COVID-19 lockdown from other causes-these may include seasonal differences (eg, lower physical activity levels in the pre-COVID-19 winter months), and other unobserved factors which we were unable to account for. If these factors affected the sub-groups we analysed equally (gender, SEP, ethnicity), our analysis of risk factors of change would however not be biased due to this. As in other web surveys, 4 response rates were generally low-while the longitudinal nature of the cohorts enable predictors of missingness to be accounted for (via sample weights), 38 39 we can't fully exclude the possibility of unobserved predictors of missing data influencing our results. Finally, we investigated ethnicity using a binary categorisation to ensure sufficient sample sizes for comparisons-we were likely underpowered to investigate differences across the multiple diverse ethnic groups which exist. This warrants future investigation given the substantial heterogeneity within these groups and likely differences in behavioural outcomes.

Conclusion
Our findings highlight the multiple changes to behavioural outcomes that may have occurred due to COVID-19 lockdown, and the differential impacts-across generation, gender, socioeconomic disadvantage (in early and adult life) and ethnicity. Such changes require further monitoring given their possible implications to population health and the widening of health inequalities.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted July 31, 2020. Data availability 2001c, 1990c, 1970c and 1958c data are available from the UK Data Archive: https://www.data-archive.ac.uk. 1946c data are available from: https://www.nshd.mrc.ac.uk/data . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted July 31, 2020. .