The effects of tactical voting sites
I find large positive effects on Lib Dem vote share of between two and four percentage points
The dispersion of the Remain vote between multiple competing parties created strong incentives for coordination in the 2019 election. Three “Remain” parties (the Liberal Democrats, the Greens, and Plaid Cymru) were able to coordinate at the elite level, agreeing a joint candidate in 60 seats, but this coordination was limited in comparison to the Brexit Party’s decision to stand aside in Conservative held-seats and recommend a tactical pro Conservative vote in 100 more seats. In order to fill this “coordination gap”, several websites offering tactical voting advice sprung up. Here, I review those sites and their effects.
It's not possible to give an exhaustive list of tactical voting advice sites, so I concentrate on four sites which, for various reasons, became important during the campaign:
getvoting.org, a site backed by Best for Britain, a campaign group “committed to finding a democratic way to stop Brexit”, chaired by former Labour minister Mark Malloch Brown, and set up before the 2017 general election.
remainunited.org, a group set up by Gina Miller, the main litigant in the UK Supreme Court case Miller v. Secretary of State for Exiting the EU (which required Parliament to authorise the beginning of Article 50 negotiations), and a former member of Best for Britain. Miller left Best for Britain shortly after the 2017 election, saying that the group had become more about stopping the Conservatives than stopping Brexit.
peoples-vote.uk, a group with a strong degree of overlap with the 2016 Remain campaign, and which saw a significant staff mutiny during the course of the campaign.
tactical.vote, “a project by the grassroots @votetools collective”, and which distinguished itself by providing a meta-analysis of recommendations made by different tactical voting campaigns.
Two sites (getvoting.org and remainunited.org) tried to estimate current public opinion using multilevel regression and post-stratification (MRP), a technique for estimating opinion in small areas using large national samples. In most cases they recommended the non-Conservative candidate who was best placed according to this analysis. Because MRP relies on polling data, and because polling data can become out of date, these sites changed their recommendations over time. For example: the first Best for Britain recommendations were published on November 3rd, and were updated again on November 28th and December 9th.
tactical.vote made its recommendations on the basis of performance in the 2017 election. They recommended the best placed non-Conservative candidate, except where the Conservatives are a distant third place (in which case they make no recommendation), or where the seat is judged “unusual”. 29 seats were deemed unusual, including Chorley (Speaker’s seat), Buckingham (the former Speaker’s seat), and seats with a credible independent candidate such as Dominic Grieve (Beaconsfield) or David Gauke (SW Hertfordshire).
peoples-vote.uk made artisanal recommendations based on “talking to candidates… consulting leading pollsters, and analysing the results of local, European and national elections”.
The recommendations from Best for Britain were greeted with considerable skepticism. This was motivated in part by the large number of recommendations to vote Liberal Democrat, concern over the freshness of the data, and some misunderstanding of how MRP works. As tactical voting advice sites multiplied over the course of the campaign, attention focused on constituencies where the sites offered differing recommendations, giving rise to articles with titles like “Can you trust tactical voting sites?”. This ignored the fact that most sites recommended the same party, and that most of the seats with differing recommendations were difficult to call by any measure.
By the end of the campaign, most sites were in agreement on a large number of constituencies. Because Labour support grew over the course of the campaign whilst Liberal Democrat support shrunk, earlier differences between
tactical.vote and MRP-based sites diminished.
tactical.vote were very much more popular than
peoples-vote.uk. Best for Britain claimed that the site was visited by 3.8 million unique visitors, and the site’s Alexa page rank on the morning before the election (332,947) was beaten only by tactical.vote (231,484). These estimates of unique visitors are probably an over-estimate of the number of unique electors who consulted these sites: the British Election Study team reported that only 6% of respondents (two and three quarter million people when expressed as a proportion of the adult population) listed tactical voting websites as a source of information they used when considering how to vote.
The way in which Best for Britain and Remain United based their recommendations on the result of an MRP model allows a neat test of their influence. If, for example, we were just to examine the swings towards the Liberal Democrat in seats where Best for Britain recommended the Liberal Democrats, we might over-estimate the site’s influence because the Liberal Democrats were always fated to do well in those seats, and the MRP model picked up on it. If, however, we compare seats in which Best for Britain just recommended the Liberal Democrats, because they were fractionally ahead their nearest Remain competitor in the MRP analysis, compared to seats where they were fractionally behind, then we can estimate the effect of the endorsement alone. (You can repeat the exercise for Labour, but it gives much smaller effects).
The problem with this kind of analysis is that it throws away a lot of data which is not close to this discontinuity. As a consequence, the confidence intervals surrounding our estimates of the effects of these endorsements are very large. For example, using rdrobust with a fuzzy regression discontinuity design, the central estimate of the effect on the Liberal Democrat vote share of being endorsed by Best for Britain is 2.05 percentage points — but the 95% confidence interval runs from -17 to 21 percentage points. The estimate of the effect of a Remain United endorsement is very large (+14 percentage points), but again very imprecisely estimated (the 95% confidence interval runs from 0 to 29 percentage points).
We can keep all of the data by running a (more straightforward) linear regression. Here we include the fact of the recommendation alongside other possible influences on the Lib Dem's vote in 2019 — factors such as their past vote share in 2017, 2015 and 2010; incumbency; candidacy decisions by other parties; the Remain share of the vote; and the original MRP estimates. It's broadly similar to what you see in the figures below.
When we do this (full details in the code), we find that the estimate of the effect on the Lib Dem vote share of being endorsed by Best for Britain is around 5 percentage points (95% CI: 4 to 6 percentage points).
The equivalent figure for Remain United is smaller (2.6 percentage points), but significantly different from zero (95% CI: 1.1 to 4.1 percentage points).
These figures are very large — and very much larger than I expected to find. By way of comparison, commonly cited figures for incumbency for Conservative, Labour and Liberal Democrat MPs are 1, 2 and 4 percentage points.
Although these figures are very large in percentage terms, they do not materially change the distribution of seats. If we suppose that Best for Britain had never existed, and remove the benefit to the Liberal Democrats (and add on the benefits/costs for other parties, estimated using similar regression models), the only two seats change hands: Tim Farron's seat of Westmorland and Lonsdale (which might otherwise have been lost), and Caithness, Sutherland and Easter Ross.
It therefore seems that tactical voting can have a significant impact in elections, particularly where the recommendation goes to a smaller party which otherwise struggles to demonstrate local viability. Whether tactical voting will return in such a notable way for the next 202X election remains to be seen. There was a reasonable case for providing tactical voting advice in this way given that there was, especially following the EP2019 elections, reason to think that this election might show significant within-block changes of support along Brexit lines. Although Brexit will continue to be important in policy terms for years to come, it may now have created the voting blocks that we'll see for the next few elections.
I've posted the data here; the code is at GitHub. It's quite possible that I've made a mistake somewhere along the lines here. If you find something odd, please let me know on Twitter (@chrishanretty).
Regression models of Lib Dem vote share using Best for Britain (B4B) or Remain United (RU) data and recommendations.