Tuesday, March 8, 2011

Election Projections and You

With election speculation ramping up, and pretty much all media saying that an election is inevitable at this point (I hold out for the NDP cave-in), Canadians need resources online that will tell them what all these polls being released actually mean for them. Will that 12-point spread lead to a Conservative majority (no), or will the NDP sitting at 12% lead to a collapse of the party (probably no), or will the Greens ever actually win a seat (definitely no). These are questions whereby ballparking doesn't simply cut it anymore. And that's where election projections come in.

Back in 2008, we didn't have a lot of election projections, except for DemocraticSpace, Paulitics, LISPOP, and ElectionPrediction, and to a point TrendLines. But the fact is, these were all relatively unknown sources that didn't get a huge amount of traffic for the amount of information they gave us. There's also the fact that they were wrong by relatively large margins.

Since 2008 though, there's been something of an explosion of projection sites dedicated to new ways of showing how Canadians voting patterns will end up. The most famous of them, 308.com, is actually a relatively new site (coming in after the federal election) and has gone through several updates and tweeks that have given the site a reputation for, if not accuracy (308 has yet to face the test of a real election), at the very least dedication to trying to make sure its accurate.

But 308.com isn't the only one that's cropped up, so below is a list of Canadian election projection sites that you should pay attention too, and keep tabs on for after the election, just to see how badly off the mark they may have ended up being:

308.com, as noted above, is the most well-known so far, thanks to author Eric Grenier's wonderfully well-presented projections and articles in the Globe and Mail, Hill Times, and etc. 308 uses a basic notional swing system with applied weights to polls (older, less accurate polls weighted accordingly, etc.), and just recently introduced a riding-by-riding system, with riding-specific weighting (historical trends, star candidacy, ministers, incumbency, regional swings, etc.) applied, giving us what is hoped to be a more accurate and specific system. Really well done, accessible, and understandable.

Canadian Election Watch
Canadian Election Watch is one of the most basic projection systems out there. It is essentially a notional swing based on some weighting, unknown as of yet, that is compounded - meaning you start off with a certain level (like the 2008 results), and the polls from there-on are added on to those results, than extrapolated into a projection result. Because its almost completely poll-based, CEW is not necessarily an election projection as it is a poll projection, if you get the difference. It is however an understandable and regularly updated site.

Too Close To Call
TCTC is a very new projection site run by Bryan Breguet whose system is one that is as confusing as it is interesting, using a term he calls 'econometrics'. "Econometrics" usually implies a statistical analysis based on economic data, trends, etc. Near as I can tell, however, Breguet's system doesn't deal in any way with economics, though he's apparently tweaked the system to work with polls. His methodology is explained here, good luck. His site though is well done, and is very accessible and thought out, if not as pretty as 308. But, for all of Bryan's boasting, like 308 he's yet to go through an election, so it remains to be seen how accurate his system really is.

TrendLines is run by a man named Freddy Hutter, who is about as Conservative as you can unfortunately get. That being said, his system is relatively well thought of, essentially using basic statistical projection methods (similar to what 308 uses) and trend analysis to project where the parties will end up in the future based on their current trajectory. However, it's not regularly updated, and he charges a fee for you to get in, starting just recently. His site's layout could also use a touch-up or two.

The Laurier Institute for the Study of Public Opinion and Policy (just say LISPOP) is a university-run projection site that is similar in model to 308.com before the introduction of the latter's riding-by-riding system. They were around back during the 2008 election, but as I mentioned, were off with their seat count by about 13 seats, with big misses in Ontario and the Prairies. Nevertheless, they're regularly updated and have been tweaking their results since then, and will continue to due so.

Democratic Space
DemSpace is a popular name to a lot of people, so not a lot of explanation is needed. It uses a system similar to LISPOP and 308, though closer to 308 because of its riding-by-riding projections. That being said, it offers a somewhat more visual cue of the ridings rather than numerical, with a basic safe, too close to call, or totally projected to go to the other guy system of classification. Like LISPOP, however, it faced a somewhat off-the-mark result in 2008, though a better result in 2006. That being said, author Greg Morrow has done a great job with the site, though he's yet to update it since August due to his new family - totally excusable.

Election Prediction Project
The EPP is not like the other sites noted above, as it actually doesn't rely on any statistical analysis. Instead, it's a site that worked based generally on what you'd more or less call "gut predictions," based on eyeballing trends and the results seen of statistical sites. It's interactive insofar as you can yell at those that disagree with you and make your own prediction, however crazy it may be. It's also been around for a longggg time, and you can go back as far as the 2000 election to see what people were saying then (it's not too different, fyi). It's interesting, but relies a lot on active participants. That said, for a site that doesn't rely too heavily on statistical terminology or analysis, it's pretty damn accurate.


  1. So what do you think of Barry Kay & LSU's only using an aggregate of Angus Reid, Ekos, Nanos and (ugh) Abacus, now... (and not Harris-Decima).


    I see the Post Media folk are running with him as the great quasher of hope to crush those uppity non-confidence folk.

    (even tho' two of those polls are either dated (Nanos, pre-scandals) or suspect, now)


  2. You know, that's an interesting point to bring up. HD doesn't have the best track record but they're a reliable poster with a history, as compared to Abacus. Why would they drop HD?

  3. Hello, I'M Bryan Breguet from Too Close to Call. I would like to add some information.

    1) Yes I use the term econometrics because this is the statistical method I'm using. Econometrics is simply the word used by economists for statistical tools. So no, my model doesn't have anything to do with economics per se. But the term econometrics never implied that. For instance, econometrics tools (such as a regression) are used in political science or psychology all the time. It is simply a method to identify the impact of various factors on one outcome (like explaining if women vote more Liberals, rich people more Conservatives, etc). The main benefit of this method is to allow me to use way more variables than 308 or DemocraticSpace.

    2) Websites' templates and layouts are a matter of personal choices. But I can't help but being a little insulted that you would actually prefer the absolutely ugly and outdated layout of 308. His presnetation is plain ugly if you ask me. On top of that, my blog has pages (FAQ, make your own projections, etc), dynamic comment system and direct access to full page pdf for riding-by-riding projections. Seems to me much better than an ugly, hard-to-read gif picture. But again, personal taste here.

    3) My model has been through a general election. I had a similar model (but simplified) for the 2008 federal election and the 2008 Quebec provincial election. The model worked ok for the federal election (it wasn't working as well as expected for the NDP, but things should go better now that I'm using more data to estimate the coefficients). But it worked great for the Quebec election. The overall projections (based on the actual percentages) were correct everywhere, except for the seat won by QS. And even there, I could have added manually 2-5 points to this party and get 100% right. Both models were on blogs in french though, so that might explain why you didn't know about them. In any case, a clear benefit of my model is that with every new election, I get more data and thuse can estimate my coefficients better. While a model like 308 which is entirely based on one assumption (proportional changes) cannot really update its model. If you want, imagine you want to estimate how much more money does a university degree gets you annually. 308 would simply assume that a college degree would get you 25% more. While for me, I would take data and estimate how much exactly do college gradutes earn more than other people.

    Thanks for the link. Keep the good work.