Tuesday 28 July 2009

Labour polling: in ICM we trust?

There are quite a few political betters who are placing their money on the basis that Labour are going to get hammered. That might very well happen and Robert Smithson's new election predictor certainly adds grist to the mill of those who believe that. But while things look grim for Labour, there is real uncertainty over just how bad things are for them. A slight discrepancy has opened up between the pollsters. An apparent gap has opened up between ICM and most of the others. The difference could be the difference between Labour losing and being slaughtered.

Let us look at the evidence since 1 April - a period of nearly 4 months now, so we have a reasonable amount of data. The ICM polling has been as follows (with the Labour rating following the date):
ICM
11 July 2009: 27
14 June 2009: 27
28 May 2009: 22
17 May 2009: 28
19 April 2009: 30

Marketing Sciences (apparently effectively ICM under a pseudonym)
16 April 2009: 26

One thing immediately stands out - with the exception of the poll of 28 May, Labour's record is very consistent. If we were judging Labour's performance by ICM alone, we would treat the poll of 28 May as a rogue or as reflecting a spasm of anger at the expenses scandal, which was at its height then. With the exception of that poll, ICM do not record Labour as having dropped below 25% and otherwise are consistently recording Labour in a 26-30% band. The average Labour rating in these six polls is just under 27%.

Now, compare that with all the other pollsters. Populus's results are closest to ICM's. It had six polls in the period, with a Labour high of 30%, a low of 21% and an average of just under 26%.

In the same period, YouGov had 13 polls with a Labour high of 34% (the first poll of the period - otherwise the high was 27%), a low of 21% and an average of just under 25% (without the first poll, the average is just over 24%). IPSOS-MORI had five polls in the period a high of 28%, a low of 18% and an average of just over 23%. ComRes had seven polls, with a Labour high of 26%, a low of 21% and an average of 23%. BPIX also polled three times in the period, with a Labour high of 26%, a low of 20% and an average of 23%. Harris's single poll in the period gave Labour a rating of 20%.

Now, all of these pollsters make grim reading for Labour: they have not exceeded 30% in any of the last 40 polls. But there is a world of difference between Labour tallying 20% or 23% at the next election and Labour getting 27%. For example, if the Tories tally 40% and the Lib Dems tally 20%, on a uniform swing Labour would score 184 seats with 23% (according to Baxter) and 216 seats with 27%. If you sold Labour today at 202 on the SPIN spreads, that's the difference between making a packet and eating cornflakes for the rest of the month. So if you are betting on the spreads or the Betfair party line, resolving this polling discrepancy is important.

The first thing to ask is whether this is just a case of being fooled by randomness. Others better qualified than me can comment on how statistically significant this is, but in the trade-off between certainty of statistical reliability and using the limited information we have available, it seems reasonable enough to me to make deductions from the sample that we have.

So with caution, we move onto the next question: who is right? And here we run into two conflicting schools of thought. Both of them have been espoused by our host.

The first is that ICM are the gold standard of polling and that especial weight should be given to their findings. Our host made this point two weeks ago here:

http://www2.politicalbetting.com/index.php/archives/2009/07/14/is-the-polling-form-book-bad-news-for-labour/

It has to be said that this post should give even the most exuberant Labour seller pause for thought. If Smithson pere is right and ICM is this accurate next time, Labour sellers sure as hell need to hope that Smithson fils's seat prediction model is accurate.

But what of Mr Smithson's golden rule? This is explained here:

http://www2.politicalbetting.com/index.php/archives/2008/05/05/introducing-the-golden-rule-seat-predictions/

"based on the results of the last four general elections and all three London Mayoral races the most accurate poll has always been the one showing Labour (Ken in 2000) in the least favourable position in relation to the Tories"

This too is based on hard data. So how are we to make sense of these?

It remains possible that both are right. The post from two weeks ago draws attention to Labour polling well before the general election. The golden rule is only tested at the point of an election. That time lag would allow both theories to remain correct.

