Table 1.
Explanatory variables selected based on existing analysis and ostensibly the ‘left-behind’ interpretation, also selected to support comparison between US and GB contexts.
Variables are expressed as a proportion of the total population in each US county or GB LAD, with the exception of household income and population density.
Fig 1.
Choropleths of the three outcome variables aggregated to GB LAD and US county level.
Left: majority Leave is brown, Remain is green. Middle: majority Trump is red, Clinton is blue. Right: shift towards Trump is red, away is blue.
Fig 2.
Multilevel models fit to each explanatory variable for net-Leave, net-Trump and shift-Trump.
The overall regression line is bold and regional slopes grey (Scotland is identified with a red line). Plots are annotated with estimates of pseudo-R2, p-values from likelihood ratio tests comparing varying slope with varying intercept models and correlation values between slopes and intercepts for the random slope models.
Fig 3.
Coefficients for multivariate models fit using elastic-net.
Each variable selected is labelled and identified with a bar; bars left of the vertical represent negative coefficients and right of the vertical positive coefficients; bar and variable label lightness varies according to coefficient size; and if the 95% bootstrap confidence interval around the coefficient does not cross zero (e.g. the coefficient is statistically significant) is accompanied with a filled dot. In the bottom right of each plot are estimates of adjusted R2. Left: summary of global models fit without subnational controls; middle: summary of global models fit with subnational controls (England, Scotland and Wales for GB, the nine Census divisions for US); right: choropleth of residuals for models fit with subnational controls. Green—Leave vote is lower than expected; Blue—net-Trump and shift-Trump is lower than expected.
Fig 4.
Semi-spatially arranged small multiples of outputs for models fit separately to each GB GOR.
The plot mappings described in Fig 3 are repeated. To the left, a choropleth and spatially arranged small multiple clarifies the geographic ordering to GORs and filled according to model fit (the darker the fill, the greater the model fit). For a GOR containing fewer than 30 LADs, too few for parametric assumptions to hold, we add observations from neighbouring LADs (based on centroid-to-centroid distances) until that GOR contains 30 observations. These special cases are identified with red text labels and fill.
Fig 5.
Semi-spatially arranged small multiples of outputs for models fit separately to each US state.
The plot mappings described in Fig 4 are repeated. For US states containing fewer than 30 counties, we add observations from neighbouring counties (based on centroid-to-centroid distances) until that state contains 30 observations—identified with red text and fill.
Fig 6.
Semi-spatially arranged small multiples of outputs for models fit separately to each US state.
The plot mappings described in Fig 4 are repeated. For US states containing fewer than 30 counties, we add observations from neighbouring counties (based on centroid-to-centroid distances) until that state contains 30 observations—identified with red text and fill.