Inside Story

Polls and the pendulum

It’s wise to take care in interpreting the two-party-preferred poll figures and the 2016 electoral pendulum, writes Murray Goot

Murray Goot 17 June 2016 3388 words

Near miss: Kim Beazley, shown here with opposition leader Bill Shorten in May this year, is one of six postwar party leaders to have won a majority of the national two-party-preferred vote but not a majority of seats. Mick Tsikas/AAP Image

Much analysis of the national opinion polls, which have Labor and the Coalition running “neck and neck,” is misconceived. It takes the figures at face value and then assumes they map on to a corresponding proportion of seats. If the polls move from a two-party-preferred 50–50 to 49–51 or 51–49, the conversation about what they mean will change dramatically, despite the fact that no statistically significant shift in support has occurred, and despite the fact that the pollsters are reporting not just something they have measured (respondents’ first preferences) but something they have had to guesstimate: the distribution of the “don’t knows” and the direction of the minor parties’ and independents’ second preferences.

Elections are not determined by the parties’ vote shares; they are determined by the number of seats they win. A party can attract more than half the two-party-preferred vote but not win the majority of seats. This is what happened to Labor under Bert Evatt in 1954, Arthur Calwell in 1961, and Gough Whitlam in 1969; it happened to the Liberal’s Andrew Peacock in 1990; and it happened to Labor under Kim Beazley in 1998 and Julia Gillard in 2010. In each case, they won a two-party-preferred majority of votes but failed to gain the largest number of seats.

It’s wrong to assume that if the two sides are level-pegging in their shares of the two-party vote then they must be level-pegging in their shares of the seats. Differences in the size of enrolments across the seats and in the shape of electoral boundaries, and the continuing phenomenon of parties piling up votes in their own safe seats, ensure that vote shares don’t necessarily translate to seat shares.

A number of other assumptions are equally mistaken or misleading. One has to do with the number of seats that will determine the outcome, with many commentators imagining the figure to be as low as the nineteen that Labor needs to add to its present tally if it is to win in its own right. Another is the nature of the seats Labor needs to focus on, with some believing that if the overall swing is towards Labor then the party will win seats while not being at risk of losing any of its own. A third assumption is about the characteristics of the voters in the seats that the parties believe to be in play, with almost everyone thinking that what will be decisive in these seats are forces that affect swinging voters in these seats rather than swinging voters overall.

Some of these misunderstandings reflect a simplistic reading of the electoral pendulum, which was devised by Malcolm Mackerras for the 1972 election and has featured in campaigns ever since. The pendulum ranks the seats held by the Coalition and by Labor in order of their two-party-preferred margin, from those requiring the lowest swings to those requiring the greatest (see, for example, Antony Green’s version). Note that the pendulum doesn’t include the five seats held by other parties or by independents – Melbourne (Victoria), held by the Greens; Fairfax (Queensland), held by the Palmer United Party; Kennedy (Queensland), held by Katter’s Australian Party; and Indi (Victoria) plus Denison (Tasmania), held by the two independents, Cathy McGowan and Andrew Wilkie.

The Australian Electoral Commission, or AEC, defines marginal seats, somewhat arbitrarily, as those requiring a swing of less than six percentage points to change hands; others, equally arbitrarily, fix on a bigger figure or a smaller figure – typically five percentage points or less. Importantly, the fact that one seat requires a smaller two-party-preferred swing than another to change hands – whether it is a “marginal” or not − is no guarantee that the seat requiring the bigger swing won’t change sides if the seat requiring a smaller swing doesn’t.

If the two-party-preferred vote on 2 July is 50–50, the pendulum would lead us to expect not a line-ball result but a Coalition win. This is because it shows that Labor needs to win 50.5 per cent of the two-party-preferred vote to gain the required number of seats (leaving aside, as any pendulum should, the seats not held by either side). At the last election, the Coalition won 53.5 per cent of the two-party-preferred vote and Labor 46.5 per cent. On a swing of 3.5 percentage points to Labor, which would leave the two sides level-pegging in two-party preferred terms, the pendulum predicts that Labor would fall short by five seats, assuming the seats that notionally switched to Labor after the last redistribution start in the Labor column; or by six seats, assuming the Coalition picks up Fairfax from the Palmer United Party, as everyone expects.

