Effect of park use and landscape structure on COVID-19 transmission rates

Introduction

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 implemented social distancing measures and a nationwide lockdown to control the spread of the virus (Gov, 2020). During this period, the general public was limited in the distances they could travel and the number of times they could leave their residence each day; with an allowance of one non-essential trip during the peak of transmission (Gov, 2020). This restriction meant that members of the public were unusually reliant on amenity spaces close to their residences for daily exercise and/or recreation. 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 in England 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 ecosystem services, benefiting both mental and physical health (Beyer et al., 2014; Cohen-Cline, Turkheimer, & Duncan, 2015). These benefits are usually considered in terms of reducing the prevalence or severity of conditions such as mental stress (Nutsford, Pearson, & Kingham, 2013) or cardiovascular disease (Seo, Choi, Kim, Kim, & Park, 2019). Green spaces could also have important effects on an acute stressor such as a disease pandemic, though, to date, these effects have not been well explored. COVID-19 has properties that make it a unique case-study of the influence of green space on health outcomes during a disease pandemic as 1) lockdown measures meant that access to green space was necessarily local, and so any relationship between green space and case-number was less likely to be confounded by travel outside of a local authority; 2) the disease was prevalent nationwide, affording a comparison between a large number of local authorities that differed in both case-numbers and green space access.

Possible effects of green space on COVID-19 incidence could be from two causes: 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, Verheij, Groenewegen, De Vries, & Spreeuwenberg, 2006; Mitchell & 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, Kivimäki, Gale, & Batty, 2020; Jordan, Adab, & Cheng, 2020), green space may, therefore, indirectly provide some level of resilience to the disease and/or reduce incidence. However, as COVID-19 is a novel highly infectious disease, the main effects of green space are likely to be related to transmission probabilities. So far, evidence suggests that COVID-19 is transmitted via droplet infection, contact with contaminated individuals or surfaces, and potentially through aerosol transmission (Bahl et al., 2020). Green spaces are amenity areas where these risks are likely minimised as they are typically less spatially confined and have fewer surfaces prone to frequent touching or contact. Consequently, incidence could be higher in areas where green space is limited as recreation, however limited, may take place in alternative amenity spaces such as retail areas, with larger densities of people and more frequent contact with surfaces.

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 & 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 & Dobson, 2002). However, 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 (Zhang, Huang, Su, Ma, & Li, 2018; Gatto et al., 2020), 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 mobilty, 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.

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 local authority cases in England, UK, we explore how both green space use and access influence COVID-19 incidence, which we define as the pre-peak rise, and post-peak drop in cases. Our approach is to first construct a baseline health 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 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).

• Download figure
• Open in new tab

Figure 1.

Mechanisms by which green space and patchiness could interact with park use to influence COVID-19 transmission. The upper two rows describe the primary predictions, whilst the bottom row explains alternate predictions. All variables (e.g. population density) except green space and patchiness, respectively, are held at a constant in these predictions. Green circles with a tree icon indicate the presence of green space. Dotted lines indicate walking routes, which becomes park use when the line overlaps a green space. The green health symbol indicates that the landscape metric and park use is beneficial, whilst the red toxic symbol indicates a risk.

Methods

Modelling

We developed three core model types (Figure 2): baseline transmission – aimed at controlling for the major mechanisms influencing case-rate changes; green transmission difference – impact of landscape structure and park use on case rates; and park use – impact of landscape structures on park usage, providing a mechansistic understanding of the green transmission difference model e.g. if park use is important, what influences park use. The baseline transmission and green transmission difference 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 affect disease transmission (e.g. population density). As a result, we structured our analyses to first account for the presence of these more key variables in a baseline transmission model, and then in the green transmission difference 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 per person are derived in similar ways), and resulted in multicollinearity issues.

• Download figure
• Open in new tab

Figure 2.

Model structure for baseline transmission, green transmission difference, and park use models, depicting the process for developing the response variables, as well as the predictors used in each model. In the green transmission difference models, the red text represents the main model predictions.

We z-transformed all predictor variables in the models to determine effect sizes and reduce multicollinearity where interatctions are present. Each of the models described below was repeated for the pre- and post-peak period, and all model assumptions passed e.g. multicollinearity, absence of spatial autocorrelation, homogeneity of variance, and normality of residuals. When summarising results, we report the mean ± standard deviation, and when describing model outputs we report the following: coef = standardised slope coefficient, se = standard error of slope coefficient, df = model degrees of freedom, t = model test-statistic, and p = significance value. We report the model accuracy using two metrics described in Mayer (2020): _wR² – a weighted r-squared which describes the proportion of deviance captured by the model; and _wmae – the weighted mean absolute error. Both of these metrics describe the accuracy of the model predicted values, and were derived within the dataset, rather than through an independent sub-sample. These metrics cannot indicate which aspect of the model contributed most to the prediction accuracy (e.g. model accuracy will be influenced by the fixed effects, random effects, the spatial correlation structure, and weighting structure).

