Can we predict individual ridings using the provincial swing?

Ok so here we are, in election! I know most people are not happy about that, but I'm sure that most of my readers are, as any political junkie. Today, I'll not post new projections based on one or two polls. Instead, I would like to justify a little bit my model and why I think it is valid. Of course, you can get more details if you read the methodology under the tab FAQ.

The point I want to address today is the following: can we really predict riding-level swings using the provincial swings only? After all, I'm sure some of you are skeptical. Simply because one party increases by 4 points in a province doesn't mean this party will increase by 4-points in every ridings. The 4-points provincial swing is an average, and an average could hide a lot of variation. For instance, it is technically possible that this 4-points increase came from an increase of 12-points in 50% of the ridings, and a decrease of -4-points in the other 50%.

I showed in this post what the swing looks like. It isn't perfectly uniform, nor is it perfectly proportional. Actually there is a lot of variations. Nevertheless, it turns out that the provincial swing is still a good predictor, especially if we interact it with other variables. But before showing you numbers, let's see what kind of information we have about each riding. I'm talking here about information publicly available during an election, so the GDP per capita is not one of them.

-The past results of every party.
-The current incumbent.
-The province and region where the riding is located (ex: Montreal West or Quebec city).

So by using the provincial swing and these other variables, we can get a lot more predictive power. For instance, it is obvious that when the Conservatives increase in the province of Quebec, they increase much more in the region of Quebec city than on the island of Montreal. We know that because we can observe the past elections' swings. We can also look if a party did better or worse in the ridings it won the last election (an incumbency effect). This is really where the use of statistical tools such as a regression (or econometrics in general) helps a lot. Other websites can't take into account so many variables, not at once. I can even differentiate the swing in function of the level of support of the other parties (for instance, when the Tories go up in Ontario, do they take votes from the Liberals or the NDP?).

At the end, the predictive power varies across provinces and political parties. It works really well for Quebec and Ontario, for the two main parties. For instance, I can explain as much as 89% of the riding-level swing for the CPC in Quebec! This number is 86% for the Liberals. The remaining 12-15% could be inputed to riding-specific events, or on an efficient organization (to get the vote out).

On the other hand, it works usually less well for the NDP. The reason being that this party has experienced less variaiton over the last couple of elections (in most provinces, the NDP stayed really flat), so I have less variation to identify (pin down if you prefer) the coefficients. Nevertheless, I can explain around 76% of the swing for the NDP in Quebec, but only 22% in Ontario. When this number is that low, I actually switch to a simple linear-uniform model for this party. This is unfortunate but there isn't much that can be done. And anyway, as shown in my methodology, the linear-uniform model already provides a good approximation.

So, here you go. It seems that we can indeed project riding-level swings using only the provincial ones as the source of new information. In average, we could explain over 70% of the swing. But it works better in big provinces and for political parties that have experience changes over the year. I would say the best scenario for the model would be to have the CPC down a little bit (going closer to its 2006 results) and the Liberals up a bit. On the other hand, if the Tories get closer to 40% and the Liberals keep falling, we could face a problem of extrapolation. Indeed, since we've never observed the Liberals so low, we don't really know what would happen. But as long as the changes are not too big (and let's face it, they usually aren't), it works just fine. In 2006, given the correct percentages, the model would have predicted correctly 286 ridings. In 2008, it was 283. And most of the mistakes were in close races (where the margin of victory was less than 5-points). Better than that, in average, the projections were accurate within around 2-points (so for instance, the model predicted the CPC at 46% and the actual result was 44%.