COVID-19 vaccination, risk-compensatory behaviours, and contacts in the UK – Scientific Reports

Data

Data are taken from the UK Office for National Statistics Coronavirus (COVID-19) Infection Survey (ISRCTN21086382; https://www.ndm.ox.ac.uk/covid-19/covid-19-infection-survey/protocol-and-information-sheets; details in Pouwels et al.³⁵). Individuals were approached from households that were randomly selected from previous surveys and address lists in England, Northern Ireland, Scotland, and Wales to provide a representative sample of the UK population. The survey consists of repeated cross-sectional household surveys with additional serial sampling and longitudinal follow-up. Data collected between 1st October 2020 and 15th September 2021 were used for this analysis. The study received ethical approval from the South Central Berkshire B Research Ethics Committee (20/SC/0195). All methods were performed in accordance with the relevant guidelines and regulations. Informed consent was obtained from all participants. Private households are randomly selected on a continuous basis from address lists and previous surveys to provide a representative sample across the UK. After verbal agreement to participate, a study worker visited each selected household to take written informed consent for individuals aged 2 years and over. Parents or carers provided consent for those aged 2–15 years; those aged 10–15 years also provided written assent. The supplementary materials provide further details of the sampling design.

Data include information on the behavioural outcomes of interest, sociodemographic characteristics, and medical information (https://www.ndm.ox.ac.uk/covid-19/covid-19-infection-survey/case-record-forms for survey questionnaires).

Vaccination status and population-level uptake

Patients were asked about their vaccination status, including type, number of doses and date(s). Survey participants from England were also linked to administrative records from the National Immunisation Management Service (NIMS)³⁵. Where available, we used records from NIMS, otherwise we used records from the survey, since linkage was periodic and NIMS does not contain information about vaccinations received abroad or in Northern Ireland, Scotland, or Wales. Where records were available in both, agreement on type was 98% and dates 95% within ± 7 days.

Data on population level uptake were taken from publicly available government sources (https://coronavirus.data.gov.uk/details/download). This was mapped to the participants’ data at the level of CIS geographical regions In England and Scotland.

Outcomes

10 behavioural outcomes were self-reported by individuals

1. (i)

   The number of physical contacts, e.g. handshake, personal care, including with personal protective equipment, with individuals aged < 18 years old in the past 7 days;

2. (ii)

   The number of physical contacts with individuals aged 18–69 years old in the past 7 days;

3. (iii)

   The number of physical contacts with individuals aged 70 years and over in the past 7 days;

4. (iv)

   The number of socially distanced contacts, worded as, “direct, but not physical, contact”, with individuals aged < 18 years old in the past 7 days;

5. (v)

   The number of socially distanced contacts with individuals aged 18–69 years old in the past 7 days;

6. (vi)

   The number of socially distanced contacts with individuals aged 70 years and over in the past 7 days;

7. (vii)

   The number of times the participant spend one hour or longer inside their own home with someone from another household in the past 7 days;

8. (viii)

   The number of times the participant spend time one hour or longer inside the building of another person’s home in the past 7 days;

9. (ix)

   Among those that reported engagement in work or studying: mode of travel to work/place of education (grouped as public transport versus other for the current analyses) in the past 7 days; and

10. (x)

    Among those that reported engagement in work or studying: main work/study location in the past 7 days.

The physical and socially distanced contact variables (i–vi) were recorded as ordered variables with 5 response categories (0, 1–5, 6–10, 11–20, 21 or more). Both types of home visits (vii and viii) were recorded as ordered variables with 8 response categories (0, 1, 2, 3, 4, 5, 6, 7 times or more). Work or study location was recorded as an ordered variable with three categories (working from home, working from a mix of home and outside of home, working from outside the home).

Sample eligibility

Behaviour changes among vaccinated individuals

Individuals that met the following criteria were included: age 18 years or older, or an individual self-reported a long term health condition (in the UK those with underlying health conditions aged 16 years and over were prioritised after those aged 65 years and over). Preliminary testing plotting difference of the smooths³⁶ indicated two groups, patient-facing healthcare workers and over 65s had different behaviours to those under 65 or that did not work in healthcare settings; results are presented in the Supplementary Figs. 1–3. Patient-facing healthcare workers were both prioritised for vaccination and also are expected to have had unusually high contacts during periods of strict mitigation measures given the nature of their job. Older individuals have much higher risk of severe outcomes, generally less frequent contact patterns, and an increasingly small number of unvaccinated older individuals remaining unvaccinated³³. Individuals were only included if there was evidence that they were vaccinated at any point during the survey to avoid that those not willing to get vaccinated, who are expected to have different behaviours than those that are willing to get vaccinated.

