Case Study 22.1: YouGov's 2017 UK MRP — How Demographics Beat the Consensus

Background

On June 8, 2017, British voters went to the polls in a snap general election called by Prime Minister Theresa May. The election had been called in April with an apparent Conservative lead of 15–20 points over Labour under Jeremy Corbyn. May's strategic calculation was clear: a landslide majority would give her a stronger mandate for Brexit negotiations.

The final YouGov MRP projection, released the day before the election, predicted a hung parliament — no party winning a majority of seats. This was a dramatic departure from the industry consensus. Every major polling firm, using standard polling averages, showed the Conservatives on course for a comfortable majority of 50–100 seats. Betting markets priced the probability of a hung parliament at approximately 15 percent.

The actual result: a hung parliament, with the Conservatives winning 317 seats (short of the 326 needed for a majority) and Labour winning 262. YouGov's model was the only major forecasting operation to predict this outcome.

What YouGov Did Differently

Sample Size

YouGov's MRP model was based on a cumulative survey of approximately 50,000 respondents, collected over two weeks in the final stretch of the campaign. Standard constituency polls might have 600–1,000 respondents and cover perhaps 30–50 constituencies. YouGov's approach provided approximately 80–100 survey respondents per constituency on average — still too few for direct estimation, but sufficient to inform the demographic regression.

The Regression Model

YouGov's regression used the following predictors: - Age group (18–24, 25–34, 35–44, 45–54, 55–64, 65+) - Education (degree vs. non-degree) - Social class (ABC1 professional/managerial vs. C2DE working class) - 2016 Brexit referendum vote (Remain or Leave) - 2015 general election party vote - Region (10 English regions + Scotland, Wales, Northern Ireland) - Urban/rural classification

The Brexit referendum variable was the key innovation. In 2017, Brexit vote was more predictive of 2017 general election vote choice than party identification in many constituencies. Constituencies that had voted heavily Leave in 2016 were moving toward the Conservatives; heavily Remain constituencies were moving toward Labour. Standard demographic variables (age, class) were correlated with Brexit vote but did not fully capture the shift.

The Poststratification

YouGov applied the regression predictions to each of the 650 UK constituencies using Office for National Statistics (ONS) Census data on the demographic composition of each constituency. Because Brexit referendum results were available at the counting area level (309 counting areas covering all of Great Britain), these were incorporated as constituency-level covariates — a technique that substantially improved accuracy.

The Conservative Vote Collapse — Not Primarily Polling Error

An important clarification about 2017: YouGov's success was not primarily about correcting a systematic polling error. The Conservatives genuinely had a large lead at the beginning of the campaign. What happened between the election call and polling day was real movement: a Corbyn surge driven by a successful campaign, the Conservative manifesto controversy (the "dementia tax"), and substantial mobilization of young voters.

The key insight from YouGov's model was not that the polls were wrong but that the geographic distribution of the surge toward Labour was highly concentrated in specific constituencies — those with high proportions of young, educated Remain voters. In those constituencies, Labour's surge was sufficient to hold or flip seats that standard polling averages were treating as Conservative wins.

Standard polling averages showed a national Conservative lead of 3–7 points in the final week. What they did not capture was that this lead was not uniformly distributed: the Conservative lead was larger in Leave-voting constituencies (which they were already winning comfortably) and smaller or nonexistent in Remain-voting constituencies (many of which were marginal). YouGov's geographic disaggregation captured this pattern; the national average obscured it.

What YouGov Got Wrong

YouGov's model was directionally correct about the hung parliament. But its constituency-level accuracy was imperfect. Several marginal seats that the model showed as Labour gains went Conservative by narrow margins; others it showed as Conservative holds went Labour. The overall seat count was within the model's stated uncertainty range, but individual constituency predictions had meaningful error rates.

Additionally, Scotland was a particular challenge. The SNP vote share and its geographic distribution created seat outcomes in Scottish constituencies that the regression model did not fully capture. YouGov's Scottish constituency projections were systematically less accurate than its English ones.

Aftermath and Influence

YouGov's 2017 success transformed the British polling industry's approach to MRP. By the 2019 general election, multiple firms had developed their own MRP models. The 2019 election saw a different challenge: accurately capturing the collapse of Labour's "Red Wall" seats in Leave-voting northern England. YouGov's 2019 model, updated to capture the post-2017 changes in the relationship between Brexit vote and party preference, performed well overall — correctly predicting the Conservative majority, though underestimating its size.

The UK MRP experience has been studied extensively as a model for applying the same technique to American congressional forecasting, with early applications by Economist/YouGov and academic researchers showing promise in U.S. House and state legislative contexts.

Discussion Questions

1. YouGov's model succeeded because it captured a demographically concentrated swing that the national polling average obscured. Does this mean that MRP is always better than polling averages? Under what conditions would a simple polling average outperform MRP?

2. The Brexit referendum vote was a key predictor in 2017. How long do you expect this variable to remain a useful predictor? What would you use to replace it as the Brexit effect fades from direct political relevance?

3. YouGov used approximately 50,000 respondents — a much larger sample than typical polls. Is MRP's accuracy attributable primarily to the large sample, to the demographic modeling, or to the poststratification? Design an experiment that could distinguish between these explanations.

4. The 2017 YouGov projection was released the day before the election, creating a public forecast of a hung parliament when every other forecaster was predicting a Conservative majority. What are the ethical and political implications of publishing such a dramatic outlier? What responsibility did YouGov have to explain the uncertainty in its projection?

5. Several YouGov Scottish constituency projections were less accurate than English ones. What features of the Scottish political landscape might explain this differential accuracy, and how would you extend the model to better capture Scottish dynamics?