Friday, October 7, 2011

Ontario election - the morning after

Ontario elected a Liberal government last night, but it took until the wee hours of the morning before it could be definitively called a minority government. The Liberals won 53 seats, one short of a majority. It doesn't get any closer than that.

I called a Liberal majority. The polls suggested that would be the case, with the gap being 3.3 points in my weighted aggregation and 4.8 points in the average of the last six polling firms to report. Instead, the gap was 2.2 points. Some pollsters did very well (Forum and Abacus, for example), others did less well. That is how it goes.

And just like the pollsters, ThreeHundredEight will learn from this election and move on. It was, by no means, a failure like that of the 2011 federal election. And it was only a few days ago that I called 98.2% of ridings, 56 out of 57, in Manitoba. Like the pollsters, you lose some and you win some. In this election, my vote projection beat out four of the seven polling firms to report in the last week of the campaign.

Before moving on to how the projection performed, I invite you to read my articles today on The Globe and Mail website about how the pollsters performed, and at the Huffington Post Canada, on where the election was won and lost. Also, a full post-mortem of the Ontario projection will follow in the coming weeks.

The projection model performed adequately, with an 85% accuracy rating on the riding calls and an average error of 5.3 seats per party. In a more clear cut election, that would have been acceptable. In a close election like last night's, that was the difference between a minority and a majority government. But it compares quite favourably to the 13.3 seats per party error in Ontario that I had in the abysmal federal election projection.

The seat ranges, dipping to 51 for the Liberals, did envision this sort of minority result.

But the polls are not to blame for the errors in the precise projection. With the actual results, I would have still projected a Liberal majority of 58 seats, with 30 going to the Progressive Conservatives and 19 to the New Democrats. Why?

There are two reasons for the discrepancy. The first is turnout. At 48%, a drop from 53% in 2007, the question of who votes becomes paramount. Polls showing between 70% and 90% of respondents as decided are going to have problems and seat projections based on past results are going to have problems. Whether this was actually one of the reasons for the errors in the projection is difficult to say, but it would seem to be one likely suspect.

The second is the regional variations in Ontario. As the polls don't divide Ontario up uniformly, I did not have a regional model for the province. In other provinces this might not be a big issue, but in Ontario it is a big problem. I will work towards developing a regional model for Ontario for the next election, whenever it comes.

That it was the regional breakdown that caused the most problems is apparent by the projection's accuracy on a region-by-region basis. The projection called every seat correctly in the Hamilton/Niagara region and 91% of the seats correctly in central Ontario and Toronto. The accuracy rating was 89% in the GTA, still acceptable, while it was 86% in eastern Ontario. In northern Ontario, 9 out of 11 were called correctly.

In southwestern Ontario, however, the projection was only 67% accurate. This is where the projection failed.

Had the polls been dead-on, my projected seat range would have been much better, at 49-60 Liberals, 28-39 Progressive Conservatives, and 18-21 New Democrats. I only would have been off for the NDP by one seat in the ranges. This is where local factors come into play - the projection (except in rare cases, as in Manitoba) cannot get everything right. The ranges are just as important, but just as a pollsters can't report only using the MOEs, I need to make a precise call.

The Electoral Track Record has been updated with these details, and you can see how it stacks up to past performances. Let no one say that I cover up past errors, they are there for everyone to see - as are my successes.

Going forward, I believe that a regional model is necessary for some provinces, particularly Quebec and Ontario. They will be developed. I also believe that ThreeHundredEight needs to be able to predict the vote and not just aggregate the polls. An aggregation needs to be maintained to give a clear picture of what the polls are saying, but a prediction should be more useful in the future. With the Conservatives being under-estimated in every election so far this year, it seems that there is a systemic problem in how people respond to polls. Is the "shy Tory" factor for real in Canada?

I don't consider it too likely that Ontario will be heading back to the polls soon, so there is time to work on these things. Newfoundland and Labrador votes Tuesday and Saskatchewan goes to the polls on November 7. There is still a lot to do! The Newfoundland and Labrador projection will be updated over the Thanksgiving weekend, but in the meantime the Campaign Polling Trends chart has been updated for the province.