If the government and opposition are tied 50–50 in the polls but the government (according to both parties) is ahead in the seat count, there is no need to conclude, as much of the analysis has, that Labor must be experiencing a bigger swing in its own safe seats and/or the safe seats held by the Coalition, while suffering a smaller swing in the seats it doesn’t hold and needs to win. On the contrary, a 50–50 split with Labor falling short by five or six seats is exactly what would happen under a uniform swing. Of course, Labor might be piling up “wasted” votes in safe seats while still benefiting from swings where it matters most, in the additional seats it’s after; in fact, it would be truly remarkable were that not the case.

What matters most is the total number of seats each side needs to win to form government, whether on its own or with the support of others. If the seats Labor picks up and those it doesn’t pick up turn out to be exactly the seats you would expect it to have won or not won, assuming there’s a swing to Labor, then the validity of the pendulum won’t be affected; statements, endlessly repeated, about the pendulum depending on a “perfectly uniform swing,” as Antony Green puts it in his blog, are simply untrue. What the pendulum requires is that the number of seats that don’t swing, even though their margin is less than the overall swing, and the number of seats that do swing, even though their margin is greater than the overall swing, are equal so that the anomalies on each side cancel out.

If we want to use the pendulum to work out which side ought to form government, what matters is the net number of seats that is required. Labor needs seventy-six seats to govern in its own right – the fifty-five it currently holds (or the fifty-seven that are notionally Labor after the redistribution, which the published pendulums acknowledge) plus twenty-one (or nineteen). If Fairfax falls to the Coalition, Labor would require an overall swing of 4.3 percentage points on the pendulum − or a two-party-preferred result of 50.8–49.2 – a substantially bigger challenge than 50–50. If the Coalition wins Fairfax but loses a seat to Nick Xenophon in South Australia, Labor would need 50.5 per cent of the two-party-preferred vote. It would also need 50.5 per cent if the Coalition wins Fairfax, doesn’t lose a seat to Xenophon, but loses New England to Tony Windsor – or loses a seat to Xenophon but doesn’t lose New England. Of course, if the Liberals lose two seats to Xenophon, Labor’s target on the pendulum might drop a touch to 50.4 per cent. And so on. As the number of such possibilities increases – possibilities that involve the minor parties or independents − so the utility of the pendulum declines.

If we look at what the pendulum said would happen in past elections, how useful has it proved to be? Over the last eighteen federal elections, it has missed the mark by no more than two seats on six occasions and by three or four seats on another ten – including the 1969 election, against which the British psephologist David Butler tested the pendulum. Only twice has the performance of the pendulum been less reliable than this in predicting the net shift in seats between Labor and the Coalition: in 1987, when Labor won six more seats than predicted, but would have won office anyway; and in 1998, when the Coalition won twelve more seats than the pendulum would have led one to expect, and the country elected a Howard government with a majority of twelve instead of a Beazley government with a majority of ten.

The closeness of an election is no guide to how good a bet the pendulum is likely to prove. Since 1969 the net difference between the number of seats the pendulum said would change sides given the size of the two-party-preferred swing and the number that did change sides has been not much greater in the eight elections where the swings were relatively large (from 3.6 to 7.4 percentage points) than in the ten elections where the swings were relatively small (from 0.9 to 2.6 percentage points).

Can the pendulum’s various errors – ranging from the 1998 disaster to those elections when it erred by just a seat or two − be explained by the superior campaigning skills of either Labor or the Coalition? It appears not. If Labor’s skill in targeting vulnerable seats were responsible for the party’s winning more seats than it should have in 1987, according to the pendulum, why was that skill not evident in 1984 when the Coalition beat the pendulum, and why was it not sustained in 1990 when the Coalition, again, did better than the pendulum suggested it would, given the size of its two-party-preferred vote? If the Coalition’s campaign superiority were responsible for the party’s winning so many more seats than it should have in 1998, why did it not win more than it should have in 1996, when it underperformed the pendulum by three seats, or in 2001, when it again underperformed the pendulum? It is to factors other than the skills of the campaigners that we should look for explanations – the (un)popularity of state governments, the impact of members’ retiring, the advantage enjoyed by candidates who are first-term incumbents, and so on, including luck.