That is a neat solution, but a little too convenient. I am deeply suspicious about the idea that future polling movements can be derived in advance from current responses. Movements in public opinion are too dependent on events - while the thought of an impending general election may concentrate minds in a partly predictable manner, that will be just one consideration among many that moves public opinion at any given time.

It comes to the same thing in practice, but for now I prefer to place the greatest faith in the pollster with the best past track record. ICM have a great record in predicting election results and it would be foolish to ignore that when deciding which polls to pay most attention to.

However, if Labour really do poll this badly at the general election, I suspect that the seat predictions implied by applying uniform national swing using the results of other pollsters will probably prove more accurate.


antifrank

15 comments:

Richard Nabavi said...

Excellent research, antifrank, thank you.

One sentence which jumped out of the page for me was "Now, all of these pollsters make grim reading for Labour: they have not exceeded 30% in any of the last 40 polls."

On the pollsters, I think the systematic differences are almost certainly caused by the different handling of 'certainty-to-vote', plus the fact that YouGov uses internet-based surveys. My take on this is as follows:

- I would largely ignore ComRes and the one Harris poll. ComRes seem to make completely arbitrary adjustments to their adjustments, and in any case have shown themselves unprofessional in the wording of questions.

- The YouGov methodology seems to be good for picking up, and perhaps exaggerating, short-term changes in sentiment. That is very useful in one sense, but I tend to add a mental smoothing factor to their results.

- ICM does seem to have the best record in predicting actual votes in real elections.

But in the end, you have to make a judgement; it is necessarily subjective. My best guess at the moment is that Labour will end up with a vote share of around 28% to 30%, and the Conservatives in the 41% to 43% range - basically what ICM are telling us, adjusted for some drift back to the main parties and a focusing of minds as the real choice presents itself.

Of course, major political upsets, or a change of Labour leader, might disrupt things. But what is very striking is the stability of the polls over quite a long period now, even allowing for the expenses scandal.

Based on that, and subject to further consideration of the Smithson fils projections, I don't see value in the spreads at current levels.

Easterross said...

antifrank, a very well researched piece. Like Richard I have faith in ICM and YouGov but worry about ComRes constantly jumping "all over the place".

In the same way that uniform swing calcuations can badly misjudge the eventual outcome due to one party doing particularly well or badly in a region/nation rich with marginal seats, I often think the polls can be skewed depending on the location of those responding to the questions.

My hunch at the General Election is that we will see the Tories at the higher end of the margin of error i.e. in the 42-44% range and Labour at the lower end of the margin of error range i.e. 25-27%. The problem is in trying to guess where the LibDems will fall, possibly in the 18-20% range but how their vote splits nationally could have major implications on the eventual make up of the House of Commons.

Ted said...

The only niggle I have on ICM is also their strength - their weighting using past voting behaviour. It could be smoothing changes in sentiment, a cause of the stability Antifrank (Alistair?) mentions.

Its rare for ICM to show the peaks and troughs that other pollsters report, change is slow over time. That may well be better at revealing the underlying strengths or weaknesses of the parties but the niggle is that it could be keeping Labour artificially high based on past voting behaviour.

ken said...

There is the smoothing of vote shares based on allocation of those who are likely to vote and who dont nominate a party - they seem to be allocated on the basis of their past party affiliation.

These folks in MORI are less likely to be counted - which may well be correct as apathetic former Labour voters desert them in the next election. ICM may turn out to be optimistic.

Of course thanks to the fact that the election isnt held now all of this is castles in the air stuff.

Nick palmer MP said...

Good analysis. If you have five or six samples of anything you will get a range of results, of course, and that applies to polling averages too. But ICM's consistency probably does reflect people switching between "Labour I suppose" and "Dunno, used to be Labour but...". ICM counts the second as equal to half the first. So does Populus, I think. Everyone else counts it as Don't Know.

