If an election was held today, who would win?

Introducing a new way of turning polling data into seat projections

Since I started Abacus Data over 13 years ago, I’ve resisted producing seat projections to go along with our polling. We’ve had a great track record producing accurate estimates of the popular vote, but I haven’t developed any models to estimate how those polls would translate into actual seats. Parties form government by the number of seats they win, not by the number of votes they earn.

I’m always looking for new ways to deliver insights to subscribers and my clients.

I’m very excited to share something new with you today.

Introducing MRP: A powerful tool for predicting elections and advocacy/advertising campaigns.

What is MRP?

Multi-Level Regression and Post-Stratification (MRP) is a sophisticated statistical technique increasingly used in election forecasting. MRP has shown remarkable accuracy in predicting election outcomes, including surprising results in the 2017 UK general election, the 2019 UK general election, and the 2016 US presidential election.

How Does MRP Work?

MRP involves conducting a large nationwide poll, focusing less on the representativeness of the sample for the entire population and more on collecting extensive data about respondents. This data is then used to develop a mathematical model predicting how different demographic groups are likely to vote.

We then analyze the relationship between these predictors and the individual's primary opinions or intentions. For instance, it may be observed that younger Canadians are more inclined to support progressive parties, while older individuals might favour conservative options, or that people in urban areas like Toronto have different voting patterns compared to those in rural parts of Alberta. These correlations, although often expected, are quantified with precise numerical values through regression.

Lastly, we apply this regression analysis to predict the likelihood of certain opinions or voting intentions across different regions or electoral districts. In the realm of political surveys, this enables us to estimate the level of support for each political party within each Canadian constituency, aiding in the prediction of electoral outcomes.

The Edge of MRP Over Traditional Polling

Traditional polls often struggle to translate national vote shares into specific parliamentary seats due to the challenges of polling every constituency with sufficient sample sizes. MRP, however, allows for detailed local predictions based on national surveys. It integrates census data, such as age and income distributions, with survey data, enabling predictions even for areas with few survey respondents.

Furthermore, regression analysis can pinpoint potential supporters based on their likelihood to resonate with a particular standpoint. This information is incredibly useful for designing targeted marketing or advocacy campaigns, whether they are deployed through digital channels or more traditional media.

One of the most significant advantages of regression is its ability to infer probable responses to specific questions within even minuscule population segments. This might include areas as small as a municipal wards with around 20,000 residents or a census output area of about a few hundred individuals.

With the capabilities afforded by regression, it's possible to make these inferences without directly polling each distinct subgroup.

A weakness of this model is that it cannot easily account for local candidate impacts like the impact of incumbency (which is why I don’t think Elizabeth May will lose her seat in BC, despite the model projecting that outcome).

Imagine being able to tell all 338 MPs what their constituents think about issues. With this methodology, we can do that.

As telephone survey response rates continue to drop and the cost to do these surveys becomes far too expensive for most advocacy groups or political organizations, this methodology provides an accurate and more affordable alternative.

How would a Canadian federal election result if today’s polls matched votes cast?

We use data from four nationally representative surveys conducted by Abacus Data from September 4 to November 5, 2023 with a total of 8,225 respondents in the surveys.

Over this period, the national vote intention estimate of these surveys by federal party is:

Conservative 40%
Liberal 26%
NDP 18%
BQ 7%
Green 4%
People’s 4%
Others 1%

When the MRP modelling is run - using key demographic, socio-economic, past vote, and regional variables - it produces a similar vote distribution:

Conservative 40%
Liberal 26%
NDP 19%
BQ 8%
People’s 4%
Green 3%
Other 1%

Seat Projection

Based on these vote distributions, we are projecting a Conservative majority government.

The Conservatives would win 204 seats in the new 343 seat House of Commons with the Liberals at 69, forming the Official Opposition. The BQ would win 43 with the NDP 27. The Greens are projected to win 0 seats, based on the model, but it’s likely that at least Elizabeth May would hold her seat if she runs in the next election.