Green blue space exposure changes and impact on individual-level well-being and mental health: a population-wide dynamic longitudinal panel study with linked survey data

Green space

Reviews have shown that exposure to green space can promote life satisfaction,24 aid psychological restoration25 and mitigate the impact of stress. 26 Views of green space from the home have been linked to measures of subjective well-being, and there are relationships between positive mental health in adulthood and overall greenness of the living environment,5 residential proximity to GBS27,28 and frequency of green space visits. 29,30 However, these relationships are not necessarily straightforward. The complexity of characterising exposure to green spaces and the pathways through which they affect mental health means that causal relationships are not always clear. 8,31 Most caution that the evidence is currently limited and often inconsistent. 8

Factors such as age, sex, physical activity, ethnicity and socioeconomic status appear to moderate relationships between exposure to green space and mental health outcomes in adulthood. 27,32,33 The associations between greener living environments and mental health appear to vary over the life course34 and there is little consistency according to gender. 35 Green space may be of greater benefit to people from disadvantaged groups and living in more socioeconomically deprived areas compared with those in less deprived areas. 8,27 Again, this is not a straightforward relationship and is not consistently statistically significant in studies investigating the effects of socioeconomic status. 31 Less is known about interactions between ethnicity and the salutogenic effects of green space exposure, but a UK birth cohort study showed that ethnicity moderated relationships between residential green space and mental well-being. 33

Blue space

The hazardous qualities of blue spaces are well known, but their salutogenic public health implications have, until recently, received less scientific attention (and less than those of green space). 36,37 There is a growing body of evidence that exposure and proximity to blue space is linked to positive mental health and well-being. 38 Living near the coast is shown to have general mental health benefits. 4,39 There is also evidence that people living near (or with views of) the coast experience lower psychological distress. 40,41 Epidemiological studies do not typically distinguish between coastal and freshwaters,41 but there is some evidence that proximity to inland water also yields positive mental health effects. 42,43

Mediators and moderators of relationships between blue space and mental health are less well understood than those for green space, but it is likely that similar factors apply. For example, a 2019 study in England showed that the mental health benefits of residential coastal proximity appear to moderate the effects of low socioeconomic status. 44 Physical activity also appears to be a mediator between positive mental health and coastal proximity. This effect is limited to terrestrial physical activity near the coast rather than ‘watersports’. 43

Quantity, proximity, quality and type of green and blue space and mental health and well-being

The amount of green space in an area has been linked to good population mental health4,45 and reduced inequalities in health in the UK. 44 However, there is insufficient evidence to dissect how type, proximity, quantity and ‘quality’ (either ecological or in terms of maintenance) or ‘qualities’ (the characteristics) of green or blue space each influence population-level mental health and well-being. 46 Most available evidence linking GBS to health outcomes considers relationships at a local area level, typically the total amount of green space around residences, with no assumption about type, quality or usage. 5

The type, quality, and qualities of GBS may be significant in the context of mental health benefits. Individual studies suggest that environment (or landcover) type is an important consideration for general and mental health. 3,39,47,48 In urban settings, gardens and allotments are notably beneficial for mental health. 49,50 The impact differs according to the type of garden, with private spaces appearing to be most strongly associated with better outcomes. 51 In other studies, proximity to specific types of environments, such as coasts, have been found to be linked to better mental health. 44

The impacts of GBS ‘quality’ on mental health and well-being are more difficult to assess, given the breadth of meaning that can be attached to this term. We consider ‘quality’ as the state, whether ecological or physical, of the green or blue space. ‘Qualities’ refers to the characteristics, or composition of the space. GBS ‘quality’ and ‘qualities’ are multidimensional concepts. They may be defined and assessed subjectively (e.g. perceived ecological quality, perceived safety) or objectively (e.g. public facilities, air pollution levels) in a number of ways. Evidence of links between indicators of GBS quality and mental health is limited. 27,52 However, there are some studies supporting ‘quality’ and ‘qualities’ of GBS as important components for mental health and well-being impacts. General satisfaction with local green space is important,33,53 and perceptions of safety and maintenance determine people’s use of green space. 54,55 Public accessibility and related urban planning influence the likelihood of the use and resulting experience of spaces, impacting potential health gain. 53,56 The relative biodiversity of environments may also play a role, particularly in terms of the types of nature people experience,57,58 but evidence is limited and inconsistent. 59

Overall, the type, proximity, quality, and qualities of GBS are likely to be important in the context of mental health benefits. However, current evidence is inconsistent for relationships between the: quantity,5 type60 and quality and qualities of GBS,27 and mental health and well-being.

Use of green and blue spaces

The studies mentioned previously often make no assumption about usage of the GBS. Usage of a space is theorised to be a factor in potential benefit;61 however, actual visits to, and time spent in, GBS are less well studied. 5 Studies have, however, shown that recreational use of GBS is associated with multiple mental health outcomes. 62,63 Further, there are indications that minimum weekly baselines of time spent in green spaces are linked to well-being and good mental health. 63 Whether or not physical activity mediates benefit has been investigated in a number of studies, with mixed results. 64

System-wide impacts of green and blue spaces

Additional pathways may link the presence of GBS to mental health. These include the mitigation or avoidance of the effects of urban heat islands, air and noise pollution, and flooding. 46 Although evidence is limited in terms of demonstrable associations of these pathways with better mental health,65 some linkages have been demonstrated; for example, green spaces were found to be associated with reduced psychological distress from local noise pollution. 66 Thus, there are likely to be multiple benefits of GBS, impacting our physical and mental health through indirect mechanisms within wider ecological systems.

Inequalities

There are inequalities in the distribution and accessibility of good quality GBS in the UK. People in deprived areas, minority ethnic communities, elderly people and those with long term poor health and disabilities often have less (physical) access to good quality green spaces. They tend to use green spaces less and are more likely to have negative perceptions regarding their usage of such spaces. 67–69 These groups also tend to have poorer health outcomes. 70,71 However, there is cross-sectional evidence suggesting that health inequalities are lessened with greater proximity to green or blue spaces. 44,72,73

National population dataset objectives

• Create a longitudinal dataset of residential GBS exposure data for all homes in Wales using UK Ordnance Survey (OS), local authority and remotely sensed satellite data.

• Create an 11-year dynamic cohort of individual-level longitudinal residential GBS exposure to answer the questions: ‘Is a greater residential exposure and access to GBS associated with the likelihood of seeking help for a CMD in general practice?’ and ‘Is the association between exposure, or access to, GBS and seeking help for a CMD modified by socioeconomic disadvantage and moving home?’

Survey dataset objectives

• Create longitudinal, environment, health and demographic individual-level data linkages between survey and routine data within the SAIL databank.

• Utilise the data-linked National Survey for Wales (NSW) to answer the question: ‘What is the impact of GBS residential exposure, modified by GBS use and multiple socio-physical modifiers, on well-being and CMD?’

The national population dataset uses routinely collected health and demographic data (data collected for purposes other than research). The survey sample dataset links routine data in the SAIL databank with in-depth survey responses from the NSW. Both consider socioeconomic health and well-being inequalities.

Synthesising objective

To translate results into policy recommendations for government and disseminate results from health outcome and economic impact assessment to academics, stakeholders, and policy makers.

Ethics considerations

We used data available in the SAIL databank at Swansea University, Swansea, UK. 93,94 All proposals to use SAIL data are reviewed by an independent information governance review panel. The panel considers each project to ensure proper and appropriate use of SAIL data. If access is approved, it is gained through the SAIL gateway, a privacy-protecting safe haven and remote access system. This study was approved by the panel as project 0562.

Cohort creation

The cohort was created using data from the WDSD linked to:

• Annual District Death Extract: mortality information including the date and cause(s) of death from the Office for National Statistics (ONS) mortality register.

• Patient Episode Database for Wales: demographic and clinical data on all inpatient and day case admissions to NHS Wales hospitals and Welsh residents treated in other UK nations.

• Welsh Longitudinal General Practice (WLGP): information on symptoms, diagnoses, prescriptions and referrals collated from clinical information systems in use at each general practice in Wales.

• Welsh Index of Multiple Deprivation (WIMD, 2011 and 2014), the Welsh Government’s official measure of relative deprivation for small areas in Wales. 95

• Rural–urban ONS classifications (2011) at lower-layer super output area (LSOA). 96

• NSW, an annual, repeat cross-sectional survey of about 12,000 adults in Wales.

The cohort, designated the GBS e-cohort, is held in the SAIL databank, a privacy-protecting trusted research environment providing secure storage of anonymised, person-based demographic, health, social and education data for the population of Wales. 93,94

Environmental exposures

Chapter 4 describes environmental exposures. In brief, we derived measures of ambient residential greenness exposure and potential access to GBS for all residences in Wales using:

1. Satellite data (Landsat™ 2008–2019)97 to create annual greenness densities of mean enhanced vegetation index (EVI) and normalised difference vegetation index (NDVI) within 300 m of each residence.

2. OS MasterMap (OSMM) Topography Layer98 (2018) to capture natural and man-made features, including the outline of homes and parks.

3. OSMM-derived Greenspace dataset (2018). 99

4. Local authority technical advice notes (TAN 16), legally required records of data on sport, recreation and open spaces managed by local authorities. 100

5. Open-source portal data from the Lle Geo-Portal (forestry, urban tree cover). 21

6. OpenStreetMap road/footpath data. 101

Ambient greenness within 300 m of each residence was measured yearly and linked to the cohort in three-component variables that represent ambient residential green/blueness:

• Mean EVI (minimum, mean, median, max)

• Mean NDVI (minimum, mean, median, max)

• Coastal and/or inland water (yes/no)

The potential for an individual to access a range of types of GBS along a network of paths and roads within 1600 m of each home was modelled for 2018. Access measures have multiple variables available at three (nested) levels of detail, following consultation with stakeholders (see Chapter 3):

• Level 1 divides green spaces into amenities, functional spaces, seminatural habitats, enclosed and linear spaces.

• Level 2 provides more detail on the type of space; for example, identifying a recreation space.

• Level 3 contains greater detail; for example, identifying play areas within a recreational space.

Ambient residential greenness, and potential to access GBS, were augmented by survey responses about leisure time visits to outdoor spaces in Wales for the NSW subgroup.

Changes in access and exposure to GBS could be either at the environment level (e.g. a change in the GBS itself), or at the individual level, such as through moving home. 83

Survey data

A cohort subgroup responded to NRW questions in the 2016–17 and 2018–19 NSW. 102 The NSW is an annual repeat cross-sectional, government-sponsored omnibus survey of a representative sample of the population of Wales (annual n ~12,000). Topics include education, culture, health and well-being, and more detailed information on socioeconomic circumstances than found in administrative data. The NRW questions102 ask whether respondents visited outdoor spaces in Wales, including time spent outdoors on leisure activities, and types of activities undertaken. The NSW respondents aged 16 years and over who consented to NSW-administrative data linkage (> 90%) were linked to the cohort.

Mental health outcomes

General practice events

A daily count of all GPE per person (e.g. appointments, tests), not limited to those relating to CMD, were extracted from WLGP and aggregated quarterly. This eliminated counting multiple test results. These events are not only mental health-related, so reflect a general need for primary health-care services.

Data linkage

A strength of the SAIL trusted research environment is the split-file process for anonymising all individuals and households in Wales. This overcomes confidentiality and disclosure issues arising in health-related data warehousing. 93 The data sources described above were split at the source organisation into demographic (identifiable) data and clinical (non-identifiable) data. A temporary system linking field ensures the data can be rejoined later in SAIL. 93 The identifiable component was sent to our trusted third party, Digital Health and Care Wales, who anonymised and encrypted the identifiable data, assigning a unique anonymised linking field (ALF), the primary key variable for record linkage, and a residential ALF (RALF) based on the unique property reference number of their place of residence. Data were rejoined in SAIL using the temporary system-linking field, and the ALF was further encrypted before being made available to project statisticians. Detailed information on the matching technique used to assign an ALF to each individual across diverse administrative datasets has been published elsewhere. 93 Briefly, a matching algorithm is applied in sequence. Records with valid NHS numbers are accepted, deterministic record linkage is then carried out on a set of matching variables and, finally, the probabilistic record linkage is carried out on the remaining unmatched records. 93

The environmental exposure indices created within a secure geographic information system (GIS) platform (the Secure eResearch Platform) were linked at household level with demographics from WDSD and NSW using the RALF for record linkage. This approach has been successful in a range of public health studies. 10,108–110 The dates of each change of address are recorded in the WDSD, providing duration of residency and the ability to link our time-varying environmental exposures to each residence and LSOA as individuals move home.