Individuals were included if they had at least one pre-vaccination and one post-vaccination observation between 1st October 2020 and 15th September 2021.

Influence of vulnerable household member vaccination

This analysis included individuals that lived in a household with a vulnerable individual (defined as either > 65 years or having a long term health condition) but themselves were not considered vulnerable and not yet vaccinated, with visits from the 1st October 2020 and 15th September 2021.

Influence of population-level uptake

All individuals with visits in the period 1st October 2020 and 15th September 2021 were included in analyses, regardless of vaccination status or reported job.

Statistical analyses

All analyses used generalized additive ordered categorical regressions that were estimated on the ordinal outcomes³⁷. In these models, the linear predictor is a latent variable with estimated thresholds that mark the transitions between levels of the ordered categorical response. For public transport for working travel, which as a binary outcome, a generalized additive logistic regression was used.

Predicted probabilities were estimated for all response categories of the ordinal outcomes. We then collapsed these into binary variables, where the values of the outcomes at zero are treated as zero, and one otherwise (posterior probabilities of being in each category, and their summation to construct the reported probability, are provided in the Supplementary Fig. 14 for physical, outside of household contacts with 18–69 year-olds). In this way, the binary variable captures the margin between “none” and “any”: for contacts, this variable is either no outside of household contacts in the past 7 days (0) or any outside of household contacts in the past 7 days (1); for home visits it is either no visits in the past 7 days (0) or any home visits in the past 7 days (1); for work location, this is either working at home every day in the past 7 days (0) or working in outside of home location (i.e. office, café, etc.) one or more times in the past 7 days (1); and for mode of travel to work, this is either private transport in the past 7 days (0) or public transport (1). Using the ordinal variables in models allows us to exploit all of the information in the data, and collapsing them into posterior binary variables eases exposition, aligns the results across outcomes with different numbers of categories, and focusses the analyses on the margin of “none” and “any” which is most useful for understanding transmission.

Behaviour changes among vaccinated individuals

This analysis examined the behavioural response to individual vaccination. Outcomes were modelled from 180 days prior to first vaccination to 120 after the first vaccination (given few visits outside of this range). To capture non-linear behavioural change, we used thin plate splines on time from first vaccination (the number of basis functions (k in the mgcv package³⁸) was the number of unique values divided by 3). The degree of smoothing of the splines was optimised using a fast implementation of restricted maximum likelihood (REML)^38,39. Regressions controlled for confounding with individual-specific characteristics of gender, ethnicity, socioeconomic status (rank of index of multiple deprivation calculated separately for each country in the United Kingdom^40,41,42,43), urban/rural classification^11,12,44, the stage of lockdown (and corresponding restrictions applied), household size, whether the household is multigenerational (household individuals aged 16 or younger and individuals aged school year 12 to age 49 and individuals aged 50 +), and if the individual ever reported a long-term healthcare condition. Calendar time by region/country (9 regions in England and Northern Ireland, Scotland, and Wales) interactions were captured using thin plate splines (the number of basis functions was the number of unique values divided by 3) interacting with region/country (factor smooth interaction) to control for non-linear confounding of region/country-specific behavioural change over time (see Supplementary Fig. 8 for an example of physical contacts with under 18 year-olds). Thin plate splines were also applied to age to control for non-linear confounding of age-specific behaviour.

Testing changes in behavioural responses to individual vaccination involved testing trends pre- and post-vaccination. More specifically, the gradient in the outcomes (over time from vaccination) in a period before vaccination (− 50 days to − 14 days) was tested against two periods (1–13 days and 14–50 days) after vaccination, arguing that longer-term changes were less likely related to vaccination. The period 50 days to 14 days was used as the comparator to avoid conflating any changed post-vaccination behaviours with any behavioural shifts leading up to having a vaccine. This was conducted using 10,000 simulations based on pseudo-random draws from the posterior distributions of the fitted GAM models using the Gaussian approximation to the posterior of the model coefficients. The pseudo-random draws are obtained from a multivariate normal with mean vector equal to the estimated model coefficients and covariance matrix equal to the covariance matrix of the coefficients. For each draw, the gradient of the outcome was computed for the pre- and the two post-periods; and the differences in the gradients between the pre-period and each of the post-periods, respectively, were taken. Across the draws, the median, 2.5th, and 97.5th percentiles of the distribution of differences were taken and compared to assess if the gradients of the outcomes changed between pre- and post-vaccination periods.

Influence of vulnerable household member vaccination

This analysis examined the behavioural response to vaccination of vulnerable household members as described above. This analysis was similar to that for vaccinated individuals, except the definition of vaccination date was altered to be the date of vaccination for the last vulnerable individual in the household to be vaccinated. In addition, this analysis was limited to the subset of (a) individuals considered to be non-vulnerable and (c) observations until individuals themselves received a first vaccination. To capture non-linear behavioural change, we used thin plate splines on time from first household vaccination⁴⁵ (the number of basis functions (k in the mgcv package³⁸) was based on the number of unique values divided by 3). Model specification was otherwise identical to the analysis of vaccinated individuals.

Influence of population-level uptake

This analysis examined the behavioural response to vaccination rates at the population level rather than to individual vaccination. In these models, the time from first vaccination variable was replaced with the percentage of individuals in the population that had been vaccinated. Daily rates of vaccine uptake were available for local areas and these were merged to the CIS data by CIS geographical area to enable modelling. To capture non-linear behavioural change, we used thin plate splines on population-level-vaccination⁴⁵ (the number of basis functions (k in the mgcv package³⁸) was the number of unique values divided by 3). In addition, a variable for whether individuals were patient-facing healthcare workers was included. Thin plate splines were also applied to age to control for non-linear confounding of age-specific behaviour (where again the number of basis functions was the number of unique values divided by 3); age was truncated at 95 to reduce the influence of outliers. A categorical variable was defined which indicated if the individual was a child, a vaccinated adult or an unvaccinated adult. This variable was interacted with the thin-plate splines on population level vaccination to allow for differential behavioural responses for each of the groups.

Specification and sensitivity analyses

The proportional odds assumption was tested in preliminary analyses by interacting the threshold parameters with a post-vaccination dummy variable. No evidence of variation in the response over the range of the outcomes was found. We tested tensor splines for time-age-region interactions (knots divided the time range at 3-day intervals and the age range at 3-year intervals) to control for non-linear confounding of age- and region-specific behavioural change over time. There was no improvement in model fit nor impact on the difference in the outcomes over time. Thus, the simpler specification, i.e. region by time interaction, was retained.

Models examined if behaviours were consistent in three groups (relative to the remaining sample): those over age 65, reporting working in patient-facing healthcare, and reporting a long-term health condition. Evidence of different patterns of behaviour were found for those over 65 years and for patient-facing healthcare workers, but not for those self-reporting long-term health conditions (results are presented in the Supplementary Figs. 1–3). Differential behavioural trends in response to ChAdOx1 versus BNT162b2 vaccines were evaluated by plotting difference of the smooths (other types of vaccine, e.g. mRNA-1273 (Moderna), were controlled for separately). No differences were observed in behaviours for ChAdOx1 versus the BNT162b2 vaccines (Supplementary Fig. 9). For the population-level vaccination models, including an interaction between the thin-plate splines for time-from-vaccination and a categorical variable indicating if the individual was a child, a vaccinated adult or an unvaccinated adult improved the fit of the model and was therefore used. A model that included both time from own vaccination and population level vaccination were estimated to ensure that the results were robust.

Models that use the date of second, rather than first, vaccination yielded similar patterns across the outcomes. Results are shown in Supplementary Figs. 10–12. In addition, we analysed responses to population level vaccination among those that had been vaccinated and consistent with the eligibility criteria described for the analysis of own vaccination. Results are presented in Supplementary Fig. 13. In contrast to the full sample, an initial peak in the probability of contacts in the first 25% of population vaccination is not observed, suggesting this peak is driven by children and the unvaccinated; otherwise the patterns of behaviour are consistent. Interval regressions were estimated to give complementary results on associations between vaccination and the number of contacts over time. These models were implemented using the ‘survival’ package and used p-splines—with the degrees of freedom based on the AIC—to allow for non-linear effects of the exposures of interest.

To evaluate whether passage of time—potentially reflective of pandemic fatigue—or increases in vaccination rates are more important for behavioural responses, we added a comparison between models with calendar time only, models with population-level vaccination uptake only, and models with both calendar time and vaccination (supplementary Figs. 14–16).

To assess whether almost perfect correlation between vaccination uptake and time when modelling both using flexible splines would hinder our inference, we also explored the use of calendar time measured in weeks (by 1 and 2 weeks), as one would expect pandemic fatigue to be rather stable within 1- or 2-week intervals while population-level vaccination uptake changed quickly within these intervals (e.g. 5% increase in the first week of February 2021).The results appear to be robust to the specification of calendar time. Results for the weekly and fortnightly models are presented in supplementary Figs. 17 and 18.