I've been thinking about how to reflect this factor in the projection model. Originally, I had thought that my inclusion of the last three election results would act as a bit of a guide wire for my projection. But seeing the Green Party at over 10% leads me to believe that I may need to look into other methods of predicting voting behaviour based solely on polls.
So, I took a look at the last three elections, and how the polling compared to the actual results. I did something like this back in October 2009, and the conclusion I drew was that there wasn't enough data. While I'm still not happy with the amount of data I have to use, I've reversed my decision not to include some method to reflect this in the projection model.
The following chart was done-up for that post, and it is still relevant.
This shows how things generally even out for the Liberals and Conservatives, but also that the NDP and Greens tend to over-perform in polls. First, I looked at the last week of polling in the three last elections. In 2008, the average result of that last week was 34.2% for the Conservatives, 26.6% for the Liberals, 19.5% for the NDP, 9.6% for the Bloc Québécois, and 9.6% for the Greens. The actual election results were 37.7%, 26.3%, 18.2%, 10.0%, and 6.8%, respectively.
In other words, on election day the Tories had 110% of their average polling result, compared to 104% for the Bloc, 99% for the Liberals, 93% for the NDP, and 71% for the Greens.
But, you say, people change their minds in a week. How about three days? The difference is negligible: 110% for the Conservatives, 101% for the Bloc, 99% for the Liberals, 94% for the NDP, and 72% for the Greens. However, we can reasonably assume those last three days to be more accurate, so these are the numbers I will be using. Nevertheless, I think it interesting that the average polling results over the last seven days is more or less the same as the results over the last three days. It makes you wonder when is the point that voters make up their minds.
In the last three days of the 2006 election, the polling average for the Conservatives was 37.1%. It was 27.5% for the Liberals, 18.5% for the NDP, 11.3% for the Bloc, and 5.5% for the Greens. The actual result was 36.3% for the Conservatives, 30.2% for the Liberals, 17.5% for the NDP, 10.5% for the Bloc, and 4.5% for the Greens.
So now the Liberals performed best on voting day, with 110% of their polling average. The Conservatives had 98% of their polling average, the NDP 95%, the Bloc 93%, and the Greens 82%. You might be noticing a trend.
In 2004 there were fewer polls, but nevertheless the result was 111% for the Liberals, 103% for the Bloc, 97% for the Conservatives, 86% for the Greens, and 85% for the NDP.
After weighting the three elections (2008 being worth three times as much as 2004, for example) the average I worked out was to increase the Liberal vote by 1.046 times and the Conservative vote by 1.037 times. I also needed to decrease the Bloc vote by a factor of 0.987, the NDP by 0.928, and the Greens by 0.778.
But why punish these parties for what the pollsters have done in the past, much of it explained by the margin of error? After all, the next election might be completely different. Polling methods could change or the pollsters could simply be luckier. So, I've halved those factors to come up with a number that I am comfortable with.
The popular vote projection, then, will now include this voting-day factor (though I haven't added it yet to the charts at the top of the page yet).
Note, this factor will not be added to the regional results, as the margins of error are much larger and the inclusion of past electoral results minimizes the importance of this voting day effect.
With this new factor weighed into the projections, the popular vote you currently see at the top of the page turns into:
Conservatives - 33.7%
Liberals - 28.3%
New Democrats - 16.3%
Bloc Québécois - 9.6%
Greens - 9.3%
I am far happier with those numbers, but I am open to suggestions as to how I can best deal with this issue.
