Model Development Diary #2: Do candidates matter?
And if so, by how much?
The limitation in any seat projection model is that it is impossible to predict how voters will behave in each individual riding with a great degree of confidence. Even though a national poll of 1,500 Canadians can only hint at the broader shifts in public opinion, for the most part ridings generally swing according to those broader shifts.
But there’s always some variation from one riding to the next.
Granted, some of that variation will always be unpredictable without more granular data, such as riding-level polling. But the rest of that variation can be predicted to some degree by analyzing what impact certain types of candidates have had in the past.
When they matter at all, that is.
This is the second installment of the Model Development Diary, in which I share my conclusions from the work I’m doing in re-designing a seat and vote projection model ahead of the next federal election. The first installment took things down to the studs and looked at first principles.
Those first principles — how to swing the polls — provided the baseline. Now, how do we refine the model a little bit more to try to take into account the exceptions and the outliers, the ridings that do not behave entirely as expected?
I identified a few common threads in the riding-level errors that emerged when testing my swing model against the results of elections held between 2006 and 2025. They were:
The presence, or lack thereof, of incumbent MPs on the ballot.
Byelections, particularly when a seat changed hands.
Exceptions, such as when a major party failed to nominate a candidate or had a floor-crosser on the ballot.
Strong Independent candidates, especially when they are incumbent MPs.
Candidates with high profiles, such as party leaders and so-called “star” candidates.
So, for the last few months I’ve been going through the results of the last seven elections to track down these special cases and determine what impact, if any, these candidates had on the outcome. Here are the results.