Impact of COVID-19 on health behaviours and body weight: A prospective observational study in a cohort of 1.1 million UK and US individuals

1 Evidence regarding the impact of COVID-19 on health behaviours is limited. In this 2 prospective study including 1.1 million UK and US participants we collected diet and 3 lifestyle data ‘pre- ’and ‘peri- ’pandemic, and computed a bi-directional health 4 behaviour disruption index. We show that disruption was higher in the younger, 5 female and socioeconomically deprived (p<0.001). A loss in body weight (-0.45kg) 6 was greater in highly disrupted individuals compared to those with low disruption (- 7 0.03kg). There were large inter-individual changes observed in all 46 health and diet 8 behaviours measured peri-pandemic versus pre-pandemic, but no mean change in 9 the total population. Individuals most adherent to unhealthy pre-pandemic health 10 behaviours improved their diet quality (0.93units) and weight (-0.79kg) compared 11 with those reporting healthy pre-pandemic behaviours (0.08units and -0.04kg 12 respectively), irrespective of relative deprivation. For a proportion of the population, 13 the pandemic may have provided an impetus to improve health behaviours.


Introduction: 25
Mandatory public health initiatives to control and limit COVID-19 disease spread 26 have led to dramatic changes in day-to-day routines, resulting in increased social 27 isolation 1,2 , employment and financial insecurities 3 , and an altered food 28 environment 4 . This is placing most of the world's population in a unique global 29 experiment on a massive scale. Small European-based studies have observed 30 exacerbation of unfavourable diet and lifestyle behaviours attributable to these 31 changes such as increased sedentary behaviour, more snacking, less fresh food 32 consumption and weight gain, although the published data are inconclusive [5][6][7] . 33 34 At an individual level, significant life events are associated with changes in health 35 behaviours such as alcohol intake 8 , sleep 9,10 , diet 11,12 and physical activity 13 . The 36 complex interrelation of these health behaviours potentially mediates increases in 37 body weight observed during adulthood [14][15][16] , impacting the number of adults living 38 with excess weight and consequent morbidities [17][18][19][20] , which is a significant public 39 health threat 21 . Therefore, understanding how health behaviours change in the 40 context of a pro-longed pandemic is critical to understanding its long-term 41 consequences, and to inform short-and long-term strategies to prevent excess 42 weight gain. 43

44
In a prospective observational cohort study of 1.1 million participants from the ZOE 45 COVID Symptom Study (ZCSS) we: i) describe the self-reported impact of the 46 COVID-19 pandemic on diet and health behaviours using a composite disruption 47 4 8 lifestyle disruption attributable to the pandemic, we computed a novel disruption 116 index (DI) including 5 metrics; Diet Quality Score 25 , alcohol frequency, physical 117 activity, snacking frequency (food consumed outside of main meals) and weekday 118 sleep duration (see Methods). These domains were selected to capture the primary 119 diet and lifestyle behaviours associated with multiple health outcomes including 120 obesity [26][27][28][29][30][31] . The DI was independent of direction of change, ranged from 0 (no 121 disruption) to 5 (change in all five health behaviour domains) and approximated a 122 normal distribution (Figure 1a). In the UK discovery cohort, most participants 123 experienced a moderate level of disruption (65.5%; DI ≥ 2), while 15% had a high 124 level of disruption (DI ≥ 4). The DI (Figure 1a)

A greater disruption is associated with a larger bi-directional change in weight: 133
The mean overall change in body weight was small in the UK discovery cohort; 134 mean change (10 th , 90 th percentile) was -0.1 (-4.1, 3.6) kg. However, weight change 135 was highly variable among individuals, with 32% of participants losing a mean body 136 weight of -4.0kg (10 th centile; -8.2, 90 th centile; -0.9) and 34% gaining a mean body 137 weight of 3.5kg (10 th centile; 0.9, 90 th centile; 6.3) during the pandemic 138 (Supplemental Table 4). A similar pattern and magnitude of change was also 139 observed for the US replication cohort (Supplemental Table 4). Owing to the 140 9 marked bi-directional weight change in the study population, we separately 141 examined associations according to weight change, loss and gain. 142 143 When we analysed the association between the DI and body weight change, we 144 observed a more pronounced and variable weight change (both loss and gain) 145 among individuals with a high DI (CV; UK discovery cohort; 1355%, US; 1750%) 146 compared with those with a low DI (CV; UK discovery cohort; 1142%, US; 1531%); 147  Table 3). 152

