Monday, 17 November 2014

Reading the entrails: a few polling observations

So far, I have made only limited attempts to try to work out how the polls might move in the run-up to the general election.  That is deliberate.  I am highly sceptical that anyone can foretell the future with any great confidence.  Things happen.

But we do have a lot of polling data, and while I absolutely agree with the pollsters like Lord Ashcroft who repeatedly advise us that an opinion poll is a snapshot and not a prediction, buried not all that deep in the data we have information that may help us predict at least some of the movements in opinion polls.

Understanding the present

Before we can work out what might happen in the future, we need to understand what is happening right now.  And unfortunately that's far from clear.  We have eight pollsters who regularly take national opinion polls: ICM, Ipsos-MORI, YouGov, Populus, Opinium, Survation, ComRes and Lord Ashcroft.  They differ quite markedly.  Are Labour on 36%, as in the most recent poll from Populus, or 29%, as in Ipsos-MORI's most recent poll?  Are the Conservatives on 35%, as Populus found today, or 29% as Opinium found in their most recent poll?  The Lib Dems tallied 6% with Survation and 11% with ICM in their most recent polls.  And UKIP recorded 11% with Populus this morning and 23% in the last Survation poll.  These are not minor differences.

How can you tell which poll is most accurate?  You can't definitively, unfortunately.  All we have to go on is past results.  Past performance is not necessarily a guide to the future and the value of pollsters can go down as well as up.

Pondering the future

But perhaps we can understand a little better the differences between the various pollsters and identify as a result how polls might move in the future.  Others are far more knowledgeable than me about the intricacies of past vote weighting, the use of certainty to vote, the differences between telephone polls and online polls, and so on.  Different pollsters use different methods to get a balanced sample.  These explain the differences in present results.

I want to look at two aspects that may affect how polls move in the future: how parties are prompted for in opinion polls and how voters adjust their thinking in their specific constituency.  We have illuminating information about both of these.

Party prompting

It would be natural to think that all pollsters would ask the same question when trying to get their headline polling figures.  But they don't.  Six of the pollsters follow essentially the same format.  Survation and ICM each have their own house style.

UK Polling Report have a handy explanation of the standard approach here:

The bit I want to look at now concerns "the issue of whether or not the question prompts by party name. While in the questions above only three of the pollsters include party names, in fact they all prompt by party name. Since Yougov survey’s are mostly selecting onscreen options, obviously the party names are there for people to tick. MORI would have done the same when they used to do their surveys face-to-face (interviewers use laptops to display the questions and record answers), I can only assume they now read out the party names afterwards.

Historically prompting by party name was a very significant difference between pollsters, as it gives a significant boost to the Liberal Democrats. A poll that includes the Lib Dems in the prompt tends to show their support about 2 points higher than one which doesn’t. These days that is no longer an issue as everyone does it, but it is still a consideration for minor parties like the Greens and UKIP. These are not normally prompted for, and prompting for them gives them significantly more support. The problem is, when they are compared to election results, polls that prompt for minor parties invariably overestimate their support. So while it seems unfair, the most accurate way of polling seems to be to prompt using the names of the three main parties (plus the SNP and Plaid Cymru in Scotland and Wales), but ignoring minor parties."

You will note that the standard approach excludes UKIP at the first stage.  This is a point not lost on ardent kippers.  Does it remain the case that the most accurate way of polling is to prompt without mentioning UKIP (or indeed the Greens)?

Uniquely at a national level, Survation's prompts include UKIP.  Its methodology is explained here:

Survation also record the highest level of support for UKIP.  Is this cause and effect?

It is important to appreciate that the other aspects of the composition of opinion polls also affect the party shares.  Opinium do not prompt for UKIP but consistently record high ratings for UKIP.  There are other things in the mix.

But we have recently had an experiment which shows the potential impact of prompting for UKIP.  ComRes conducted an opinion poll on its standard basis recording UKIP at 19%.  It simultaneously conducted another where UKIP was prompted for, which increased UKIP's rating to 24%, a jump of 5% (almost entirely among men, interestingly):

Now it is important to appreciate that you can't just add 5% onto the polling of every other pollster to get a "with prompting" figure for UKIP.  Polling methodologies are fully interlocking and no doubt those pollsters that do not prompt for UKIP would claim to account for UKIP through different means.  But it is an illustration of how prompting for UKIP could make a substantial difference.

We must envisage the possibility of a point being reached where UKIP are sufficiently established that to fail to prompt for them would risk under-recording their ultimate vote share.  We may already have reached that point.  If so, we should expect UKIP to do better than they otherwise might be expected to do from the results of most pollsters. 

One observation at this point: UKIP draw a disproportionate amount of their support from 2010 non-voters.  I suggest that if such supporters need prompting before remembering to support UKIP, their support needs to be downrated to a considerable degree.  The likelihood of them getting to a polling booth to have the opportunity to be pleasantly surprised by a purple option is substantially diminished.

Local constituency dynamics

Seven of the eight pollsters who regularly conduct national polls do not direct the respondent to consider their own constituency.  The eighth is ICM, which asks the following question:

"The Conservatives, Labour, the Liberal Democrats and other parties would fight a new election in your area. If there were a general election tomorrow which party do you think you would vote for?"

So as you can see, ICM do not prompt for UKIP, but uniquely prompt respondents to think about their constituency.

What difference might that make?

Lord Ashcroft's constituency polling

We now have quite a large body of constituency opinion poll evidence, thanks primarily to Lord Ashcroft. 
He has conducted 102 constituency polls, with more no doubt to come.  One aspect of that polling is potentially very helpful in understanding how voters will react when they focus on the question of constituency polling.
When he conducts national opinion polls, Lord Ashcroft does not prompt for UKIP.  But when he conducts constituency polls in England, he does so asking two questions in succession, as follows:

"1. If there were a general election tomorrow, which party would you vote for? Would it be Conservative, Labour, Liberal Democrat, UKIP, or another party? If 'another party’: Would that be, the Green Party, the British National Party (BNP), or some other party – or do you not know how you would vote?