WDSD contains dates of change of address provided by patients to their GP. We aggregated these data quarterly, so individuals can have multiple RALFs within a three-month period. We selected one RALF per quarter using detailed rules to identify and rank RALFs (Appendix 1, including Table 22) to assign one set of environmental metrics to an individual each quarter. Briefly, a RALF confidence score (low/intermediate/high confidence in address match) is provided within the WDSD. To determine the most highly ranked RALF within the quarter, this measure of linkage confidence is combined with the total duration (start date to end date) within that RALF. RALFs with higher linkage confidence and which are assigned to the individual for longer are given priority over RALFs with lower confidence and shorter durations. When more than six RALFs are linked to an individual within a three-month period, the quality of data is considered unreliable, and no RALF is assigned during that period.

Cohort variables

The variables available for the main cohort include: (1) sociodemographic and economic characteristics; (2) CMDs/well-being; (3) comorbidity index; (4) social environment and life events (births/deaths in the household); (5) environmental metrics; and (6) other administrative cohort information (Table 1). NSW variables are described in Chapter 7.

Analysis platform

The SAIL databank operates on the UK Secure Research Platform. 110 All analyses were conducted in RStudio (version 1.4.1103) with R (version 4.1) and MLwiN (version 3.05) through a remote desktop protocol within the SAIL gateway. 111

Cohort description

The cohort inclusion criteria were all persons aged 16 years and over registered with a GP providing patient records to the SAIL databank between January 2008 and October 2019, giving a cohort of 2,800,483 individuals. Exclusion criteria were not being registered with a GP providing data to the SAIL Databank, not having a Welsh residential address between January 2008 and October 2019, and not having sex or week of birth recorded in WDSD. The WDSD contains demographic characteristics of people registered with a free-at-the-point-of-care GP in Wales providing data to the SAIL databank (80% population coverage). 112

We intentionally removed people who did not fit with study criteria from the total people recorded in WDSD (Figure 1). This comprised those who did not reach the age of 16 years during the study period, who died or moved to a home outside of Wales before January 2008 or moved out of Wales before turning 16 during the study period. Additionally, we excluded those missing key demographic data (sex, week of birth; n = 12,210). We further excluded 839,063 individuals who had missing data; for example, they were not registered with a GP providing data to the SAIL databank, did not have a Welsh residential address between January 2008 and October 2019 or did not have sex or week of birth recorded in WDSD. Those who meet the criteria after the study start are censored until their qualifying entry date and those who exit before the study finishes (through death or migration) are right censored.

FIGURE 1.

Cohort construction of the 2,800,483 individuals using the demographic dataset (WDSD) and linkage to the WLGP dataset.

Responses to NRW questions in the 2016–17 and 2018–19 NSW were available for a cohort subgroup of 5312 individuals.

The cohort has 24.9 million person years of follow-up. An additional average of 30,238 people joined the cohort annually through migration into Wales, and an average of 34,709 people by reaching age 16 years, totalling 710,570 (25%). Annually, an average of 22,987 people died and 1603 moved out of Wales permanently, totalling 294,437 (10.5%). The cohort sample was broadly representative of the Welsh population compared with 2011 Census figures (Table 2). 113 More than one-fifth (22.6%) of the adult population in the most deprived quintile moved home at least once during the cohort period 2008–19, with fewer moving in the least deprived (18.7%) and next least deprived (18.2%) quintiles (Table 2). Younger people (< 30 years) and those living in the most deprived areas had the highest prevalence of moving at least once during their time in the cohort (44.9% and 22.6%, respectively; Table 2). To check for potential bias, area-level characteristics of residences linked and unlinked to the cohort were compared. The percentage of unlinked homes did not increase with deprivation. However, a higher proportion of unlinked homes were in rural areas (Table 3).

Literature review

We reviewed academic publications, grey literature and policy documentation to collate GBS definitions and typologies. We searched ISI Web of Science databases up to 8 August 2018 using used the same search criteria as a review of green space definitions between 2009 and 2014. 23 We excluded conference proceedings, reports and planning policy documents. We extracted GBS definitions used by UK studies to align with our study setting and noted the geographic location of the study. Where we could not identify an overt definition, we included any examples used in the papers. In addition, we reviewed the main policy literature used in Wales, including the Fields in Trust guidelines, Recreation Opportunity Spectrum, Future Landscapes and WHO.

Stakeholder workshop

In October 2018, we held a semistructured, in-person public involvement workshop for stakeholders that manage GBS and encourage GBS use in Wales. The workshop focused on challenges of existing definitions and typologies, potential solutions and implications for working with GBS and data, defining GBS and use of evidence. We invited planning professionals from the 22 Welsh unitary authorities (‘local government’), representatives from Welsh Government and arm’s length bodies, and individuals involved with third-sector charities. A project summary was shared prior to the event. All participants were offered remuneration in line with INVOLVE guidelines.

Attendees were purposefully allocated to four groups to provide a mix of job roles and expertise. We introduced the project aims and three main discussion topics: (1) challenges and solutions, (2) defining GBSs, and (3) how evidence is used in policy development. Each discussion was designed to elicit feedback that could be used to develop a GBS typology meeting our objectives. Two project team members joined each group, one to chair and one to document the discussions in writing. Those unable to attend the workshop were invited to provide feedback via email.

Literature review

We identified eight articles published between January 2014 and August 2018 to add to the 125 papers identified by Taylor et al. for the period 2009 to 2014. 23 We found that policy literature included little information on different GBS types. Definitions of GBS incorporated one or more of availability, accessibility (legal access and perceptions of accessibility), usage or characteristics of a space. Natural England defines accessible green space as ‘places that are available for the public to use free of charge’ and natural green space as ‘places where human control and activities are not intensive’. 123 Definitions differ in what is counted as green space, with certain types of landcover, domestic gardens for example, included by some definitions but not others. 23 This is often related to (private) ownership. NRW defines green space as ‘shorthand for both vegetated land – parks, community gardens, trees, woodlands and hedges, informal spaces, allotments and food growing sites, areas of water, such as rivers, canals, lakes and ponds and the seashore’. 22 Blue space is a more recent concept, often defined as visible water bodies ‘rivers, lakes and the sea’. Blue space is usually considered a subtype of green spaces, with green spaces potentially including blue features. 114

We reviewed national planning/policy typologies used in Wales or the UK. Many documents did not define GBS. Where definitions were made, national-scale GBS typologies often classified GBS into broad habitats (e.g. semi-natural grasslands) based on landcover;124,125 urban areas were considered one landcover type with no further breakdown. Other typologies used limited categories based predominantly on land use/function, such as recreational, amenity and public open spaces. 126,127 Rarely, typologies incorporated the intended age range of users (e.g. provision for children and young people). 120 Many typologies focused on urban GBS. 117,121

Several policies and reports promote the value and use of GBS in Wales or the UK, typically providing recommendations or standards for availability and accessibility. 128,129 One example is Natural England’s Accessible Natural Greenspace Standards, which provides a four-level typology based on distance/size and also considers the quality of the spaces. 130 Fields in Trust has a long-standing recommendation of 6 acres (2.4 hectares) of accessible green space per 1000 head of population. 131 Several typologies originating from the policy and planning sector also consider the availability of different forms of GBS. NRW has sought to implement principles from the Recreation Opportunity Spectrum, which offers a framework to support GBS managers to provide diverse GBS settings for public recreation. 132 Policy guidelines that are widely applied to promote the availability of GBS for health and well-being are often based on arbitrary distance/size cut-offs (e.g. the percentage of the population living within 300 m of a green space of at least 1 hectare). 123,131,133

Public involvement workshop

Twenty-seven professionals from local government planning departments and third-sector charities, and two individuals from local park engagement groups participated in the workshop. Participants represented 12 different unitary authority planning departments, Welsh Government, the Office for Future Generations, NRW, Sport Wales and Public Health Wales. The third-sector charities Keep Wales Tidy and the Canals and Rivers Trust were represented. Five members of North Wales unitary authorities participated via email. All 22 unitary authorities were engaged via the workshop or email.

Many participants reported that there is no consistency in definition or data sources between agencies. Participants from third-sector charities highlighted that variation in GBS definitions and not knowing which spaces relate to which health outcomes were barriers to advocating for and implementing change. Lack of information related predominantly to social indicators such as safety and usage, rather than to the specific type of GBS. Such factors are rarely integrated into typologies or classifications, nor are relevant data available at scale. However, participants felt that it was important to think about how people perceive places and to ask questions such as: ‘Is a park considered to be safe?’ or ‘Is the space maintained?’ It was recommended that the co-created typology classifications should provide a framework to gather information on peoples’ perceptions of specific spaces or types of spaces.

Workshop participants noted the need for nuanced consideration of available GBS using more comprehensive typologies. Currently, using available data and applying relevant – if crude – typologies, it can be possible report achievement of guidelines at a city-scale on availability of GBS (e.g. whether they conform to the common standards from Fields in Trust, Accessible Natural Greenspace Standards or WHO). However, they noted that using GBS definitions which overlook finer-scale inequalities can lead to inaction because the guidelines appear to have been met.

Participants considered that differentiating between types of GBS was important for evaluating how different spaces influence health and behaviour. They suggested that the typology should include more diverse forms, or understandings, of GBS such as green corridors. Multifunctionality was also considered important, and the limitations of existing use-related typologies were noted. Further, participants highlighted that different spaces have value for different demographics of the population. For example, one participant suggested that intended age ranges for spaces could be included in the framework. Participants extensively discussed quality as an important element for the framework, but acknowledged that quantifying quality is complex and, in many cases, subjective, individual and context dependent. One suggested solution was that the typology could identify spaces that are nationally recognised (as high quality, e.g. green flags). Finally, several participants raised whether and how the revised typology could capture GBS change over time.

A point of significant discussion was how to manage differences between urban and rural areas. Participants reported that GBS typically differs between urban and rural settings, as does availability and accessibility. They felt that ‘countryside’ as one ‘type’ was too broad and that this should be expanded to consider the diverse types of countryside activity, and the specific needs and usage by people who live in rural areas.

Typology synthesis

The results of the literature review and workshop were combined to develop a new typology with a hierarchical structure enabling different levels of aggregation. The revised typology has three nested tiers (Appendix 2, Table 23). Tier 1 divides GBS into amenities, functional spaces, seminatural habitats, enclosed and linear spaces; tier 2 provides greater detail on the type of space (e.g. identifying a recreation space); tier 3 contains even greater detail (e.g. identifying play areas within a recreational space). We included GBS transport corridors as a specific category based on workshop feedback. The classification has been designed so that additional planning designations (e.g. green corridors and wedges)134 can be overlayed to the GBS categories. ‘Quality’ (e.g. biodiversity, upkeep) and ‘qualities’ (presence of infrastructure such as toilets or natural features such as lakes) relate to each type of GBS, so we did not add green flags as a category. We recommend that the inventory of GBS resulting from applying the typology be combined with information on characteristics, quality, qualities and environmental value where data are available. The detailed methods used to derive GBS exposure metrics from these data are described in Chapter 4.

Strengths and limitations

Green space is a predominantly urban definition that is more challenging to apply to rural areas. This is partly because we tend to conceive green space as discrete units of space. In rural areas, linear features such as rights of way are more common. We include rural green space but exclude enclosed farmland due to the homogeneity of land classifications in rural areas and a lack of information on accessibility via rights of way. This typology is grounded in published literature and the perspectives of key stakeholders from across Wales. Although this typology is widely applicable, nuances of the Welsh policy and practice landscape should be considered when implementing it elsewhere. The typology is designed to be operationalised using GIS data available at a national scale to fit the requirements of this project. This led to some pragmatic choices. For example, not all GBS have a defined access point(s) and there is a paucity of data on GBS qualities, so these were not included. Where data on these qualities and characteristics are available, users of the framework can overlay them for the different GBS types. 135 We included all GBS which met the criteria for each category, regardless of size, as the outcomes for the overarching project were not dependent on any specific use of a space, such as for physical activity.

