Examining the spatially varying and interactive effects of green and blue space on health outcomes in Northern Ireland using multiscale geographically weighted regression modeling

Previous studies have mainly examined the independent effects of green or blue space on health from a perspective of spatial homogeneity, which neglects their interactive or spatially varying effects. Here, we examined the spatially varying and interactive effects of green and blue space on health using open access data in Northern Ireland (NI). Aggregate health data was collected from 2017 Northern Ireland Multiple Deprivation Measure at the Super Output Area (SOA) level. Green and blue spaces were extracted from Land Cover Map data. The proportion of grassland and the proportion of woodland for each SOA were calculated as proxies for green space, while the proportion of water bodies was calculated for measuring blue space. Spatially varying effects of green and blue space were modelled using multiscale Geographic Weighted Regression (MGWR). Interaction terms between green and blue spaces were added into the MGWR models to test the interactive association of green and blue space on different health outcomes (e.g., preventable death ratio and cancer registrations). Results indicate that associations were distributed zonally, with green and blue spaces in eastern areas of NI more strongly associated with health outcomes than in western areas. Within these large regional zones, further spatially varying effects of different green and blue spaces were observed. Grassland was generally positively associated with some health outcomes (e.g., less preventable death ratio, cancer registrations ratio, multiple prescriptions ratio, and long-term health problem or disability ratio), while the results of woodland and water body were mixed. Water bodies were found to strengthen the effect of woodland and grassland. The above results indicate that green and blue space have independently and interactive spatially varying associations with different health outcomes in NI. It is also important to combine both green and blue space elements to enhance health impacts in future interventions.


Introduction
Existing evidence has suggested that natural environments such as green and blue space have beneficial impacts on health outcomes such as mortality (James et (Bloemsma et al 2018), prescriptions for mental illness (Chang et al 2020, and long term health problems or disabilities (Huang et al 2021).There are three main mechanisms linking green and blue space to health including instoration, restoration and reducing harms (Wang et al 2019). The first mechanism (instoration) indicates that green and blue space can help people build capacities for health behaviors such as encouraging physical activity and facilitating social cohesion, which both have health benefits (Zhong et al 2020, Wang et al 2021. The second mechanism (restoration) highlights that certain features of green and blue space may help people reduce their stress (Wang et al 2019). The third mechanism (reducing harms) suggests that green and blue space may mitigate environment hazards such as heat waves and air pollution, which are harmful to health (Arghavani et al 2020. Previous studies have confirmed that green space can influence health through all three mechanisms (Markevych et al 2017), but less attention has been paid to blue space, and only the instoration and restoration mechanisms have been confirmed for blue space (Georgiou et al 2021). Therefore, focusing on the effect of blue space on health can enhance our knowledge regarding the beneficial effect of outdoor natural environment. Also, different types of green and blue space may influence health (Markevych et al 2017). For example, Reid et al (2017) suggested that trees outside of parks, may be associated with better health, but such findings were not found for grass.
Several systematic reviews regarding green and blue space and health revealed inconsistent findings across different countries and regions (Britton et al 2020, de Macedo et al 2021, Shuvo et al 2020, White et al 2020, Jinguang Zhang et al 2020. Some scholars argue that such inconsistency may be explained by the spatial landscape heterogeneity of different places (Houlden et al 2019). Landscape heterogeneity refers to the differences in certain characteristics of research areas such as environmental contexts, urban-rural disparities and socio-economic status (SES), which may have influence on the relationship between green and blue space and health outcomes (Houlden et al 2019). For instance, existing literature suggested that people living in more socio-economically disadvantaged areas may be more influenced by natural environments ('equigenesis') (Mitchell et al 2015). Acknowledging the spatially varying effects of green and blue space may further help us to understand whether and how they contribute to urban-rural and SES disparities in health outcomes. However, to date, few studies have shed light on the spatially varying effects of green and blue space (Houlden et al 2019). For example, Houlden et al (2019) found that the association between public green spaces and health varies spatially, with stronger association detected in the outskirts of London than the center.
Most existing studies have considered different types of vegetation as a whole (Houlden et al 2019). However, recent studies indicated that different types of green space (e.g., grassland versus woodland) may have different influence on health outcomes , Yao et al 2022, so it important to specify the type of vegetation to better understand the health benefits of green space. Also, green and blue space commonly co-exist in urban landscapes, but most previous studies only focused on the independent effect of green space (Britton et al 2020, Donati et al 2022, Shuvo et al 2020, Sörensen et al 2021, White et al 2020, Jinguang Zhang et al 2020. In recent years, scholars have begun to realize the importance of blue space, but its effects are still largely examined independently (Britton et al 2020, White et al 2020. Environmental psychological theories such as Stress Reduction Theory and Attention Restoration Theory highlight the importance of naturalness of natural environments in producing restorative effects (Ulrich et al 1991, Kaplan 1995. However, in order to reveal restorative effects, the natural environments must have a high level of naturalness (i.e., like what human ancestors have viewed), which means vegetation and water bodies need to be present together. Thus, it is likely blue space may strengthen the effect of green space on health by enhancing its restorative effect. Also, the independent effects of green or blue spaces have been well explored (Britton et al 2020, Shuvo et al 2020, White et al 2020, Jinguang Zhang et al 2020, so theoretically it is likely that they can have cumulative effects on health.

Aims & research questions
Therefore, the objectives of this study were to: 1) describe the association between green and blue space and different health outcomes, 2) explore the spatially varying association between green and blue space, and health outcomes, and 3) examine the interactive effect of green and blue space. The study was set in Northern Ireland (NI), which is representative of other regions across the UK and Europe and has a range of open-access data. Figure 1 presents the conceptual framework underpinning our objectives, based on the narrative evidence reviewed above.

Dependent variables
Previous studies indicate that green and blue spaces are associated with a variety of health outcomes (Britton et al 2020, Shuvo et al 2020, White et al 2020, Jinguang Zhang et al 2020. Small Areas are the basic geographical units for collecting primary census information in the NI. They are aggregated to create Super-Output Area SOAs, each of which contains 400 people and 155 households on average. Data are commonly made available by SOA in NI. Therefore, in this study we obtained a series of aggregated health outcome data at the SOA level in NI. In total, we used eight variables including preventable death ratio, physical health-related burden ratio, cancer registrations ratio, emergency admission ratio, low birth weight ratio, children's dental extractions ratio, multiple prescriptions ratio, long term health problem or disability ratio (detailed description in table 1). First, based on our literature review (Houlden et al 2019), there is initial evidence that these variables might be associated with natural environments. Second, these variables were the components for the health domain of the Northern Ireland Multiple Deprivation Measure (NIMDM) (Ijpelaar et al 2018). The deprivation measure, NIMDM, is the main metric for reflecting inequalities in different domains in NI including income, employment, health and disability, education, skills and training, access to services, living environment, and crime and disorder domains. Therefore, focusing on these variables from health and disability domain can also help us understand the effect of green and blue space on health inequalities in NI.

Independent variables
Green space and blue spaces were measured based on Land Cover Map 2017 (LCM2017) vector data (Morton et al 2020). It represents a suite of geospatial land cover dataset describing the UK land surface in 2017. The LCM2017 was produced by the UK Centre for Ecology & Hydrology through the classification of satellite images in 2017. The vector datasets of the LCM2017 product were derived from the 20-m pixel imagery. This dataset gives land use information in 21 classes based on UK Biodiversity Action Plan broad habitats (Morton et al 2020).
Green space was divided into grassland and woodland, since previous studies indicate they may have different influence on health outcomes , Yao et al 2022. Grassland includes all subcategories of grassland in LCM2017 datasets, while woodland includes all subcategories of woodland (table S1). Blue space refers to water bodies, both saltwater and freshwater (table S1). The proportion of grassland, woodland and water body were calculated for each of the 890 SOAs in NI as proxies for green and blue spaces. The mean and median size of SOAs is 15.9 and 1.5 km 2 respectively, while the size ranges from 0.1 to 250.1 km 2 .   Indirectly age and sex standardised ratio of nonoverlapping count of (1) Income Support claimants in receipt of disability premium, (2) State Pension Credit claimants in receipt of severe disability premium, (a) Attendance Allowance, (b) Severe Disablement Allowance, (c) Disability Living Allowance-physical health, (d) Incapacity Benefit, and (b) Employment and Support Allowance.

