"the blogger whose youthful effusions have won him bookmarks all over Whitehall ... horribly compelling" - The Guardian
Wednesday, October 13, 2004
Diary: Forecasting the 2005 Election - David Sanders
Yesterday Professor David Sanders gave Essex University's Government Department's weekly lecture, entitled "Forecasting the 2005 Election".
Sanders has a substantial reputation in this field. Most famously, eighteen months in advance of the 1992 election - at a time when John Major was Chancellor of the Exchequer - he predicted the result, forecasting the Tory majority correctly to within one seat. In 1997 and 2001 he was able to produce similarly reliable forecasts. His presentation yesterday looked back to the models he had used then and the models he was using now to predict the result in 2005. It really was interesting, and the room was crowded with attendees interested to hear what was still quite a technical and very much an academic speech. I thought I'd blog it for the sake of my own memory, and because I suspect plenty of readers of this site will be interested in what he had to say.
Sanders began by explaining the basis for his forecasting - what he called, simply enough, "empirically-informed common sense". Look at opinion poll data and match it to what ordinary expectations would be. He also worked on the principle that once one identifies the variables that cause something to happen, then when those variables change, one ought to be able to explain how the effect changes.
In a foretaste of the models he was about to introduce, he explained that the economic perceptions model (the "Essex model") had served him well in 1992, 1997 and 2001. But that now it seemed a reasonable view that political factors were becoming more important and economic factors less so - perhaps because the economy had been in a relatively good state for so long and so is declining as a priority for voters.
On most models for 2005, Labour does best. Worryingly for Conservatives, he couldn't find any model that gave a Tory victory, no matter how the assumptions were altered. (And Prof. Sanders made clear enough later that he did not himself lean to the Left.) The Lib Dems and "fourth parties" would have a good election, with the result that the share of the vote taken by the two main parties would be low by historical standards.
As a very rough guide to his findings, he said the figures for the three main parties' shares of the vote would correspond to about 40%/30%/20%. He presented the poll of polls from 1997 to 2004, which showed a downward trend for Labour mirrored by a modest upward trend for the Tories and, more recently for the Liberal Democrats, over the period.
Then the models were explained, with the basis for them and the empirical evidence he drew in support of them telling one a great deal about the determining factors of how people have voted in recent elections.
According to the old model, economic perception is what matters. People confident of their economic position would be ill-disposed towards changing the government. Significantly, economic reality matters far less. By this model it is not objective economic facts which determine voting intention but subjective economic perceptions. When questioned on this later, he explained that whenever he attempted to map objective economic factors such as inflation and unemployment against voting intention it was very difficult to find a link. Of course, he explained, the economic reality may well still influence voting intention by changing subjective economic perceptions.
So Sanders had predicted the last three elections with the results of the Gallop economic competence question, for the insertion of which he thanked his colleague Anthony King.
The new model takes account much more of political factors. He referred to the Danish academic Martin Paldam's "costs of ruling" thesis that merely by being in power a government loses popularity as it is blamed more and the previous government is blamed less for what goes wrong. Paldam's own estimate is that this cost of ruling reduces the governing party's poll position by about 3% each parliament. Sanders later explained his more modest assumptions of the cost of ruling for New Labour. The new model also took account of the Iraq War, whose significance was also later explained in more detail.
In one sentence (which I may not have copied exactly), Prof. Sanders explained the voting assumptions of the first 'old' model:
If I am optimistic (pessimistic) about my economic future I will not (will) want a change of government.
He then showed opinion poll support for the Conservatives and economic confidence mapped on the same graph. The correlation was obvious throughout the period 1979 to 1997. However, there was a significant difference from about 1994. While economic confidence rose and rose from then until 1997, Tory support was very distant. Although in those three years Tory support rose also, it was far behind economic competence, rather than in step with it, as it had been for the previous decade and a half. But the jump itself was significant - from about 25% to 35% - leading him to ask rhetorically how much worse the Conservatives would have done in 1997 without the recovery of the previous three years.
At this point I was reminded of my own interview at Essex in November 2002, which Prof. Sanders had conducted. When I asserted the importance of economic success to electoral attitudes, he had challenged me to explain why the British economy had done so well in the five years after leaving the ERM even as the government had remained in the doldrums. I had answered that when people were asked to whom they attributed this success, they resolutely refused to credit the government with the recovery. I felt quite pleased after last night to have made the all important distinction between objective economic reality and subjective economic perceptions.
