Sunday, January 25, 2015

Introducing the 2015 federal election projection model

It is hard to believe that 2015 is now finally upon us. The 2011 federal election seems like eons ago. At the time, the prospect of a majority government, after seven years in which four elections were held, seemed like an endless expanse. But we're now finally in the election year, so it is time to get ThreeHundredEight.com's projection model up and running.

I'll get into analyzing the numbers themselves in the coming days. For now, you can click on the link above or the table above to see the breakdown. For now, let me go over the features of the model, the changes that have been made, and how we got here.

What's been learned since 2011

The roots of the model that will be used for the 2015 federal election, and which has been used in 12 provincial elections that have occurred from coast-to-coast, was first employed in the 2011 federal vote. In some ways, that election was a successful proof of concept. In other ways, it was a failure.

Those who doubt the usefulness of projection models and other critics will undoubtedly point to that failure in 2011. They would not be wrong in doing so, but it would be dishonest. Why? Because the projection model worked very well in 2011.

At least, the seat projection model did. After the election was over, I plugged the results in each region of the country into the model, as if the polls had nailed the call. The outcome was exceedingly close to what actually happened:

The model would have missed the Conservative count by just five seats, and pegged the Liberals and New Democrats to within three. Even the huge breakthrough in Quebec for the NDP would not have thrown the model for a loop - the New Democrats would have been awarded 60 seats and the Bloc Québécois just four (the actual result was 59 and four, respectively).

So the seat projection model proved its ability to translate regional support levels into accurate numbers of seats. This has been proven again and again in provincial elections since. But the call in 2011 was nevertheless missed.

There were two reasons for that. The first was entirely my fault. At the time, I had been basing my model on past elections like those in 2006 and 2008. In each of those elections, the polls in the last weeks of the campaign hardly budged. The swing that occurred over the holidays in 2005-2006 had settled in for some time by the end of that campaign. With those elections as my guide, I calculated that I needed to decay the weight of a poll by just 7% with each passing day.

That was a grave error. It is a system that would have worked well in 2006 or 2008, but utterly failed to capture the late swing away from the Liberals and towards the New Democrats that occurred in 2011. What happened was that, by election day, the projection model was still roughly one week behind where it should have been. Accordingly, it projected a Conservative minority government with 78 seats going to the NDP, 60 to the Liberals, and 27 to the Bloc.

After that disaster, I changed the way the weighting system worked and increased the decay to 35% per day. This has been employed in the past dozen election campaigns and has had no trouble capturing the shifting mood of the public. But that is, of course, if the polls capture it as well. And that is the second reason why 2011 did not go well.

The polls did a moderately good job of things in 2011. They were on the money in Quebec, where the electorate had shifted so dramatically. But at the national level, the polls missed out on the Conservative majority government. With the system in place now, the vote projection model would have given the Conservatives 36% of the vote, instead of the 39.6% they actually got. Both the NDP (31.5% instead of 30.6%) and the Liberals (20.1% instead of 18.9%) would have been over-estimated. The seat projection would have put the Conservatives about a dozen seats shy of a majority government.

This is why, starting with the 2012 Alberta provincial election, I developed a system to estimate likely ranges of support. This system takes into account the potential for the polls to be wrong, and is reflected in the table at the top of the page by the low/high and minimum/maximum ranges. Nineteen times out of 20, the result should fall within the maximum/minimum ranges. There are varying levels of likelihood of the results falling within the other ranges, as explained on the methodological page.

These ranges allowed me to be one of the only people to consider the possibility that Wildrose would not win that 2012 vote.

Starting in 2013, I added another element to the projection model. I found that, too often, people were taking the projected results in individual ridings too literally, as if they were a real indication of support. For that reason I have long thought about not publishing riding projections. But I think it is important to 'show my work', and that is what the riding projections do. They also include ranges that should give people a decent idea of who is and who isn't in play in a given riding.

Still, I wanted to make sure people considered the potential for error. I added probabilities to the projection model, measuring the likelihood that the call the projection model makes in each individual riding will be right.

To take London North Centre, where the Liberals held their two-day caucus meeting last week, as an example, the model currently says that the Liberals would win it with between 43% and 50% of the vote. The Conservatives, at 29% to 34%, are apparently not in play. But the model also says that the Liberals have an 82% chance of winning it. That means that, despite the ranges projected by the model, there is still a roughly 1-in-5 chance that the Liberals will not win the riding.

These probabilities were the last major addition made to the basic projection model, which has been employed with very few changes in the most recent provincial elections in British Columbia, Nova Scotia, Quebec, Ontario, and New Brunswick. A few new minor features, however, have been added for 2015.

What's new in 2015

While none of the basic mechanisms of the model have changed (see the methodological page for complete details on how it works), there are some differences between the 2015 model and the model that was used in prior elections.

First is something that is gone. In past elections, I've applied a "floor-crosser" factor in ridings where a sitting MP has crossed the floor to run with another party. It has had a mixed record. In some cases, the "floor-crosser" factor was quite effective. In others, it was greatly off the mark. I have decided to drop it as a factor that is taken into account. In its place, the party losing a floor-crosser suffers a 'no incumbent' penalty, and the party picking up the floor-crosser gets a 'star candidate' bonus. In past elections, this would have worked better than the system that had been used.