The voters that parties target in the seats they think they can win but might lose – regardless of whether they hold these seats at the time – are sometimes described as “softly committed” voters, and more often as “swingers,” labels that cover voters of diverse kinds. While exact definitions differ – some of the parties’ market researchers, certainly in the past, have refused to disclose their definition on the grounds that it’s their intellectual property – what these voters have in common is their willingness, if not propensity, to change or to consider changing from their current preference or from their past vote.

How different from swingers in every other seat are the swingers in the seats that count? While the efforts of the parties are focused on a relatively small number of seats that appear vulnerable, it doesn’t automatically follow that the swingers in these seats are very different from the swingers in other seats. According to the Australian’s Phillip Hudson, Bill Shorten needs “to convince just 30,000 voters” in “exactly the right twenty-one marginal seats… to switch from the Coalition to Labor.” Statements of this kind, as much a feature of elections now as are pendulums, are quite misleading. Apart from overlooking Labor’s need to defend its own seats and not just win others, it rests on three shaky foundations: the concept of the “right” seats (a sufficient number of any of the seats Labor doesn’t hold will do); the number (twenty-one includes two seats that boundary changes have made notionally Labor); and, above all, the fact that it is virtually inconceivable that 30,000 voters in the “right” seats would shift without several times their number also shifting in seats Labor already holds or doesn’t need to win. If this weren’t the case, Australia would be a very different place. Election campaigns would be quite different as well.

It is true that the election will be won or lost, by and large, in the marginals, however defined; this is a characteristic of the single-member electoral system we have in the lower house. But the outcome will not necessarily be determined by factors that are relevant only in those seats or even by factors that are more relevant in those than in seats of other kinds. If that were not the case, the pendulum, contingent as it is on the national swing, would almost never work. Pork-barrelling, intense campaigning and a range of other factors may be peculiar to marginal seats and other seats that the parties think of as being in play. (Among the latter are a number of Liberal seats in Western Australia, which would normally be considered fairly safe; Mayo and Sturt in South Australia, where Xenophon poses a threat to Liberals in seats that in other circumstances would be safe; and some Labor seats.) But the parties’ images, policy promises, leaders’ appeal, perceived competence and records shift votes in every seat, as do state and local factors. The electoral significance of marginal seat politics, including claims of other voters being “disenfranchised,” is easily exaggerated.

The more accurate the pendulum in predicting the net shift in seats nationally, the less important the special characteristics of the marginal seats to any explanation of the result. Historically, as we have seen, the pendulum has predicted overall outcomes fairly well; one disaster in eighteen isn’t a bad record. Former Labor senator John Black’s claim that pendulums “are useful only as retrospective devices to rank votes” is mistaken. But to maximise its utility the pendulum requires a reliable estimate of the overall swing. And for that, election watchers follow the national polls.

How good are the polls? Not as good as one might imagine – certainly not as good as the close analysis and breathless interpretations of their every report on the state of play might suggest. Over the seven elections held between 1993 and 2010 − when the polls with the longest continuing records were conducted by interviewers using telephones connected to landlines − the mean difference between the pollsters’ final estimates of the two-party-preferred vote and the election-day figure recorded by the AEC was 1.4 percentage points for Newspoll, 1.8 percentage points for Roy Morgan Research and 2.0 percentage points for Nielsen. In other words, if Morgan’s estimate of the two-party-preferred vote had been 50–50 at some election, the true two-party-preferred, on average, would likely have been 49.1–51.9, either in Labor’s favour or in the Coalition’s.