Baseline transmission models

To control for the baseline health and transmission mechanisms influencing COVID-19 case rates, we developed linear mixed effect models, with case rate (pre- or post-peak) as the response. We included the following predictor parameters: health, demography, economy, population density, population clustering, and mobility change. We also included interactions between population density and clustering, population density and mobility change, and population clustering and mobility change. Lastly, in the post-peak case-rate models, we wanted to control for instances where the case-rate drop was steeper because it had a higher starting point. As a result, we included the maximum daily case count within the model, as well as pre-peak mobility change which was found to be influential in the pre-peak case rate model.

We allowed the model to vary by region with a random intercept to account for the non-independence of some regions sharing mobility data. We also set an exponential spatial correlation structure (using the local authorities’ latitude and longitude centroid) to account for the spatial autocorrelation of neighbouring local authorities sharing similar case rates. Case rates were also weighted according to their variance (i.e. the variance of the slope in the log-linear case rate regressions), such that areas where pre- and post-case rates were less certain contributed proportionally less to the model fit. We extracted the residuals from these models for the green transmission difference models.

It should be noted, even though variance inflation factors were suitably low for the baseline transmission models, the precision of the coefficients may be reduced as many predictors were correlated. Our justification for retaining these predictors in the baseline model, is that the primary goal was identifying the impact of green space use after controlling for other factors, and multicolinearity has no impact on predictive performance. However, to support inference and explanation, we also report a simplified version of these baseline transmission models in Table S1, where the health, demography and economy variables were removed.

Green transmission difference models

To assess how landscape structure and park use influenced residual case rates, we developed linear mixed effect models, with pre- or post-peak residual case rates form the baseline transmission models as the response, as well as the following predictor parameters: park use (inverse hyperbolic sine transformed), green space (log10 transformed), patchiness, an interaction between park use and green space, and finally an interaction between park use and patchiness. These models also included region as a random intercept, but we did not control for the spatial structure of the data, which was addressed in the earlier baseline transmission model. We did, however, still include the variance weighting to specify that residuals from local authorities with a smaller variance in case rates should contribute more to the model.

Park use models

To understand how landscape structure influences park use and in turn, clarify the mechanisms by which park use influences case rates, we developed linear mixed effect models, with pre- or post-peak park use (inverse hyperbolic sine transformed) as the response, and the following predictor parameters: gardens, green space (log10 transformed), patchiness, and an interaction between green space and patchiness. We also included the region random intercept and a spatial correlation structure for the same reasons as in the baseline transmission model.

Results

Across the 98 local authorities, the pre-peak cases rose 9.8% (± 4.6) per day and peaked between 24^th March and the 8^th May. The peaks ranged in size from 9 to 132 daily cases, and then post-peak, the daily case rate declined at −4.8% per day (±2.2). Mobility reduced substantially in the pre-peak period (−28.3% ± 8.5) and further again in the post-peak period (−37.0% ± 6.4), but not equally across England, with mobility increasing in 20.4% of local authorities between the pre- and post-peak period (Figure 3). In the pre-peak period, park use ranged from −19.7% to 22.4%, but post-peak this range increased from −43.3% to 30.8%. Between this pre- and post-peak period, some authorities altered their park use, for example, park use decreased by 38% in Peterborough, whilst park use increased by 22% in Luton.

• Download figure
• Open in new tab

Figure 3.

Change in a) pre-peak case rate, b) post-peak case rate, c) pre-peak mobility, d) post-peak mobility, e) pre-peak park use, and f) post-peak park use, across 98 local authorities in England. X and Y axes indicate longitude and latitude in decimal degrees.

Baseline transmission models

Pre-peak case rates were lower in areas with a reduction in mobility (coef = 2.43, se = 0.51, df = 80, t = 4.56, p < 0.001; Figure S1). Population density had no effect on pre-peak case rates. Population clustering had a weak effect when interacting with mobility change, where the effect of mobility change diminished when population clustering was low (coef = 1.00, se = 0.56, df = 80, t = 1.76, p = 0.08; Figure 4a) - this interactive effect was lost in the simplified baseline transmission model (Table S1). Given the variation in pre-peak case rates (Figure 3a), where daily case rises ranged from 2 – 21%, the model error was reasonably low (_wmae = 2.45% the _wR² = 0.4).

