Monday, July 20, 2009

It's a Launch! Projection Methodology

Welcome to Canadian Election Watch, a blog where I will make federal seat projections based on the most recent publicly available polls. Of course, several websites like this exist, so in addition to politically neutral seat projections, I will also occasionally provide my own biased view on political issues in Québec, Canada and the United States.

The backbone of the projections is a uniform arithmetic swing model within each region: if polling average says that the Greens have gained 4% in Ontario relative to the last general election, I assume that they have done so in every riding in Ontario. The six regions are: Atlantic Canada, Québec, Ontario, MB/SK, Alberta and BC.

Of course, while the underlying projection method is simple and objective (though its selection is of course subjective), judgment calls are needed* to deal with the following issues, among others:

1. Poll inclusion: I define a poll's age as the time between its midpoint date and the most recent poll's midpoint date. In normal times, a poll up 7 days old is included at full value (unless it is not the most recent poll by a firm), and a poll up to 27 days old is given some weight. These periods may be lengthened when polling is infrequent (e.g. during summers), and will be shortened during election campaigns. The point is to get enough polls to make the projections consistent, but not so many that they become unresponsive.

2. Poll weighting: National polls of n respondents will receive a weight of sqrt(n). I do not use full proportionality because part of the appeal of using multiple polls is to prevent one firm's methodological biases from driving the results. For the same reason, polls that are not the most recent by a firm are discounted by half (in addition to the time discounting).

3. Intra-provincial breakdowns: I will generally ignore these, since it is often hard to figure out what these regions mean (does "905" mean the Greater Toronto Area outside Toronto, the Toronto Census Metropolitan Area outside Toronto, or the actual area of the 905 area code?). Even when we know what the regions refer to, they may not correspond to riding boundaries, and I do not have the desire to go through results by polling station. However, in special cases where there is evidence that a region might be moving (not just voting) very differently from the rest of the province, I will make use of these breakdowns. For example, if the Conservatives are weak in Québec, numbers specific to the Capitale-Nationale area will be very helpful.

4. Ad hoc intra-provincial/regional patterns: a) Tight Montréal ridings move much less than the rest of Québec. So if the numbers tell me, for example, that the Liberals will gain a Montréal riding from the Bloc by a narrow margin, I will likely override that change. b) New Brunswick normally behaves very much like Ontario. So when projecting New Brunswick ridings, I will give consideration to both Atlantic Canada and Ontario numbers. Have you noticed any other (useful) patterns? Something that can improve projections in British Columbia would be especially helpful...

5. Riding polls: These will be given very high, but less than proportional, weight. The main issue here is that I will likely miss quite a few of these, so if you know of one, please inform me by posting a comment about it!

6. Special circumstances: Northern ridings, scandals, by-elections, open seats (especially when the incumbent was popular), seats held by independents, etc. These will be considered on an ad hoc basis, but adjustments for these factors will be minimal due to my lack of local knowledge about most of the country, and the fact that most electors in most ridings vote for a party rather than a candidate.

*Yes, if I built a sophisticated statistical model, I'd be able to push back where the subjectivity comes in. But it is unlikely to help that much, time is a finite resource, and I don't enjoy programming...

Note: Edited on August 28, 2010 and March 2, 2011

No comments: