I've got a lot of comments for one of my previous post. Those comments were mostly about the Atlantic provinces and whether or not I was right in my forecasting. So I went back to the drawing board and made some (actually needed) adjustments. At the end, I believe my projections are more realistic. But at the end, because of various reasons, forecasting those 4 provinces is still less reliable or accurate than Ontario or Quebec for instance. Here is why:
- One of the problem is that the polls are conducted at the regional level. But the variation is usually different in every province (for instance, it is totally possible that one party increased in one province but decreased in the other one). So right from the start, we are facing a problem hard to overcome. On top of that, even at the regional level, the sample size is usually quite small.
- The big, big problem is the 2008 election where so much things were going on. In NF-L, you had the ABC (Anything But Conservatives) campaign of the former Premier. A campaing that was indeed very successfull. Thus, while the Conservatives did pretty good overall in the Atlantic, they dropped by a lot in this province; Other riding-specific things happened. For instance, the fact that Elisabeth May was running without a Liberal opposition. Or still in NF-L, there was the return of a NDP candidate in the riding of St-John-East which boost this party.
My model (as opposed to the other ones) is able to take into account and estimate those riding-effects. But the real question is more: are those effects still present? For instance, is the ABC campaign gone or not? Will the NDP still enjoy such a high level of support in St-John-East? Or are those effects gone?
- The answer is not clear and almost impossible to really assessed. So at the end, after many tries, I decided to go with a compromise: the effects are halfway gone. Therefore, if the Conservatives were (say) down 30 points in NF-L, they are now down only 15 points. The only effect that is completely accounted for is the Elisabeth May effect, for obvious reasons (for the Liberals, I use the level of support that the model would have predicted if they had run a candidate). Those are arbitrarily choices, I know it. But it seems to work better than assuming that all the effects are still there or that they are completely gone. If you think I'm mistaken, please feel free to let me know by commenting.
- While I use my estimated coefficients for those effects, I decided to go with a uniform-swing model for the remaining of the projections. After many tries, it appeared to me to actually work better. On top of that, for past elections, the Atlantic is one of the region where the uniform model actually performs quite well. I might switch back to a fully-estimated model later on, but for now on, I'll stick with this hybrid model.
So check the lastest projections (there will be new ones later today, using the latest Ipsos poll) and let me know if you think the percentages are more realistic.