Data source selection

Our primary exposure is defined as temporal changes in ambient green space within 300 m of each household in Wales. Multiple data sources (e.g. aerial photography, land use data and satellite imagery) capture greenspace, but very few have the spatiotemporal coverage required for this study. Satellite imagery is suitable in terms of spatiotemporal coverage and availability of data which are free at the point of use. We exploited the rich satellite imagery data to derive longitudinal measures of greenness as our primary outcome. These satellite-derived measures have long-established methodologies and have been widely applied in research studies. 142

We chose to use Landsat satellite data (Appendix 3, Table 25) since they offered the best compromise between cloud-free images, appropriate imagery resolution and temporal coverage (for detailed rationale, see Appendix 3, Figures 30–40) downloaded satellite data from Landsat projects 4, 5, 7 and 8 using the US Geological Survey EarthExplorer online tool. 143 Owing to the longitudinal design of this study, we required satellite data from 2008 to 2019. Our project straddled different operational lifetimes of remote sensing platforms meaning that no single open-data remote sensing platform had continuous coverage across our study period. To overcome this issue, we downloaded and created composite images from the Landsat project.

Exposure metric selection

Using these data, we created two measures of greenness density – EVI and NDVI for every year during the study period. We selected EVI as the primary exposure outcome instead of NDVI because EVI reduces both atmospheric and soil background noise. NDVI is sensitive to topographic and meteorological effects; this sensitivity was an important consideration for this study because we do not have satellite imagery for the same time of day and year. Appendix 4 (Figures 41–42, Tables 30–33) documents the sensitivity analysis between the NDVI and EVI measures that informed our decision to define the primary exposure as mean EVI within a 300 m buffer of the home location.

Methods

We used Landsat platform versions 5–8 to create a yearly measure of EVI for every household in Wales (Figure 1). We selected images from springtime to minimise the number of poor-quality images due to cloud cover. Cloud-free images taken between May and July were downloaded for each year. We preprocessed the images using the Semi-Automatic Classification Plugin tool in QGIS. 144 We applied DOS1 (Dark Object Subtraction 1) atmospheric correction to each image as recommended by Young et al.,145 and calculated EVI for each image using the vegetation index GRASS (Geographic Resources Analysis Support System) tool in QGIS. 146 We also created cloud masks for each image using the Cloud Masking for Landsat Products plugin. 147 We used the cloud masks to set pixels covered by cloud in the satellite imagery to NULL to prevent these values from influencing the final greenness density. We produced annual composite images of Wales in QGIS by mosaicking EVI images together for the same year (see Appendix 3).

For each household location in Wales, we created a linear (as the crow flies; Euclidean) buffer of 300 m. For coastal households the buffer was clipped to the coastline to avoid underestimates of green space (blue spaces were captured as described later). Using this buffer, we performed an intersection analysis with the EVI calculation to estimate the density of vegetation at yearly intervals during the study period. EVI theoretically records landcover as values ranging from −1 to +1, with healthy vegetation values found in the 0.2–0.8 range. Values below 0 represent water bodies and were generally excluded by clipping coastal areas.

Data sources

We used vector data from multiple sources to calculate access to GBS (Appendix 3, Table 26). Sources included data from OS, local government audits and open data sources. The following section describes the sources of data.

Harmonising geographic information system data

We harmonised the GIS data described above to create a baseline dataset of all potential GBS in Wales. We brought together different spatial data types, reprojecting data to consistent geographic reference systems and spatial scales and classifying data into broad types of GBS. The data were managed and harmonised in PostGIS.

Operationalisation of the green and blue spaces typology

The GBS typology is a theoretical framework to categorise land parcels in Wales into GBS which people can potentially access, both visually and physically, to understand the impacts and benefits to mental health and well-being. The typology does not directly map to predefined GIS data or categorisation found within these data. To operationalise the typology, we applied a set of criteria-based processes and pragmatic choices.

First, we applied a set of criteria-based rules to rationalise each of the land parcels in the GBS data using the typology (Figure 4). To avoid duplication when combining multiple data sources, only vectors that did not overlap with the OS dataset were added from TAN 16 local authority data and the open data sources. We defined both GBS as land parcels to maintain a consistent definition of access to GBS.

FIGURE 4.

Conceptual diagram of applying typology to longitudinal dataset using a criteria-based approach.

We next applied the typology to the data using a three-tiered approach, based on feedback from the stakeholder group, with tier 1 being the least detailed and tier 3 the most. Figures 5 and 6 show the differences between tier 1 and tier 3 for the area around Swansea University’s Singleton campus. SQL code for operationalising the GBS typology is available at https://gitlab.com/envhe/BGS.git.

FIGURE 5.

Tier 1 typology for the area around Swansea University’s Singleton campus.

FIGURE 6.

Tier 3 typology for the area around Swansea University’s Singleton campus.

Once each land parcel which met the typology criteria was categorised into the typology subgroups, we created potential access points to measure distances from each home to all GBS. Most GBS captured in our typology are not managed and do not have specific access points. We developed a method which accounted for nested GBS and made it feasible to calculate access to GBS distances and counts. Owing to the variety of types (e.g. managed and unmanaged) and morphologies (e.g. nesting of GBS within other GBS) of GBS within the dataset intersections with road or footpath features with a GBS did not always produce an access point (e.g. a linear beach feature with a footpath running parallel). We developed a method for defining access points based on feature types which would act as proxy for accessing each green space. The following criteria needed to be met:

• For each regularly shaped green space a proxy access point for each side of access needed to be defined (e.g. north, south, east, and west).

• For linear features (e.g. rivers, canal, coastlines) an access point will be defined at set intervals to reflect multiple points of access along the length of the space.

• Each access point needs to be snapped to a footpath or road network node to define a point along the network at which the GBS can be accessed.

Figure 7 shows the result of implementing the access point criteria on a sample area of the GBS dataset. It shows how far each access point was moved in metres; this snapping distance was included in the final distance calculations. The resultant data were a set of point locations snapped to the road and footpath network representing potential access points to all GBS in Wales.

FIGURE 7.

GBS proxy access points snapping including distance moved in metres.

circle = original access point, triangle = snapped location – figure distance moved in metres from local road network = blue lines, to an access point within the GBS = green shaded area.

Methods

We created an origin–destination (OD) network (roads and footpaths) distance matrix for each household in Wales (n = 1,498,120 households) to estimate the distance to all GBS access points (n = 3,242,803) potentially accessible within 1600 m. This resulted in an OD matrix which paired each GBS with each household in Wales to create unique combinations of distances for 772,557,144 pairs of homes and GBS. We de-duplicated each OD pairing so that each household only accessed the nearest access point of an individual GBS. The final distance calculation was a sum of: (1) the snapped distance from household to nearest network node (d_i); and (2) the total network distance (d_j) between d_i and the network location of the snapped GBS access point (d_k) for all GBS within 1600 m of each home.

For use with the health data, we created three GBS access measures: summed total counts, nearest GBS, and average distance to a GBS. We also grouped each OD matrix per household to count by type, resulting in a secondary GBS access measure; a count of all GBS types within 1600 m (Equation 1).

Equation 1: Access to green and blue space calculation where d_i = snapped household distance, d_j = network distance, d_k = snapped GBS access point distance.

Typological classification of GIS data and calculation of proxy access points were performed in PostGIS using SQL. Subsequent summary counts were calculated in Python 3.6 using a high memory environment (128GB RAM with 8 cores). We managed the data in a secure, dedicated GIS repository to ensure best practice with regards to household-level data and because we produced very large volumes of longitudinal, spatial data for linkage with individual-level health data. The final data were anonymised through our trusted third party using a split-file process. 157,158 Identifiable details were removed from the dataset and replaced with an ALF or a RALF. 159 An ALF is assigned if the data are individual-level records or a RALF assigned if the data are address-level data. We linked the GBS access measures with our study cohort (Chapter 3) where we also describe the GBS exposures by population characteristics (e.g. deprivation, age, sex).

Primary exposure: enhanced vegetation index

The spatial distribution of ambient greenness across Wales is consistent with the theoretical principles of an EVI estimate, with rural areas appearing to have higher ambient exposures. We found some variation over time in the spatial distribution of EVI estimates (Figure 8) but spatially they appear to be random effects with no obvious clustering. We captured average EVI scores per household every year during the study period (see Appendix 3, Table 27 for annual average EVI aggregated to LA level).

FIGURE 8.

Aggregated EVI measures 2008–2018.

Rural regions have a greater EVI score every year compared with urban areas (Figure 9, Appendix 3, Table 28); however, there has been increased greening in both settings over the study period. In 2012 there is a noticeable dip in the EVI in rural areas, this corresponds with the lower mean EVIs visible in Figure 8 for the northwest area of Wales. Cloud cover meant that every satellite image taken in 2012 had poor visibility.

FIGURE 9.

Temporal variation in household-level mean EVI by rural/urban location.

Spatial distribution and access to green and blue spaces in Wales

Figures 10 and 11 show the distribution of potential access to GBS by household, aggregated to hexbins to maintain disclosure controls. Areas where there are too few households to aggregate data are shown in white. Figure 10 shows the distribution of access to green space across Wales with rural areas experiencing access to less formal green spaces (defined by the typology) than those found in coastal and more populated areas. The northwest of Wales has some of the highest access to green space but is also very rural and has formal access to national parks and coastal areas for the population living in the vicinity. Areas which are managed as national parks or sites of special scientific interest (e.g. Snowdonia, Brecon Beacons, Pembrokeshire and the Gower Peninsula; formal conservation designation in the UK, of particular interest to science due to rate fauna, flora or important geological features) are more sparsely populated and have poorer household levels of access. Appendix 3, Table 29 highlights that 97% of land cover does not change in Wales from 2007 to 2015. Increased access to blue space is found in the more populous regions of the South Wales valleys, cities and coastal towns.

FIGURE 10.

Aggregated access measures to green spaces per household within 1600 m.

FIGURE 11.

Aggregated access measures to blue spaces per household within 1600 m.

Capturing changes in potential access to green and blue spaces

The typology we developed to classify land parcels into different categories of GBS can be applied to the same data sources from different years, where they exist, due to the standardised method for capturing and storing information. To assess the feasibility of detecting change between two epochs of data, we applied the typology to OSMM data to capture greenspace availability in 2012 and 2018. We found that only limited changes in the access profile to GBS at a population level using different epochs of the GBS access data. There do, however, appear to be small regional variations which would impact at a household level (see Figure 12).

FIGURE 12.

Density plot showing distribution of number of green spaces potentially accessible in 2012 and 2018 for ~1.5 million households in Wales.

At an individual household level, changes are more apparent, with some households recording substantive changes in the numbers of GBS available.

Table 4 gives an example of a household which has shown substantive changes between 2012 and 2018.

Upon exploration of the data, it became apparent that these changes did not necessarily represent changes in the built or natural environment around a home, we discuss this in further detail in the next section.

Study design

Our time-aggregated analysis comprised 47 quarters, from January 2008 to October 2019. Only exposure data prior to the first recorded CMD were used. For instance, if an individual first had a CMD in the second quarter of 2013, then data on their GBS metrics up to and including the first quarter of 2013 were aggregated. Ambient greenness (EVI) and GBS exposure were normalised to between 0 and 1 to allow comparability across exposure measures. Results are presented for increments of 0.1 unit.

The time-aggregated cohort analyses used a subset of the longitudinal cohort described in Chapter 2. Individuals from the longitudinal cohort with a CMD in the first quarter in the study were excluded since there were no prior environmental exposure data. All other individuals in the longitudinal cohort were eligible for inclusion in the time-aggregated analysis. Individuals were removed if it was not possible to derive a suitable set of covariates and factors from the longitudinal data due to missing data.

Variables and data aggregation

The primary outcome for this time-aggregated analysis is binary. For each quarter we assigned every individual an indicator for either a high likelihood or low likelihood of having a CMD.

We defined the primary exposure as ambient exposure to greenness, the average EVI value of a 300 m buffer around the household location. Our secondary exposures describe potential access to GBS: number of GBS within 1600 m of the home location; distance to nearest GBS within 1600 m; and average distance to GBS within 1600 m. Individual-level GBS exposure and access summaries were modelled quarterly.