Department for Communities
The higher the value, the more allowance is needed for a specific area to support its health-related disadvantaged groups, so it is a proxy for burden of general health conditions. Numerator: Observed non-overlapping claimant count Denominator: Expected non-overlapping claimant count Cancer registrations (ratio) Indirectly age and sex standardised ratio of people registered as having cancer, excluding non-melanoma skin cancers.

Northern Ireland Cancer Registry
Numerator: Observed cancer incidence Denominator: Expected cancer incidence Emergency admission (ratio) Indirectly age and sex standardised ratio of emergency admissions resulting in a stay of four nights or more.

Department of Health
Numerator: Observed emergency admissions Denominator: Expected emergency admissions Low birth weight (ratio) Proportion of singleton births of low birth weight. Low birth weight is defined as a birth weight of less than 2.5 kilograms. Standardised proportion of people on multiple prescriptions on a regular basis.

Business Services Organisation and
Numerator: Observed number of people on multiple prescriptions on a regular basis

Department of Health
Denominator: Expected number of people on multiple prescriptions on a regular basis Long term health problem or disability (ratio) Standardised proportion of people with a longterm health problem or disability.

Census
Numerator: Observed number of people with a long term physical health problem or disability Denominator: Expected number of people with a long term physical health problem or disability including population density, sex, age, urbanity, income, crime and disorder, active travel, marriage status, ethnicity, religion, education, employment, working duration, and industry of employment (table 1). These variables are potential confounders of both green and blue space exposure and health outcomes, so we controlled for them in the final model. . In addition, we ran OLS and GWR model as the reference and compared their goodness-of-fit.

Spatial regression models
In this study, we applied MGWR to explore the spatially varying effects of green and blue spaces on health outcomes. We built a spatial regression model for each health outcome (table 1). For each model, we included the independent variables, Grass, Wood, and Water. To assess the interactive effects of green and blue spaces on health outcomes, we also inserted the interaction term between water body and grassland (Water * Grass), and the interaction term between water body and woodland (Water * Wood). As for the moderation analysis for the interactive effects, we mainly focused on the interaction terms. If the interaction terms are significant and the direction is the same as green space indicators, then it indicates that blue space may enhance the effect of green space, while if the direction is opposite to green space indicators, then it indicates that blue space may weaken the effect of green space. Finally, to ensure the right estimation of the effect of blue and green spaces on health we added all covariates as listed in table 1. Variance inflation factors (VIF = 2.33) suggested no severity of multicollinearity among predictors.
The MGWR provided local R 2 (R-squared) for each SOA observation, and also reports local regression coefficients, local residuals, local intercepts, local t-tests, and local p-values. Therefore, we visualized the local R 2 in ArcGIS 10.2 (Esri 2013) to examine the spatially varying effects of green and blue space. The MGWR was implemented in an open access and free software (mgwr v2.2, 2020, Spatial Analysis Research Center (SPARC), Temple, USA) (Oshan et al 2019). The results of model fit metrics for OLS, GWR and MGWR models (table S2) indicated that MGWR model performed better than OLS and GWR models across all dependent variables.

Results
The average local R 2 was relatively high for preventable death (0.7), physical health-related burden (0.9), emergency admission (0.5), children's dental extractions (0.7), multiple prescriptions (0.9) and long-term health problem or disability (0.9) models (table 2). In contrast, the average local R 2 regarding the cancer registrations (0.4) and low birth weight (0.2) were relatively low. The results of local residuals were provided in figure S3, suggesting local residuals of MGWR model are not influenced by spatial autocorrelation problem. The results of local intercepts are also provided as a reference ( figure S4). The percentage of SOAs whose p-values were 0.05 for the relationships between health outcomes and the independent variables varied for each health outcome (table 2). Emergency admission, low birth weight, and children's dental extractions models, had p-values that were >0.05 for all independent variables across NI at the SOA level. Results of the preventable death model indicate that the proportion of woodland was negatively associated with preventable deaths in a minority of SOAs (1.3%), while the proportion of water body was positively associated with preventable deaths in 46.6% of SOAs. Second, although the proportions of grassland and water body were negatively associated with physical health-related burden in a majority of SOAs (80.7% and 55.7%, respectively), the results of woodland were mixed. Among the 54.4% of SOAs whose p-value was 0.05, the relationship between proportion of woodland and physical health-related burden was positive in about 80% of the SOAs, while negative relationship was found in the rest of the SOAs.
The interaction term between water body and grassland was negatively associated with physical healthrelated burden in 56.0% of the SOAs, while the interaction term between water body and woodland was positively associated in 41.3% of the SOAs (table 2). Third, the proportion of grassland was negatively associated to multiple prescriptions ratio in 45.4% of the SOAs. In contrast, the proportion of woodland, the proportion of water body, and the interaction term between water body and woodland were all positively associated with cancer registrations ratio (in 92.6%, 40.2%, and 8.0% of the SOAs, respectively). Fourth, the proportion of woodland was positively related to multiple prescriptions ratio in 12.2% of the SOAs. Last, both proportion of woodland (in 19.6% of the SOAs) and the interaction term between water body and woodland (in 81.9% of the SOAs) were negatively associated with long term health problems or disability ratio.
To simplify the results, in the next sections only the spatial distribution of MGWR parameter estimate surfaces for green or blue space indicator with at least one observation with p-value 0.05 will be displayed. That is, if a green or blue space indicator attained p-value > 0.05 across all SOAs, we did not visualize its MGWR parameter estimate surface. Therefore, results of emergency admission ratio, low birth weight ratio, and children's dental extractions ratio were excluded from the following analysis. The maps for coefficient of MGWR estimate surfaces for all variables and outcomes were shown in figures S1 and S2 (binary coefficient) as the reference.

Preventable death model
The local R 2 for the preventable death model ranged from 0.612 to 0.716 (figure 2). Geographic variations were found, with higher values in east NI and lower values in south NI. The negative local regression coefficient values of the proportion of woodland were only significant in Armagh City, Banbridge and Craigavon areas (figure 2(b)), just south Belfast area. Positive local regression coefficient values of the proportion of water body were only found in east NI.

Physical health-related burden model
The distribution of the local R 2 of the physical health-related burden model had a clear spatial heterogeneity, with local R 2 values varying from 0.861 to 0.967 ( figure 3). Overall, the model fit across all areas was high, as demonstrated by the large local R 2 values (figure 3). Physical health-related burden model had higher explanatory power in east NI, especially in Belfast city and its surrounding areas, demonstrated by the zonal distribution of the local R 2 values ( figure 3(a)). Figure 3 also shows the spatial differences present in the coefficients of the physical health-related burden model. In the spatial distribution of local regression coefficients, the association between the proportion of grassland and physical health-related burden was negative and significant in middle and east NI ( figure 3(b)). The absolute value of coefficient increases as it gets closer to the Belfast city centre, indicating the closer to Belfast city centre, the stronger the association of the grassland with physical health-related burden.
The local regression coefficient values of the proportion of woodland were negative and significant in west NI, while a positive and significant association was found in east NI ( figure 3(c)). The proportion of water body was found to have a positive and significant association with physical health-related burden only in the east ( figure 3(d)). Such association strengthened closer to Belfast city centre.
The interaction term between water body and grassland was found to maintain a negative and significant relationship in areas located in the east ( figure 3(e)), which indicates that water body strengthens the effect of  grassland. In contrast, the interaction term between water body and woodland was found to have a positive relationship with physical health-related burden in the east (figure 3(f)), which suggests that water body strengthens the effect of woodland. Both associations strengthened closer to Belfast city centre.

Multiple prescriptions model
The local R 2 of the multiple prescriptions model were relatively high in NI, ranging from 0.744 to 0.890 (figure 4). Geographic variation showed higher values in east NI and lower values in south NI. The local regression coefficient values of the proportion of woodland were only significant in the south. The positive value of coefficient increases as it gets closer to the southern areas, indicating the closer to southern NI, the stronger the woodland-multiple prescriptions association.  Grassland, (c) Woodland, (d) Water body, (e) interaction term between water body and grassland, (f) interaction term between water body and woodland. A positive coefficient indicates the metric may contribute to worse health (e.g., higher health burdens), while a negative coefficient indicates the metric may contribute to better health.

