key: cord-0822194-9mqprimv
authors: Johnson, Thomas F.; Hordley, Lisbeth A.; Greenwell, Matthew P.; Evans, Luke C.
title: Effect of park use and landscape structure on COVID-19 transmission rates
date: 2021-06-02
journal: Sci Total Environ
DOI: 10.1016/j.scitotenv.2021.148123
sha: fa47f693080728a57f60c92b2f513c3eddaf7366
doc_id: 822194
cord_uid: 9mqprimv

The COVID-19 pandemic has had severe impacts on global public health. In England, social distancing measures and a nationwide lockdown were introduced to reduce the spread of the virus. Green space accessibility may have been particularly important during this lockdown, as it could have provided benefits for physical and mental wellbeing. However, the effects of public green space use on the rate of COVID-19 transmission are yet to be quantified, and as the size and accessibility of green spaces vary within England’s local authorities, the risks and benefits to the public of using green space may be context-dependent. To evaluate how green space affected COVID-19 transmission across 299 local authorities (small regions) in England, we calculated a daily case rate metric, based upon a seven-day moving average, for each day within the period 1st June - 30th November 2020 and assessed how baseline health and mobility variables influenced these rates. Next, looking at the residual case rates, we investigated how landscape structure (e.g. area and patchiness of green space) and park use influenced transmission. We first show that reducing mobility is associated with a decline in case rates, especially in areas with high population clustering. After accounting for known mechanisms behind transmission rates, we found that park use (showing a preference for park mobility) was associated with decreased residual case rates, especially when green space was low and contiguous (not patchy). Our results suggest that a reduction in overall mobility may be a good strategy for reducing case rates, endorsing the success of lockdown measures. However, if mobility is necessary, outdoor park use may be safer than other forms of mobility and associated activities (e.g. shopping or office-based working).

The COVID-19 pandemic has had severe impacts on public health (Mahase, 2020) and remains an emergency of international concern. In response to the pandemic, the UK government has implemented social distancing measures and nationwide lockdowns to control the spread of the virus (UK Government, 2020a) . During these periods, the general public were limited in the distances they could travel and, at certain points, the number of times they could leave their residence each day; with an allowance of one non-essential trip during the peak of transmission (UK Government, 2020a) . Though social restrictions have fluctuated in response to case rates, social distancing has been constant and there has been a general message throughout England of reduced movement and staying local where possible for much of 2020. These restrictions have meant that members of the public became more reliant on amenity spaces close to their residences for daily exercise and/or recreation (Geng et al., 2021) . Green spaces may provide a comparatively safe place for these activities, though the amount and structure of green space available for public use differs widely across the UK. Here we evaluate if differences in the availability and structure of public green space within local authorities (local government bodies responsible for public services within a specified area) in England, and their usage, influenced the local rate of incidence of COVID-19.

Green spaces, which we define as vegetated non-arable areas -see Taylor & Hochuli (2017) for further details -provide important cultural and recreational ecosystem services, benefiting both mental and physical health (Beyer et al., 2014; Cohen-Cline et al., 2015) . These benefits are usually considered in terms of reducing the prevalence or severity of conditions such as mental stress (Nutsford et al., 2013) or cardiovascular disease (Seo et al., 2019) , and some of these benefits have continued throughout the pandemic (Slater et al., 2020; Soga et al., 2020) . However, the influence of green space use on disease transmission rates has received less investigation, but is of great J o u r n a l P r e -p r o o f Journal Pre-proof importance as green space use has increased rapidly during the pandemic (Venter et al., 2020) . Furthermore, it is unclear how 'safe' green spaces are during periods of higher incidence especially in densely populated areas (Shoari et al., 2020) .