Data from quarters up to and including that before a recorded CMD, or until the individual moved out of Wales, were used to derive a normalised single time-aggregated ‘average exposure’ value for each individual. All covariates were derived from data in the same time interval (Appendix 4, Table 30).

Descriptive statistics

Ninety-one per cent of the longitudinal cohort (n = 2,801,482) were eligible for inclusion in the time-aggregated cohort. Following exclusion of 247,588 individuals due to missing data (Figure 1); the final time-aggregated cohort comprised 2,553,894 people; 10 or 11 years’ worth of data were available for almost half (45.4%) of the cohort (Table 5).

Of the 2,553,894 individuals in the time-aggregated cohort, 21.9% (n = 513,239) had a CMD recorded at least once. A higher proportion of females had a CMD compared with males (26.8% and 17.2%, respectively; Table 6). More than one-quarter of those in the 22–50 years age groups recorded a CMD (‘cohort with CMD’, Table 6). This was approximately twice as high as the proportion in the oldest age group (80+ years: 13.3% with a CMD, Table 6).

Common mental health disorders and variation in exposure

Individuals with a CMD had, on average, a slightly lower ambient greenness exposure than those who were not identified with a CMD (EVI: 0.27 and 0.29, respectively; Figure 15 and Table 6). The average count of GBS spaces was also greater for people without a CMD (counts: 430 and 445, respectively; Figure 16 and Table 7).

FIGURE 15.

Distributions of average ambient exposure (EVI) within the time-aggregated cohort subgrouped by those with and without a CMD.

FIGURE 16.

Distributions of average access to a GBS within the time-aggregated cohort subgrouped by those with and without a CMD. Density represents the density of the population as a function of exposure value within the subgrouped population (with or without a CMD).

There were differences in the extent of primary and secondary exposure with age. Younger people (age 22–30 years) had an average ambient greenness of 0.26 with access to an average of 523 GBS. In contrast, older people (age 61–70 years) had a higher EVI value of 0.30, but access to an average of 382 GBS.

Differences were seen in the distribution of access to GBS and ambient exposure when grouped by CMD and rurality (Figure 17) with rural areas having fewer publicly accessible GBS (i.e. mainly privately owned farmland with varying rights of access).

FIGURE 17.

Density distributions of the subgrouped populations (with or without a CMD) population of average exposure (EVI and access to GBS) by deprivation and rurality of the time-aggregated cohort.

Types of green and blue spaces

In the time-aggregated cohort, deciduous woodland was the most common nearest GBS type within 1600 m of home (Appendix 4, Table 31). This was recorded for 30% of the cohort (n = 701,966). The next nearest types were grassland, 8.5% (n = 199,458), and other grounds, 8.5% (n = 198,256). The most common nearest type of blue space was river, 5.1% (n = 129,239) followed by beach, 3.9% (n = 92,225). For individuals with a CMD, the most common nearest types of GBS were allotments and sports pitches (both 24.5% of the CMD cohort). For individuals without a CMD, the most common was deciduous woodland (n = 549,558). Living near a lido was associated with the largest proportion of the cohort without a CMD (83.3%; note, however, n = 670).

Research question 1: Is greater residential exposure and access to GBS associated with the likelihood of seeking help for a common mental health disorder in general practice?

We used the framework of our time-aggregated analysis to address each of our research questions. The results to each question are presented using pairs of tables that contain results for associations of CMD with (1) our primary exposure of ambient greenness, and (2) secondary exposure of counts of GBS access.

Predicted probabilities

Using fully adjusted model results, we predicted the probability of having a CMD as a function of each GBS exposure type, while holding all other variables constant. The predictive probabilities are based on the models detailed in Appendix 4, Table 33 and Table 34. These are adjusted for age, sex, death and birth in household, comorbidity, deprivation, and house moves and used covariates associated with reference groups. Typical normalised EVI exposure values range from 0.25 to 0.80. Our model predicts an approximately 30% lower probability of a CMD at the bottom of this range compared with the top. Typical normalised GBS access values range from 0 to 0.55; CMD probability is approximately 10% lower at the bottom of the range compared with the top (Figures 18 and 19).

FIGURE 18.

Predicted probabilities of a CMD, stratified by ambient greenness (EVI). EVI is defined as normalised annual ambient exposure to EVI within 300 m of home. GBS access is defined here as a normalised count of all GBS within 1600 m of home.

FIGURE 19.

Predicted probabilities of a CMD, stratified by potential GBS access. EVI is defined as normalised annual ambient exposure to EVI within 300 m of home. GBS access is defined here as a normalised count of all GBS within 1600 m of home.

Research question 2: Is the association between exposure or access to green and blue spaces and seeking help for a common mental health disorder modified by socioeconomic disadvantage and moving home?

We defined subgroups within our cohort based on categories of area deprivation, history of CMD, settlement type (rurality) and moving home (migration). Through separate analyses of these subgroups, we calculated AORs for associations between ambient greenness or potential access to GBS and CMD (Table 10). We present AORs for 0.1-unit increments in EVI and in GBS count.

Types of green and blue spaces

Individuals with specific types of nearest GBS including play areas, quarries, sports pitches, and allotments were associated with 12% to 19% greater odds of CMD compared with those living nearest to parks (Figure 20). In contrast, people living nearest to religious grounds were associated with 5% lower odds of CMD relative to people living closer to parks.

FIGURE 20.

Adjusted logistic regression model results for nearest type of GBS compared with park as nearest GBS.

Models adjusted for age, sex, birth/death in household, WIMD, comorbidity index and moving home.

Main findings

Question 1: Do people with different GBS exposure through time have different associated risks of having a CMD?

We found that ambient greenness (EVI) and all measures of GBS access were associated with lower OR of having a CMD. The strongest association was seen for ambient greenness compared to more potentially accessible GBS (20% lower odds of seeking help for CMD compared with 7%, respectively).

Question 2: Is the association between changes in exposure or access to GBS, and CMD modified by multiple sociophysical variables, migration, and socioeconomic disadvantage?

When we stratified our models by area-level deprivation indices, we found that people were less likely to have a CMD when they lived in a home surrounded by more ambient greenness and had increased access to GBS, across deprivation quintiles. There was some variability, but no pattern in EVI–CMD associations across deprivation quintiles. However, there was some suggestion that the inverse association between GBS counts and CMD was stronger in more deprived areas. Focusing on the results for counts of GBS, we found that people living in the most deprived areas had higher odds of a CMD compared with those living in the least deprived areas (AOR 1.81, 95% CI 1.79 to 1.83). The ambient greenness–CMD association was stronger among people who had not moved home during their time in the cohort than among those who had not moved. Ambient greenness was associated with lower odds of a CMD in all ‘moving home’ groups but the effect was strongest for those who did not move home (AOR 0.77, 95% CI 0.77 to 0.78) compared with those who moved (moved once: AOR 0.86, 95% CI 0.86 to 0.87; moved more than once: AOR 0.92, 95% CI 0.91 to 0.93). The magnitude of association between ambient greenness and CMD increased around two-fold in people with an historical CMD (32% lower odds for a standard deviation increase in EVI vs. 16% for those without).

When stratified by rurality, we found that increasing ambient greenness is beneficial to people living in both urban and rural areas. However, increased access to GBS was only significant for individuals living in urban areas. In town and fringe and rural areas, a greater number of GBS within 1600 m was associated with higher odds of having a CMD. For individuals living in urban areas, there were reduced odds of having a CMD with greater EVI.

Comparing people whose nearest GBS was a park with those nearest to another type of GBS gave mixed results. Some comparisons indicated a higher likelihood of having a CMD, while others indicated a lower likelihood. It should be noted that these do not indicate the absolute association between these different types of GBS and CMD. However, this analysis suggests that different qualities of GBS have different impacts on CMD outcomes and that not all GBS are equally beneficial.

Wider context

One fifth of our time-aggregated cohort had a CMD, as did more females than males. This is consistent with survey results published on the Welsh population. 164 For adults aged 16 years and over living in Wales, we found that living in an area with greater ambient greenspace was inversely associated with having a CMD. This is consistent with previous studies in the Netherlands and Spain that found greenness in the immediate environment around the home location supported good mental health. 160,165

Gascon et al. 166 recommended that satellite-derived measures of greenness should be interpreted with care. Our ambient greenness exposure is based on satellite-derived EVI values, where raw values from 0.2 to 0.8 represent healthy vegetation. A larger EVI score does not necessarily equate to more greenness by area but instead represents a larger volume of green; there is increased biomass. A small forest, for example, could produce the same EVI score as a large grass-covered area. Combining multidimensional aspects into a single EVI figure makes the relative qualities of greenspace difficult to interpret. This echoes the main conclusion of a 2018 systematic review. 82 Future research should consider the function of a space and dimensions such as biodiversity.

Greater potential access to GBS was also associated with better metal health for all metrics of GBS. Again, this finding is consistent with previous studies where poorer access to GBS (defined by distance or travel time) was associated with worse mental health outcomes. 138,167,168 It is useful to have this finding confirmed based on household measures of GBS access. Our study is based on potential to access GBS but studies have found that people tend to engage with GBS within 1600 m of their home. 169 Some GBS may be appropriate to most of the local population, however, the type of space may be most important to some subgroups. Evidencing the benefits of GBS types is likely to be helpful for policy makers and planners designing policies and interventions to include vulnerable groups who would benefit the most from these spaces. Current WHO indicators are based on simply having a half-hectare GBS accessible within 300 m of home, without thought of the type of space needed. 114,170 This study sheds new light on which types of GBS may be more important, with the likelihood of a CMD varying by type of nearest GBS. This builds on previous research that states that the relationship between GBS and mental health is complex, with the specific function of a GBS important to individuals. 171

Our results for different population subgroups suggest that future research should investigate the reasons why each subgroup is impacted differentially by greenness exposure and potential GBS access. This agrees with the work of others that there is not a ‘one size fits all’ green space, with different population groups likely to benefit from specific GBS tailored to their needs. Our work provides evidence toward the development of more specific public health policies and interventions.

Primary outcome: common mental health disorders

The primary outcome was likelihood of seeking GP help for CMD, determined using the algorithm described in Chapter 2 to produce a binary outcome. The algorithm identified people seeking GP help for a CMD each quarter using read codes in their health records held in the WLGP dataset (Report Supplementary Material 1). 103 People were defined as ‘seeking help’ in a particular quarter when historical diagnosis(es) were treated, and/or the GP recorded current diagnoses or symptoms (treated or untreated). A CMD drug treatment was identified as at least one prescription for an antidepressant, anxiolytic or hypnotic in the previous 12 months. 103 This same algorithm is used throughout this study, however, in these longitudinal analyses, individuals could have periods of seeking, and of not seeking help for a CMD in general practice.

Secondary outcome: general practice events

The total number of GPE each day were available for each person in the WLGP dataset. ‘Events’ can include general practice interactions (e.g. with a practice nurse, or a GP) as well as investigations/tests, prescribed medication and referrals. These events reflect the general need for primary health services. We converted the daily count of GPE to a binary variable (yes/no on any given day) and then aggregated to quarterly counts (minimum = 0 if no events in the quarter, maximum ~120 if events occurred every day in the quarter). Daily aggregation of events prevented multiple test results on one day being given more ‘weight’ compared with interactions with a nurse or GP.

Environmental exposures

Changes in the environmental metrics could occur at the environment level (e.g. a change in the GBS itself such as building on a local park, reduction, or a new community garden on vacant ground, gain) or at the individual level, such as through moving home. 83 We obtained start and end dates of periods lived at each address for all cohort participants to enable individual level analysis. The dates allowed us to allocate dynamic measures of the environmental exposures specific to all individuals living in the household, for the relevant quarters.

Covariates

Covariates included sex (male, female), age group (years, continuous), deprivation (quintiles of Welsh IMD), home moves (yes, no), birth(s) in household (yes, no), death(s) in household (yes, no), Charlson Comorbidity Index (continuous). 195 In preliminary analyses, we also included rurality (‘village, hamlet and isolated dwellings’, ‘town and fringe’, ‘urban > 10 k’, defined using ONS settlement type categories). 96 However, this could represent another measure of the primary environmental exposure, EVI, so rurality was excluded from our final models to avoid overadjustment. 196 Covariates used in each analysis are listed in the relevant table.

