Participation in electoral processes is a central element of any liberal democracy. However, it also acts as a strong marker of inequalities and exclusion. A crucial indicator of democratic vitality is the number of eligible electorate that actually cast a vote in the ballot, also known as the electoral turnout (Franklin, 1996). A low turnout threatens the legitimacy of the results, and is especially worrisome when a population category broadly chooses not to exercise their civil right to participate in the most quintessential democratic process (Denver, 1974). If political science has extensively researched the socio-economic determinants of electoral turnout, the question is seldom approached from a geospatial perspective (Mansley et al., 2015).
Here, using a data-driven approach, I focus on a set of determinants of socio-economic inclusion to explore how turnout at the 2012 local elections varied across Greater London in relationship to these variables.
For this analysis I chose to work at the ward level. They are indeed the electoral unit and therefore the most relevant scale to study voter turnout determinants. The London Ward Profiles, retrieved from London Datastore, was the main dataset I used. It consists in a CSV file of demographic and socio-economic data (including 2011 Census data) sorted by ward. The dependent variable of interest is Mayoral election turnout at the ward level. The independent variables used are the following:
I also used the London Datastore London Ward and City Merged Boundaries shapefile to draw my maps.
1. Simple OLS regression
I first run an OLS multiple regression using a full set of eight variables that are potentially useful to explain electoral turnouts. The initial regression, with a R-squared value of 43%, shows that only four variables are statistically significant at 5% significance level. These are namely i) average age ii) proportion of home ownership iii) proportion on social rented housing and iv) proportion with no qualifications. These are the variables presented in Table 1. I removed insignificant variables and re-ran the regression on the four variables according to the equation below:
The only surprise from the output is that the coefficient on proportion of households renting social housing is positive. This means that a 1% increase in the proportion of households living in rented social housing increases voter turnout by 0.18%. Social housing has been a major political hot potato for the Conservative party since the last election with the Labour party campaigning strongly on this issue. This regression analysis supports the Labour strategy of gaining votes through a call for more affordable social housing. Lack of education seems to have the largest impact on political turnout. A 1% increase in proportion of populace without qualification reduces voter turnout by 0.6%. This is unsurprising since voting requires a basic understanding of policies and issues which can only be gained with a good level of education.
The R-squared value of my regression is 43%, meaning that the explanatory variables could only explain 43% of the variations in voter turnout. There are many research studies on voter turnout, especially at the local level, that showed that voter turnout is also largely driven by psychology. Hence, it is difficult for my model to capture the general level of political aloofness and degree of laziness among the populace.
2. Spatial autocorrelation
The fundamental assumption for OLS is that observations should be independent of one another. However, since I am looking at observations spread out spatially, there is a possibility that there is spatial dependence between the observations. Wards adjacent to one another might be affecting each other in influencing voter turnout. This has implications for the OLS regression performed above. If the residuals from my model did indeed exhibit spatial autocorrelation, my results would be both biased and inefficient. This is because the dependent variable in one ward would be influenced by independent variable from the neighbouring ward.
I perform the Moran test for spatial autocorrelation for the residuals in our OLS model. The result is tabled below:
From the statistical output, I can see that the Moran statistic is 0.49, meaning that there is indeed a positive spatial autocorrelation for the OLS residuals. The p-value is also less than 5%, showing that the result is significant. I plot the residuals on the map to show where there are significant spatial autocorrelations:
3. Geographically-weighted Regression (GWR)
The solution around spatial autocorrelation is to run a Geographically Weighted Regression (GWR). This is an extension to the OLS regression run previously in that I add weights to the observations depending on how close they are from each other spatially. The result of this regression is tabled below:
The results are very similar to the previous OLS regression but the GWR allows me to estimate the coefficients for each ward. Once again, low level of qualification is the most significant variable in this model. The distribution of each coefficient under GWR can be displayed using a map:
The final geographically-weighted regression enables me to identify differentiated effects of each of the 4 variables on voter turnout. Interestingly, the difference in coefficients display different geographical patterns for each variable; there is no clear East/West, North/South or Inner/Greater London divide simultaneously present in two variables. Age for instance is negatively correlated to turnout in an area that surrounds central London, whereas it is positively correlated in the West and far East parts of Greater London. Owning a house is linked to higher turnout mostly in Inner London. Living in council housing is correlated with higher turnout in the Centre and South, but the relation is negative in the outskirts of the metropolitan area. Finally, the effect of low level of education, our most important variable, is more important in South London and much less in the centre and West.
In this exploration of London Census data, I discovered that out of 8 socio-economic variables centred around social inclusion, the voter turnout at 2012 local elections could mainly be understood by looking at each ward’s median age, proportion of house-owners, proportion of council housing and most importantly the level of education. These determinants influence turnout at various extents across Greater London wards. It is also important to bear in mind that local election turnout is usually much lower than for general elections in most democracies, and this gap seems to be much wider in the UK (Orford et al. 2008). It would be interesting to use this model to study other recent British elections such as the Brexit referendum of general elections organized since 2010.
- Denver, D. T. and Hands, H. T. G. (1974) “Marginality and Turnout in British General Elections,” British Journal of Political Science. Cambridge University Press, 4(1), pp. 17–35.
- Fieldhouse, E., & Cutts, D. (2008). Mobilisation or marginalisation? Neighbourhood effects on Muslim electoral registration in Britain in 2001. Political Studies, 56(2), 333-354.
- Franklin, M.N., (1999). “Electoral engineering and cross-national turnout differences: what role for compulsory voting?”, British Journal of Political Science, 29(1), pp.205-216.
- Mansley, E., & Demšar, U. (2015). Space matters: Geographic variability of electoral turnout determinants in the 2012 London mayoral election. Electoral Studies, 40, 322-334.
- Orford, S., Rallings, C., Thrasher, M., & Borisyuk, G. (2008). “Investigating differences in electoral turnout: the influence of ward-level context on participation in local and parliamentary elections in Britain”, Environment and Planning A, 40(5), 1250-1268.
- Páez A., Le Gallo J., Buliung R.N., Dall’Erba S. (2010) Progress in Spatial Analysis: Introduction. In: Páez A., Gallo J., Buliung R., Dall’erba S. (eds) Progress in Spatial Analysis. Advances in Spatial Science (The Regional Science Series). Springer, Berlin, Heidelberg