This is not to criticise how particular polls are conducted; rather, it is to alert us to the limits not just of one poll but of all polls. Since the technologies that most pollsters (including Morgan and Newspoll) now use – the internet, mobile phones, automated telephone calls on landlines − have been around for only a short time, it’s impossible to say whether the accuracy of the polls is likely to be any better now than when they used the older technologies. Given the nature of surveys, however, a better record in the making seems unlikely.

Polls that report their findings in terms of a two-party-preferred vote – as all polls do − allocate respondents’ second preferences on the basis of assumptions that may or may not be justified, regardless of whether the second preferences are nominated by respondents or imputed by the pollster based on the preference flow at the last election. Pollster John Utting notes that the method by which individual pollsters allocate preferences can change the two-party-preferred result by as much as 2.5 percentage points. With up to a quarter of respondents saying that they intend voting for minor parties or independents – some that will not be on the ballot paper − there are a lot of second preferences to guesstimate and a lot of scope for non-sampling error.

Preferences apart, the sampling variance associated with a poll means that no one can be confident about the polls’ published figures beyond their “margin of error” − the variance associated with sample surveys. The differences, from 1993 to 2010, between the two-party-preferred vote predicted by the final pre-election polls and the final two-party-preferred figures, differences already noted, speak to this if to nothing else. When pollsters report a 50–50 split on the basis of 2000 interviews – some sample sizes are considerably smaller – what they are saying with great (95 per cent) confidence is no more than that Labor’s two-party vote is somewhere in the range of 48 to 52 per cent; and, correspondingly, that the Coalition’s two-party vote is also somewhere in the range 52 to 48 per cent. Note that if 50–50 is the result of rounding up from as low as 49.6 and rounding down from as high as 50.4, then 48 to 52 is the result you can get from rounding up 47.6 and rounding down from 52.4. While pollsters sometimes note the “error” associated with the size of the sample, it is easy to understand why they – and, more particularly, the newspapers, radio stations and television channels that carry their reports − are reluctant to spell out what this really means. A statement that support for Labor (or the Coalition) could be anywhere between 48 per cent, a clear loss, and 52 per cent, a comfortable win, is not what the media that pay for this intelligence usually want to hear.

And while much is made of the parties’ being “neck and neck,” a conclusion one can reach with greater confidence by pooling the polls, as some analysts have done, the chances that any polling organisation will come up with the same two-party figure – 50–50, for example – on two successive occasions, as a number of pollsters have done, are small. The chances of producing the same 50–50 two-party figure three times in a row, as the Essential poll did a couple of months ago, are even smaller. This is true, it should be stressed, even if the electorate’s support for the parties hasn’t changed. What is widely accepted in the media as evidence of no net shift might be taken, at best, as evidence of the pollsters enjoying extraordinary luck.

Rather than relying exclusively on the national polls, it might be useful to listen out for the whispers from the party bunkers. If the word spreads that Labor’s gains in the marginal seats, where the parties poll relentlessly, are not sufficient for Labor to win, it is unlikely that Labor has lifted to a two-party-preferred figure of 50 per cent nationally; on the pendulum, falling short by ten seats translates to a two-party preferred of 49.6 per cent for Labor and 50.4 for the Coalition. Some of the national polls, keen to boost their claims for accuracy, may be adjusted to reflect this, too. If Labor’s two-party-preferred figure doesn’t exceed 50 per cent but it goes on to win, it will be at least as big a blow to the credibility of the pendulum as the result in 1987. If Labor’s two-party vote gets to 51 per cent but the party falls well short of winning, it will be the biggest setback the pendulum has suffered since 1998.

Neither a Labor two-party result of 51 per cent nor a Labor victory seems likely. The Coalition may well finish up winning more seats than the pendulum predicts. And a number of seats may change hands − to or from the minor parties and independents – for which the pendulum doesn’t allow. Whatever the outcome, the polls need to be read with their sampling and non-sampling errors in mind, the nature of the pendulum properly understood, and the difference between the parties’ share of the vote and the parties’ share of the seats carefully noted. •