We anticipate that green space could impact COVID-19 incidence in two ways: general health and wellbeing, and transmission. It is conceivable that general health and well-being is greater in areas with more green space, as higher levels of green space are associated with healthier populations (Maas et al., 2006; Mitchell and Popham, 2007; van den Berg et al., 2015) . As COVID-19 has a greater impact on those with underlying health conditions and sedentary lifestyles (Hamer et al., 2020; Jordan et al., 2020) , green space may, therefore, indirectly provide some level of resilience to the disease and/or reduce incidence. However, our key focus here is on transmission, as it is likely that the major benefits of outdoor recreation in green space are related to a lower risk of infection.

Current evidence suggests that COVID-19 is spread via droplet infections, contact with contaminated individuals or surfaces, and through aerosol transmission (Bahl et al., 2020) . These risks are likely minimised in green space areas, as generally, they are less spatially confined, and with fewer surfaces prone to frequent touching or contact. Consequently, green space use may represent a safe form of recreation by minimising risk of infection.

In England and Wales approximately 87% of the population are within a 10-minute walk of public parks and gardens (Shoari et al. 2020) . However, both the structure and amount of green space vary between local authorities, and both could influence COVID-19 incidence. Generally, it has been found that greater health benefits are derived from larger areas of green space (Ekkel and de Vries, 2017) . In the context of disease transmission, larger areas may offer more space per individual, lowering transmission risk. However, smaller fragmented areas of green space are common in many residential areas and are, therefore, more accessible to much of the population and may be used more frequently. Further, if public use is distributed across fragmented green spaces, the wider effects of a transmission incident could be reduced, as contacts would be isolated to the members of a neighbourhood or community adjacent to a particular green space. This process can be seen in animal diseases where habitat fragmentation reduces transmission due to limiting interactions between groups in different patches (Mccallum and Dobson, 2002) . However, J o u r n a l P r e -p r o o f Journal Pre-proof fragmentation also typically results from reductions in the total area of green space (Fahrig, 2013) , leading to less overall space per individual, possibly increasing transmission rates.

Whilst the effects of green space on COVID-19 transmission are currently unclear, other environmental and social factors are known to influence both the spread and severity of the disease. For example, human mobility drives the spread of infectious diseases (Kraemer et al., 2019) and studies have shown that reducing social interactions by restricting mobility can lead to a decrease in transmission rates of COVID-19 (Chinazzi et al., 2020; Gatto et al., 2020) . Furthermore, as diseases are often spread along transport links and in offices (Gatto et al., 2020; Zhang et al., 2018) , enforcing lockdown situations that curtail movement, such as requiring people to work from home, can have a great effect on reducing transmission rates. In addition to mobility, health and social factors have been associated with increased severity of the disease such as age, underlying health conditions, and deprivation (Richardson et al., 2020; Williamson et al., 2020) . Consequently, any possible effects of green space must be considered after attempting to account for factors that could increase recorded incidence.

Given the stated benefits of green space, it is important to attempt to evaluate using available evidence, the impact of green space use on transmission rates. In addition, understanding the influence of green space on COVID-19 incidence could provide an estimate of the value of green space for maintaining public health if subjected to a resurgence of the COVID-19 pandemic. And, in the longer term, indicate the potential benefits of local green space on future pandemics of comparative severity. Here, using time series of COVID-19 cases within local authorities in England, we explore how both green space use and access (i.e. availability of green spaces) influence COVID-19 incidence. Our approach is to first construct a baseline transmission model to attempt to control for factors likely to influence recorded COVID-19 incidence and then to explore how green space influenced case rates above or below this baseline. We predict that green space and the way it is structured, in itself, will have no effect on case rates. However, we expect that an increase in park use (i.e. spending time in green space) will make the structure and availability of green space important ( Figure 1 ). Specifically, when green space is low, park use will likely represent a safer form of movement (e.g. compared to shopping), unless the green space becomes J o u r n a l P r e -p r o o f Journal Pre-proof a congregation zone that inflates transmission risk. Furthermore, we predict that case rates will be lower when green space is fragmented, as the disease will be contained in more localised areas.

