Motivation and Research Questions
Global warming has emerged as a top crisis in today’s world. Fossil fuels are a major contributor to increased levels of 〖CO〗_2. Although there has been a decline in total energy consumption in the UK over the past 10 years, 2015 saw an increase 3.7% in gas consumption, with domestic sector being the largest contributor towards the increase [1]. In order to devise better policies to reduce our carbon emissions, we need to understand the underlying factors of energy consumption across UK. Past research [2]–[4] focuses on explaining the socioeconomic factors behind the total energy consumption in households with particular light on electricity consumption. For this study, we wish to move the focus towards gas consumption in households in England & Wales and study the underlying socioeconomic factors at a disaggregated geographical level to assess if gas consumption can be explained geographically.
For our analysis, we used the sub-national gas consumption data [5] at Local Authority (LA) level which provided us with domestic gas consumption. We combined this with the 2011 Census data for England and Wales to get demographic and socioeconomic factors. Since we were mainly interested in exploring the characteristics that explained gas consumption per household, proportion of households were calculated at LA level.
Our aim for this study is to explore how household gas consumption is related to the socio-economic and demographic factors across England and Wales and identify key variables that appear to distinguish geographical differences. Our analysis tries to answer the following research questions:
- What socio-demographical factors best explain the variation gas consumption across England and Wales?
- Do these factors vary geographically and are they able to explain the gas consumption equally well across the region?
Tasks and Approach
Understand spatial variation in gas consumption per household in England and Wales using chloropleth maps
In order to investigate how gas consumption varies across area, we plotted chloropleth maps at LA level using a sequential colour scheme that helped in differentiating areas with low gas consumption (pale yellow) and those with high gas consumption (red). From Fig 1 (left), a cursory look reveals that metropolitan cities like London, Birmingham, Liverpool, Newcastle-upon-Tyne and Cardiff seems to be the highest consumers of domestic gas.
Fig. 1: Maps showing the distribution of domestic gas consumption per household (KWh) in England and Wales. A traditional chloropleth and region faceted chloropleth is shown from left to right respectively.
Identify spatial variations in relevant household categories using statistical geovisualisation techniques
A survey [6] reviewed different factors that explain household energy consumption. Themes broadly revolve around building characteristics, lifestyle and affordability which helped in choosing explanatory variables. To identify associations, we explored spatial variations in the variables by calculating the geographically-weighted mean and standard deviations using GWModel package [7] and visualising the results using chloropleth maps.
Identify associations between gas consumption and household characteristics using scatter plots
Scatter plots along with Pearson’s correlation coefficients were used to assess the associations between socio-economic factors and gas consumption. In order to improve the interpretability of these associations, regression lines were plotted and the data points were coloured by region. This enabled us to visualise the dispersion of data around the regression line and to identify there were any outliers concentrated within regions.
Explore geographic variation using local statistics
From the scatter plots, it was evident that for some factors, outliers were concentrated within regions. Using the GWmodel package [7], we calculated geographically-weighted correlation coefficients and plotted them using chloropleth maps to identify the extent of spatial variation of these factors.
Build explanatory models that control for socio-economic characteristics using regression
We built a multivariate regression model to investigate the variables that could possibly explain gas consumption. Multicollinearity is a known issue in this domain [2] and in order to preserve the assumptions of a linear model, collinearity was assessed visually using a correlation matrix. Computationally, Variation Inflation Matrix (VIF) was used in a step-by-step fashion to eliminate variables with high inflation. We chose our threshold of 2 for our VIF. Our model was built using the following steps:
- Identified distinguishing variables using correlation
- Explanatory variable with the highest VIF was removed and the impact on the model fit R^2 was assessed. This was done until we achieved a VIF of 2 or less for all variables
- Assessed geographic variations in the model by plotting model residuals as chloropleth maps
Analytical Steps
Spatial variations in gas consumption
Fig. 1 shows gas consumption per household across England and Wales. It reveals that larger cities across England and Wales are major consumers of gas. This is not surprising given the large population size of these cities. The faceted view reveals interesting insights. Major cities of Wales like Cardiff and Newport are not the highest consumers but areas like Rhondda valley and Caerphilly show highest consumption per household. In major cities like London and Manchester, there are variations within the region as well. For example in London, the periphery districts consume more than the central districts.