Long-term health problem or disability model
As shown in figure 5, the distribution of the local R 2 of the long-term health problem or disability model shows spatial heterogeneity, with local R 2 values varying from 0.834 to 0.942. The local R 2 were found to be higher in east NI and lower in west NI. The relatively higher R 2 values demonstrate a better regression fit in NI's east, especially in Belfast city and its surrounding areas. The results in figure 5 indicate that there are spatial differences regarding the distribution of local regression coefficient values across the independent variables. The proportion of woodland was found to have a negative and significant effect on long-term health problem or disability only in west NI. Such association gets stronger as it gets closer to west NI. In contrast, the interaction term between water body and woodland was found to have a negative and significant association with long-term health problem or disability only in east NI, which indicates that water body strengthens the effect of woodland. Such association gets stronger as getting closer to east NI.

Cancer registration model
As shown in figure 6, the distribution of the local R 2 shows some spatial heterogeneity, with local R 2 values varying from 0.302 to 0.382. The local R 2 were found to be higher in east NI and lower in west and north NI. The relatively higher R 2 values demonstrate a relatively more accurate model in the east of NI. Figure 6 also shows the spatial differences present in the coefficients of the cancer registrations model. In the spatial distribution of local regression coefficients, the association between the proportion of grassland and  Composite maps for MGWR parameter estimate surfaces of ratio of people with a long-term health problem or disability for (a) Local R 2 , (b) Woodland, (c) interaction term between water body and woodland. A positive coefficient indicates the metric may contribute to worse health (e.g., more long-term health problem or disability), while a negative coefficient indicates the metric may contribute to better health. cancer registrations was negative and significant in north and west NI. The absolute value of coefficient increases as it gets closer to the northwest NI, which indicates the closer to northwest NI, the stronger the grassland-cancer registrations association is. The local regression coefficient values of the proportion of woodland were positive and significant in middle and east NI. Also, the proportion of water body was found to have a positive and significant association with cancer registrations only in east NI. Such association strengthens closer to the eastern edge of NI. The interaction term between water body and woodland was found to maintain a positive and significant relationship in areas located in the Causeway Coast and Glens, Derry City and Strabane, and their surrounding areas in northwest NI, which suggests that water body strengthens the effect of woodland.

Discussion
Our results indicated that the local R 2 were spatially varying, with higher values in eastern areas higher than in western areas. Spatially varying associations of green and blue space were observed with different health outcomes. Grassland was generally beneficially associated with health outcomes (e.g., physical health-related burden, cancer registrations). Results of woodland and water body were mixed. Water bodies were found to strengthen the potential beneficial effects of woodland and grassland.