The most striking observation Sanders made regarding the first of the old models, which he had used to predict the result in 1992 and 1997, was the pertinence of interest rates on the overall result. And this was the problem with the model post-1997: it assumed government control of interest rates. Specifically, governments would cut interest rates in advance of an election, ensure far more money in the pockets of everyone with a mortgage, and increase their economic confidence. The electoral effects would follow. I began to understand better why it was that Margaret Thatcher and John Major had both been so reluctant to support Bank of England independence even as Chancellors Geoffrey Howe, Nigel Lawson, Norman Lamont and Kenneth Clarke in succession expressed their support for this move - whatever the economics of setting interest rates centrally, the politics was obvious.
So this model had been a good predictor for two of the the last three elections, but with control of interest rates passed to the Bank of England in May 1997, something new was needed. For 2001, Sanders used the tax index instead of the interest rate. The tax index was simply the government's Price and Tax Index (PTI) with the price figure taken out.
If anyone had felt inclined to give great credit to Gordon Brown for giving away what Sanders showed to be a hugely beneficial electoral tool in surrendering control of interest rates, the tax index revealed the reality. The chart showed that the considerable upward trend in taxes from 1997 suddenly took an enormous dive in advance of the 2001 election. "Now, what's that?!", Sanders asked, pointing to what looked like a steep cliff as the negative effect of tax changes dropped so far, so quickly. He predicted that we would see the same drop in the effect of taxes in time for the next election, even though the policies that would have this effect will already have been implemented by now. He thought the fall would be from +0.6 to -0.7.
The second 'old' model was the Economic Valence model. Again placing great value on economic management perceptions, it assumed (i) that economic performance matters most to voters whatever they may tell opinion pollsters about how they prioritise issues like the NHS and (ii) that the best economic figure is not perceptions of government economic competence but the net value of that figure minus the equivalent figure for the main opposition party (eg. If Labour is on 48% and the Conservatives 43% the net figure would be 5%).
These figures had been collected since 1991 and from then until 1997 they showed a huge correlation between the overall support for the Conservatives and the lead the Conservatives had (or didn't have) over Labour in perceptions of economic competence.
The graph of 1997 to 2004 told the same story about as clearly for Labour's lead.
This second 'old' model had two variations: two-stage and three-stage.
Two stage is:
Tax -> Expectations -> Labour support model
While three stage is:
Tax -> Expectations -> Economic Management Expectations -> Labour support model
Both of these variations took special account of the events of September 2000, where fuel protests caused a sudden plummet in support for Labour and an equal rise in Tory support, at one point having the effect of an eight-point lead for the Conservatives. This is taken into account by noting the effect it has, and the declining importance of that effect for each month that passes. What was so interesting about Iraq, he went on, is that so far it seems that the polls do not show the passing of time to be ensuring the same decline in the negative effect on the polls. While Labour recovered from the fuel protests because the protests' relevance declined every month, that is not what is happening over Iraq.
The key problem with all these models, and the reason they have been abandoned for 2005, is that there is no test for exogeneity. Events that do not exist within the model may have effects, but the old model does not account for them.
The new models, then, differ considerably. Politics matters more, economics less. The new models incorporate the 'cost of ruling', the honeymoon Labour had after coming to power, the fuel crisis, and the 9/11 'bounce'. It takes account of the effects of the Iraq War on Labour support and of other perceptions - Blair's popularity, the quality of public services and views on government honesty.
Examining the issue of Tony Blair's popularity revealed a fairly strong correlation between the share of voters who think Blair is the best PM (as opposed to the Tory or Lib Dem leader) and those who favour the Labour Party. The significance of September 2000 in this regard was that prior to then, Blair was ahead of Labour. Since then he has been behind his party. But the correlation remains.
Graphs of government dishonesty and trustworthiness ratings also showed a very close correlation. The correlation with public service delivery was fairly close.
Prof. Sanders said that this new core model works very well for all three parties. Labour and the Conservatives have always been mirrors of one another support for one tracks inversely support for the other. That remains the case. Though it works for the Lib Dems, it does so less well than for the other parties. For assessments of public services, the model failed to achieve statistical significance. When questioned about this later, Sanders suggested that the relative novelty of this sort of polling may explain it. All other measures go back to 1997. This measure was only introduced in July 2000.