For example, if the local authority has one large park the presence of an infected individual puts more people at risk than an infected individual attending one of many parks. Further, we predict, as others have found (Kraemer et al., 2020) , that increased mobility will increase incidence, but that park use (measured as relative use of parks) is the safest form of mobility (e.g. preferable over shopping). The green health symbol indicates that the landscape metric and park use is beneficial, whilst the red toxic symbol indicates a risk.

We compiled daily lab-confirmed cases (incidence) of COVID-19 in England from February 15 th 2020 up to 30 th November 2020 (available from https://coronavirus.data.gov.uk/). We only included cases until November, as in December England began an aggressive vaccination campaign and the more infections COVID B1.1.7 variant began to spread widelyfactors that could confound our models (see below). Cases were recorded at the lower tier local authority (administrative areas for local government) level (N = 299). These local authorities vary in size (3 -26,000km 2 ) , demographics, cultures, and in socio-economic circumstances. Incidence over this time was highly variable with periods of rapid increases, which were then relatively controlled by periods of national lockdown ( Figure 2 ). To determine factors influencing COVID-19 transmission, we estimated case rates for each day in each local authority. Case rates were derived by fitting log-linear models, regressing the natural log of daily cases against date (days). To reduce the effect of noise, we fit these regressions over 7-day moving windows (Figure S1) e.g. to estimate the case rate on March J o u r n a l P r e -p r o o f Journal Pre-proof between March 2 nd -8 th . The coefficients of these models provided a daily case rate. We converted these coefficients into a daily percentage change in cases. indicated with red shading. b) Google mobility trends (Google, 2020) , describing change in mobility over time for five different categories, relative to a baseline period (3 rd of January to the 6 th of February 2020). We excluded the sixth category 'residential mobility' as it is measured differently to all other categories. Each line within the mobility trends represents a local authority. All plots extend from the 15 th of February 15 th to the 30 th of November 30 th 2020. For the 'parks' plot, we limited the y-axis at 300% to exclude a small number of extreme observations with high park use.

We compiled variables which describe the mechanisms considered to influence case rates (Table   1 ). Firstly, we derived two variables which describe the structure of the local authority population:

population densityresidential population density (controls for green space in the green transmission model below); and population clustering -Moran's I spatial autocorrelation of residential population density (controls for patchiness in the green transmission model below). National lockdowns, and the resulting reduction in people's mobility, were an important tool for reducing transmission within England during the COVID-19 pandemic. We used Google Community

Mobility Reports to track human mobility and its effect on case rates (Google, 2020) . These reports chart movement trends over time across six categories: retail and recreation, groceries and pharmacies, transit stations, workplaces, residential, and parks. These trends describe how visitors to, or time spent in, each of the six categories changed compared to a pre-pandemic 5-week period (the median value from 3 rd January to 6 th February 2020). As the mobility data contained missing values (c.12%) for some local authorities and dates ( Figure S2 ), we were conscious these missing values may lead to statistical inference errors within the models below. As a result, we filled missing mobility values using mice: multiple imputation chained equations R package and '2l.pan' imputation approach, which is a hierarchical normal model within homogenous within group variances (Van Buuren and Groothuis-Oudshoorn, 2011). This hierarchical structure allowed us to model mobility trends accounting for differences in local authorities. We included the following terms within our imputation model: all six Google mobility timeseries, as well as a 1-day lag period for each timeseries, the number of days along the timeseries since February 15 th with a cubic polynomial term, an indicator variable to describe whether each day was a weekend or not, and the timeseries of daily COVID-19 cases within the local authority. We also included terms that didn't vary through time, including: the latitude and longitude of the local authority, and all local authority covairates within the baseline and green transmission models below (population density, population clustering, health, demography, economy, green space, and patchiness). Finally, we also included some national metrics that could infleunce local mobility, including: a timeseries of daily COVID-19 cases measured at the national scale, as well as the mean daily temperature and precipitation within Central England. We ran this model through 10 chains, each with 20 iterations, and 20 pan

iterations. The imputation model converged.