Statistical analyses

This longitudinal analysis was carried out using a panel design with repeated observations for each participant. A multilevel model structure with quarterly observations over time (level 1) nested within individuals (level 2) allowed us to take account of clustering of observations over time using random effects. This structure also enabled the handling of unbalanced data, where the number of observations can vary for individuals, which is an artefact of a dynamic cohort. 104 It was not possible to nest within LSOAs, as individuals could change LSOA when moving home, resulting in LSOA multi-membership. For these longitudinal analyses, we tested the association of change in ambient greenness and in potential GBS access around the home through time with CMD and GPE. We did not incorporate spatial differences between people at a single point in time into our longitudinal analyses.

For the outcome CMD, we used multivariate, multilevel logistic regression models. Where GPE days was the outcome, we used multivariate, multilevel Poisson regression models to reflect the count nature of the outcome data. For ease of interpretation, Poisson regression coefficients were exponentiated to incidence rate ratios (IRRs). We first explored associations between potential confounders and our exposures and outcomes separately. Covariates associated with both the outcome and exposures were added in a stepwise fashion to the multivariate models. The step-by-step iterative construction of the multivariate models involved adding potential explanatory variables in succession and testing for statistical significance after each iteration. Stratified analyses were conducted by deprivation and moving home. For deprivation, stratified analyses considered those living in urban areas, with the multivariate, multilevel regression models run separately for two WIMD quintiles of area-level deprivation, the least and the most deprived. For our primary outcome (EVI), stratified analyses by moving home were conducted using three categories of the number of home moves (none, one, more than one). For our secondary outcome (potential to access GBS), where individuals could only experience a change in access (in our data) by moving home (section 0), stratified analyses by moving home were conducted using two categories of the number of home moves (one, more than one). We calculated the variance partitioning coefficient for a null model without covariates for those who did not move LSOA, confirming negligible variance in our primary outcome at the LSOA level (< 1%).

We also conducted preliminary modelling of the potential impact of changes in ambient greenness on NHS costs associated with GPE via the impact on the Welsh GP budget over our study period (2008–2019). We used the total spend on GPs to estimate the Welsh general practice budget. This includes the reimbursement of drugs dispensed in general practices in Wales in 2017/18, totalling £543.9 million. 197

Cohort description

Everyone aged 16 years and over with at least one address recorded and registered with a general practice providing records to the SAIL databank between January 2008 and October 2019 (the ‘study period’) were included. The cohort for these longitudinal analyses was 2,801,483 individuals (female: 51%) with a total of 99,682,902 observations (see Chapter 2 for cohort inclusion flow chart).

As described in Chapter 2, the cohort was broadly representative of the Welsh adult population compared with 2011 Census figures. 113 We identified 816,242 individuals as seeking help for CMD in general practice at least once during the study period (29.1% of the cohort: of those, 62.1% were female). One in five people aged 16–21 years sought help for CMD in general practice at least once between 2008 and 2019 (Table 12; see also Appendix 5, Table 35), the highest percentage of any age group. The percentage of those who had sought help for CMD in general practice at least once was lower in each successive age group. There was a social gradient in seeking help for CMD, with 17.2% in the least and 25.8% of residents in the most deprived areas having sought care for CMD in general practice. The mean number of GPE days per person per quarter was 3.8 (median 2, SD 4.6) quarter increased by 1 day, from 3.3 to 4.3 days.

Description of change in environmental metrics

Change in enhanced vegetation index

The primary environmental exposure was change in mean ambient greenness within a 300 m buffer, centred on each residential location. For brevity we refer to EVI, rather than ambient greenness here. Overall, there is very little or no change in EVI compared with the previous quarter for most of the cohort (mean change 0.00; Figure 21, left-hand side). Change in EVI compared with the previous year shows a normal distribution centred around the no-change value (mean 0.01), with the majority of those in the cohort experiencing change in EVI within 1 SD (±0.12; Figure 21, central panel). Change in EVI compared with baseline (defined as the first period an individual enters the cohort) shows a slight skew towards positive change (mean change 0.05; Figure 21, right-hand side). This represents the ‘greening’ of Wales over the study period (2008–19) described in Chapter 3, where mean EVI increased by 0.13 in rural areas and 0.09 in urban areas.

FIGURE 21.

Distribution of change in mean EVI from left to right compared with previous quarter, previous year, and baseline.

Change in access to green and blue spaces among those who moved home only

The secondary environmental exposure was change in the number of potentially accessible GBS. In this study, only those people who moved home (42.5% of the cohort; Table 2) had the potential to experience a change in their access to GBS (see Chapter 4). People who moved home more than once experienced, on average, an increase in the number of potentially accessible GBS compared with baseline (change in mean number of GBS 0.65), while those who moved home only once experienced a decrease in the number of potentially accessible GBS (change in mean number of GBS −7.48; Table 14). Of all home moves, 31% resulted in an increase in GBS access and 31% resulted in a decrease (and 38% remained the same). Similarly, 35% resulted in an increase in EVI around the home and 33% in a decrease in EVI (Appendix 5, Table 36).

Research question 1: Is greater residential exposure and access to green and blue spaces associated with the likelihood of seeking help for a common mental health disorder in general practice?

Research question 2.1: Is the association between EVI exposure and seeking help for CMD in general practice modified by socioeconomic disadvantage?

Increases in EVI around the home did not significantly affect seeking help for CMD in general practice for those living in the most deprived urban areas, or those living in the least deprived urban areas (Table 15).

The odds of seeking help for CMD in general practice were higher in the quarter when an individual moved home, regardless of whether that move resulted in an increase in EVI (AOR 1.15, 95% CI 1.14 to 1.16 for those who moved once, AORs 1.19 to 1.20 for those who moved more than once, Table 16). We found the same pattern of higher odds of seeking help for CMD in general practice in the quarter an individual moved, regardless of whether the move resulted in an increase in the number of potentially accessible GBS (AOR 1.15, 95% CI 1.14–1.16 for those who moved once, AORs 1.19 to 1.20 for those who moved more than once, Table 14).

Research question 2.2: Is the association between EVI and seeking help for CMD in general practice modified by moving home?

Increases in EVI around the home affected seeking help for CMD in general practice for those who moved home more than once, with an increase of 0.1 EVI around the home associated with slightly lower odds of seeking help for CMD in general practice (change in mean EVI relative to baseline: AOR 0.99, 95% CI 0.99 to 0.99, Table 16). An increase in EVI (+ 0.1 EVI) around the home did not significantly affect the likelihood of seeking help for CMD in general practice (Table 17).

Research question 3: Do people with different GBS exposures through time have different associated risks of using general practice services?

Findings by research question

Findings in context

Living close to GBS such as parks, woodlands, trails, ponds, lakes, rivers and beaches is associated with positive impacts on physical and mental health. 3,4,23,198 However, the majority of evidence has not unpicked associations between the type, proximity, quantity and ‘qualities’ of GBS, and changes in mental health/well-being. 30,73 As a result, existing evidence to inform policies shaping our environment is limited. 60,82 In Chapter 5, we found that those with overall higher levels of ambient greenness around the home, or an overall greater number of potentially accessible GBS throughout their time in the study had lower odds of seeking help for CMD in general practice at any point in the study.

Those time-aggregated results add to growing evidence suggesting that GBS have positive impacts on mental health. 1–3,139 However, there is only limited evidence for a causal relationship between the ‘greenness and blueness’ of environments and adult mental health,8 with a relatively limited amount of robust, longitudinal evidence. 8,45,82 This chapter reports our national, longitudinal evaluation of the impact of changes in ambient greenness and availability of potentially accessible GBS around the home on seeking help for CMD in general practice among adults in Wales.

We found no evidence that changes in EVI or the number of potentially accessible GBS over time impacted on people’s propensity to seek help for a CMD in general practice. We suggest these findings are indicative of the limitations of the data and modelling approach. In Chapter 7, we examine associations between leisure use of GBS with well-being, and seeking care for CMD in general practice, by linking data from the NSW to the cohort.

Strengths

This is the first e-cohort with national household-level longitudinal annual environment metrics linked to longitudinal EHRs. Our work has several strengths due to the available data and the analytical approach. First, many studies of the association between the environment and mental health have focused on ‘greenness’ (measured by EVI or NDVI). We have supplemented this with measures of GBS availability, based on a large, integrated, national-scale spatial dataset of potentially accessible spaces.

Many studies have used environmental data aggregated over small areas. By using the residential data linkage available in the SAIL databank, we have linked the environmental data for each home to the full cohort of 2.8 million people and their routinely collected, administrative health data. This enabled objective assessment of environmental changes, with no research burden for individuals. 104,108 Our approach reduces ecological fallacy by using accurate data for each home and allows us to anonymously follow home moves, capturing associated larger changes in GBS exposures.

The large adult population of the cohort provides sufficient power to examine variations between subgroups to investigate inequalities. The capability for anonymous follow-up also reduces the risk of unintentionally capturing ‘green gentrification’ effects when improvements to GBS are made in an area. This is a risk to area-level study designs; the original population may be replaced with a healthier population as an area gentrifies. 199 This is a particular issue over the longer timescale needed to evidence the impact of smaller, cumulative environmental changes. Place-based intervention studies investigating area-level health effects over long periods of time following environment improvements risk recording health outcomes of a different population. 200 Our ability to stratify analyses according to whether individuals had moved reduced our exposure to these health selection effects.

Limitations

There are several limitations of the longitudinal analyses presented in this chapter. The primary limitation is that we observed high levels of ambient greenness across urban and rural areas, and smaller greener conurbations, in Wales. Households did not generally experience large changes in EVI values. This limited the amount of change in EVI observed across the period using Landsat remotely sensed imagery, impacting our ability to find associations between change in EVI and future chance of having a CMD.

Additionally, EVI is a measure of vegetation, which we used to estimate ambient greenness (as for many other health research studies). 64 EVI values below 0 represent water bodies and, for the most part, were excluded from our modelling by clipping imagery to coastal areas. While we clipped buffers in coastal areas to historical high-tide marks, our EVI metrics are still likely to have been impacted by the relatively high concentration of coastal and inland waters in and around Wales. Negative EVI values produced by surface water coupled with the potential inverse associations between blue space and CMD, mean that the lack of association observed between an increase in mean EVI and CMD may have been impacted by lower levels of CMD in ‘bluer’ areas. That said, the results for GBS accessibility did include blue spaces and were materially no different to those using EVI.

There are additional considerations when using change in EVI in longitudinal analyses. As described in Chapter 3, we were restricted to using images from Springtime (May–July) in Wales each year to maximise image quality because cloud cover. However, natural fluctuation in EVI with the seasons in temperate zones can result in changes as large at 0.2–0.3 EVI between winter and spring. 201 Combined with the three-month May to July window we applied to maximise image quality, this means that EVI may not have been captured at peak greenness every year. When EVI is used to measure change in ambient greenness in longitudinal studies, these issues may introduce noise to EVI data. This noise might exceed real variation in ambient greenness, resulting in potential exposure misclassification particularly when the observed changes are small. However, confidence in our results is strengthened by the comparable findings between the two GBS exposure measures (change in mean EVI and number of potentially accessible GBS).

The OSMM Topographic Layer used to create the GBS access measures is a rich dataset, but it was not designed to capture discrete longitudinal change in land use. During our study, we uncovered limitations such as changing land use classifications in the database being due to continual improvement programmes rather than real changes in land use (described in detail in Chapter 3). This meant we were unable to use these data to reliably measure temporal changes in GBS access. The results for change in GBS access and future chance of having a CMD were therefore restricted to those who moved home. This resulted in a smaller cohort (n = 622,025) compared with the cohort available for analyses of the impact of change in mean EVI (n = 2,801,483), our primary exposure. People who moved home may differ from those who did not move home in other ways that may affect the chance of having a CMD. This residential self-selection bias can result from unmeasured neighbourhood selection factors related to environmental exposures and CMD. Neighbourhood selection factors could include preference for greener areas or those with high-quality outdoor spaces, which could affect neighbourhood choice and CMD. 202

Our analyses of GBS availability also do not distinguish between different types, sizes or qualities of spaces, and therefore we would expect heterogeneity in what this exposure measure reflects, and therefore the mechanisms through which these spaces might be expected to affect mental health.