A new factor that has been added to the model, however, takes into account party leaders. My research has suggested that leaders are far more difficult to defeat than normal incumbents (though it can always still happen, of course) and that an MP running as a leader for the first time perform especially better than other incumbents. The loss of a leader as a candidate in a given riding, in addition, is far more penalizing than the loss of the average MP. So, for example, Gilles Duceppe's old riding will be harder to win back for the Bloc Québécois than the average riding they lost in 2011.

The particularities of the 2015 election

As always, an election has its own particularities that need to be taken into account. The idea behind the projection model is that it can be applied uniformly in all jurisdictions in Canada. But the fact is there are always oddities in individual elections that have to be considered.

The new federal riding boundaries is one thing that will make the election potentially harder to call than would have otherwise been the case. But the model has dealt with new boundaries before. In the 2011 Manitoba provincial election, for example, 56 of 57 ridings were called correctly despite the shifting boundaries. For this election, the model merely uses the transposed results that were calculated by Elections Canada. MPs that have shifted ridings are treated as incumbents. Ridings with no incumbents are left unadjusted by the incumbency factor.

The presence of Forces et Démocratie complicates matters. Normally, new parties have been handled by the model without much difficulty. Their support is awarded equally in all ridings within a given region (with the correct regional results, the model would have been within one seat of Wildrose's result in 2012, despite the party fielding candidates for the first time in many ridings).

But no pollster has included FeD in its surveys yet, so we cannot gauge their support. What I have done, until better data becomes available, is to treat Jean-François Larose in Repentigny as I would any sitting MP running as an independent (he retains a portion of his 2011 vote). For Jean-François Fortin, I have simply decided to adjust his 2011 support levels with the Bloc exactly as Jean-Martin Aussant's PQ support in the 2012 election shifted when he led Option Nationale. He seems like an appropriate example to use as a guide.

The floor-crossing of Bruce Hyer from the NDP to the Greens (via a stint as an independent) is also hard to model. Floor-crossing between the 'establishment' parties is relatively straight forward. Not so when it is the Greens, because their base in Hyer's riding was so small in 2011. Treating him like any other floor-crosser would, I think, under-state his likely support. What I've done for him, then, is to adjust his support levels in the same way that Blair Wilson's support shifted when he crossed to the Greens before the 2008 election. Hopefully a riding poll will be done in the riding to give us a better idea of what is going on, as I fear the system in place is also inadequate.

The last particularity of the 2015 election is the presence of an independent candidate like Inky Mark. In 2011, the model did not make any adjustments for the independent candidacy of Hec Clouthier, a former Liberal MP, in Renfrew-Nipissing-Pembroke, who captured a respectable share of the vote. Mark used to be an MP for the riding of Dauphin-Swan River-Marquette (now known as Dauphin-Swan River-Neepawa) and is running as an independent. So, based on past cases of former MPs attempting a comeback as an independent, Mark has been awarded a portion of the support he had in the last election.

Don't miss the forest

As always, I urge readers to exercise caution when looking at the projections. Consider the ranges very carefully. Don't take the individual riding projections as fact. The model simply cannot take into account the effects of local issues in all 338 ridings, and some will be wrong.

Take, for example, the riding of Outremont. It is currently projected to go Liberal, with the NDP in range of winning it. That is Thomas Mulcair's riding. Do I think he is actually going to lose it? Not for a second. But the model does, simply because of the gains the Liberals have made province wide in Quebec.

And that brings up something I am concerned about in the 2015 election. Polls have shown strong gains for the Liberals in Quebec. They have roughly doubled their support in the province since 2011. The model, then, doubles their support in every riding. That means the party is in a good position to sweep Montreal, where they had a decent base in 2011.

But regional polling by CROP has shown disproportionate growth for the Liberals outside of Montreal, meaning that the model may be inflating the party's numbers too much in Montreal and not enough outside of it. This is why, for example, it could be very wrong about Outremont.

The overall seat count for the province, however, is likely to be accurate. Using the polls from CROP, I made projections for the province at the sub-regional level. The end result was not much different. At the riding level, there would be some important differences - more Liberal wins outside of Montreal, fewer around the city. But the sum total of the seat count remains the same - particularly when we consider the ranges. This is why I beg readers not to miss the forest for the trees. The individual riding calls are not nearly as important as the overall regional and national projections.

You might wonder why I don't adjust the model to project sub-regionally in Quebec. The issue with that is that no other pollster is releasing sub-regional numbers for the province. And in 2011, CROP exited the field over a week before the end of the campaign. It would seem unwise to turn the Quebec portion of the model into a sub-regional one, when in all likelihood I won't have information to plug into it.

That about covers everything. I reserve the right to make some adjustments between now and the start of the campaign, though I don't suspect anything but tiny tweaks (perhaps to the values of the 'factors', for example, if I have time to plug some more data points into my calculations). If you notice anything that looks fishy, or even typos, please do alert me to them as errors could have crept in.

Though the model is fully automated, the graphics aren't. They can take quite some time to pull together, so at this stage I'm not sure how frequently I will update the projection. At least weekly, certainly, but updates with every individual poll, as I was doing for the running poll averages, may be too much. And just to make things clear: no, I am not changing the name of the site to ThreeThirtyEight or some variant of that. Consider the name of the site as a historical monument to when it was launched, during an era of seemingly endless minority governments.

Are we heading back to that this year? The current projection, with the Conservatives at between 117 and 155 seats and the Liberals between 107 and 144, would seem to say yes. But more on that later this week.