Quantifying the Impact of Scenic Environments on Health

Few people would deny an intuitive sense of increased wellbeing when spending time in beautiful locations. Here, we ask: can we quantify the relationship between environmental aesthetics and human health? We draw on data from Scenic-Or-Not, a website that crowdsources ratings of “scenicness” for geotagged photographs across Great Britain, in combination with data on citizen-reported health from the Census for England and Wales. We find that inhabitants of more scenic environments report better health, across urban, suburban and rural areas, even when taking core socioeconomic indicators of deprivation into account, such as income, employment and access to services. Our results provide evidence in line with the striking hypothesis that the aesthetics of the environment may have quantifiable consequences for our wellbeing.


Scenicness ratings and basic characteristics of the photographs
As the degree of the quality of the photographs may in itself affect their scenicness ratings, we also evaluate the relationship between color characteristics of the photographs and their scenic rating. We investigate whether brighter or more colorsaturated images correspond with higher ratings. We also investigate whether images with warmer colors, which contain more red, tend to receive higher ratings than images with cooler colors, which contain more blue. We calculate the warmth of an image pixel by extracting its Red, Green and Blue (RGB) values and defining it as warm if the red value exceeds the blue value. We calculate the warmth of an image as the proportion of warm pixels over the total number of pixels in an image. Figure S2 depicts the relationship between scenicness ratings and the brightness, color saturation and warmth of an image. We build a simple linear regression model to check to what extent higher scenicness ratings can be explained by higher saturation, brightness and warmth values. We find that images with greater color saturation tend to be rated slightly more highly than images with lower color saturation (β = 0.027, t(206869) = 74.14, p < 0.001). However, it is unclear whether saturation is a property of the scenic areas themselves or of the photographs. For instance, the sample of images with high scenicness ratings and low scenicness ratings presented in Fig. 1 suggests that images with higher ratings may contain fewer low-saturation grey manmade structures. Furthermore, although the linear regression analysis suggests that both brightness (β = 0.010, t(206869) = 23.02, p < 0.001) and warmth (β = 0.004, t(206869) = 18.74, p < 0.001) significantly increase with scenicness ratings, visual inspection suggests brightness and warmth do not have a simple linear relationship with scenicness, where warmth in particular appears to be -3 -highest for pictures with a medium scenic rating of around 5 or 6, and thus do not steadily increase or decrease with each consecutive scenic rating (Fig. S2).

Analyzing pollutants using Principal Component Analysis (PCA)
Strong collinearity between predictor variables can make it impossible to identify which predictor variable best explains the dependent variable in a regression model.
We therefore investigate to which extent collinearity exists between modeled estimates of concentrations of the following pollutants: sulphur dioxide (SO 2 ), oxides of nitrogen (NOx), particles and fine particles (PM 10 and PM 2.5 ), benzene (C 6 H 6 ), carbon monoxide (CO) and ozone (O 3 ). Following the method proposed by Belsley, Kuh and Welsch 1 , we find that high collinearity (condition number: 285.03) exists between six of the pollutant variables: sulphur dioxide (SO 2 ), oxides of nitrogen (NOx), particles and fine particles (PM 10 and PM 2.5 ), benzene (C 6 H 6 ) and carbon monoxide (CO). We therefore reduce these six correlated variables into three uncorrelated variables using Principal Component Analysis (PCA). The three PCA variables chosen each explain more than 5% of the variance of the original variables, and cumulatively account for 95.74% of the variance of the original variables.    We also control for the following pollutants: sulphur dioxide (SO 2 ), oxides of nitrogen (NOx), particles and fine particles (PM 10 and PM 2.5 ), benzene (C 6 H 6 ), carbon monoxide (CO) and ozone (O3), using the measures introduced in Table S2. A range of socioeconomic deprivation variables are controlled for, and the analysis is carried out at the level of Lower Layer Super Output Area. In this model, we find that more greenspace is significantly associated with reports of worse health across England as a whole. However, this effect does not hold in urban or rural areas when they are analyzed separately.  We also control for the following pollutants: sulphur dioxide (SO 2 ), oxides of nitrogen (NOx), particles and fine particles (PM 10 and PM 2.5 ), benzene (C 6 H 6 ), carbon monoxide (CO) and ozone (O3), using the measures introduced in Table S2.

), benzene (C 6 H 6 ), carbon monoxide (CO) and ozone (O3).
Akaike weights (AICw) can be interpreted as the probability of the model given the data.
Further details on how a model's AICw is calculated can be found in the Methods section. In all cases, the models that include scenicness perform better than the model with only greenspace.   We investigate to what extent geographic differences in health can be explained by scenicness and greenspace, by creating CAR models in which we also control for socioeconomic deprivation using data from the 2010 English Indices of Deprivation. We also control for the following pollutants: sulphur dioxide (SO 2 ), oxides of nitrogen (NOx), particles and fine particles (PM 10 and PM 2.5 ), benzene (C 6 H 6 ), carbon monoxide (CO), ozone (O3) using the measures introduced in Table   S2. To determine which model provides the best fit for predicting poor health, we calculate Akaike weights (AICw) which can be used to interpret the probability of each model given the data. Further details on how a model's AICw is calculated can be found in the Methods section. In all cases, we find that models that include scenicness (denoted by the color purple or by purple and green stripes) perform better than the model with only greenspace (denoted by the color green).