I'm going to take a stab at predicting the results of the 2015 General Election in Scotland. There are two aspects which make doing so extremely difficult:
- how many undecideds are there, and how might they vote?
- to what extent will tactical voting take place?
The undecideds are the undead zombies that lurk at the edges of election polling. There's no good way to deal with them so they are usually excluded from headline polls results. If you think about it, this is the same as assuming that either they won't vote or will vote with the same split across parties as the decideds.
The best way to detect undecideds (or don't knows) seems to be through face-to-face polling. The most recent such poll is this one from TNS. It shows that 29% of their sample remain undecided as of mid-April, which is the same as was found in their March poll. The undecideds in the TNS poll say they are as likely to vote as the decideds.
Although we cannot reliably predict how the undecideds will vote, we can assess whether they might have a significant effect. TNS and other polls suggest a 75% turnout from a registered electorate of 4,000,000 people resulting in 3,000,000 votes. This means that 870,000 voters currently remain undecided but are very likely to vote.
Dividing these numbers by 59, the number of seats, tells us that the average number of votes per constituency is 51,000, of which 15,000 are undecideds. Of the 36,000 decided voters, John Curtice's poll of polls suggests that 18,000 will vote SNP and 18,000 will vote for the other parties, of which 9000 will go to Labour.
This tells us that undecideds are important because at 15,000 they significantly exceed the average SNP lead of 9000.
Recent TNS polls contain evidence that undecideds are less likely to be past SNP voters. However, I'm reluctant to rely solely on this because it depends on voter recall of the 2010 and 2011 elections and also there have been big changes in both turnout and voting intentions since then.
A better indication can be found in YouGov's poll conducted between 16-20 April. This poll (page 3) shows that 70% of the undecideds voted No in the referendum, and 30% Yes. A similar split can also be found in a YouGov poll conducted in March.
Let's suppose that undecided Yes voters back the SNP and No voters are spread amongst the non-SNP parties. This assumption could be wrong in either direction and it's important to realise that the prediction hinges on it, but in the absence of better data we have to assume something and this seems most plausible.
If we apply the assumption to the numbers above, we can conclude that the 15,000 undecideds split 5000 for the SNP and 10,000 for non-SNP, which brings the totals to 23,000 and 28,000 respectively. This split is 45%/55% and so tells us that the assumption just made is at least consistent with the referendum result.
Next we need to consider how No voters might distribute their votes amongst parties, and in particular assess tactical voting.
The 16-20 April YouGov poll gives results for questions about tactical voting (from page 7 onwards). For example, if respondents are asked to imagine voting in a constituency where only Labour or the SNP can win, the SNP/Labour vote share changes from the headline 49%/25% to 50%/38%. It's clear from this that tactical voting can have a significant effect on the vote share, but its impact on the number of seats won by each party requires analysis at the constituency level.
The polls conducted by Lord Ashcroft between 10-16 April in Scottish constituencies show that some element of tactical voting is taking place, but the amount of it appears to vary between the eight constituencies polled. The first question (Q1) asks for a general voting intention, but the second question (Q2) asks you to think about your own constituency; differing answers suggest tactical voting is in play. The figures quoted below are for Q2 with undecideds excluded.
Let's now explore what level of tactical voting is needed to overcome the SNP lead in these seats. I'm not making a prediction (yet), just exploring the numbers to see what's possible.
Four of the constituencies show SNP leads of around 11pp (percentage points), ranging from 9pp to 13pp. Let's treat the poll as a snapshot of what decided voters in the constituency are currently thinking. As we saw above, an average constituency has 51,000 voters, of which 36,000 will be decided and so represented in the poll's headline results, whereas the 15,000 undecideds will not be (because they were excluded or because the poll failed to find some of them). Therefore the SNP lead equates to 11pp of 36,000, which is about 4000. If 5000 undecideds split to the SNP by the day of the election, then 9000 of the other 10,000 non-SNP voters would need to vote tactically to overcome the SNP lead. This seems a bit of a stretch, but let's say this happens in just one of the four seats, mostly like the one that Ashcroft shows has a 9pp lead - this is for Jim Murphy's East Renfrewshire seat, which we'll come back to later.