The goodness-of-fit of MGWR models
The spatial distribution of local R 2 indicated that our MGWR model fits better in east NI, especially in Belfast city and its surrounding areas. These areas are mainly urban areas in NI, with larger population density, indicating our models have better explanatory power in more urbanized areas. This finding is consistent with previous studies in Italy (Carrus et al 2015), England (De Vries et al 2003, Houlden et al 2019, Mitchell and Popham 2007 and Netherlands (Maas et al 2006), which also focused on the associations between natural environments and health. A possible explanation is that the green and blue space in more urbanized areas may be better maintained and have more essential facilities (Jingwen Zhang et al 2021). It could be speculated that these green and blue spaces are usually within the parks, but they may get larger and less equipped (i.e., fewer amenities) as the area becomes less urbanized in the UK context (Mell 2014). Existing studies indicated that infrastructure around and Grassland, (c) Woodland, (d) Water body, (e) interaction term between water body and woodland. A positive coefficient indicates the metric may contribute to worse health (e.g., more cancers), while a negative coefficient indicates the metric may contribute to better health. within green and blue space is associated with local residents' use, so urbanized areas are likely to be more frequently visited (Žlender and Thompson 2017) and thus have a better predictive power for health. Another possible explanation is that in rural areas, vegetation and water bodies are much more abundant, so any slight increase in vegetation or water bodies in rural areas may have an insignificant influence on health. However, in urban areas, there is a general lack of vegetation and water bodies, so any small increase may have a greater influence on health. The relatively high level of R2 for different health outcomes may be due to the reason that some of the predictors are absorbing the variability in the health outcome that is truly causal for other unobserved factors that are highly related to green and blue spaces.
5.2. The spatially varying effects of green and blue space on health outcomes Spatially varying effects of green and blue space on health outcomes were detected. We found that the association between green and blue space and health varied across different exposure and health metrics. First, both green (grassland, and woodland in some areas) and blue (water body) space were negatively associated with physical health-related burden, and such associations were stronger in Belfast city and its surrounding areas. Physical health-related burden can reflect regional healthcare costs, and existing literature suggested that natural environments were negatively associated with regional healthcare costs, since people living in such regions may be healthier and spend less on healthcare (Becker et al 2019, Van Den Eeden et al 2022. Also, as discussed before, green and blue space in Belfast city may have more health-related (e.g., physical activity) facilities and located in population dense areas, so they have more visits and have stronger effects on health.
Green space (woodland) was positively associated with multiple prescriptions, and its effect was stronger in southern NI. It is important to acknowledge that this study is an ecological cross-sectional study and therefore prone to reverse causality. It could be that people with more multiple prescriptions are more likely to live closer to green space (residential self-selection bias), since they rely more on having a better environment (de Keijzer et al 2020). Existing studies have found that woodland may cause different kinds of allergies and thus increase local demands for prescriptions (Stas et al 2021). Another explanation is that some unobserved risk factors (e.g., alcohol consumption and cigarette smoking) are also high in places with more woodland.
Green space (woodland) was inversely associated with long-term health problems or disabilities (i.e., green space contributes to less health problems), and such association was stronger in western NI. Existing studies in other European countries (e.g., Germany) had similar findings, especially for older adults, clarifying that green space may encourage people to take more physical activity and improve their quality of life (Huang et al 2021). Existing studies indicated that woodland is more flourishing in western NI than other areas (Thomas et al 2017), and previous studies found that the density of woodland is crucial for its purification impacts (Nowak et al 2014), which may explain the stronger effect of woodland in western NI.
Green space (woodland) was negatively related to preventable death, while blue space (water body) was positively linked to preventable death. Existing literature also indicates that green space may be inversely associated with all-cause mortality (James et al 2016, Vienneau et al 2017 (i.e., green space contributes to lower all-cause mortality), cause-specific mortality (e.g., cardiovascular diseases mortality) (James et al 2016, Vienneau et al 2017 and accidental mortality (e.g., suicide mortality) , Helbich et al 2020. Also, the positive association between blue space and preventable death can again explained by the residential selfselection bias (de Keijzer et al 2020), which means unhealthier people may have chosen to live closer to blue space (Helbich et al 2022). Another possible explanation is that waterbodies in these areas may not be of good quality, which may be harmful for health in NI context (Bunting et al 2007). It is also likely that some unobserved factors confound the association between green and blue spaces, and preventable death.
The association between green space and cancer registrations were mixed, while such association was positive for blue space (water body). The spatially varying effects of green and blue space on cancer registrations were quite heterogeneous, which could be suggested that the mechanisms linking green and blue space to cancer vary across different metrics and different areas. Existing evidence suggests that green space can promote physical activity and social engagement which may decrease the risk of some cancers (Porcherie et al 2021). Hence, patients with cancer may be more willing to live closer to nature (e.g., coastal areas) (Thomas et al 2004), and it is likely that there are some unobserved cancer risk factors that are more frequent where natural coverage is more abundant in NI, which both may explain the positive association between blue space and cancer registrations in this study.