• Download figure
• Open in new tab

Figure 4.

Marginal effects of important interaction parameters in the baseline transmission models. Marginal effects are held at zero for all other parameters as variables were z-transformed. In panels a and b, population clustering was set at 0.35 (Low) and 0.7 (High), where 0 indicates a random distribution of clustering, and 1 indicates a complete separation in clustering. Error bars represent the 95 confidence intervals.

The post-peak case rates were also lower in areas with a reduction in mobility (coef = 0.81, se = 0.15, df = 78, t = 5.52, p < 0.001; Figure S2a), and we observed a similar interaction between population clustering and mobility change (coef = 0.24, se = 0.12, df = 78, t = 2.0, p = 0.048; Figure 4b), as in the pre-peak model. A similar result was found for the interaction between mobility change and population density, where the effect of mobility increasing post-peak case rates diminished at low population densities (coef = 0.59, se = 0.14, df = 78, t = 4.18, p < 0.001; Figure 4c). Population density also interacted with population clustering, with case-rates declining as clustering increases, (coef = −0.29, se = 0.14, df = 78, t = −2.11, p = 0.04), but only when population density was low (coef = 0.42, se = 0.15, df = 78, t = 2.81, p = 0.006; Figure 4d). Post-peak case-rates were also influenced by pre-peak activity, as rates were lower in areas with a reduction in pre-peak mobility (coef = 0.34, se = 0.14, df = 78, t = 2.40, p = 0.019; Figure S2b), and areas with a larger peak in cases had a faster decline in post-peak case-rate (coef = −0.53, se = 0.10, df = 78, t = −5.16, p < 0.001; Figure S2c). The model error in post-peak case rates was lower than in the pre-peak case rates, with a moderately high post-peak prediction accuracy (_wmae = 0.69% the _wR² = 0.61).

Green transmission difference models

Park use decreased residual pre-peak case rates (coef = −0.57, se = 0.29, df = 84, t = 1.99, p = 0.049; Figure 5a), but there was an interaction between park use and green space, with the effect of park use diminishing when green space was high (coef = −0.91, se = 0.35, df = 84, t = 2.59, p = 0.01; Figure 5b). There was also an interactive effect between park use and patchiness, as park use decreased the residual pre-peak case rate when patchiness was low, but increased the case rate when patchiness was high (coef = −0.66, se = 0.32, df = 84, t = 2.03, p = 0.04;; Figure 5c). Green space and patchiness had no general effect on residual pre- or post-peak case rates, whilst park use and the interactive effects also had no effect on residual post-peak case rates (Table S2). Nevertheless, incorporating these park use and landscape structure effects led to a modest decrease in case rate prediction error; _wmae decreased from 2.45% to 2.34% in the pre-peak period, and from 0.69% to 0.655% in the post-peak period. This equated to more dramatic improvements in the _wR², where accuracy jumped from 0.42 to 0.5 in the pre-peak period, and from 0.61 to 0.68 in the post-peak period, which represent a 19% and 11% accuracy increase, respectively.

• Download figure
• Open in new tab

Figure 5.

Marginal effects of a selection of parameters in the green transmission difference (a-c) and park use (d-f) models. All other parameters are held at their mean (zero) for the marginal sffects. In panels d-f, the y-axes are inverse hyperbolic sine scaled, as are the x-axes in panels a-c. In panels e & f, the x-axis is log10 scaled. Error bars represent the 95 confidence intervals.

Discussion

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 increased both pre- and post-peak case rates, especially when population clustering and density were high (this mobility-density interaction effect was only important in the post-peak period). After accounting for these variables, we found that higher park use, compared to other amenity areas, led to a reduction in pre-peak case rates, especially in areas with low and contiguos green space. These results highlight that utilising green spaces rather than carrying out other activities (e.g. visiting shops and workplaces) can reduce the transmission rate of COVID-19, especially during an exponential phase of transmission.

From our baseline transmission model results, population density (Figure 4c) and clustering (Figure 4a, b) were only important when considered in an interaction with mobility, and population density was only important in the post-peak period. This is surprising, but is consistent with person-person contact as the major mechanism of transmission i.e. even a very dense and clustered population will have slow transmission rates if there are few interactions between people. This 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). However, we found mobility had less impact in low density, 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. However, even in these conditions, mobility still slightly increased post-peak case rates and so lockdown still appeared to have some effect on reducing case rates even in low density and low clustered areas.

