From time to tine I will post comment and graphical images on the state of the Canadian horserace for 2019. Some of the comments will be on substantive aspects of current politics. Some will ponder the institutional and historical background, often in line with arguments in my 2017 book, The Canadian Party System: An Analytic History.
The 2019 Horserace
This comment is updated from time to time as new polls come in. See below.
Setup — General
The plots in this comment use nonparametric regression with bootstrapped standard errors. The estimation uses an Epanechnikov kernel and settles on the bandwidth that minimizes the integrated mean squared error of the prediction. The plot is relatively insensitive to short-term dynamics, except to the extent that shifts are sustained. The lines that flank the moving estimate of the mean are its bootstrapped 95% confidence interval. This is not the same as the “margin of error” that polling firms claim for their results. It reflects not just the reliability of individual polls but also the sheer number of polls inside the bandwidth boundaries.
The plot uses all published polls since the 2015 election, as logged on Wikipedia. The date for each poll is for the median day of fieldwork, a truer representation of its place in the flow of events than the last day in the field. Unfortunately, the last day is the only consistently reported date at that site. In the absence of more information I lag most polls three days. The exception is the most frequently reported poll, by Nik Nanos. Nanos uses a weekly rolling poll, such that for any report only 25% of the sample is fresh and observations trail back four weeks. For Nanos I impute the median day as two weeks before the last one and, to avoid overlap, I report every fourth observation. I will modify the median-date imputation as Nanos shortens its tracking cycle and as other firms join it in high-frequency tracking. I do not weight polls for credibility, unlike Éric Grenier at the CBC. Also, unlike Grenier, I do not impose a fixed weighting function for the recency of polls. My smoother does downweight polls as a function of distance in both directions from a given date, and so the closer a poll’s date is to the present the greater its weight in the imputation for the most recent date in the plot. Despite the differences, my estimates are broadly similar to his but with rather less sensitivity to short-term, mainly house-specific, bumps and wiggles.
The Current Situation (posted 29 December 2018)
The first plot is for the time path of Liberal vote intentions, with the last date as 18 December 2018. (The fact that the last date is almost two weeks old testifies to the remarkable sparseness of Canadian commercial polls. See further comment here.) It confirms that the party enjoyed a “honeymoon” for nearly a year after the 2015 election. The advantage began to dissipate in late summer 2016 and by early 2017, the Liberal share had dropped to its October 2015 level, around 40 percent. The plot suggests a gradual further erosion, such that the party fell to the mid-30s. After mid-2018 the Liberal share moved back up close to 37 percent. It appears to have fallen back slightly in the last month or so, to just under 36 percent.
The story for the Conservatives is less dramatic over the long run, but indicates that the horserace has become very close. The Conservatives began a slow climb in mid-2016, as the Liberals began their first drop. Now the Conservatives sit about two points behind the Liberals, with a share just under 34 percent.
Much is being made of how close the race has become. In fact it has been close for months, and was closer in June and July than it is now. Eric Grenier at the CBC Also shows the race to have been a dead heat at mid-year but he also shows the Liberals to have pulled ahead more than I do. He still sees them further ahead than I do, an estimated lead of almost three points.
With this three-point spread he projects that the Liberals have a 51% chance of winning an outright majority. His projection methodology is rather elaborate. A simpler, historically based prediction model for point estimates of seat shares is a harsher on the Liberals. This model regresses the party’s seat share on the the square of its own vote and the square of its chief rival’s vote. The rationale for the setup was laid down by the late, much lamented Duff Spafford* and was used in our book Letting the People Decide. Although it is extremely parsimonious, this simple setup applied to elections since 1960 explains 96 percent of the variance in seat shares for each party and gets most recent elections broadly correct. If current intentions run as I estimate them, the model predicts a Liberal seat share of 40 percent (135 seats) and a Conservative share of 36 percent (121 or 122 seats). At Grenier’s slightly more generous estimate of the Liberal lead, the predictions are a Liberal share of about 41 percent (139 seats) and a Conservative share slightly under 36 percent (121 seats). Either way, the landscape seems very competitive.
* I can’t supply a reference, as Duff–typically of him–sketched the game theoretic argument on the back of an envelope. It seemed churlish to ask for the envelope and this profound insight, like too many of Duff’s, went with him to the grave. RIP.