153
In the UK, we showed that the DI was associated with weight changes after adjusting 154 for potential confounders (age and sex) in which the association was similar but of 155 greater magnitude among individuals living in areas with low deprivation (IMD; 8-10, 156 ß -0.041, 95%CI: -0.194, -0.137) compared with high deprivation (IMD; 1-3, ß -0.024, 157 95%CI: 0.-13, -0.057). When stratified according to DI groups and deprivation index, 158 there was a similar magnitude of weight loss between levels of deprivation, but a 159 moderately higher weight gain in the UK group residing in more deprived areas 160 (4.4kg) versus the group in less deprived areas (3.8kg) within the highly disrupted 161 group (Supplemental Table 3). This observation suggests that community-level 162 deprivation factors may not impair the potential positive behavioural effects of 163 disruption, but may also exacerbate the negative effects to a small extent, with 164 individuals living in more deprived areas being more susceptible to weight gain.       adherence (all p<0.001) ( Table 3). The reverse pattern was observed for those most 284 and least adherent to the 'unhealthy 'pattern. A broadly similar pattern was observed 285 in the US replication cohort (Supplementary Table 9). 286 287 We next examined the association between pre-pandemic diet and lifestyle 288 behaviours with the peri-pandemic change in health behaviours. We found that 289 individuals most adherent to the 'unhealthy 'pre-pandemic lifestyle pattern lost more 290 weight than those reporting higher adherence to the 'healthy 'lifestyle pattern (by 291 1.1kg) ( Table 3 These findings suggest that participants with unhealthy behaviours in the pre-296 pandemic phase were more likely to implement healthy changes. This may be a 297 consequence of a greater scope for improvement in the 'unhealthy 'group, whilst 298 those classified as having a healthy pre-pandemic lifestyle pattern, tended to retain 299 their 'beneficial 'health behaviours and experienced minimal change. This 300 observation was also reflected in the lower DI in those highly adherent to the 301 'healthy 'pattern and the higher DI in those adherent to the 'unhealthy 'pattern (both 302 p<0.001). A similar pattern of response was observed in the US replication cohort, 303 but with a greater magnitude of favourable change in those most adherent to the 304 20 unhealthy lifestyle, possibly due to minimal differences in data collection (see 305

Methods). 306 307
To assess if the population demographics were driving this finding, in the UK, we 308 stratified individuals most adherent to the 'unhealthy 'pattern according to their 309 community-level deprivation, age, sex and geographical location. When stratified Most surprisingly, those participants who were identified to have an 'unhealthy ' 358 pattern pre-pandemic were more likely to experience improvements in diet quality 359 and greater weight loss peri-pandemic, irrespective of deprivation status. Whilst this 360 may reflect some bias due to the users of the app being self-selected and typically of 361 a higher socio-economic demographic than the UK and US average citizen, the 362 NutriNet-Sante study also observed that positive health behaviour changes were 363 associated with less healthy pre-pandemic behaviours 36 . These findings suggest that 364 the pandemic may not have had the detrimental impact on diet and lifestyle 365 behaviours for a significant proportion of the population, as had previously been 366 speculated and reported in the media 37,38 . 367 Excess body weight has been linked with an increase of COVID-19 severity 39,40 and 368 chronic disease mortality 41 , and is known to be interrelated with social determinants 369 of health, including systematic racism 42 , and disparities in food security and socio-370 economic status 43,44 . Bi-directional weight changes were observed between groups 371 stratified by levels of deprivation, with those reporting unhealthy behaviours residing 372 in more deprived areas experiencing more weight gain than for those residing in less 373 deprived areas. However, irrespective of deprivation, physical activity and diet quality 374 were identified as key determinants of both short-term weight gain and loss. Mean Whilst this study had multiple strengths including sample size, an independent 383 replication cohort, and longitudinal data, we note several limitations. Firstly, the self-384 reported nature of the data collected with the potential of recall bias with 385 retrospective data collection. Data collection methods were modified part way 386 through the study to minimise participant burden and decrease attrition (see 387

Methods). Although we observed no significant difference in baseline characteristics 388
of individuals who had their information collected prior to and after the change, a 389 sensitivity analysis displayed minimal population characteristic differences. 390 Therefore, the UK sample population was divided into a discovery and validation 391 cohort to reduce possible error. 392 Other limitations include the absence of data on change in smoking behaviour. 393 Additionally, the DI is a crude index designed to determine positive or negative 394 changes in each health behaviour domain, further work is needed to determine how 395 direction of change in the individual components may contribute to health outcomes. 396 Further, only 5 domains were included in the DI and 6 in the SEM, and therefore 397 other health behaviours associated with the pandemic disruption and subsequent 398 weight change are not accounted for. For example, we did not include level of 399 isolation, mental health or comorbidities due to time restraints on participant 400 25 questionnaire completion. Other considerations such as job role and furlough status 401 may also be relevant. This should be a consideration for future studies to assist in 402 interpretation of lifestyle behaviour change in response to  Regression to the mean can also be problematic in studies that focus on assessing 404 change in behaviours in subgroups of the population, such as ours, where only a 405 single follow-up assessment is made. This is because the assessment of health 406 behaviours is prone to error, and error tends to be greatest at the extremes of the 407 distribution of the variable being assessed, such as the most and least healthy 408 participants. Thus, to some extent, the improvements in health behaviours seen in 409 people recording the most unhealthy behaviours at baseline could be artificial 410 improvements. Nevertheless, the same degree of change was not observed in 411 people who reported the most healthy baseline behaviours, which reinforces the 412 argument that changes in health seen in either group are likely to be real. week period), and via one app flow in the US (over a 9 week period). In the UK, the 504 questionnaire was released to 1% on 31/07/2020, 2% 04/08/2020, and 100% of 505 users on 06/08/2020. In this first app flow 425,657 participants completed the 506 questionnaire. After reviewing feedback from UK users, the app flow was altered to 507 limit participant burden and to reduce attrition. On 07/08/2020 a new question was 508 added after the participants had completed the first 'peri-pandemic 'section asking 509  Table 7). Participants were 525 divided into cohorts relative to the flow of the app (Figure 5), and by country. The 526 discovery cohort (n=425,657) from the first flow, the UK validation cohort 527 (n=422,285) from the second flow ('Has your diet changed'), and the US replication 528 cohort (n=14,366). A sensitivity analysis was performed to assess significant 529 population differences between the discovery and validation cohort. Researchers 530 defined 'change 'and ' no change 'groups within both cohorts relative to a >1 point 531 multi-directional change in diet quality score (no change; ≤1 and ≥ -1). Demographic 532 characteristics were compared between the 'change 'and 'no change 'groups 533 Table 2). The interaction was plotted between the primary outcome 534 (BMI) and the primary exposure (DQS). To minimise bias, all analyses in this study 535 were performed using the discovery cohort. Disruption index as well as diet and 536 lifestyle pattern analyses were replicated in the UK validation and the US replication and alcohol intake frequency were selected based on commonality in previous 548 published healthy lifestyle scores 48,49 . A bi-directional 'change 'was determined as 549 one point for any change (either +/-) for each variable, with maximum disruption 550 score as 5, and the minimum as 0 (Supplementary Table 3). 551 552 Pre and peri-pandemic comparisons: In the discovery cohort, and validated in the 553 UK validation cohort, the mean, SD, and 10th and 90th percentiles for 'pre-554 pandemic 'and 'peri-pandemic', percentage change (between the two time-points), 555 and the number of participants who increased/decreased was applied to the 556 continuous and categorical variables (Supplementary Tables 4, 5, and 6). For 557 32 continuous variables, the quantity of increase/decrease was also described. Change 558 in body weight, BMI, disruption index, physical activity, alcohol intake (units), 559 snacking, fruit and vegetable intake (combined) and diet quality was visualised 560 across UK geographical regions using python package geopandas v. 0.7.0. IMD was 561 estimated according to small geographical location, or neighbourhood, ranking areas 562 according to multiple deprivation parameters (least deprived = 32844, most deprived 563 = 1). Further reference to deprivation throughout is related to IMD status 24 . Deciles 564 for England, Wales and Northern Ireland were pre generated by the official data 565 source 24 . For Scotland, the same deciles were applied by the research team. 566 567 Factor analysis: Factor analysis with orthogonal transformation (varimax procedure) 568 was used to construct a distinct 'healthy 'diet and lifestyle pattern and an 'unhealthy ' 569 dietary pattern (loadings and input variables depicted in Supplementary Table 8). 570

(Supplementary
The dietary patterns were stratified into quartiles to compare demographic 571 characteristics and body weight changes (Supplementary Table 9 Table 11). Age was included in the model owing to its established 577 relationship with BMI 32 . Change in body weight was categorised as; (1) an absolute 578 bi-directional change, (2) an increase in body weight, and (3) a decrease in body 579 weight from 'pre-pandemic 'to 'peri-pandemic 'weight status. Further, stratified 580 models were also developed based on a low and high IMD. The model was fitted 581 under a maximum likelihood framework using covariance matrices. Relative model fit 582 33 was assessed using the comparative fit index (0 no fit; 1 perfect fit) 50