Oz Politics

What are opinion polls

Opinion polls use statistical techniques to interview a small number of people in order to predict what the nation is thinking on a particular issue.

When it comes to predicting a federal election outcome, the most important opinion poll is the preference voting intention between the two major parties. These polls seek to predict the party grouping that will get the majority of the two-party preferred vote nationally under Australia’s preferential voting system for the House of Representatives. While it is not always the case, the party group that wins the majority of the two-party preferred vote usually wins sufficient seats to form a government.

Morgan and Newspoll

Roy Morgan Research and Newspoll typically publish two-party preferred opinion polls every fortnight. These polling organisations use different methodologies. Morgan uses face-to-face interviews; Newspoll uses telephone interviews. Morgan asks a direct question on the person’s two-party preference; with the exception of the election campaign period, Newspoll derives a two-party result based on preference flows at the previous election. Morgan typically has a sample size double that of Newspoll.

Prior to the November 2001 election, Morgan had the edge in reputation, with Morgan polls historically outperforming Newspoll. At the 2001 election Morgan picked the wrong result and Newspoll over-estimated the size of the Government’s victory. The 2001 two-party preferred outcome was 51.0% to 49.0% for the Coalition over Labor. Immediately prior to the election, Morgan predicted 45.5% to 54.5% (out by 5.5 percentage points). Newspoll predicted 53% to 47% (out by two percentage points). ACNielsen predicted 52% to 48% (out by one percentage point).

Since January 2002, more often than not Morgan has found higher support for Labor compared with Newspoll. Nonetheless, Morgan and Newspoll typically identify the same trends in broad support for the two major political parties at the federal level. Furthermore, all opinion polls show greater variability in voting intention than actually occurs at federal polls: two-party preferred election results are usually much closer to 50/50 than the range evidenced in opinion polls.

Note: the graphs to include the monthly ACNielsen results since December 2003.

2001 and 2004 compared

2004 started out much as 2001 for the Government: well behind in the opinion polls. The question this year is whether it end the same way for the Government: ahead in the poll that matters. So far the trends look similar. The graph below takes the two party preferred voting intentions from Newspoll for 2001 and 2004.

Sampling error

The media and polling organisations often present polling results as an absolute fact. Headlines that support for a party has leapt from 52 per cent to 53 per cent are not uncommon. However, for a statistician, this sort of reporting of a two-party-preferred poll result is largely meaningless.

The statistical reality is that polling data is only a population estimate. The actual population parameter will vary from the estimate. One of the critical factors that affects the accuracy of the population estimate is the number of people polled. Statisticians call this the sample size. The larger the sample size, the more accurate the estimate will be. The following table sets out the confidence intervals for typical polling sample sizes.

For example, using the above table, we can see that a random sample of around 400 respondents will produce a population estimate with a confidence interval of some plus or minus 5 percentage points. That is to say there is a 95% probability that the actual population parameter lies within plus or minus 5 percentage points of the poll estimate.

Thus with a random sample of 400, if the population estimate suggests that 51 per cent of the population have a two-party preference for Labor, there are 19 chances in 20 that actual population parameter falls somewhere between 46 per cent and 56 per cent. With this result and this sample size, it is difficult to translate the population estimate to a confident election result prediction.

Typically, Roy Morgan uses a sample of around 2000 people. This yields a confidence interval of around plus or minus 2 and one quarter per cent. Typically Newspoll uses a sample of between 1000 and 1250 people. This yields a confidence interval of around plus or minus 3 per cent. The actual sample sizes vary from poll to poll, and sometimes the polling organisations will publish results based on a sample as small as 500 people after a critical event such as the Federal Budget.

Sampling error is not the only factor that can affect the accuracy and reliability of polling data. This is an enormous literature on whether people are more likely to “lie” in face-to-face interviews or telephone interviews. With telephone interviews, there is an inherent bias against those people who do not have a land-line (an increasing factor as more young people use mobile phones exclusively). There are also issues with question design, and so on.

Want to find out more?

• Opinion polls: Roy Morgan Research, Newspoll, A C Nielsen
• The Parliamentary Library’s Research Paper no.13 2003-04: Less tax or more social spending: twenty years of opinion polling
• The Parliamentary Library’s Research Note no. 25 2003-04: Opinion Poll Report: Comparison of Voting Intentions as at 1 December 2003
• The Parliamentary Library’s Research Note no. 24 2003-04: Opinion Polls: Important Issues and Preferred Party
• Parliamentary Library Research Note 43 1995-96 The Political Significance of Opinion Polls
• Antony Green’s ABC election site
• Peter Brent’s Mumble election site
• Adam Carr’s Psephos election site
• William Bowe’s The Poll Bludger