Conventionally, as part of a multiple imputation framework, these 10 chains should then be modelled seperately and coefficient standard errors should be inflated with Rubin's rules (Little and Rubin, 2002) . However, given the small percentage of missing values, and that there are currently no well defined steps for using Rubin's rules in genralized additive models (see our models below), we instead averaged mobility values across the 10 chains to produce mean estimates of mobility for each category, day, and local authroity i.e. conducting single impuation. We ensured the imputations produced plausible values ( Figure S3 ). From this mobility dataset, we derived a variable which described overall mobility change for each date in each local authority, which is the average mobility change across five of the six categories (excluding residential) for each day in each local authority. We excluded the residential mobility category as it is inversely correlated with all other categories and is measured differently (Google, 2020) . However, as there is a delay between a mobility reduction and a case rate reduction, we lagged the overall mobility change metric by linking each case rate with the mean mobility change from 2 -12 days prior.

We compiled two variables which describe the structure of green spaces in each local authority:

patchinessmedian frequency of parks within a 1km 2 radius around households in the local authority (Office for National Statistics, 2021c); green spaceavailable green space per person (m 2 ) within the local authority, derived by dividing the green-cover area by the local authority population size. Green-cover area was calculated from the UKCEH 2015 25metre land cover raster (Rowland et al., 2017) and included the following landscape categories: broadleaved woodland, coniferous woodland, improved grassland, neutral grassland, calcareous grassland, acid grassland, fen, marsh and swamp, heather, heather grassland, and bog. For this green-cover area calculation, we set a 1km buffer around the local authority, to represent green space access of households on the local authority border.

Using the mobility dataset, we also produced a park use variable, which represents how parks are used relative to overall mobility. This park use metric is derived by fitting a linear model between park use and overall mobility within each local authority, and extracting the residual park use, where J o u r n a l P r e -p r o o f Journal Pre-proof positive values represent a preference for using parks over other forms of mobility for a given date (in the original percentage units). Parks include public gardens, castles, national forests, campsites, observation points, and national parks, but exclude surrounding countryside in rural areas. As a result, the Google (2020) definition of parks differs slightly to the landscape categories used in our green space metric but are our best available representation of how green space was used during the pandemic. As in the overall mobility change metric, park use represents the mean use of parks in the prior 2 to 12 days.

We developed two core models ( Figure 3 ): baseline transmissionaimed at controlling for the major mechanisms influencing case rates; and green transmissionimpact of landscape structure and park use on case rates. The baseline and green transmission models are both focussed on case rates, but we anticipated that any effects of green space on COVID-19 case rates were likely to be much smaller than variables known to influence disease transmission (e.g. population density). As a result, we structured our analyses to first account for the presence of these more influential variables in a baseline transmission model, and then in the green transmission model, we explored how green areas (the focus of this study) can alter the residuals of these case rates.

Conventionally, it is advised to include all variables within one regression instead of analysing the residuals separately (Freckleton, 2002) . However, variables were highly correlated (e.g. population density and green space are derived in similar ways), and resulted in multicollinearity issues.

Dealing with the major mechanisms first (e.g. population density) mitigated these multicollinearity issues. To control for the baseline health and transmission mechanisms influencing COVID-19 case rates,

we developed a generalized additive model within the mgcv R package (Wood, 2021) , with case rate as the responseinverse hyperbolic sine transformed to address heavy tailed residuals. We included the following parameters as linear predictors: health, demography, economy, population density (log 10 transformed), population clustering, and mobility change. We also included J o u r n a l P r e -p r o o f Journal Pre-proof interactions between population density and clustering, population density and mobility change, and population clustering and mobility change. In model development, it was clear that the residuals were experiencing extreme positive temporal autocorrelation, where case rate values were very similar to values from the previous day. As a result, we also included the previous days case rate (one day lag) as a linear predictor in the model. We included random intercept smoothing over the local authorities to account for the non-independence of multiple case rates within the same local authorities. Due to working hour restrictions in the England, case counts on Saturdays and Sundays were largely underestimated, and then over-estimated on Mondays and Tuesdays. As a result, we also included a cyclic smoothing term (with up to 7 knots) over day of the week to capture reporting biases and control for daily variation (days within a week) in case reporting. We extracted the residuals from this model for the green transmission model.

