Silver’s Model on the British Election

Following five days in Chicago “working” at my day job, and then a couple days on the Oregon coast celebrating my birthday with my fiancée, it’s time to get back to serious business.  Barring that, I’ll write a general post on the British election.  This was meant to be an omnibus post addressing several topics only tenuously related as having something to do with the British election, but the more I dug into the first topic, the more I had to say.  I’ll leave the rest for a post hopefully later today.

Nate Silver, Renard Sexton, and Daniel Berman have developed a sophisticated, nuanced predictive model.  It has many strengths, so I’ll concentrate instead on a few critiques.[1]  I’ll preface this by saying that I only just read the post on this model, so these are off-the-cuff thoughts.[2]  First, he presents the “uniform national swing” as a bit of a straw argument.  Silver is correct to suggest that its value lies in its simplicity, and that it does on occasion fail miserably — but then too have the very probability sampling based opinion polls in the British context (e.g. 1992) upon which such vote -> seat translations are based.  When it tends to work, even in its lack of precision, it does so because it is assumed that the error for any given deviation in swing is randomly distributed across the entire sample.  In other words, across the 650 constituencies, deviation from a uniform swing average out.  Furthermore, there are sophisticated models, such as that used by the Elections Centre at the University of Plymouth, do factor in a number of additional variables in making seat projections.  Indeed, and somewhat ironically, in the introduction to “Step 2”, he is simply proposing a uniform national swing based on their sophisticated matrix rather than simple polling data: “Originally, this step had been simple: we assumed that the proportion of voters changing hands from one party to the next was the same in each constituency.”  They didn’t reject this approach because it was a uniform swing, but rather because using their matrix, it produced some illogical results.

Usually, the limitations of the uniform swing model are largely irrelevant as it usually does predict who will form the government, and that’s all that really matters.  Much like while we’ve known all along that vote counting is simply another statistical routine infected with concomitant error, that error is usually a hell of a lot smaller than the margin between the 1st and 2nd placed candidates.  Only in extraordinarily close elections does error become salient, the best known example of which was Florida 2000.  In the 2010 British election, the limitations of the uniform swing model are likewise laid bare, as Silver argues.

Second, while Silver believes that this model “embodies what I believe to be sound logic about voter behavior”, he does undersell the notion of the tactical voter, which is a central concept to what we as political scientists understand about voter behavior from both theoretical (e.g. Downs, Duverger, Rae, Lijphart, et al. ad nauseum) and empirical perspectives.  There is quite a bit of empirical literature on tactical (aka strategic) voting to suggest that it does happen, often, and when it does happen, those voters who employ it damn well understand it.  My British students, admittedly while taking political science classes at a university from a prof who does voting behavior , understand the concept of tactical voting intuitively — much better than my American students ever did, or presumably the students at the university in the Netherlands where I used to work.  It is part of the political culture in the United Kingdom in ways not approached in the United States (due to the lack of a strong national third party or strong regional parties) or the Netherlands (due to their use of a single national constituency and list proportional representation, which largely eviscerates the incentives to vote tactically).  Indeed, while Silver suggests that “my impression is also that the phrase “tactical voting” has somewhat too much currency is often used as a band-aid to cover up flaws with the uniform swing model”, that has not been my impression at all, either in conversation or from reading the empirical literature.

Third, regarding retiring incumbents, I’d need to see his regression data before commenting with any depth, but here I’ll offer an impression of my own.  Incumbency simply doesn’t matter anywhere near the way it does in the American context.  Granted, Silver allows for this, so I’m intrigued by the finding, and also by the deviance between the estimate of 1.5% for Labour and the Tories, and 3% for the Lib Dems and others.  What accounts for the effect to begin with?  What explains the deviance between the two sets of parties?  I don’t doubt the finding, but I’d like to better understand the causal mechanism behind it.  I suspect that there is an effect similar to retirements in House (and Senate) elections: members of the party expected to lose seats, especially those in marginal seats to begin with, are far more likely to opt out, so we’d see a lot of Labour retirements in 1979, Tory retirements in 1997, Labour in 2010.  What they take away with retirement from the electoral arena isn’t name recognition the way it’s understood in the US, but rather, I suspect, their mobilizing organizations.

I have one minor and one major point remaining.  The former is that Northern Ireland should be taken out of the equation entirely, and indeed in every “swingometer” and every matrix that translates vote % into seats, this is done.  The three “national” parties in British politics simply do not exist in Northern Ireland, thus the concept of swing to / away from any of the three major parties is irrelevant.  Northern Ireland will elect its 18 MPs from Sinn Fèin, the SDLP, the UUP, and the DUP.  There is an outside chance that there will be a new unionist party in the mix (the TUV), and likewise a chance that the UUP will be electorally eliminated — but this is partially due to the controversy surrounding their “alliance” with the British Conservative Party.  However, it’s as illogical to discuss Labour, Conservative, or Liberal Democrat for these 18 seats as it is for the SNP to gain vote share in Watford.

I assume that the major point is factored into their model, but I’d like to know how — with 2010, constituency boundaries are changed, in some cases quite radically.  In US parlance, it’s redistricting.  Therefore, without precise data on how the new constituencies map onto the old, any model of swing based on 2005 constituencies is open to considerable error.  These data do exist, indeed the media guide on the new constituencies (and how they map onto the old) was produced by my colleagues in my department.

Finally, I would like to highlight one strength in the model: that of regional adjustments.  Yes, Scotland is different than the Southeast, and yes, any variation in swing away from Labour matters where it is located.  This is, to my knowledge, already accounted for by the more sophisticated academic vote% -> seats matrices (I can’t speak for the various media outlets), but it’s good to see this acknowledged.

Regular readers of LGM know I have admired Silver’s work going back to his BP days, and this is no different.  In all, its an impressive, nuanced model, and honestly, I would like to see it succeed as the limitations of assuming a uniform national swing are plain.  However, the uniform swing is not as dire as suggested, some if not a lot of this work has been done, and there are several points, both in the narrative and in the functionality of the model itself that need to be addressed in future iterations of this model.

[1] Because that’s what I do for a living.  In fact, what I should be doing right now is reviewing a manuscript for Electoral Studies, which is two weeks late and slipping later.  Hopefully the editor of ES doesn’t read LGM, and thankfully, for a change, this is the only manuscript I have ‘on my desk’.  (Which means of course once I publish this post and check my work email . . . )

[2] When I review a manuscript, I like to read it, set it aside, and read it again before submitting my review.  The above represents my take on a first read only, and are thus limited in nature, possibly erroneous in places, so consume and dispense with appropriately.