By Jeremy Blackwell
You’re lucky this time you won’t have to go through the ‘A few words about me’ and the ‘Hits and misses’ parts. You already got them. So let’s go directly to the heart of the matter.
How does my model work?
Like all election projection models, my model works mostly on ‘Uniform National Swing’ (UNS). With UNS you have all seats moving in exactly the same direction as the national polling average—which never happens in real life.
So I have a second algorithm on top of the first one which I call ‘relative swing’. It factors in not just the usual UNS, but the way parties’ vote shares change relative to each other.
For example, you might start with the SNP on 50%, Labour on 24%, the Conservatives on 15%, and the Liberal Democrats on 8%; then later have the SNP on 45%, the Conservatives on 30%, Labour on 15%, and the Liberal Democrats on 5%. With UNS you simply subtract 5% from the SNP vote, 9% from the Labour vote, 3% from the Liberal Democrat vote, and add 15% to the Conservative vote, in every seat. (And get negative votes for some parties sometimes.)
But the truth is that it’s more complex than this. That’s what ‘relative swing’ takes care of, by taking into account the ‘multiplier effect’. With the above numbers you multiply the SNP vote by 0.9, the Labour and Liberal Democrat votes by 0.625, and the Conservative vote by 2.0 in every seat. (And sometimes get a vote total over 100%.)
So my model is a mix of both, and currently tuned on 80% UNS and 20% ‘relative swing’.
The first step as described above will sometimes deliver negative votes in a few cases. And in some seats the vote total doesn’t add up to 100%. The second step is to eliminate all the negative votes. This is done simply by automatically switching from the model parameters to ‘100% relative swing’. This being a multiplier, it guarantees there can’t be any negative results. Third and last step is to deal with seats where the vote total is not 100%. This is also done automatically by recalculating votes proportionally from what the second step delivers.
Is the model foolproof? Of course not. No statistical model is, especially when you consider tactical voting. That’s why we can get wrong projections even with accurate polls (see ‘Hits and misses’ in my previous article). But I’ve compared my results with those you can get using Electoral Calculus or ScotlandVotes and they’re pretty similar. Not identical, but pretty close (more on that later). So I guess their underlying algorithms are pretty similar to mine.
How do I feed the model?
The source data are all the polls that I can find. Ideally I would rely on full Scottish polls (that is polls fielded in Scotland only and with a sample size of over 1,000). There were plenty such polls before the 2015 General Election and they proved accurate, unlike UK-wide polling. In the graph below the small dots are individual polls. The large dots at both ends are the actual 2010 and 2015 results. The trendlines show the evolution of vote shares and how polls correctly predicted the SNP landslide.
We are not that lucky this year. The snap election took everyone, including the pollsters, by surprise and there are many fewer full Scottish polls: nine so far and only five in 2017.
I therefore also have to rely on Scottish subsamples of UK-wide polls. A subsample size is typically 100 to 150, so they have a much larger margin of error (MOE) than full polls. This is the reason why many people think they should be discarded from any analysis. In principle I agree with this point of view, but this year subsamples are the bulk of the data we have. And there is a simple way to deal with the larger MOE and still get valid projections when interpreting the data. Which I will explain in the next section.
Continue reading General Election 2017: which way will Scotland go?