To assess how landscape structure and park use influenced residual case rates, we again developed a generalized additive model developed, with residual case rates form the baseline transmission models as the response, as well as the following linear predictor parameters: park use, green space (log 10 transformed), patchiness, as well as interactions between park use and green space, and park use and patchiness. These models also included random intercept smoothing over local authorities, but we did not control for the smoothing over days of the week, which was addressed in the earlier baseline transmission model.

We opted to fill missing mobility values with imputation instead of using complete-case analyses,

where any observations with missing mobility data are removed. However, given the small percentage of missing values, and that the mobility data is averaged across five categories, and then again through time, we wanted to ensure model coefficients did not change drastically under imputation, which could be a sign of a statistical inference error (Johnson et al., 2021) . As a result, we repeated the analyses using only complete-case observations and compared model coefficients between the missing value approaches. Given the similarity in the complete-case and imputation J o u r n a l P r e -p r o o f Journal Pre-proof coefficients ( Figure S4 ), we continued using the coefficients from the imputation model which covered a greater array of local authorites.

In both the baseline and green transmission models, we were conscious that some parameter effects may have varied through time. For example, some covariates may have been particularly influential prior to mandatory mask wearing in shops on July 24 th 2020. As a result, we extracted the first four weeks of data from our case rate dataset and ran the models on this subset. We then shifted the data forwards one week and re-ran the models, repeating this procedure (moving window), creating 40 replicates of the coefficients each representing a different-overlapping period of time between February and November 2020. From this, we established that the majority of coefficients were stable over time ( Figure S5 ), but the mobility covariates (mobility change and park use) were highly variable, and mobility effects were greatest when cases were at their highest. As a result, we amended the baseline and green transmission models to include an interaction between the mobility variables and the number of cases (averaged over the nearest 7 days) in the local authority at a given moment in time (see Equation S1-2 for the final model structures). We also noted that effects were greatest in the first lockdown but also fluctuated substantially. We suspect the large effects are genuine, but given there were temporal and spatial issues with case-testing availability in the first lockdown, we opted to remove this section of data to limit biases. The final dataset covered June 1 st to November 30 th .

We standardised (subtracting values from their mean and dividing by their standard deviation) all predictor variables in the models to determine effect sizes and reduce multicollinearity where interactions are present. All model assumptions passed e.g. multicollinearity (variance inflation factors less than 3 within both the baseline and green transmission model), concurvity (observed and estimated concurvity less than 0.1), absence of spatial (Moran's I = 0.1) and temporal autocorrelation ( Figure S6 ), homogeneity of variance, and normality of residuals. When summarising results, we report the mean ± standard deviation, and when describing model outputs

Journal Pre-proof we report the standardised slope coefficient and 95% confidence intervals. We also report the R 2 for each model. All analyses was conducted in R 4.0.3 (R Development Core Team, 2020).

To understand how mobility patterns have influenced cases, we projected cases using the baseline and green transmission models under three scenarios: 1) cases under observed mobility patterns;

2) cases after a 20% reduction in each day's overall mobility; 3) cases after a 20% increase in each day's park use. We ran the baseline and green transmission models through each of the scenarios for every local authority between March 1 st and November 30 th 2020. We standardised all authorities so they had the same starting number of cases (10), community cases (10), and lagged case rate (0.58%; the mean case rate across local authorities on March 1 st ). These cases, community cases, and lagged case rate were updated and iteratively informed by the new model predictions, instead of the observed data. As a result, the projected case rates are solely influenced by the landscape structure and mobility patterns in the local authority. We converted the projected case rates into projected cases, against the starting case value of 10.

Across the 299 local authorities, case rates fluctuated substantially through time (Figure 4a ).

Mobility declined substantially during the first national lockdown in March to May, and in the run up to winter (Figure 4b ). During the summer months, mobility and the variance in mobility increased, and in some local authorities these increases were close to 100% (doubling mobility). In contrast, park use increased during the first lockdown and remained high (approximately 25% above baseline) until winter approached in October (Figure 4c ). There was less variation in park use trends between local authorities than in the mobility change metric. with park mobility regressed against overall mobility within each local authority i.e. are people visiting parks more than we would expect on a given date.

A reduction in mobility was associated with a decline in case rates, and changes in mobility had a larger impact when there was a higher number of average cases and when the population was more clustered (Table 2 ; Figure 5c , d). Population density and population clustering had no significant impact on case rates. Increases in the health index and proportion of the population over the age of 70 were both associated with significant decreases in case rates (Table 2 ; Figure 5a , b). This baseline transmission model had an R 2 of 0.45. 

Park use was associated with decreased residual case rates (Table 2; Figure 5e ) but the size of the effect depended on the availability of green space and how patchy it was. When patchiness was high and when there was a large amount of greenspace, park use had less of an impact on case rates, though was still associated with a significant reduction in cases. The green transmission model had a small R 2 of less than 0.01, despite the significant effects.

J o u r n a l P r e -p r o o f

Reducing mobility is a far more effective measure of limiting COVID-19 transmission than increasing park use ( Figure 6 ). Across local authorities between March and November 2020, a 20% reduction in mobility is projected to have led to 33% fewer cases on average, and in some cases more than 50% fewer cases (Figure 6b) . In contrast, a 20% increase in park use rarely exceeds a 10% reduction in cases (Figure 6c ). So whilst park use is associated with reducing COVID-19

transmission, these benefits would only be relatively small. However, these is spatial variation in these findings, with some areas potentially benefitting more than others from a reduction in mobility or increase in park use (Figure 7) . between March and November 2020 under each scenario by the cases in the observed mobility projection, multiplying this value by 100, and then subtracting by 100 (see Figure 5 ).

In this study, we attempted to quantify the effects of local green space on COVID-19 case rates after accounting for mechanisms known to influence pandemics in our baseline transmission model.

We found that high overall mobility was associated with increased case rates, especially when population clustering was high. After accounting for these variables, we found that higher park use, compared to other amenity areas, was associated with a reduction in case rates, especially in local authorities with low green space and with contiguous green space. These results suggest that utilising green spaces rather than carrying out other activities (e.g. visiting shops and workplaces) J o u r n a l P r e -p r o o f may reduce the transmission rate of COVID-19, but these benefits are limited compared to reducing mobility more generally.

From our baseline transmission model results, case rates were lower in local authorities with healthier populations and older populations (Figure 5a-b) . These results are logical, firstly as previous evidence has shown COVID-19 has a greater impact on those with underlying health conditions (Hamer et al., 2020; Jordan et al., 2020) and more severe cases may be more likely to be tested and reported. Secondly, whilst the eldery are more at risk of mortality from COVID-19 (Williamson et al., 2020) , this fact was widely reported in public health guidance and older people may have reduced contact with other individuals (Canning et al., 2020) Our baseline transmission model also shows that reducing mobility is most valuable when community cases are high and in areas with high population clustering (Figure 5c-d) . This is consistent with person-person contact as the major mechanism of transmission and appears to demonstrate the general effectiveness of lockdown measures in reducing case rates, as others have demonstrated previously (Davies et al., 2020; Lau et al., 2020) . Mobility had less impact in low clustered areas, which again may be expected, as people are more likely to be able to maintain distance and the potential number of interactions is reduced.

Once we had accounted for known drivers of case-rates, we investigated how landscape structure and park use (i.e. mobility in green spaces) affected residual case rates using the green transmission model. Here we found that using parks, relative to other types of mobility, was associated with a reduction in case rates ( Figure 5-6) . However, reducing overall mobility (i.e. mobility to all amenity areas) led to a far more substantial decline in case rates. For example, a 20% reduction was projected to reduce cases by c.35%, whilst a 20% increase in park use was projected to reduce cases by 5% to 10% ( Figure 6 ). This suggests that the use of parks may have modestly helped in reducing transmission rates in some areas during the pandemic, but reducing overall mobility is substantially more beneficial than maintaining mobility at pre-pandemic levels and spending that mobility in parks.

Journal Pre-proof the context of the local area was considered as using parks was beneficial in authorities with low green space and authorities with contiguous green space (Figure 5e-f and Figure 6 ). That park use has a minor beneficial effect overall seems to support our hypothesis that recreation in green space and parks may be safer than in other amenity areas. This is probably because it is easier to maintain distance and green spaces have fewer surfaces which could result in transmission if contaminated. However, the limiting impact of this when green space is high and accessible seems to suggest diminishing returns in how park use can impact COVID-19 transmission. This is perhaps not surprising if the main value of parks in this context is as an alternative to other relatively more hazardous amenity areas. Consequently, if there are other safe options outside of public parks then parks will likely have little impact. However, our findings do suggest that the use of public parks in a highly urbanised area may be advantageous, though as noted above the strongest effect was from the reduction of all forms of mobility. Therefore, cautiously, and given that it corresponds with common sense, we suggest that reducing mobility is a successful strategy for reducing case rates but given a need for some non-essential time outside of a home, using green spaces such as local parks may be the next best thing, particularly in highly urbanised areas.

A major limitation of the work is the difficulty in comparing across local authorities that vary simultaneously in many different variables likely important to case rates. This makes inference about the importance of their individual effects very difficult and so effect sizes should be interpreted cautiously and with caveat. Another challenge is that pandemics are rare events, consequently, our analysis covers only a snapshot of time for each local authority. During this period, many different factors not included in the analysis (e.g. chance super spreading events) may have explained much of the variation between local authorities. Despite this, the model fits are reasonably high. An additional limitation in our analyses is the absence of complete Google mobility data in some local authorities. We handled these missing values with imputation and attempted to ensure models were robust by comparing imputed models with complete-case models. Encouragingly, our completecase and imputed results are very similar, which suggests the imputation has not introduced any missing data bias (Johnson et al., 2021 )although both the imputation and complete-case analysis J o u r n a l P r e -p r o o f Journal Pre-proof could just be equally wrong. Given this uncertainty, and the further limitations we have identified above, our mobility findings should be interpreted cautiously.

Understanding the risks of different amenity areas could be important for longer-term management of COVID-19 and the landscape-dependency of this advice could be important for developing 'locallockdown' guidance. In particular, access to green spaces has been shown to have benefits for mental and physical well-being (Slater et al., 2020; Soga et al., 2020) , and consequently, understanding the relative risks of using these areas is important. Our results show that COVID-19 case rates may be reduced with individuals spending time in parks, relative to other amenity areas, especially in urbanised, high-density areas. Although further research is needed, these findings suggest that the use of parks for recreational activity in these contexts could be advisable, demonstrating a possible additional utility of these green spaces in addition to the known benefits to health and wellbeing (de Vries et al., 2003; Mitchell and Popham, 2007; Nutsford et al., 2013) in normal non-pandemic conditions. J o u r n a l P r e -p r o o f Green space per person (m 2 ). Derived by dividing total green space area in each local authority by the local authority's population size. We consider green spaces as any area meeting the following land cover types: broadleaved woodland, coniferous woodland, improved grassland, neutral grassland, calcareous grassland, acid grassland, fen, marsh and swamp, heather, heather grassland, and bog. We excluded agricultural land cover types as these were deemed a largely inaccessible/private land cover area. Source: Rowland et al. (2017) Static Patchiness Median frequency of parks within a 1km buffer around local authority houses. Source: Office for National Statistics (2021a) Static