That category of voters is huge - at least 10% of the electorate, I'd say. Howe they'll actually vote and whether they will is really anyone's guess.

ChristinaD said...

antifrank, great article (she says in a gentle voice).

I think that you have picked up on a serious point, one that those betting hard cash need to balance. Its turnout that will be the key to this conundrum. We are going to have lots of less dyed in the wool voters who previously voted Labour switching to other parties.

But what about Labour's core vote? Is it like that of the Tories, one that would settle for a while into the 30/33% bracket under a Tory government?

But come the next GE, will they turn out and vote in large numbers anywhere at all? And if not, will it not matter because its in their heartland seats? That is the question that will decide whether Labour gets either side of 30% if they maintain their present polling figures.

When we look at the current strategy of Brown and his government, its clear that they are aiming their pitch at the core vote rather than the retaining the floaters. Can Labour do enough to push them into the polling booth on election day?

rogerh said...

ICM tends to have slightly higher Labour figures than other pollsters.How has that delivered on election day?
taking the alst 3 elections and looking at the polls in terms of av for months GE-4 to -1,Av for month -1,and final poll,and then actual gives for labour share
-4to-1 -1 final actual
1997 47 44 43 44
2001 47 46 43 42
2005 38 39 38 36

The conclusion is that although the final poll is reasonably accurate,the poll sduring the preceding months exaggerate Labours position.
This falling off is partly the result of voters becoming aware of who is actually standing and therfore able to vote for.Given more candidates from minor parties this effect is lileyb to be bigger in 2010 with Labour as party of government affected more than the other two main parties.
The other fly in the ointment in terms of constituency share is the fact that local elections are likely to be taking place on the same day as the GE-in London Metropolitan and a selection of Unitary and Districts.Now will this be an advantage or disadvantage?Will the GE votes boost the local vote or vice versa? Labour usually records a local share below its national sahre ,the Tories around the same and the Lib Dems better.One is tempted to conclude that Labour will get no boost from its local voting and therefore may show a further slight reduction in its GE vote.

Alastair said...

Thank you all for your kind comments (this posting lark is a lot more daunting than I'd realised). Nick Palmer's comment has made me think hard - it seems to me to capture an important difference between the pollsters and crystallises a central question to be resolved.

antifrank

Mark Senior said...

A couple of comments on the pollsters :-
Mori Rather than take their headline absoluteky certain to vote " figures it would be better to take their figures for those with 6-10 certainty to vote which are in their detailed data tables .
Comres appear to have now reviewed and stabilised their past vote weighting calculation and their figures should now be more stable .
Populus after a 2 poll experiment with their past vote weighting have returned in their last poll to the same weighting as ICM .

MichaelK said...

Thanks antifrank, interesting stuff.

Regarding the conflict between ICM-as-gold-standard and the Smithson Golden Rule (lot of gold here :)), I believe ICM tends to produce less extreme numbers than anyone else. That being so, while Labour are dominant, the gold-standard and the Golden Rule are not in conflict as ICM will be the pollster giving the lowest Labour share.

So my question is, does the Golden Rule hold now that Labour are in the basement? Well, it did in the Euros, so, maybe.

I agree with Richard about the lack of value in the spreads atm.

Financier said...

I have not studied the results of each pollster, but in view of Norwich North, is the importance of "certainty to vote" an even more vital factor, if it is the portent of a trend that will continue until 2010, even if it is then in a milder form.

It would appear at NN that the Labour vote melted away. Were any polls done just prior to that recent election that showed "certainty to vote" by party loyalty and if so how did that poll correlate with the actual result?

MG said...

I think the final result will be somewhere in the region of:

Cons 41%
Lab 27%
Lib 19%

MG said...

Oh and Tory majority at about 65 to 75 'ish.

Plato said...

Great analysis - have nothing of any wisdom to add.

Anonymous said...

Interesting article, Antifrank, thanks.
AnneJGP