The analytical approach employed here considers the relatively immediate impacts of a change in exposure. Previous evidence suggests that mental health benefits may arise during a longer time period following a move to a greener area. These analyses may miss mental health impacts that take a longer time to arise following a change in GBS exposure. 45

The validity and reliability of research using data from routinely collected EHRs depends upon its quality and completeness. Overall, the validity of primary care diagnoses in the UK tends to be high. 203 Case finding for CMD in routinely collected administrative health data can unobtrusively identify patients for mental health research, including on the effects of intervention. 204 Data derived from EHRs are predominantly routinely recorded and lack data on behaviour, some potential confounding factors, and outcomes such as well-being. For example, there is no health-related quality of life instrument routinely used to assess changes in health status in general practice in Wales. Because of this, the longitudinal analyses in this chapter are restricted to detecting changes in outcomes that involve health service utilisation. However, through linkage to NSW survey data, a subset of the cohort has information on well-being. We examine associations between GBS exposures and well-being in Chapter 7.

Data

Data linkage

Individual (CMD) and household (GBS metrics, WIMD) level data were anonymously linked to NSW datasets within SAIL using ALFs (individual level) and RALFs (household level). Linkage was carried out through standard anonymised processes in SAIL (see Chapter 2), and permission to access selected variables from the linked survey data was obtained from the Welsh Government. The final sample size was 5791 after removal of missing data at each stage of the linkage process (Appendix 6, Table 38).

Statistical model development and analyses

Descriptive statistics

The total sample size of participants included in modelling was 5971. Nearly 60% (weighted) of participants lived in areas with an EVI of 0.2–0.4, with only 11% with an EVI of < 0.2 and 4% with EVI ≥ 0.6 (Table 21). The vast majority of people (82%) lived less than 300 m, and only 4% lived more than 500 m from the nearest green or blue space. A total of 40% of the sample reported spending more than two hours outdoors for leisure per week (Table 21). In terms of potential modifiers, 64% lived in urban areas and 13% were classified as living in material deprivation (Table 21, Report Supplementary Material 2, Table 1). Owing to the reduced sample for physical activity, the sample size for analyses of mediation by duration of walking physical activity was smaller at 3310 (Table 21). Of this sample, 63% spent no time walking at an intensity qualifying as moderate-vigorous physical activity.

Modelling results

Unless otherwise stated, modelling results described here are for fully adjusted models including urban/rural status.

Question 1: Is greater residential exposure to green and blue spaces associated with higher subjective well-being, and is this mediated by intentional use of green and blue spaces for leisure?

Results presented in Figure 22 (full results in Report Supplementary Material 2, Table 2) show that, partially supporting our hypothesis, in unadjusted and adjusted models, EVI was associated with WEMWBS (EVI beta −10.2, 95% CI −17.13 to −3.17; EVI² beta 12.49, 95% CI 3.02 to 21.97). However, rather than the hypothesised positive linear relationship (greater exposure associated with better well-being), we found a U-shaped relationship such that WEMWBS decreased as EVI increased to around 0.4 but then increased with increasing EVI at values greater than 0.4. For example, an increase in EVI from 0.2 to 0.4 was associated with an average decrease in WEMWBS of 0.53, while an increase in EVI from 0.4 to 0.6 was associated with an average increase in WEMWBS of 0.47 (not considering covariates). This pattern and coefficients were similar whether or not urban/rural category was included in models (fully adjusted without urban; EVI beta −9.69, 95% CI −16.58 to −2.79; EVI² beta 12.49, 95% CI 3.33 to 21.66).

FIGURE 22.

Predicted subjective well-being (WEMWBS and life satisfaction), based on fully adjusted GLMs. Note: Predicted values are based on fixed covariates where WIMD = middle quintile, urban status = urban, age = 16–24 years, sex = female, economic status – employed, material deprivation = not in material deprivation, car use = yes, season = autumn, wave = 2016/17 and (GBS only) local authority = Cardiff.

A similar pattern of association was found between EVI and life satisfaction (Figure 22; EVI beta −1.16, 95% CI −2.51 to 0.19; EVI² beta 1.69, 95% CI −0.15 to 3.53). Proximity to nearest GBS was not related to either WEMWBS or life satisfaction (Figure 22 and Report Supplementary Material 2, Table 3). There was some suggestion of decreasing WEMWBS with increasing distance to nearest GBS, but CIs were very wide. Time outdoors was positively related to WEMWBS, with evidence of diminishing marginal returns after about five hours (positive linear coefficient and negative quadratic term; Figure 22). For example, an increase in time spent outdoors from zero to one hour was associated with an average increase in WEMWBS of 0.83, while an increase from six to seven hours was associated with an average increase in WEMWBS of 0.12. Time outdoors was positively and linearly related to life satisfaction, although with no indication of diminishing returns (Report Supplementary Material 2, Table 4).

EVI was positively related to time outdoors, with individuals living in greener areas spending more time on visits to GBS than those living in less green areas. An increase of 0.1 EVI was associated with an average increase in weekly time outdoors of 4.34 minutes (EVI beta 43.41, 95% CI 6.85 to 79.97; Report Supplementary Material 2, Table 5). There was some suggestion that those living further away from their nearest GBS spend less time outdoors than those living closer (300–500 m vs. < 100 m beta −16.72, 95% CI −29.42 to −4.03) but no evidence of an overall association across distance categories (Report Supplementary Material 2, Table 5). This result has high uncertainty due to relatively low sample size.

Wellbeing associations with covariates

Higher well-being scores on both WEMWBS and life satisfaction were generally indicated for: males (vs. females); 65–79 years (vs. 16–24 years, with lowest scores among 45–64 years); those employed (vs. economically inactive or unemployed); individuals classified as not living in material deprivation (vs. in material deprivation); and those with access to a car (vs. no access) (Report Supplementary Material 2, Tables 2–4).

Moderation by urban/rural status

Results of models including interaction terms provided some evidence of modification of the residential GBS exposure and subjective well-being association by urban/rural status. This was particularly the case for the association between GBS proximity and life satisfaction, where, for example, living further than 500 m from the nearest GBS was associated with lower life satisfaction for town and fringe residents compared to urban residents (500–1100 m × town and fringe beta 0.66, 95% CI 1.32 to 0.00).

No other significant differences were observed by urban status between EVI and both measures of subjective well-being and between GBS proximity and WEMWBS. However, likelihood ratio tests (LRTs) comparing models with and without interaction terms indicated that models with interactions were significantly different for both GBS proximity models and for EVI predicting life satisfaction (e.g. LRT; EVI and WEMWBS, χ² = 11.59, p = 0.0.41; Report Supplementary Material 2, Table 6). In common with results of models without interaction terms, confidence intervals around coefficients are very wide, and it is difficult to discern clear patterns (Figure 23; full results in Report Supplementary Material 2, Table 6).

FIGURE 23.

Moderation of the relationship between GBS exposures and subjective well-being by urban status. Note: T&F, town and fringe; U, urban; V/H/I, village, hamlet and isolated dwellings. Predicted values are based on fixed covariates where WIMD = middle quintile, urban status = urban, age = 16–24, sex = female, economic status – employed, material deprivation = not in material deprivation, car use = yes, season = autumn, wave = 2016/17 and (GBS only) LA = Cardiff. Significance is where p < 0.05 for model terms, for interaction categories this indicates that they are significantly different to the reference category.

Mediation by leisure time outdoors visiting green and blue spaces

Based on the results presented above, the prerequisite conditions for testing the mediation of GBS exposure and well-being by time outdoors are met for the relationships between EVI and both WEMWBS and life satisfaction (Figure 24).

FIGURE 24.

Diagram showing the prerequisite relationships for potential mediation between environmental exposures and subjective well-being outcomes when relationships are modelled separately (based on fully adjusted models; Report Supplementary Material 2, Tables 2–5).

To facilitate interpretation and comparison, EVI, WEMWBS, life satisfaction and leisure time outdoors were all standardised (by subtracting the mean and dividing by the standard deviation), as such the units in the results that follow are in standard deviations.

Results of the full models including terms for mediation by time outdoors are presented in Report Supplementary Material 2, Tables 7 and 8. There are non-linear relationships between both EVI and WEMWBS (U-shaped) and time outdoors and WEMWBS (diminishing returns). However, the coefficients for EVI do not change when time outdoors is added to the model (excluding time outdoors: EVI beta −0.02, 95% CI −0.05 to 0.01, EVI² beta 0.02, 95% CI 0.01 to 0.04; including time outdoors: EVI beta −0.03, 95% CI −0.05 to 0.00, EVI² beta 0.02, 95% CI 0.00 to 0.04). Therefore, these results provide no evidence of mediation of the relationship between EVI and WEMWBS by intentional use of outdoor spaces for leisure.

EVI was associated with life satisfaction with a similar effect size as WEMWBS. Similarly, the coefficient for EVI is not diminished when time outdoors is added to the model (when urban status is not included). With urban status excluded, the coefficient is reduced slightly but it is not significant with time outdoors excluded (excluding time outdoors: EVI beta −0.00, 95% CI −0.03 to 0.03; EVI² beta 0.02, 95% CI 0.02, 95% CI 0.00 to 0.03; including time outdoors: EVI beta 0.00, 95% CI −0.03 to 0.03, EVI² beta 0.01, 95% CI 0.00 to 0.03). These results therefore provide no evidence of mediation of the relationship between EVI and life satisfaction by intentional use of outdoor spaces for leisure.

Question 2: Is greater residential exposure to green and blue spaces associated with lower likelihood of common mental health disorders, and is this relationship mediated by intentional green and blue space use for leisure?

Neither EVI nor proximity to nearest GBS were associated with CMD (e.g. EVI 0.4 to 0.8 vs. < 0.2, OR 1.06, 95% CI 0.80 to 1.42; GBS proximity OR 1.15, 95% CI 0.97 to 1.35, GBS proximity² OR 0.97, 95% CI 0.95 to 1.00; Report Supplementary Material 2, Table 9), suggesting that mediation via time outdoors is unlikely. However, greater time outdoors was associated with a reduced likelihood of having a CMD, such that an increase of one hour per week is associated with a reduction in the odds of a CMD of approximately 4% (OR 0.96, 95% CI 0.93 to 0.99; Figure 25; Report Supplementary Material 2, Table 10). Associations between EVI, GBS proximity and CMD, and mediation by time outdoors, are depicted in Figure 26.

FIGURE 25.

Predicted CMD by time outdoors, based on the fully adjusted model. Note: Covariates fixed where WIMD = middle quintile, urban status = urban, age = 16–24, sex = female, economic status – employed, material deprivation = not in material deprivation, car use = yes, season = autumn and wave = 2016/17.

FIGURE 26.

Summary of the prerequisite relationships for potential mediation between GBS exposure and CMD status by leisure time outdoors when modelled separately.

Question 3: Are associations between residential exposure to and use of green and blue spaces, and subjective well-being mediated by walking physical activity?

Residential exposure

EVI was positively associated with weekly duration of walking physical activity (PA), with a non-linear relationship indicating walking physical activity increasing from EVI values above 0.3 (EVI beta −103.28, 95% CI −260.35 to 53.79, EVI² beta 211.23, 95% CI 1.13 to 421.33; Report Supplementary Material 2, Table 11). For example, an increase in EVI from 0.2 to 0.3 was associated with an average of 0.23 minutes more walking per week, while increasing from 0.3 to 0.4 was associated with 4.46 more minutes of walking per week on average. The greatest effect was found with an increase in EVI from 0.7 to 0.8, which was associated with an average of 21.36 more minutes of walking per week. Greater distance to nearest GBS was also associated with lower walking physical activity (500–1100 m vs. < 100 m beta −34.96 minutes, 95% CI −65.05 to −4.87; Report Supplementary Material 2, Table 11). Walking physical activity (hours) was positively associated with WEMWBS and life satisfaction (WEMWBS beta 0.29, 95% CI 0.17 to 0.40; life satisfaction beta 0.06, 95% CI 0.04 to 0.08; Report Supplementary Material 2, Table 12). The prerequisites for mediation as summarised in Figure 27 (EVI related to time walking and time walking related to well-being) are therefore met for the relationship between EVI and well-being but not for proximity to nearest GBS.

FIGURE 27.

Summary of the prerequisite relationships for potential mediation between GBS exposure and subjective well-being by walking as at least moderate physical activity.