Spatial patters in household characteristics
Household composition characteristics were derived from UK Census 2011 data. Based on the themes mentioned before, variables were broadly selected based on intuition and limited by the information available from the dataset. Table 1 shows a list of variables selected and the justification behind them.
Variables Justification/Theory
Household size Building/Household characteristics
Rooms per household
Detached house
Elderly households Lifestyle/Time spent at home
Lone parents with dependent children
Independent children households
Long-term health issues
Socially rented Affordability
Socially high
Table 1: Proposed variables for explaining gas consumption per household and their justification
The maps in Fig. 2 and Fig. 3 display the geographically-weighted means and standard deviations of these household characteristics respectively. We observed that London and parts of South East and East of England which surround London are quite unique. We see some variation in Wales and South West which have the highest proportion of households with people who have a long-term disability or health issue. South West Wales, Yorkshire and the Humber along with East of England have a higher proportion of detached dwellings. After London, North East and West appear to have the highest proportion of social rented households.
Fig. 2: Chloropleth maps displaying geographically weighted means for household characteristics derived from census 2011.
Fig. 3: Chloropleth maps displaying geographically weighted standard deviations for household characteristics derived from census 2011.
Association between household characteristics and gas consumption
An important task for this study was to identify how well these household variables relate to the gas consumption of a household. Fig. 4 shows the gas consumption per household against the different variables. The data points are sized by density and coloured by Region. It also shows Pearson’s correlation coefficients and regression lines.
Fig. 4: Scatter plots showing the correlation between gas consumption per household and different variables. Data points are sized by people density and coloured by region.
Geographic variations in household characteristics
Fig. 5 displays geographically-weighted correlation coefficients for each variable against the gas consumption per household. The maps show that there spatial variation exists in the coefficients and is more varying across average rooms, social rented, elderly, lone parents with dependent children and detached households. Correlations in these factors diverges across regions with Wales, South West and east coast of England showing similar relationships while the North West, Midlands and London seem to display similar relationships. For long-term health and social high variables, we see that part of South East seems to display a strong correlation however; there is almost no correlation with the rest of the region.
Fig. 5: Chloropleth maps showing geographically weighted correlation for household characteristics
Build explanatory models that control for household characteristics
Based on the themes in Table 1, we wanted to build a multivariate model that explains the consumption of gas per household. The scatter plots in Fig. 4 suggest that gas consumption per household is correlated with independent children households and average household size. However, other correlations are loose and scattered. Collinearity among variables is a known issue [2] in this domain. So in order to assess multicollinearity amongst variables, we visually analysed the correlation matrix (Fig 6). Next, we computed their VIF to remove variables with the highest inflation in a step-by-step manner. Our final model after eliminating collinear variables was:
gas consumption per household=7.87+0.23×avg household size +7.45×indep children+1.55×social high
However, the explanatory power of our model is only 31%.
Fig. 6: Correlation matrix of our dependent and independent variables
Evaluate generalisability of the model
Our aim was to develop a model that logically explains the consumption of gas geographically across England and Wales. In Fig. 7, we plotted the residuals of gas consumption as a function of average household size versus a multivariate model control for household size, socially high and households with independent children. We see that the residuals across England and Wales do vary especially in some areas in East, South West regions of England and Wales containing strong residuals with areas in red forming clusters, however, the generalisability of our model is rather low. When controlling for explanatory factors, it fails to explain the variation across regions and the residuals are not very different to our first plot.
In order to test for global spatial autocorrelation, we use Moran’s I statistical test [8] which measures the degree of similarity between our dependent variable and location. Although we observed some spatial non-stationarity amongst the variables independently (Fig. 5), when testing for spatial autocorrelation using Moran’s I, we get insignificant results (Moran’s I 0.45, p-value 0.08) and therefore fail to reject our null hypothesis that gas consumption is not spatially correlated.
Fig. 7: Chloropleth map showing model residuals for linear regressions explaining gas consumption per household as a linear function of avg_household_size and a separate multivariate model (avg_household_size, indep_children and social_high).
Findings
We wanted to analyse factors that explain geographic differences in gas consumption across England and Wales. From the scatter plots (Fig. 4), it is evident that there is weak correlation among the variables and distribution is quite dispersed. Households with independent children and average household size seem to have the highest correlation. Socially high households and households with lone parent and dependent children seem to be positively correlated. It is worth noticing that some factors like elderly households and detached houses seem to be negatively correlated although we expected them to have a positive relationship.
