key: cord-0811160-9j25jwp4 authors: Gershuny, J.; Sullivan, O.; Sevilla, A.; Vega-Rapun, M.; Foliano, F.; Lamote de Grignon, J.; Harms, T.; Walthery, P. title: Using time use diaries to track changing behavior across successive stages of COVID-19 social restrictions date: 2021-02-01 journal: nan DOI: 10.1101/2021.01.29.21250766 sha: f41367427b1e6193796081a45a6f8b18207499c9 doc_id: 811160 cord_uid: 9j25jwp4 We present analyses of 24-hour time use diary records of the UK populations activities, their social context and their location, collected in real time from four successive waves of representative UK samples, both before, and during, the current COVID-19 pandemic. We show changes in behavioral response to social restrictions at three time-points during the pandemic, including across two nationwide lockdowns. Comparing samples across the four waves, we find evidence of substantial changes in the populations activities, locations and social interactions, which may be interpreted as direct responses to changes in regulations. One of our main motivations was to compare risk-related behavior across the first and second lockdowns. We find suggestive evidence that the second lockdown (recorded late November 2020) was less strictly adhered to than the first (recorded May-June 2020). Holding constant gender, age and social grade, the population spent on average 32 more minutes per day in high risk activities in late November 2020 than in first lockdown. Disaggregating by gender and age group, the difference was found to be statistically significant for all groups, with the exception of those aged 55 and over. What do people actually do in response to restrictions on their activities and social contacts imposed over the course of the pandemic? Epidemiological surveys of behaviour during the pandemic have focused on tracing social contacts in order to determine risk [1, 2, 3, 4] , while social science surveys have focused on asking respondents to estimate the quantity of time they spent in certain activities at particular times [5] . Yet neither of these sources provides a complete record of daily behavior. In this article we present analyses of 24-hour time use diary records of the UK populations' activities, their social context and their location, collected in real time from four successive waves of representative UK samples, both before, and at three points during, the current pandemic. At a time when capacity is still limited both in respect both of immunization and track-trace technology, governments must continue to rely on changes in people's daily behaviors to contain the spread of the COVID-19 virus. Time use diary information of this type, collected in real time across the course of the pandemic, will provide policy-makers with information to assess and quantify changes in daily behaviors following the imposition of social restrictions, and the impact they are likely to have on overall behavioral-associated risks. As far as we are aware this ongoing data collection is unique in providing information on changing behavior in response to each major change in social restrictions. Comparing samples across these four waves, we find evidence of substantial changes in the UK population's activities, locations and social interactions, which are interpretable as direct responses to social regulations. One of our main motivations was to compare behavior across successive lockdowns. We find suggestive evidence that the second lockdown (November 2020) may have been less strictly adhered to than the first (starting March 2020). Complete spatio-temporal accounts of the activities (including their location and social context) and socio-demographic characteristics of representative samples are key to understanding populations' changing behavioral risks of infection [6] . It is clear from the epidemiological literature that infection follows social proximity, in physical and temporal terms, to infected individuals. Time use is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint diary information offers the potential to identify levels of risk associated with daily behavior through attribution of levels of infection risk to certain combinations of activity, location and social context, and to compare these between population subgroups, such as by age and gender. When time use diary-derived patterns of daily behavior are linked to infection risk in this way, it enables the identification of those changes in behavior which are likely to contribute to subsequent changes in infection rates. It appears that the short second (November 2020) lockdown had, overall, a smaller effect on the prevalence of high-risk daily behavior (such as doing paid work in the workplace) than did the first, starting in March 2020. We disaggregate these changes to examine potential differentials by two characteristics known to be associated with the risk of infection; gender and age [7, 8] . There are various approaches to measuring what people do through the day: (a) Human observation; a labor intensive, and hence costly, technique, with dangers of contamination associated with the presence of the observer (e.g. from social desirability effects and privacy concerns). (b) Mobile devices recording individuals' activities continuously (e.g. worn body cameras, accelerometers) provide the highest level of accuracy, but involve heavy costs related to the interpretation of real-time records. (c) Survey questions about activities or behavior (e.g. "how often/how many times per day/week did you ….?", and "how long do you spend…?") are subject to problems of accuracy relating to recall, self-representation (desirability bias), and generalization (as the same activity undertaken at different times of the day or week may have varying durations, so that some attempt at 'averaging' is required of the respondent). (d) Continuous diary accounts recording successive activities are regarded as superior to survey questions because they reduce recall issues (being generally recorded on the diary day), deter misrepresentation (since activities merely over-or under-represented in response to survey . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint questions must be replaced by a complicated construction of entries in continuous diary records); and enhance generalizability (as different durations of the same activity through the day or week may be recorded). Time use diary data has been validated through small-scale comparisons with more expensive approaches using objective instruments (cameras, motion sensors) worn through the diary observation period [9] . Time use diaries have been used before in the context of investigating behavior related to COVID-19 [10, 11, 12] , but not to report changing behavioral-related infection risks at successive surveys reflecting periods of changing social restrictions (the UK Office of National Statistics collected a pilot online time-use diary only once during the pandemic, in March-April 2020 [13]), and did not use it for combining multiple diary fields to estimate behavior-related infection risks. The data collection instrument is a low-respondent-burden (12-15 minutes per day), low-cost, online diary (the Click and Drag Diary Instrument, CaDDI), developed by the Centre for Time Use Research, and suitable for rapid deployment in real time in situations such as the current pandemic [14] . The is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint May and October of 2016 (representing pre-pandemic behavior patterns, N=1011 diary days); May-June 2020 (during the first UK lockdown, N=1007); August 2020 (during the post-lockdown summer relaxation of restrictions, N=987); and November 2020 (the second lockdown, N=1358). Weights were calculated both to yield the correct numbers of days-of-the-week for each sex and (10-year) age group in the UK population, and to correct for an under-representation (of about 16%) in the 2020 diary samples in the 65+ age group. All analyses in this article were conducted using these weights. We assign the level of risk for each diary episode by combining three of the simultaneous diary fields: activity type, location and co-presence. These assignments are made taking cognizance of the literature on COVID-19 infection transmission, which considers time at home alone or with members of the same household as lowest-risk, with the main focus for transmission being contact with nonhousehold members, both at or away from home. The virus is more likely to be transmitted indoors, in crowds, and through personal contact of over 15 minutes [15, 16] . Table 1 shows assignments for each combination of the three diary fields to one of five risk categories, ranging from lowest (1) to highest risk level (5) . Activities are shown down the first column, and assigned a risk category according to copresence information (alone or with other household members/with non-household members), nested within location (at home/away from home). Estimates of risk for location and copresence status vary according to the activity (e.g. cinema implies the presence of other, nonhousehold, individuals), and are influenced by its characteristic location (e.g. indoors enclosed, vs open-air). Co-presence information may be supplemented from the activity fields, so 'using public transport' can be taken to imply current or recent presence of other, non-household, individuals. We assigned activities lasting only one 10-minute timeslot to the lowest risk level (level 1) [16] . For further details on these assignments and their rationale see Gershuny et al 2020 [6] . ********* Table 1 about here******************* . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint We used multivariate OLS regression models to investigate the statistical significance of differences in the time spent in different risk categories across the waves of the data. For this purpose we divided risk categories into three groups: low risk (category 1 in Table 1 above); medium risk (categories 2-4) and high risk (category 5). In these models the dependent variable is the time spent at each wave in each of the three risk groups, and the independent variables were survey wave, gender, age group and social grade. To estimate statistical significance we used robust clustered standard errors estimated from single-respondent cross-day clusters (stata vs 16) , to take account of the varying number of diaries per respondent. Table 1 . Changes across the four columns of Figure 1 indicate larger changes in the UK population's time allocation across the successive phases of the COVID-19 epidemic than were observed in the UK over the fifteen-year period 2000-2015 [17] . The initial lockdown period (collected in May-June 2020) was associated with a substantial shift of time away from the high risk behaviors characterizing pre-pandemic behavior (in particular, paid work in the workplace and out of home leisure) towards home-based activities involving lower levels of contact with non-household members (and hence lower risk of infection). Partial relaxation of the lockdown regulations in August 2020 produced the expected partial return to previous patterns of daily activity. Finally, the subsequent re-imposition of lockdown regulations during November 2020 produced, as we show below, only a partial return the patterns of behavior observed during the first lockdown. *********** Figure 1 about here******************* . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint Are these changes statistically significant? Table 2 shows results from an OLS regression model with robust clustered standard. Three models are shown in the columns of the table, each with time (minutes/day) spent in the respective risk level (risk levels 1; 2-4; 5) as the dependent variable. The models include dummy variables for survey wave, sex, age group and social grade. The results provide support for the straightforward reading of Figure 1 . The substantial reduction of time per day in high-risk, and increase in time in low-risk, activities associated with the movement into the initial lockdown (May-June 2020 -the reference category) are both clearly statistically significant. Firstly, the models for high and low risk activities both account for 10-11% of the explained variance Of particular note is that the return to lockdown in November 2020 elicited significantly less risk-avoidant behavior than that associated with the first lockdown. 32 more minutes were spent in high risk activities (P<.005), and 31 minutes less in low risk activities (P<.05) compared to first lockdown. This may be to some degree associated with differences in the detail of the official regulations. But we cannot ignore the possible interpretation that it also reflects a decline in the UK population's willingness to comply with government regulations. ********** Table 2 about here **************** Table 3 provides more detail on the changes in the detail of behavior that are associated with these changes in risk levels, again expressed as contrasts with behavior during the first lockdown. Nine is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint models are shown across the rows of the table, each showing coefficients for the minutes/day spent in each activity/co-presence/location combination shown in Figure 1 . Each model includes the same list of independent variables as the models shown for Table 2 (see Supplementary Table S1 for the full set of coefficients from the regression models for each activity category). The first four rows of the table show the much lower levels of low risk activity categories done at home pre-COVID in 2016 (including personal care, unpaid work, home leisure, and paid work). Correspondingly, much more time was spent away from home with people from outside the household (in paid work, education, out-of-home leisure or caring), or in enclosed areas (e.g. cafes or cinemas). Almost all these activity categories were statistically significantly different from the time spent in these activities at first lockdown (May-June 2020), which saw reductions of over one hour per day in out-of-home leisure with non-household members, and nearly two hours' reduction in paid work or study away from home. ********** Table 3 about here********************* The intermediate period (relaxation of restrictions, August 2020) produced reversals in these behavior patterns, again, in most cases highly statistically significant, but around half the size of the pre-pandemic to first lockdown changes. In particular, less time was again spent in activities done at home, with the exception of paid work at home, which, interestingly, remained unchanged from its first lockdown level -a lack of change that persisted into the second lockdown. At the other end of the scale, the higher-risk return to the workplace during this period, though highly statistically significant, is only one quarter of the size of the previous shift of paid work from the workplace to the home during first lockdown (half an hour more, compared to nearly 2 hours). And as the analyses of Table 2 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint out of home, activities (paid work in the workplace, up 33 minutes, and leisure away from home, up 12 minutes) compared to the first lockdown, but these differences are on the margin of statistical significance. Disaggregation of response to changing regulations by gender and age Figure 2 disaggregates behavioral change across the pandemic by age and gender, two factors which have previously been shown to be related to important differences in risk of infection [7, 8] . The graphs display regression coefficients for time spent (minutes per day) in the 3 broad risk categories (high, medium and low) outlined in Table 1 , expressed as contrasts to the UK first lockdown in May-June 2020. They are based on separate regression models for the two genders and three age groups (sample sizes were unfortunately not sufficiently large to permit modelling age within gender). The same combination of dependent and independent variables is used as in the regression models for Table 2 (although omitting the gender variable in the analyses disaggregated by gender, and the age group variable in the analyses disaggregated by age group). Full regression coefficients are provided in Supplementary Tables S2 (by gender) and S3 (by age group). ************* Figure 2 about here******************* The first thing to note is that the pattern of the coefficients across the stages of social restrictions for most gender and age groups broadly reflects the overall pattern shown in Table 2 . Specifically, for the top pair of panels, showing the models for high-risk activities, there were substantial and strongly significant differences between much higher levels of high-risk activities pre-pandemic compared to the first lockdown period across all age groups and both genders. Subsequently, in August 2020, there was a less dramatic reversal, in the direction of more time being again spent in high-risk activities. This difference is again, for most gender and age groups, highly significantly different from first lockdown, but lesser in scale than that observed between the pre-pandemic to lockdown period. Women and the oldest age group (55+) spent somewhat less time than, respectively, men and younger age groups, in high risk activities both pre-pandemic and again during the second lockdown period (indeed for the oldest age group there was no difference between first . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint and second lockdown in the time spent in high-risk activities). Reduction in high risk-activities associated with the second (November) lockdown was considerably smaller for men and the two younger (18-34 and 35-54) age groups than observed during first lockdown; consequently the November lockdown coefficients are statistically significantly different for all groups from those at first lockdown (with the exception of ages 55+). Four out of the five coefficients for high-risk activities during the November lockdown are substantially higher (by 50 or more minutes) than those for the first lockdown, indicating a decrease in reduction of risky behavior in successive lockdowns (except for the oldest age group). The middle panels refer to the mid-level risk activities. Differentials here are at first sight similar in pattern to those for high-risk activities: an initial substantial reduction in time devoted to medium risk level activities during first lockdown, followed by a smaller reversion during the intermediate period, most notably among the 35-54 year age group, and women. These groups devoted significantly more time in the intermediate period to mid-risk-level activities (in particular, leisure, caring and other socializing activity) with non-household members than men and other age groups, respectively. The November coefficients show no differences between first and second lockdown for this level of risk activity. The bottom pair of panels, describing changes in time devoted to relatively low-infection-risk activities, show the converse of the top set of panels. The initial very substantial increase in low risk at-home activities (home leisure taken with other household members; unpaid housework, childand elder-care at home) during first lockdown, particularly for ages 18-34 and 35-54 (who started from a lower base), may reflect the reduction of paid work commitments, either though home working, furlough or redundancy. First lockdown was followed by a smaller reversal, again particularly for women and 35-54 year olds, who increased their low-risk activities during this period to a greater extent than, respectively, men or other age-groups. As in the case of the medium risk activities, there is no statistically significant difference in the time spent in these low-risk activities during first and second lockdown (with the possible exception of women, who spent marginally statistically significantly less time in these low-risk activities during second lockdown). is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. Figure 1 . One of our primary motivations was to investigate differences in population behavior between the first and second lockdown, and we were able to demonstrate the more limited impact of the November restrictions on the UK population's risk avoidance behavior (see Table 2 ). Holding constant gender, age and social grade, the population spent on average 32 more minutes per day in high risk activities in late November 2020 than in first lockdown (recorded May-June 2020). Looking in more detail at the types of activities (Table 3) , the November lockdown involved more time spent both in leisure or caring activities done away from home with nonhousehold members (12 minutes/day more on average), and in paid work at the workplace (33 minutes/day). Disaggregating the changes by gender and age group shows that these differences in time spent in high-risk activities between the two lockdown periods applied across all groups, with the exception of the oldest age group, aged 55+, whose behavior during the second lockdown was not statistically significantly different from that during first lockdown (see Figure 2 , upper pair of graphs). This suggests that there was perhaps greater awareness of risk among this group, which was reflected in their behavior. Women and those aged 35-54 in the intermediate period between lockdowns were more likely to return to medium-risk activities, spending less time in low-risk activities, than other groups. At a time when capacity is still limited both in respect both of immunization and track-trace technology, governments must continue to rely on changes in people's daily behaviors to contain the spread of the COVID-19 virus. However, the potential for formulating and evaluating the imposition of regulations in this way have scarcely begun to be considered. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint Among the limitations of our current study is the use of a commercial quota-sampling base. However, it is encouraging that the sample size of approximately 1000 diary days and 350 respondents at each of the four sample time points appears to provide sufficient statistical power to discern the effects of changes in regulation. In future research, sampling randomly from a national population frame, and more frequently-either monthly, or, preferably, continuously-would enable the tracking of changes in behavior alongside changes in rates of infection. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. Table S1 . Models for 9 activity risk categories Table S2 . Activity risk categories modelled by gender Table S3 . Activity risk categories modelled by age group . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint Notes: Risk category 1=lowest risk; risk category 5=highest risk. 'Write-in' represents the small proportion of time as yet unallocated; each of the four waves of data is weighted to represent 2016 age and sex distributions in the UK population. . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted February 1, 2021. ; https://doi.org/10.1101/2021.01.29.21250766 doi: medRxiv preprint Social contacts and mixing patterns relevant to the spread of infectious diseases A systematic review of social contact surveys to inform transmission models of close-contact infections Contagion! the BBC four pandemic-the model behind the documentary Quantifying the impact of physical distance measures on the transmission of COVID-19 in the UK Survey in Five National Longitudinal Studies: Waves 1 and 2 User Guide (Version 2). London: UCL Centre for Longitudinal Studies and MRC Unit for Lifelong Health and Ageing P. Walthery (forthcoming). 'A new perspective from time use research on the effects of lockdown on COVID-19 behavioral infection risk'. PLOS ONE Demographic science aids in understanding the spread and fatality rates of COVID-19 Gender equality in work and Covid-19 deaths Testing self-report time-use diaries against objective instruments in real time Using time-use data to parameterize models for the spread of close-contact infectious diseases Social contact structures and time use patterns in the Manicaland province of Zimbabwe The effect of social distancing on the spread of novel coronavirus: estimates from linked state-level infection and American time use survey data Time use diary design for our times -an overview, presenting a 'drag-and-drop' diary instrument for online application The relative importance of frequency of contacts and duration of exposure for the spread of directly transmitted diseases What we really do all day: Insights from the Centre for Time Use Research (Chapter 1: Fifty years of change in UK life at a glance). London: Penguin (Pelican series)