Park use

Contribution of park use to the overall mobility metric, derived by extracting the residuals of a linear model between park mobility (response) and overall mobility (predictor) within each local authority. A positive residual value indicates park use exceeds what we would expect given park and overall mobility trends within the local authority. As with the mobility change variable, park use is averaged over the 2 to 12 days before each case rate. Source: Google (2020) Timeseries J o u r n a l P r e -p r o o f 

Exposure to neighborhood green space and mental health: Evidence from the survey of the health of wisconsin

The association between age, COVID-19 symptoms, and social distancing behavior in the United States

The effect of travel restrictions on the spread of the 2019 novel coronavirus (COVID-19) outbreak. Science

Global, regional, and national estimates of the population at increased risk of severe COVID-19 due to underlying health conditions in 2020: a modelling study

Access to green space, physical activity and mental health: A twin study

Effects of non-pharmaceutical interventions on COVID-19 cases, deaths

Natural environments -Healthy environments? An exploratory analysis of the relationship between greenspace and health

Nearby green space and human health: Evaluating accessibility metrics

Rethinking patch size and isolation effects: The habitat amount hypothesis

On the misuse of residuals in ecology: Regression of residuals vs. multiple regression

Spread and dynamics of the COVID-19 epidemic in Italy: Effects of emergency containment measures

Impacts of COVID-19 pandemic on urban park visitation: a global analysis

Google COVID-19 Community Mobility Reports

Lifestyle risk factors, inflammatory J o u r n a l P r e -p r o o f mechanisms, and COVID-19 hospitalization: A community-based cohort study of 387,109 adults in UK

Handling missing values in trait data

Covid-19: Risk factors for severe disease and death

Utilizing general human movement models to predict the spread of emerging infectious diseases in resource poor settings

The effect of human mobility and control measures on the COVID-19 epidemic in China

The positive impact of lockdown in Wuhan on containing the COVID-19 outbreak in China

Statistical Analysis with Missing Data, Statistical analysis with missing data Second edition

Green space, urbanity, and health: How strong is the relation?

China coronavirus: WHO declares international emergency as death toll exceeds 200

Disease , habitat fragmentation and conservation

Greenspace, urbanity and health: Relationships in England

An ecological study investigating the association between access to urban green space and mental health

Local authority ageing statistics, based on annual mid-year population estimates

Population estimates for the

R Development Core Team, R: a language and environment for statistical computing

Presenting Characteristics, Comorbidities, and Outcomes among

Association between urban green space and the risk of cardiovascular disease: A longitudinal study in seven Korean metropolitan areas

Accessibility and allocation of public parks and gardens in England and Wales: A COVID-19 social distancing perspective

Recommendations for keeping parks and green space accessible for mental and physical health during COVID-19 and other pandemics

A room with a green view: the importance of nearby nature for mental health during the COVID-19 pandemic

Defining greenspace: Multiple uses across multiple disciplines

UK Government, 2020a. The Health Protection (Coronavirus, Restrictions) (England) (Amendment) (No. 4) Regulations 2020

UK Government, 2020b. UK gridded population 2011 based on Census 2011 and Land Cover Map

Unemployment by ethnicity

Unemployment by ethnicity: Unemployment by local authority

MICE: Multivariate Imputation by Chained Equations in R

Health benefits of green spaces in the living environment: A systematic review of epidemiological studies. Urban For

Urban nature in a time of crisis: Recreational use of green space increases during the COVID-19 outbreak in Oslo

Factors associated with COVID-19-related death using OpenSAFELY

A human behavior integrated hierarchical model of airborne disease transmission in a large city

Baseline

☒ The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work

☐The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: No competing interests of park use and green spaces on COVID

Modelled transmission against mobility and landscape variables in England

Reducing mobility and shifting mobility outdoors may reduce transmission