From the chloropleth maps in Fig. 5, we see some regional differences in correlation across London, Midlands and North West forming one group and parts of east coast, North East and Wales forming another group. However, across all variables, there were some areas with zero correlation. This suggested that we may have missed some crucial variables in our analysis.
Most importantly, the key finding of our analysis is that the factors identified as being able to distinguish between gas consumption are insufficient to explain the variation. Our linear model presented in section 3.5 does not explain the variation very well. It can currently explain only 31% variance in the data. Even when we included all the variables available to us, the explanatory power of our model was no more than 49%. We also find that the spatial autocorrelation evident in our residuals (Fig. 7) is insignificant and therefore we cannot say that variation in gas consumption is location specific.
Given the low explanatory power of our model and the absence of spatial autocorrelation, it appears that our model may be misspecified and lacking explanatory variables crucial to our analysis. This indicates that there is a case for further research and better quality data to improve our understanding of socio-demographic factors influencing gas consumption.
Critical Reflection
Implications of findings for domain
Our aim was to explore if gas consumption could be explained by different socioeconomic factors and whether these factors differed geographically. Our study indicates that socioeconomic factors do not have a major influence the consumption of gas, although they do vary spatially. Average household size is by far the most significant explanatory variable identified in our analysis which suggests that house characteristics play a larger role in explaining gas consumption in a household. These results seem to be in line with previous research in energy consumption [2] where socio-demographic factors were only able to explain 29% of the variation. One important conclusion to draw from here is that we need to reflect on how we define our socioeconomic variables that could affect gas consumption.
On the basis of our findings and past research [2], [4], [6], [9], [10] in the sector, there may be some implications for policy making. Building characteristics have consistently emerged as major contributors to energy consumption, however studies suggest that family size and family formation could be an indirect, but large contributor in reducing carbon footprint of the country [10]. Therefore policies designed around these factors could help in lowering energy consumption of a household.
We identified two key limitations to our analysis. Firstly, our study was highly influenced by the socio-demographic themes uncovered in research survey. Although this led to a smaller pool of variables, it also led us to formulate conceptually similar variables introducing redundant variables and missing out on key variables. A better set of conceptually-diverse variables could have been chosen to capture a richer set of factors that could explain the phenomena.
Secondly, our analysis looks at an aggregated area level to explain the consumption of gas based on local household characteristics and is not representative of individuals. Some error or bias about household composition is very likely, leading these empirical conclusions are approximate representations of an area [11]. One alternative could be to use a richer and more relevant consumption dataset like the English Housing Survey which measures detailed housing circumstances and energy efficiencies across England.
How well the data and visual analytical approaches enabled to answer the research questions
Although we could have used a richer dataset, the aggregation of area level data was suitable to our study since we wanted to understand gas consumption in a household as opposed to consumption per person. The variables used in our study were proportions of households and conclusions were not representative of individuals.
Linear regression has been commonly used in previous research to model energy consumption [2]. Some basic assumptions need to be preserved like linear association and collinearity among variables for validity of our model. We used conventional visualisation techniques like scatter plots and correlation matrix which are an effective way to visually assess the associations of variables and interdependence between them. These, in conjunction with statistical techniques like VIF were used to eliminate collinearity and select relevant variables for our model.
We were also interested in understanding spatial variation of different variables and therefore heavily used chloropleth maps to display and compare geographically weighted means, standard deviations and correlations across regions. All these local statistics were computed using the GWmodel package, which enabled us to explore local spatial context. Finally, Moran’s I value was computed to assess the randomness of our residual plots.
Applications of approach to other domain
For this study, we wanted to explore the different factors that may explain the obdurate consumption of gas at the higher global level and a more localised spatial level. The series of analytical tasks conducted in this study are not new and have been formulated previously to study spatial variations in Brexit vote [12]. Since our objective was a similar spatial exploration, this approach was quite suitable to our purpose.
A visualisation is measured on how well it answers the analytical questions being asked of the data and how well the information is displayed [13]. Overall, we believe this approach could be used across a number of domains. For example, it could be used to understand spatial differences in tourism, weather change and insurance sectors. However, it is important to keep in mind the analytical questions being asked of the data. For example if we are interested in analysing temporal aspects of spatial data like seasonal trends across a country in tourism, then we may need to seek a different range of visualisation techniques.
4.1.2017