2. And thinking specifically about your own constituency and the candidates who are likely to stand there, which party's candidate do you think you will vote for in your own constituency at the next general election? [Prompts as at Q1]"

As you can see, he actively prompts for UKIP, and then gets the respondent to think about the specifics of the constituency.  We can therefore see, through a mechanism that kippers will be happy with, how respondents factor in the dynamics of their own constituency.

I have therefore looked at these polls and analysed how each party performs in each constituency, tallying up the difference between their poll ratings between the two questions:

Note, I am not taking a view at present on the absolute accuracy of Lord Ashcroft's constituency polls (I am wary of constituency polls), but rather I am relying on their ability to measure movements between generic questions about voting and considering constituency-specific considerations.  I see no reason to doubt the aggregate picture that these constituency polls demonstrate.

Labour and the Conservatives

In the Conservatives and Labour battlegrounds, the effect is relatively minor.  While there is a lot of variation by constituency, overall the Conservatives put on an average of just over 1/2 % in constituencies that they hold when voters are asked to think about the local considerations.  However, they are outgunned by Labour, who put on over 1% on average in such constituencies.  (I'm aware, incidentally, that some constituencies have been polled twice and that there is the theoretical risk that they would unduly skew the sample, but in practice they don't particularly).  Labour look as though they will get a small lift in most of the national polls when voters in Conservative/Labour marginals start thinking about their own constituencies.

The effect is more marked and still more in Labour's favour in constituencies that Labour hold, with Labour on average advancing by 1% and the Conservatives falling back by closer to 2% on average.  This will make the chances of many Conservative gains from Labour still harder than the polls currently suggest.

The Lib Dems

These movements pale into insignificance when one looks at the movements in constituencies where the Lib Dems are in contention.  In Lib Dem seats with a Conservative challenger, the Conservatives drop by over 4% on average when respondents are asked to think about their own constituency and the Lib Dems rise by nearly 12% on average.  In some seats, the movements are much greater: in Eastbourne, the Lib Dems rose 22% on the second question and in Sutton & Cheam the rise was 19%. 

Lib Dem seats with a Labour challenger show nearly as large movements: Labour drop by 3% on average while the Lib Dems rise by 8% on average. 

It should be noted that even such large movements in the Lib Dems' favour will not by themselves prevent the Lib Dems from having a torrid time.  Their poll ratings are so low that if Lord Ashcroft is in the right ballpark even such large movements in their favour on the constituency question would leave them with substantial projected losses.

As I have already noted, seven out of eight national pollsters do not direct respondents to think about their constituency when getting their view on how they would vote.  If those polling movements are replicated across just those seats that the Lib Dems hold, their poll ratings would rise by 1% nationally without any other movements in public opinion when voters come to consider their local options.  

But in fact, the Lib Dems do better not just in constituencies where they are in contention but also in constituencies where they are making up the numbers.  They put 1/2% on average in Conservative held constituencies with Labour challengers and over 1% on average in Labour held constituencies with Conservative challengers.  It seems that when voters are asked to think about their own constituencies, the Lib Dems benefit.  They appear to be the local party for local people.

All of this goes a long way towards explaining why the Lib Dems poll particularly well with ICM in the periods between general elections.  It also suggests that ICM's methodology on this point at least is superior to their rivals, given that all voters will be thinking about their local constituency at some point.

What is unclear is whether the local effect is yet fully factored in.  Lord Ashcroft's polls do not name candidates (in some cases they are not yet known), so we don't know what impact name recognition might have.  It seems reasonable to assume that may help the Lib Dems further (particularly where they are the incumbents), given that focussing on local factors already gives them a boost.


UKIP on the other hand suffer when things get local.  When polling respondents are asked to focus on their constituencies, UKIP lose votes in every category of marginal constituency.  Overall, they average losses of 2% per constituency pretty much evenly across the board.  UKIP gained support in only six out of 102 constituency polls between the first and second question, and in each case by just one percentage point.  This looks most unlikely to be a coincidence.

UKIP presumably benefit under the first question from the prompting for them.  But when respondents consider their own constituency, UKIP's star appreciably dims.  It appears from the data tables to be a general dimming rather than specific to either sex, any age group or any class.

If we put together the ComRes experimental poll with this finding, it appears that prompting for UKIP encourages more men to select UKIP, but prompting respondents to think about their constituencies leads to a drop in UKIP support across all groups.  It isn't as simple as a single group of voters first being encouraged and then being discouraged.  It appears that different groups wash into and out of supporting UKIP by these two mechanisms.

We know that respondents will ultimately start thinking about their constituencies before the election, because they will have no choice.  What we don't yet know is whether the general election process will do the equivalent of prompting voters to think about supporting UKIP.


We shall see further polling movements caused by events, and possibly by voters starting to concentrate more seriously on the coming general election.  But we can infer from the above that we may see some polling movements caused solely by latent biases in the polls.  The Lib Dems are likely to rise with most pollsters at least towards the levels currently recorded by ICM because of their strength at local level.  Labour may also see their vote share bolstered by local strength in Labour/Conservative marginals.

UKIP, as so often, remain the big imponderable, with two competing forces alternately potentially boosting and dampening their chances.  My guess - and it is no more than a guess - is that UKIP will remain in the public eye right the way to polling day, having the effect of prompting voters to think of them.  If so, UKIP may have a small net boost to their national polling figures (independent of other voting movements).  But their weakness at the local level is likely to make their gains harder to make.

No comments: