(Note: For those of you interested in strategic voting (polls suggest ~30% of you are), see my guide for the advanced polls. It was written Thursday night, and does not take into account polls released subsequently.)
No EKOS update today (apparently there has been virtually no change), so instead of a projection update, here's a post to help you understand why projections differ.
This is the first post of a two-part series on understanding projections. The second post, on how to interpret projections in terms of probabilities,will be available in a few days is available here.
Building a projection model involves many decisions: how to weigh polls (sample size, recency, pollster accuracy, etc.), whether/how to incorporate sub-regional breakdowns and riding polls, whether/how to account for incumbency and star candidates, what to use as a baseline (just the last election or also give weight to historical results), etc. But most of the variation across different publicly available models for this election comes down to one thing: whether the model is based on uniform swing or proportional swing.
What is uniform swing?
Uniform swing assumes that if a party goes up/down by X points in a region, then it goes up/down by X points in every riding in that region. For example, if the Liberals increase from 13.4% to 26.8% in BC, uniform swing would add 13.4% to the Liberal vote share in every BC riding. (Note that the Liberals are, in fact, currently a bit higher than 26.8%.)
What is proportional swing?
Proportional swing assumes that if a party goes up/down by fraction X in a region, then it goes up/down by fraction X in every riding in that region. For example, if the Liberals increase from 13.4% to 26.8% in BC, proportional swing would double the Liberal vote share in every BC riding.
OK, I get the difference. Does it really matter?
It does not matter much when no party went from really low to competitive (or vice versa) in any region. But that's not the case this time: the Liberals more than doubled their support in the West relative to 2011, and almost doubled it in Québec. In such cases, it can matter a lot. Take the BC example from above, and consider a riding where they had 20% in 2011. With uniform swing, they would be projected to get 33.4% of the vote - sufficient on win only in a three-way race. With proportional swing, they are instead projected to get 40% of the vote - sufficient to win most races.
Here is a list of the most recent projections:
No EKOS update today (apparently there has been virtually no change), so instead of a projection update, here's a post to help you understand why projections differ.
This is the first post of a two-part series on understanding projections. The second post, on how to interpret projections in terms of probabilities,
Building a projection model involves many decisions: how to weigh polls (sample size, recency, pollster accuracy, etc.), whether/how to incorporate sub-regional breakdowns and riding polls, whether/how to account for incumbency and star candidates, what to use as a baseline (just the last election or also give weight to historical results), etc. But most of the variation across different publicly available models for this election comes down to one thing: whether the model is based on uniform swing or proportional swing.
What is uniform swing?
Uniform swing assumes that if a party goes up/down by X points in a region, then it goes up/down by X points in every riding in that region. For example, if the Liberals increase from 13.4% to 26.8% in BC, uniform swing would add 13.4% to the Liberal vote share in every BC riding. (Note that the Liberals are, in fact, currently a bit higher than 26.8%.)
What is proportional swing?
Proportional swing assumes that if a party goes up/down by fraction X in a region, then it goes up/down by fraction X in every riding in that region. For example, if the Liberals increase from 13.4% to 26.8% in BC, proportional swing would double the Liberal vote share in every BC riding.
OK, I get the difference. Does it really matter?
It does not matter much when no party went from really low to competitive (or vice versa) in any region. But that's not the case this time: the Liberals more than doubled their support in the West relative to 2011, and almost doubled it in Québec. In such cases, it can matter a lot. Take the BC example from above, and consider a riding where they had 20% in 2011. With uniform swing, they would be projected to get 33.4% of the vote - sufficient on win only in a three-way race. With proportional swing, they are instead projected to get 40% of the vote - sufficient to win most races.
Here is a list of the most recent projections:
C-L-N-B (all models give the Greens one seat)
Date = last day of polling included
Links to all these sites are in the blog roll on the left.
Links to all these sites are in the blog roll on the left.
129-115-86-7 (Canadian Election Watch, unadjusted, 10/9)
131-113-91-2 (Too Close to Call, 10/9)*
127-115-88-7 (Le calcul électoral, 10/8)
127-115-88-7 (Le calcul électoral, 10/8)
123-113-98-3 (LISPOP, 10/4)
(The Globe's Election Forecast, which provides probabilities, is also based on uniform swing.)
*Includes slight turnout adjustment (smaller than this blog's).
*Includes slight turnout adjustment (smaller than this blog's).
Proportional swing models
122-131-80-4 (Three Hundred Eight, 10/9)
130-136-67-4 (The Signal, 10/9)
126-129-80-2 (Election Almanac, 10/8)
123-125-81-8 (Election Atlas, 10/8)
Methodology not found
125-133-77-2 (Predictionator by David Akin, 10/9)
131-110-96-0 (CVM Election Model, 10/7 - this one is almost certainly uniform)
All uniform swing models have the Conservatives leading. All proportional swing models have the Liberals leading. (Update 10/12: The Globe's Forecast now appears to have the Liberals marginally ahead, partly due to its model now assigning zero weight to polls more than one week old.) Variations within each category exist, but they are slight compared to the differences across categories.
Which method is better?
Obviously, I believe in uniform swing since that's the method I chose. On actual election results, uniform swing has historically performed very well. In 2011, uniform swing (clarification: that is, my version of uniform swing, which included an adjustment for the Toronto area) would have, given the actual regional vote shares, projected every party's total seat count outside Québec within 3. It's hard to do any better than that! (In Québec, the NDP would have been underestimated, but I believe that proportional swing would have done so as well, perhaps by more.)
However, in most elections, proportional swing would have also done very well, so it's very hard to say definitively that one method is better than the other. This election may provide a good test, particularly with respect to Liberal results in Western Canada.
I would encourage Liberal/Green supporters to take uniform swing seriously, and Conservative/NDP supporters to take proportional swing seriously. This way, you won't be too disappointed come election night!
All uniform swing models have the Conservatives leading. All proportional swing models have the Liberals leading. (Update 10/12: The Globe's Forecast now appears to have the Liberals marginally ahead, partly due to its model now assigning zero weight to polls more than one week old.) Variations within each category exist, but they are slight compared to the differences across categories.
Which method is better?
Obviously, I believe in uniform swing since that's the method I chose. On actual election results, uniform swing has historically performed very well. In 2011, uniform swing (clarification: that is, my version of uniform swing, which included an adjustment for the Toronto area) would have, given the actual regional vote shares, projected every party's total seat count outside Québec within 3. It's hard to do any better than that! (In Québec, the NDP would have been underestimated, but I believe that proportional swing would have done so as well, perhaps by more.)
However, in most elections, proportional swing would have also done very well, so it's very hard to say definitively that one method is better than the other. This election may provide a good test, particularly with respect to Liberal results in Western Canada.
I would encourage Liberal/Green supporters to take uniform swing seriously, and Conservative/NDP supporters to take proportional swing seriously. This way, you won't be too disappointed come election night!
3 comments:
Many thanks for this. Much appreciated. I look forward to the next one. May we also have an explanation as to why some polls differ so much from others? Best wishes. JA
No problem, glad you found this informative! As for polls, I don't have anything better than the stuff you hear others say: random variation, methodology (IVR, online, live), sampling procedure, time of calls, etc.
On 308.com's CBC anaylsis, in addition to the debateable choice (appears biased) poll weighting that commentor Jeremy Akerman mentioned an another of your blog posts, I also note the following:
There is a "simplistically" labelled, low, middle, avg, high distribution for the seat projection. The conservatives "likely" bubble has always included the low, middle, and avg points, with the "high" projected as an extreme outlier.
The NDP and Liberals, (its shifted over time), have their "low" points projected as extreme outliers, but their highs within the shaded "likely" outcome window. By what justification does he skew these likely seat projections to push the cons high to an outlier, but the left leaning parties only have their lows as extreme outliers....
Post a Comment