Seat projections
In predicting the result of the election on Thursday, getting the levels of party support is only half the battle (if anything, it's probably the easier bit). The harder bit is translating that into actual seats - especially given that this looks likely to be an election where there are large shifts of support between the parties.
To put my own cards on the table, uniform national swing has some obvious logical flaws (like occassionally predicting a negative vote in some seats) and a naive application of it - like the one in the corner of my own site :) - has obvious flaws, such as assuming Scotland will have the same swing as the rest of Great Britain when we know from past elections that it normally doesn't.
However, that aside the swing at general elections does tend to be broadly uniform - there are not systemic, predictable patterns of different sizes of swing in different types of seats, and therefore my view remains that the best way of predicting seats is to take uniform national swing as your basis and build upon it - treating Scotland separately, taking account of random variation of swing between seats and uneven distribution of marginals, and adjusting for any likely divergences from the UNS that regional or marginal polling has suggested will occur. This is essentially the approach that Rob Ford et al are using for the PoliticsHome prediction model here (and it is in turn based on a model developed by John Curtice and David Firth for the highly successful 2005 exit polls).
You regularly get other seat calculators cropping up, often based upon variations of proportional swing. The attraction of these is that the Proportional Loss hypothesis doesn't have a bad record of actually predicting number of seats won at elections. It also gets rid of the logical absurdities that UNS produces - no negative votes with proportional swing! Political academics however tend to recoil in horror from it because it doesn't actually reflect the pattern of change in vote - if you graph changes in vote in seats against the level of support parties had at the previous election, the change is broadly uniform. Other than at the extremes, it is not proportional to the level of support the party started out with. This election we have one such model from Nate Silver, the lauded proprietor of 538 in the USA.
There is an article by Rob Ford on Pollster.com here arguing the case for his model over a proportional one, and Nate's response is here. My sympathies lie with Rob Ford, though Nate raises some interesting points (obviously the 2% boost to Con-v-Lab marginals is crude, though the reason Rob's team haven't done the same for the Lib Dems is that there was solid polling evidence for Conservatives doing better in marginals, and at the time no good polling evidence that the Lib Dems were. In terms of whether there is data to support an incumbency effect, a crude analysis of the data from past elections would suggest there is one. But if Rob's team have done proper regression analysis and found it was illusionary...I am not a statistician, so am not in a position to judge between their claims).
I've no idea who will end up closer - and is doesn't necessarily tell you the theory is right anyway. Despite the fact it doesn't actually model reality very well, Proportional Loss has often got closer than Uniform Swing, so you can be right for the wrong reasons. Equally, any model is only as good as the data you shove into it and while I agree wholeheartedly with the principle of Rob's team's approach, I wouldn't necessarily stick the same numbers into his model (and would probably make different assumptions about how much better the Conservatives will perform in their marginals, and how well the Lib Dems will perform in their targets.) The figures Nate is putting into his transition matrix also seem to be largely a matter of judgement, rather then directly taken from the polling figures.
While I'm on the subject, do have a look at this paper by David Voas that Rob links too, which discusses how uniform the swing has been at recent elections and what sort of deviations we've seen from UNS.
UPDATE: Rob's response to Nate Silver's response is here.
--
NB: Please bookmark the backup site to UK Polling Report here - http://ukpollingreport.wordpress.com - as we get closer to the election I expect the main site to increasingly crumble under the weight of traffic. All posts will be mirrored on the backup site, and at very busy times (especially election night itself) I will close comments on the main site and move discussion over to the backup.