But the core model plus Blair's popularity plus Iraq etc. all do work.
So the results for 2005? Under the old model, which he used in 1992, 1997 and 2001, the shares of the vote will be as follows:
Almost exactly what the parties won back in 1997.
But Professor Sanders does not believe this model any more.
The best model in his opinion - which corresponds very well to the 40/30/20 summary - gives the following:
Or, with the same model but assuming that the effect of the Iraq War on party support fades:
Finally, one can take the new model elucidated above with all the assumptions taken together. It is more optimistic for the Tories:
This would leave Labour as the largest party, but with no overall majority.
As above, but assuming again that the effect of Iraq fades:
In examining the full tables, one finds an average effect of the Iraq War of -3% on Labour support and +1% on Conservative support. Some show -4% for Labour, others -2%, but that three point drop is the mean. So it would appear that if the effect of Iraq continues up to polling day there will be a swing of an extra 2% from Labour to the Conservatives. Interestingly, the Liberal Democrats do not appear to benefit (or lose out) from Iraq's continuing relevance. If Sanders' models are correct, either Iraq will not have an effect on the next General Election, or the Conservatives will be the chief political beneficiaries of the unpopularity of the Iraq War, which they supported.
Prof. Sanders answered lots of questions afterwards. One lecturer challenged him that the new model, with all its assumptions and 'dummy values' such as for Iraq and September 2000, involved the essential abandonment of the traditional Essex model. Sanders accepted this, noting that without the dummy values the real trends were missed, and said that where he differed with his colleague was that he saw this as a good thing, that elections could no longer be predicted without taking account of exogeneity and he was therefore happy to take this into account.
Another colleague asked about how he had dealt with turnout, specifically noting that while Labour may get the shares of the vote mentioned, if middle class floating voters in marginal constituencies switch to the Lib Dems in anger over Iraq, then Labour will still not have a majority. How could he be sure this wouldn't happen? Prof. Sanders replied that as yet he knew no way to incorporate this sort of turnout into the model, except in terms of the way some polling firms do, by weighting less heavily (or not at all) those who say they are less likely to vote.
When asked about the claim made at the Conservative Party Conference last week that by the Tories' private polling the Conservatives were six points ahead of Labour in 130 marginal constituencies, Sanders said he was very surprised when he heard of this. On Thursday, dining with two senior Tories, he had keenly asked them what they knew about it. One claimed ignorance, the other indicated that the polling was actually done by Conservative activists. Sanders had then asked if it was being done professionally, with these activists disguising their own party allegiances, with no reassuring answer forthcoming. So he is deeply sceptical of these claims. [UPDATE: Apparently (see comments) the work was done by polling firm Opinion Research Business, and as Anthony Wells notes "no pollster would stand for made up figures being published in their name, so the actual figures are probably true".]
The most fundamental question was whether or not the values could truly be judged to be independent of one another. When a person says they think Blair is the best PM and that they will be voting Labour, is the latter caused by the former, or do people say both to appear consistent, or because the two go together? How does one detect causation?
This is a point that was made well by Christopher Montgomery in a sceptical account of a C-Change lecture:
Statistics drawn from a generally low level of Conservative support are reverse-engineered to extrapolate a teleological explanation of that low support. When, in fact, the 'statistics' are always and in every instance the consequence of that low support rather than the cause of it. So it's no more than chicken-and-eggery to exclaim, 'look! the Tories are at 30% in the polls, and only 12% of first time Prospect readers/users of W2 drug-rehabilitation clinics/Surrey supporters support them'. It's precisely because the Tories are only at 30% in the polls that their scores are so relatively low in any of the selected sub-groups. One does not, in isolation, cause the other: both sets of statistical observations, the party's overall poll standing, and its ghettoised break-down of support, are part of the same problematic whole.
Prof. Sanders said only that it was very hard to answer the question of whether these values are given independently of each other.
Dr. John Bartle, the academic hosting the lecture, noted the significance of Sanders making his prediction in advance of the release of the publication of party manifestos and in advance of the election campaign. Did these factors not also determine how people vote? Sanders said that he believed voters decided long before an election how they would vote, based on the images they had of the parties contesting it. It was indeed reasonable on that basis to make such predictions long in advance of polling day.