Previous research has shown the importance of health, deprivation, demography and economic prosperity on COVID-19 case rates (Abedi et al., 2020; Atkins et al., 2020; Dowd et al., 2020). However, in our baseline health transmission models, these variables did not explain differences in COVID-19 case rates found in our dataset. This is likely due to the variables being somewhat correlated with population density, with consequent reductions in the information available in the data to assess their effects (as described in Methods). For example, an older population is correlated with decreased population density and though older people seem to be more vulnerable to COVID-19 (Jordan et al., 2020), they are more likely to reside in safer (low population density) areas; consequently separating these effects may simply not be possible given the data used. Another consideration between findings for demographic/social groups at a national scale and local case rates, is that the assemblages being compared contain varying proportions of the groups in question i.e. we ask ‘does an older population lead to higher local case rates?’, not ‘are there differences in the per-capita rate of cases between different demographic, economic, and health groups?’. We suggest that our analysis is unsuited to resolving questions of the second type and that the baseline transmission model is best viewed simply as controlling for possible factors in an attempt to isolate the effects of green space.

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 other types of mobility, led to a reduction in pre-peak case rates (Figure 5a). However, reducing overall mobility (i.e. mobility to all amenity areas) led to a far more substantial decline in both pre- and post-peak case rates, especially in dense and clustered populations as found in the baseline transmission model. For example, continuing mobility at pre-pandemic levels led to 15.4% daily increase in cases when park use was at its highest, compared to a 17.4% daily increase when park use was at its lowest - a 2% case rate reduction. In contrast, halving overall mobility reduced case rates by 13.6%. This suggests that the use of parks may have modestly helped in reducing transmission rates in some areas during the increasing phase of the pandemic, but reducing overall mobility is substantially more beneficial than maintaining mobility at pre-pandemic levels and spending that mobility in parks.

Whilst park use, overall, had a relatively small effect, we did note stronger effects of park use when the context of the local area was considered as using parks was beneficial in authorities with low and contiguos green space (Figure 5b and 5c). 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 or shopping 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.

As we demonstrated the importance of park use over other types of mobility in reducing the exponential rise in cases, we then aimed to understand what green space parameters (area and patchiness of green space and access to gardens) influenced pre- and post-peak park use. We found that, overall, accessible green space marginally predicted an increase in pre- peak park use (Figure 5e), but not post-peak (Figure 5f). We also noted an interaction between use of parks and the amount and patchiness of the green space (Figure 5d – but should be f). We found that if an area had high green space, but it was not patchy, then it was used less, but that in low green space areas patchiness had little effect. A possible explanation for this is that in less green areas overall, large areas of concentrated green space act as a stronger attractor for a certain kind of experience for members of the public. For example, a large park in a heavily urbanised area provides a walking experience that may not be readily substituted by walking through other local areas, whereas a walk in an overall greener area may be more comparable to that in a park. This is supported by evidence that suggests greater phychological benefits when walking in green parks as opposed to urban streets (Johansson, Hartig, & Staats, 2011), as well as evidence highlighting the combined health benefits of physical activity within green spaces (Pretty et al., 2007). Consequently, as would be predicted from basic considerations of scarcity, local green space provision seems to have more consistent value in areas where there is less green space independent of its structure, however, overall, accessibility seems to be a driving factor of usage.

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. As mentioned previously, this makes inference about the importance of their individual effects very difficult, or simply not possible. Therefore, we reiterate that our results do not provide evidence that the demographic and social groups included are not more or less affected by COVID-19, and we suggest that any findings from studies directly addressing questions about these groups are given priority. 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, especially after incorporating the green transmission models. However, the modest beneficial effect of park use on COVID-19 transmission could be useful in the general attempt to develop guidance for which spaces to use during an exponential phase of transmission.

Understanding the risks of different amenity areas could be important for longer-term management of COVID-19 and in particular, the landscape-dependency of this advice could be important for developing ‘local-lockdown’ guidance. Our results show that spending time in parks, relative to other amenity areas can reduce COVID-19 case rates 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, Verheij, Groenewegen, & Spreeuwenberg, 2003; Mitchell & Popham, 2007; Nutsford et al., 2013) in normal non-pandemic conditions.

Data accessibility

Code and data to repeat analysis are presented in https://github.com/GitTFJ/COVID19_parks_landscape

Author contributions

All authors contributed to project design. Analysis was lead by TFJ and LCE, but all authors contributed. TFJ and LCE co-wrote the first draft and co-authors contributed to revisions.