The interactive effects of green and blue space on health outcomes
Our results suggest that water bodies in combination with woodland and grassland strengthen the beneficial associations with health. Stress Reduction Theory indicates that natural elements such as vegetation and water played an important role in human ancestors' daily life during the evolution process, so viewing natural elements may awaken modern humans' ancient memories and make them feel relaxed, which benefits health (Ulrich et al 1991). Attention Restoration Theory suggests that natural environments have four types of restorative features (Kaplan 1995). In ancient times human ancestors usually live close to both water and vegetation, so the coexistence of both green and blue space may strengthen their restorative features (e.g., fascination) and give modern people more sense of naturalness (Egner et al 2020, Liu et al 2022, Menzel and Reese 2022. The presence of both green and blue space increases the diversity of natural scenes, which is important for the restorative effect of natural environment exposure (Lange et al 2008, Stoltz andGrahn 2021). Existing studies pointed out that the more mixed the natural elements are, the more aesthetically pleasing is the environment (Stoltz and Grahn 2021). Aesthetics is a measure of how people perceive the beauty and tastefulness of the natural environment, so when natural environment has higher level of aesthetics, it usually means more usage and visitation. Therefore, the presence of blue space may enhance the effect of green space, and vice-versa.

Limitations
There are some limitations to be noted. First, the cross-sectional and ecological study design in this study may limit us in identifying the causation between green and blue space and health outcomes in NI. Future studies should consider using longitudinal panel data with a time series adapted GWR. Second, the health data was mainly from 2017, while the covariates were from 2011 census data, so this mismatch and the potential changes of this information may further lead to bias regarding the effect of green and blue space on health outcomes. However, our validation analysis (supplement file) indicates that although absolute sociodemographic information may change from 2011 to 2017 in NI, the relative sociodemographic differences between regions in 2017 is similar to that in 2011 in NI, which means such mismatch may not lead to significant bias. Third, the health outcomes were aggregated at the SOA level, which means our study was based on an ecological study design. This may lead to ecological fallacy, and the results may not be valid for individuals (Piantadosi et al 1988). Therefore, future studies should further test our hypothesis using individual-level longitudinal data, which can also further eliminate the effect of residential self-selection bias (de Keijzer et al 2020). Fourth, previous studies indicated that quality of green and blue space is important (Wheeler et al 2015, Brindley et al 2018, Venter et al 2020, Teixeira et al 2022. In this study, we were only able to access measures of the quantity of green and blue space exposure. This may prevent us from obtaining a comprehensive understanding of the effect of green and blue space. Hence, we were only able to distinguish between grassland and woodland due to the data available, and unable to distinguish between different type of water bodies such as rivers and lakes. Fifth, the observations in this study are super output areas (SOAs), which are defined by administrative boundaries. Such operation may cause the modifiable areal unit problem (Stewart Fotheringham and Wong 1991), so future studies should use more precise geographical units such as postcodes or households (Rodgers et al 2009, Rodgers et al 2012 to see how green/blue space -health associations vary across scales. Also, the size of SOAs varies across urban and rural areas, which indicates that using SOAs may not precisely reflect people's daily exposure environment in some areas (e.g., rural areas where the size of SOAs is large). Sixth, there may be some important variables missing in the model such as residents' actual visits to green and blue space, and landscape preferences, which are important for green or blue space-health associations (Van den Berg et al 2019). Hence, due to the data availability issue, we did not include health-related behaviors such as alcohol consumption and cigarette smoking. Future studies should give a more comprehensive consideration of the covariates, which can help us better understand the relationship between green, and blue space exposure and health. Lastly, we did not distinguish between public and private green space which may be important for further understanding the health benefits of green space (Verheij et al 2008).
Therefore, it is important to noted that this study does not provide conclusive evidence that green or blue space influences health via the hypothesized and tested pathways. Our analysis is exploratory, focusing on the spatially varying and interactive association between green and blue space and health. We have only provided some possible explanations for our findings, and do not rule out any other possible explanations. In addition to self-selection bias, any other limitations listed above could also be the explanation for the unexpected finding for a certain area.

Conclusion
This study indicates that the associations between green and blue space, and health outcomes may be spatially varying and interactive in NI. The association with health outcomes with spatially varying green and blue spaces should be noted before co-designing interventions with local communities to identify areas and communities that are likely to benefit. Interactions between green and blue space can be considered to enhance the effectiveness of future interventions and policies.