For WEMWBS, the magnitude of the coefficient with EVI was not attenuated when walking physical activity was added to the model (excluding walking physical activity: standardised EVI beta −0.01, 95% CI −0.05 to 0.03, standardised EVI² beta 0.03, 95% CI 0.00 to 0.05; including walking physical activity: standardised EVI beta −0.01, 95% CI −0.05 to 0.03, standardised EVI² beta 0.03, 95% CI 0.00 to 0.05; Report Supplementary Material 2, Table 13). Similarly, for life satisfaction, EVI coefficients were not attenuated. In both cases, p-values for EVI coefficients increased with the addition of walking physical activity, but there was no clear evidence of mediation (i.e. negligible or no change in coefficients).

Time outdoors

We tested the hypothesis that people who spend more leisure time outdoors visiting GBS each week tend to report higher WEMWBS scores, and that this relationship is mediated by walking physical activity. Prerequisite conditions were met for potential mediation as leisure time outdoors was positively related to walking physical activity (1 minute increase in weekly time outdoors associated with 0.25 minutes greater walking physical activity on average, 95% CI 0.22 to 0.29; Report Supplementary Material 2, Table 15) and walking physical activity was associated with both WEMWBS and life satisfaction (Report Supplementary Material 2, Table 12). The time outdoors–WEMWBS coefficient was not substantially attenuated when walking physical activity was added to the model (beta reduction from 0.21 to 0.19; Report Supplementary Material 2, Tables 16–17). The time outdoors-life satisfaction coefficient was attenuated when time walking was added to the model (beta reduction from 0.07 to 0.05). As such, there is evidence of partial mediation of the relationship between time outdoors and life satisfaction by time walking but not for WEMWBS (Figure 28).

FIGURE 28.

Summary of the prerequisite relationships for potential mediation between leisure time outdoors and subjective well-being by walking as at least moderate physical activity.

Question 4: Are associations between residential exposure to and use of green and blue spaces, and subjective well-being modified by individual-level deprivation?

Testing for interaction, we found no evidence of modification of the association between EVI/GBS proximity and WEMWBS or GBS proximity and life satisfaction by material deprivation (EVI predicting WEMWBS, LRT comparing model with interaction vs. without, χ² = 6.86, p = 0.076; Figure 29; Report Supplementary Material 2, Table 18). There were no significant differences by material deprivation for associations between EVI and life satisfaction, although likelihood ratios tests indicated that a model with interactions was improved (EVI predicting life satisfaction, LRT comparing model with interaction vs. without, χ² = 10.60, p = 0.014).

FIGURE 29.

Moderation of the relationship between GBS exposures and subjective well-being by deprivation status. Predicted values are based on fixed covariates where WIMD = middle quintile, urban status = urban, age = 16–24, gender = female, economic status – employed, material deprivation = not in material deprivation, car use = yes, season = autumn, wave = 2016/17 and (GBS only) local authority = Cardiff. Significance is where p < 0.05 for model terms, for interaction categories this indicates that they are significantly different to the reference category.

We did, however, find consistent evidence of a difference in the relationship between leisure time outdoors and both WEMWBS and life satisfaction by material deprivation (Figure 29; Report Supplementary Material 2, Table 18). For those in material deprivation, the association is stronger (more time outdoors is associated with higher WEMWBS) compared with those not in material deprivation up to approximately four hours/week. As Figure 29 indicates, the deprivation gap in well-being decreases with increasing time outdoors up to that point. For people spending zero time outdoors, WEMWBS scores are, on average, 7.15 points lower among those in material deprivation compared to those not (calculations exclude covariates); at four hours per week, this gap is 4.45. Above four hours per week, the time outdoors–WEMWBS association for those in material deprivation then decreases; further increases in time outdoors are associated with lower WEMWBS relative to four hours per week. At seven hours spent outdoors per week, the difference in WEMWBS between those in material deprivation compared with those not similarly deprived is greater again at an average of 6.30 WEMWBS points. For those in material deprivation, an increase from zero to one hour of time spent outdoors is associated with an average increase in WEMWBS of 1.89, an increase from four to five hours is associated with 0.13 points increase in WEMWBS on average and an increase from six to seven hours is associated with a 0.74-point reduction in WEMWBS on average. For those not in material deprivation, WEMWBS scores increase steadily with increasing time spent outdoors but, at durations less than four hours per week, the slope is shallower than that for those in material deprivation (Figure 29).

The association between time outdoors and life satisfaction is linear, but again is stronger among those in material deprivation; that is, there is disproportionately greater benefit of time outdoors for those in material deprivation versus those not deprived. The gap in life satisfaction score between deprived and not deprived at zero hours per week is 1.53 points, whereas the deprivation gap at seven or more hours per week is 0.96 points.

Summary of results

Findings in context

Strengths and limitations

This study has a number of strengths and novel aspects. The survey data provide the opportunity to investigate not only what environmental conditions are present in people’s local residential area, but their visits to green and blue spaces, a key mechanism by which health benefits may actually be obtained, while the NSW sample providing data on visits to GBS is considerably smaller than that for previous, similar work for England (e.g. the data linkage with SAIL resources adds considerable value and strength to the analyses here that are not possible otherwise). 63 First, the linked data permit analysis of residential GBS exposure at much higher spatial resolution than is typically possible, with address-level geolocation coupled with road/footpath network proximity analysis. The GBS data are based on a large, integrated, national scale spatial dataset of potentially accessible spaces. Second, linkage permits assessment of CMD likelihood based on validated case-finding approaches using health service data. The data linkage adds substantially to the power of the cross-sectional survey data, and future work will be able to capitalise on this linkage for longitudinal follow-up.

Exploratory analyses using GAMs permitted the nature of associations to be assessed before implementation of more traditional GLMs, developing robust, but interpretable model results. Assessment of effect modification by socioeconomic status was able to make use of a composite individual/household-level indicator, an advantage over studies using small-area deprivation indicators.

The primary limitation of these analyses is that they are cross-sectional. Being based on recent survey data (some collected during the project), they cannot yet capitalise on the longitudinal power of the SAIL databank. Causal inference is limited, and we cannot rule out the role of reverse causality, selection bias or selective migration effects on observed associations. Our analyses of GBS proximity do not distinguish between different types, sizes or qualities of spaces, and therefore we would expect a fairly substantial degree of heterogeneity in what this exposure measure reflects, and opportunities for recreation afforded. The survey does not include data on dog ownership, which was found to be important in other studies. 246 By definition, leisure visits to GBS here include visits within Wales both from home and while on holiday, which are likely to be quite heterogeneous experiences (e.g. everyday dog walking vs. a walking holiday).

As discussed above, the EVI metric, intended to estimate ambient greenness exposure, is likely to have been impacted by the relatively high concentration of coastal and inland waters in and around Wales. Negative EVI values produced by surface water coupled with the potential for positive associations between blue space and well-being, mean that the U-shaped association observed between EVI and WEMWBS could actually indicate better well-being in ‘bluer’ and ‘greener’ areas at either end of the scale.

The sample of survey participants available for analysis was smaller than the originally planned sample specified in the protocol (c. 12,000), determined by changes to the survey data collection that evolved during the project. This impacted on the uncertainty around model estimates, with very wide confidence intervals for some key results.

In terms of specific elements of the analyses, testing for mediation was limited in several ways. Testing for mediation of residential GBS exposure associations with well-being and CMDs by intentional leisure visits, assumes that these visits are made to GBS in the immediate local area – in the case of EVI, effectively visits within 300 m. These data do not permit this assumption to be tested, but we do know from other data that a reasonably large proportion of visits are made to GBS further away from home. For example, the Monitor of Engagement with the Natural Environment for England 2018–19247 showed that 32% of visits were to places more than 2 miles (3.2 km) away from home, and the mean distance travelled was 5.1 miles (8.2 km).

Mediation by physical activity was restricted to considering self-reported moderate-vigorous walking physical activity as the most appropriate option available. While most visits to GBS involve walking, some do not. Also, the questions used to derive walking physical activity included walking for reasons other than leisure and in other locations, such as active travel and walking in city streets. This mediation variable is therefore an imperfect representation of what we were intending to capture (i.e. physical activity during visits to GBS).

Mediation analyses carried out here adopted a relatively simplistic, sequential modelling approach. 239 Given the complexities of the associations under investigation and potential involvement of multiple mechanisms, more sophisticated approaches such as structural equation models, or the use of g-computational procedures to deal with complex mediational relationships hypothesized248 would be warranted to study mediation more thoroughly. 64

Mental health status itself may act as a modifier in relationships between GBS and well-being – for example, those with depression and/or anxiety may be less likely to make visits to GBS. This is not investigated here, primarily since the relatively small sample size of the survey subsample results in a small absolute number of CMD cases. Further, this interaction/subgroup analysis was not specified a priori in the study protocol. This potential interaction is not well understood, although evidence from a 2020 18-country cross-sectional study indicates that people with depression are as likely to visit, and those with anxiety more likely to visit, than those without. 212 This evidence also indicates that our finding of lower likelihood of CMD with increasing time outdoors is unlikely to be due to reverse causality (i.e. that people with CMDs visit GBS less often).

Mental health

Our time-aggregated exposure analysis for the adult population of Wales found that people living in homes surrounded with 0.1-unit greater ambient greenness had an associated 20% reduction in the likelihood of seeking care for CMDs. People living in homes with 10% more potentially accessible spaces had an associated reduction in likelihood of seeking care of 7%. Those who lived in urban areas benefited most from increased access to GBS, potentially indicating the more acute the deprivation of ‘greenness/GBS’, the greater the importance to mental health. Associations between ambient greenness and potential GBS access and seeking help for CMD were not found in cross-sectional analysis of the much smaller subsample who were also in the NSW. However, when we used the linked NSW cohort to investigate more detailed behavioural data, we found that an additional hour of outdoor leisure time per week was associated with a 4% reduction in likelihood of seeking help for CMD.

The NSW analysis found a protective effect on CMD of time outdoors, consistent with previous research, which has typically investigated the relationship between GBS visits and self-reported measures of symptoms of depression and anxiety. 30,252 For the NSW subgroup we investigated the relationship using a robust, validated measure of CMD likelihood based on recorded engagement with health services. While this approach undoubtedly misses less severe cases and those that do not result in health service use, it is useful in suggesting that the benefits of spending more time visiting GBS may well be sufficient to prevent a small proportion of CMDs that require health service intervention, although this was not replicated in the time-varying analyses using the full cohort.

Well-being

For participants who completed the NSW, there was a U-shaped association between ambient greenness and subjective well-being, but no association with potential access to GBS. Time spent outdoors in GBS for leisure was associated with better subjective well-being and life satisfaction but did not appear to mediate the association with ambient greenness. People living in material deprivation showed disproportionately positive well-being benefits of spending more time outdoors. For example, among people living in material deprivation, those spending four hours per week on leisure visits to GBS scored 4.9 points higher on the WEMWBS compared with those spending no time on such visits. The difference in score for those not in material deprivation was 2.2 points. However, people in material deprivation appeared to receive diminishing benefits the longer they spent in the outdoors, while the same was not true for those not in material deprivation. This difference could be due to a range of factors, such as lower GBS quality in more deprived areas.

Economic impact

We investigated the total use of general practice services rather than restricting to seeking help for CMDs. We found no evidence that increases in GBS exposure over time and subsequent help seeking in general practice varied by socioeconomic disadvantage. Increases in greenness around the home affected seeking help for CMD in general practice for those who moved home more than once (with very slightly lower odds of seeking help for CMD). This result was not replicated for people who only moved home once, or who did not move. Therefore, we were not able to calculate any potential cost savings as the result of increased ambient greenness or potential GBS access around the home for people using general practice services.

A natural experiment?

We proposed a natural experiment using routine health and environment data, with a cross-sectional subcohort linked to in-depth survey data. There is an ongoing debate about the definition of a natural experiment. 255 According to the UK Medical Research Council guidance, ‘By natural experiments, we mean events, interventions or policies which are not under the control of researchers, but which are amenable to research which uses the variation in exposure they generate to analyse their impact’. 256 We set out to study the impact of greening and changing GBS access on population health in a natural experimental design consistent with this perspective. Exogenous factors influence the distribution of greening and GBS access; those factors are typically external to the efforts of local communities and individuals. These would qualify our study as a natural experiment under certain definitions. 257,258 We were, however, unable to distinguish between variation in GBS due to ‘events’ (e.g. regeneration) or due to natural changes in vegetation or blue spaces. Moving home, a self-selection or endogenous factor, is also a component of our study. This is a larger component than we originally intended, due to the lack of viable environmental change data for GBS access. Residential mobility has been cited as solving a self-selection issue, but although we have these data, we disagree that this is a solution without understanding the reasons for selecting to move home. 259 We conclude that our study overall is observational because it does not fully meet definitions of a natural experiment. Having the right exposure/intervention data is critical for this kind of study; a different study could have developed a more typical natural experimental approach if the right data were available. However, we note that a component of our study fulfils some criteria of a natural experiment. We separated the ‘movers’ and ‘stayers’; for the subset of people who did not move home, we have an endogenous change in naturally occurring greening that is unlikely to have been influenced by individuals or a local community but rather but wider political and economic activities. Our study is thus a longitudinal pre–post non-experimental study, with a time series of exposures, plus an embedded subcohort for whom we have behavioural data comprising actual visits to outdoor spaces. 260

We optimised causal contrasts in our whole adult population by minimising eligibility criteria. We avoided sharp artificial boundaries that occur between sites of different exposures based on statistical areas, but instead generated individual exposures centred on their home. We strengthened our study using causal contrasts between those who moved and those who did not. 255 When designing the study, we considered adding a negative control; an additional outcome that would not be affected by the intervention. However, we were unable to identify a physical or mental health condition that could not in some way be potentially impacted by GBS. For example, the development of cancer could be influenced by the microbiome, which could be related to the presence or absence of greenness in the local environment. 261,262

There is one randomised controlled trial where people were randomised to have adjacent vacant lots greened. 263 This rare example found that for 342 participants, depressed feelings decreased by 42% and self-reported poor mental health reduced by 63%, for those living near greened vacant lots, compared to control participants. In a commentary on the trial, Jerret and van den Bosch (2018)259 queried whether the effect was partially due to litter clean up contributing to a tidier appearance, rather than the greening itself. 259

For our large national study, we prepared the dataset to ensure CMD outcomes were either after the single time-aggregated exposure, or in the subsequent quarter along a series of exposures in our longitudinal analysis. Ambient greenness exposure was captured annually, and we used the data at this frequency in the longitudinal analysis, updated quarterly for people who move home, to consider outcomes for individuals who subsequently used general practice services.

In our time-aggregated exposure analysis, we examined differences between individuals in space, rather than being only restricted to changes within the same individuals through time. With careful data preparation, we ensured exposures preceded outcomes in a time-to-event/survival analysis-type approach. This provides reassurance that exposures did not come either at the same time or after the outcome, thus reducing the potential for reverse causality. This has only previously been possible in a small number of studies. For example, a 2019 study from Denmark used NDVI at age 10 years, augmented with a cumulative NDVI measure, prior to a mental health outcome. 165 Additionally, a cohort study based in southwest England used average greenspace metrics from four follow-up waves to predict lung function at 24 years. 264 For our exposure measure we did not assume that people remained in a distinct ‘allocation group’. Instead, we purposefully captured people as they transitioned between different exposures; effectively they have moved between intervention and control groups. We used a longitudinal or panel analysis, to take advantage of these transitions with the aim of assessing the impact of GBS on mental health.

We need to mention that it was not possible to establish measures of ambient greenness or access to GBS that were as robust as we would have liked; however, our analyses added significant robustness through the longitudinal approach and the time-aggregated data reduced the potential for reverse causality while maximising the spatial-temporal differences in exposure to GBS between different people at a highly granular level.

Routinely collected data

The main benefit of routinely collected health data is its historical existence at suitable granularities to enable a decade-long time series for a very large sample. The important mechanism of joining individuals to their home environment is facilitated by the presence of ‘from to’ dates of addresses in the demographic dataset. This allowed us to attribute individuals to home addresses for appropriate durations and therefore link exposures to each person retrospectively. Data linkage mechanisms made it possible to link health and environment data at the same spatial and temporal scales. Many previous studies on similar topics have not had coincident spatio-temporal data; many studies measuring green space exposure at a single time point, assuming that the amount of vegetation does not change over time. 265 For example, a previous longitudinal study used health data from 1991 to 2008 but had green space data from 2005 only. 4 A further study using a time series of health data from the same panel dataset, used Land Cover Map data from 2007. 4 In our study, having spatiotemporally aligned data made possible a more robust exploration of the longitudinal impact of changing environments on health.

The harmonisation of routinely collected environmental data for the whole of Wales presented challenges. We needed sufficient spatial and temporal detail to estimate exposure as precisely as possible, and to maximise the variation in GBS to give sufficient contrasting exposures through time to analyse their impact on health. We set out to generate annual or even quarterly exposures of (1) ambient greenness from satellite imagery, and (2) annual GBS access from vector data supplied by Ordnance Survey, augmented with local council map data. We scaled back our ambitions due to cloud cover over Wales, and uncertainties present in the map data from inconsistent temporal updates and changes to land use classification. We produced an annual exposure measure based on satellite imagery (using an EVI), and used only the 2018 baseline GBS access measure, capturing temporal variations in access as people moved home rather than from any in situ change in potential access. How these national data represent the reality of changing exposures, and our lack of knowledge about why people selected to move to a particular location, are important pieces of missing information that are needed to dissect these complex associations. The increase in seeking help for a CMD for people with more ambient greenness in rural areas suggests there may be a ‘tipping’ point where too much ambient greenness is detrimental without sufficient access to (good-quality) GBS.

To define our health outcome, we used a validated algorithm with high specificity and positive predictive value for detecting CMD (anxiety and depression) from EHRs. This validated algorithm was designed to capture the prevalence of CMD in a population. It has been successfully adapted for previous longitudinal environmental studies. 104 In the previous and current environmental health studies, we removed more severe mental health conditions, such as psychosis, before operationalising the algorithm. The algorithm uses a combination of treatments, symptoms and diagnoses as recorded by the GP to avoid counting people who use anti-anxiety or depression medications for alternative reasons. Thus, the assignment of a CMD to a person in a particular quarter was not only based on a clinical diagnosis of a CMD. Many studies have used routinely recorded data to determine a health outcome, when obtaining a formal clinical diagnosis of a CMD would be intractable. A validation study comparing EHR algorithms (including the one we have used) to the MHI-5 demonstrated that changes in GP recording behaviour mean using only diagnosis and current treatment to identify CMD in general practice data would miss a number of true CMD cases. 103 Including historical diagnoses with current treatment, and symptoms, as in the analyses presented in this report, increases sensitivity to 0.3; we may still be missing about one-third of true cases of CMD. 103 The combination of three types of data here is a strong addition, albeit with known lack of sensitivity.

It is plausible that the severity of CMD symptoms differs between those who seek help in general practice and those who do not. The CMD measure therefore may be insufficiently sensitive to capture less severe CMD symptoms. These less severe symptoms may be more amenable to treatment or prevention from exposure to ambient greenness or access to GBS around the home than the CMDs that were captured by the algorithm. If those with milder CMDs are less likely to seek care in general practice but are also those whose CMD symptoms are more likely to respond positively to GBS, this could have diluted associations between change in GBS and milder CMDs in our study. Cognitive behavioural therapies and other non-medication treatments were not included in our CMD case definition as this information is not available in the WLGP dataset. Low sensitivity and lack of data on non-drug treatments, may have reduced the number of cases of CMD we were able to identify, with associated reductions in the precision of estimates. This may also affect the ability of EHR algorithms to capture changes in numbers of those seeking help for CMD over time, since referrals to talking therapies are increasing. 266 In future, having these data may improve the PPV for the CMD case algorithm and provide a more sensitive test of the association between GBS and CMD. Using mental health and well-being surveys linked to environmental metrics may capture CMDs where care has not (yet) been sought. This may provide greater sensitivity to detect associations between changes in the ‘greenness or blueness’ of the environment around the home and CMD. Cohort studies, or longitudinal panel studies, will be required to move beyond the existing evidence provided by cross-sectional studies.

The immense task of coding and data management within this project should be highlighted. It is important to ensure that each project has sufficient time to properly document the data preparation, and to ensure that code written as part of the project is made publicly accessible, as highlighted by the 2022 Goldacre and Morley review. 267

Health policy and delivery

Our findings suggest that GBSs should be considered an ‘asset’ in efforts to promote mental health. GBSs contribute towards healthy living environments that can support population mental health, but can also be considered as a tool for health improvement strategies. For instance, as a setting for promotion of physical activity such as walking. The evidence adds to that which has justified the involvement of the natural environment sector within integrated care approaches.

Perhaps one of the most important contributions of the study to informing health policy and delivery, is the additional evidence supporting the ‘equigenesis hypothesis’;269 that GBSs have a role to play in reducing socioeconomic inequalities in mental health. We showed that the well-being benefits of spending more time outdoors for people living in material deprivation were disproportionately positive compared to people not living in material deprivation. These findings suggest that the equitable accessibility of GBSs should be integrated in future strategies to reduce socioeconomic and other inequalities in mental health. Evidence from our research supports the need to develop public health interventions to promote more leisure time spent per week visiting GBS, particularly among those in material deprivation. Care needs to be taken to ensure any strategy is equitable and recognises the interaction between the personal circumstances and attitudes of the individual, and the availability of good quality and safe accessible green or blue space resources. The evidence we provide here can be used within strategies to support the government’s stated ambition: ‘to narrow the gap between the richest and poorest’. 270

Spatial planning and green infrastructure

We showed, using the single time-aggregated exposure measure based on each person’s residential location, that people with more ambient greenness around their home, or greater access to GBS in terms of proximity, had a reduced likelihood of seeking care for any CMD. People living in homes with 10% more potentially accessible spaces had an associated reduction in likelihood of seeking care of 7%.

These findings support the argument that green spaces should be protected where they currently exist in neighbourhoods, and integrated, at a suitable scale, into new developments. This is particularly important in urban areas, where the association was the strongest. Green infrastructure plans and strategies should recognise the associations of greener living environments and mental health, integrating appropriate metrics of provision and accessibility. Green infrastructure decision-makers should consider the potential equigenic benefit to the mental health of those living in homes with more ambient greenness or potential access to more GBS.

Environmental management and nature conservation

Our evidence also supports the ongoing efforts to promote the use of the natural environment for benefits to mental health and other outcomes. Consistent with previous studies, the NSW analysis found a protective effect on CMD of time outdoors. We showed that each additional hour of outdoor leisure time per week was associated with a 4% reduction in likelihood of seeking help for CMD. And as noted above, people living in the greatest material deprivation benefited the most. This suggests that efforts to enable and support people to use the outdoors for leisure, including to be physically active, should continue.

Data collection and infrastructure

This study demonstrated the significant benefits of data linkage of environmental data to general practice records at a large scale. Decision makers within the health and social sector who are setting up new data infrastructures should include data linkage that is based on a flexible combination of identifiers such as name, date of birth and/or address, rather than relying solely on the assumption that each person will have an NHS number. This not only has the potential to reduce bias, but also expedites the linkage of environmental data,271 and data from other systems outside of the health care sector, that are important to consider as part of a wider systems perspective. 272 This allows evidence to be generated to include mechanisms that are further upstream and may help to improve population health.

Data collections from surveys provide rich behaviour data for a limited number of participants. Here we demonstrated the utility of having consent to data link participant responses to their routinely collected health utilisation data. The ability to nest the ‘deep, rich’ data for several thousand participants within collections of ‘wide, shallow’ health utilisation data for the whole population is valuable. We suggest that data infrastructures should include mechanisms to allow these types of linkages. Additionally, we recommend that organisations collecting primary data request participants to consider consenting to have their responses linked to their health records. Data providers such as GPs, researchers and the public should be apprised of the value of these data linkages so they may help advocate for data infrastructures in their region to facilitate linkages, making the most of complementary collections of primary and routine data.

Access to detailed health usage data in this study was sufficient in terms of its temporal granularity. We could have studied changes in health utilisation as often as monthly if we had deemed this an appropriate temporal scale. However, we faced challenges in the availability of suitable environmental data. The temporal granularity of the environmental data was recorded insufficiently and annual changes in the vector map data could not be included in our analyses, instead we were restricted to using a single year of vector data and relying on raster data captured using satellite imagery.