Rain and the referendum
tl;dr version: no association between rainfall and turnout when controlling for past turnout, region
On the day of the referendum, severe thunderstorms affected London and parts of south and East England. Given that the result of the referendum was expected to be close, given that lower turnout was presumed to favour one side in the campaign, and given anecdotal and academic evidence in favour of the idea that heavy rain depresses turnout, there was some speculation that this heavy rain might affect the result.
The link is intuitive. People dislike rain. Turning out to vote normally requires braving the rain. People engage in actions at lower rates when they have consequences which are in part negative. Hence, rain should cause lower turnout.
The most cited paper on this topic argues that rain increases the (psychological) costs of turning out to vote, and finds that an inch of rain reduces turnout by almost one percent. Others authors have found a negative effect of rain on turnout in the Netherlands and Spain.
However, other authors have found either that there is no effect, or that there are mixed effects, so that rainfall depresses turnout in some places and increases turnout in other places.
So, I decided to test the link in the EU referendum. Were people right to be worried about the rain?
What are my measures?
My dependent variable is the percentage turnout in each local authority area. Due to a number of small counting areas with high turnout, the average turnout across counting areas was slightly higher than turnout nationwide, at 73.8 percent (SD = 5.07).
The independent variable in the analysis is the average rainfall in each local authority area between 7am and 10pm, measured in millimetres. The estimates are made using rainfall radar data, rather than ground-based rain gauges. A raster map of rainfall across the UK (excluding Shetland and Northern Ireland) was provided by Weatherquest, a private weather analysis firm. Using this map, I obtained, for each local authority, the average value of all the pixels contained in the local authority area.
The mean value is 3.25 millimetres (SD = 4.34). The regions with the highest rainfall were East (10.2 mm), London (7.1 mm) and South East (5.5 mm). The original map, and the area-by-area values, are shown below.
In examining the association between rainfall and turnout it is useful to control for other factors which might also affect turnout. Here, I consider three factors: turnout in the 2014 European Parliament election, an ex ante estimate of how likely each area was to vote Remain, and region.
The 2014 European Parliament elections were the last national elections for which turnout figures are available at the local authority level. By including these figures, we allow for the possibility that different regions may have levels of turnout that are generally higher or lower. Turnout in the European Parliament elections was much lower than turnout in the referendum: the average across counting areas was just 35.24 percent (SD = 3.9).
Different parts of the country also had different prior dispositions to vote for Leave or Remain. I include these in the model because commentary prior to the referendum suggested that turnout might influence the outcome, with Leave areas more motivated, and because subsequent analysis has shown that turnout was generally higher in areas which voted Leave. I use estimates I produced based on data from the British Election Study. These estimates — which were scaled in line with a 50:50 split nationwide — were published before the referendum, and represent ex ante estimates of how likely each region was to vote Leave or Remain.
Finally, I include regional dummy variables to allow for the possibility that different regions — in particular regions which had recent elections — might have levels of turnout systematically lower or higher than might be expected on the basis of the factors already considered.
What do I find?
I begin by showing the bivariate association between rainfall and turnout in Figure 2. Each plotted point represents a local authority area. The solid line indicates the best-fitting ordinary least squares regression line, whilst the dashed red line indicates a local regression. As the plot shows, to the extent that there is a relationship between rainfall and turnout, it is a positive one: the more rainfall, the higher the turnout.
We can check whether this association is significant and/or robust to the inclusion of additional controls by running a series of regression models of turnout. These are gathered in Table 1.
As the table shows, the bivariate association between rainfall and turnout is positive, and significant. On this model, each extra millimetre of rainfall increases turnout by 0.15 percentage points.
This finding may be artefactual — it may simply be that the rain fell in regions which ordinarily vote at high rates. Indeed, the moment we control for past turnout, the relationship ceases to be either positive or significant. Rainfall now has an effect one-tenth of the magnitude in the opposite direction — and an effect which is not statistically significant at the five or ten percent levels.
This relationship is slightly stronger (i.e., more negative) when we control for the expected gap (Leave-leaning places voted at higher rates) and region, but never regains statistical significance. The findings from the final model suggest that each extra millimetre of rainfall would decrease turnout by between -0.2 and 0.07 percentage points, or somewhere between an effect which is negative and five times stronger than the effect found by Gomez, Hansford, and Krause (2007), and an effect which is positive.
Even if you accept these analyses, might you still conclude that rain really ought to have some effect? There are two things you might look to.
First, these results have been obtained using whole-day rainfall data: it is possible that figures which place greater weight on times during which turnout is high (7–10am; 5–9pm) would show different figures.
Second, the data I have used are not as detailed as I would have liked, and it would be good to have figures on turnout and rainfall at, say, ward level. I am not sure, however, that the effort involved in obtaining better data would be worth the return in terms of our knowledge about turnout and its determinants.