(I've used the constituency average in the last paragraph for clarity's sake, but the argument remains valid if 51,000 is replaced with any constituency size as long as all the numbers are kept in the same relative proportions).
In another two of the eight constituencies, the SNP has little or no lead, so even a small amount of tactical voting will put a non-SNP party in front. So out of eight seats, tactical voting can give three seats to non-SNP parties.
If we scale the results from these three of eight seats up to all 59, i.e. multiply by three-eights, then we find 22 seats could be won by non-SNP parties. However, as these seats are probably not typical, I'll drop the fraction to a quarter, which gives 15 non-SNP seats. Remember, I do not mean this to be a prediction, just an illustration of a possible scenario.
First, vote share. The current poll of polls tells us that 50% of votes will be for the SNP. However, because it (quite rightly) excludes the undecideds, these percentages are likely to shift as they make up their minds. The evidence above suggests that most of them will not vote SNP, and so the SNP vote share will decrease, and the vote for other parties will increase. For this reason, I'd expect the actual election result to be 43-48% for the SNP. Accounting for tactical voting, I'd say 30-35% of votes will go to Labour, with around 20-25% mainly going to the Conservatives and Lib Dems.
The distribution of seats amongst parties is highly sensitive both to the vote share and how tactical voting occurs in each constituency. This makes it very hard to predict, but based on what I outlined above, I expect that the SNP will win between 40 and 55 seats.
If you held a gun to my head and insisted on specific numbers then I'd say the SNP/Labour vote share will be 46%/32% and the SNP will win 45 seats.
There are a few other imponderables that might affect the result.
One is that the turnout in this election is different from previous ones: 63% in Scotland in 2010; 50% in 2011; 85% for the referendum. This might distort the weighing schemes applied in polls, but I don't see an obvious way to assess its impact.
The spiral of silence effect might be operating. If people feel more comfortable saying they're voting SNP, and reluctant to admit they're voting Labour, then it's likely to affect the polls. This is the effect behind the "shy Tories" in 1992 and "silent No voters" in the 2014 referendum. From the referendum experience, I might guess that the polls shortchange Labour by a few percentage points to the SNP's benefit. Also, instead of completely excluding undecideds, some polls account for this by "returning" a small portion of them back to parties they voted for in a past election. Although I made no attempt to correct for this above, Lord Ashcroft's polls do make a spiral of silence correction in which a percentage point or two is transferred from SNP to other competing parties.
The Lib Dems have published private polling (performed by Survation) in East Dunbartonshire and found that their candidate and MP for the last decade, Jo Swinson, has a slender lead. In contrast, Ashcroft's poll found her to be 11% behind the SNP. This can be explained by the so-called incumbent bonus effect that arises because only the Lib Dem's poll mentions candidates by name. However, Lord Ashcroft countered the incumbent bonus effect by arguing that it's unnecessary to name a locally well-known MP because people will know who you mean in answering Q2. I am a resident of this constituency, and having some local knowledge, I'm inclined to believe the incumbent bonus is real. As surprising as it may be to those heavily engaged with politics, there are many people who cannot name their MP, nor the party they represent. Currently, I believe the result in East Dunbartonshire is too close to call and that a late swing in either direction is likely.
Just as I was writing the Comments section, Lord Ashcroft published new polls for two of the eight constituencies I referred to above.
For East Renfrewshire he finds the SNP's lead over Jim Murphy has shrunk from 9pp to 2pp, mainly as a result of Conservatives switching their vote to Labour. I also noticed evidence in the tables that some undecideds have made up their minds, which could also explain the change.
In Dumfriesshire, the SNP have increased their lead from 2pp to 11pp. Ashcroft attributes this to some of the Labour vote going to the SNP. Although it's true that the Labour vote drops by 3pp and the SNP increases by this amount, the "Wouldn't votes" have also decreased by 3pp.
For polls of 1000 people such as these, the sampling error is often quoted as +/-3pp, but this only applies in cases where there's a roughly 50%/50% split of the vote. In these polls most of the vote share is spread unevenly across three parties. It's therefore possible that a sampling error larger than 3pp is distorting the results.
What's remarkable is that both these constituencies have shown quite dramatic changes in the space of two weeks. The undecided zombies are coming and are likely to have a decisive effect in many seats!