Mellman: Poll failure, again

Mellman: Poll failure, again
© Getty Images

The big news in polling this week: the Australian election!

In our Netflix world, I’ve become a fan of Australian TV, but don’t recall anything from down under being big news in the U.S. since 2006, when “Crocodile Hunter” Steve Irwin was allegedly killed by a stingray that stabbed him “hundreds of times.”


Just as improbably, the Australian Labor Party lost last weekend to the Liberal-National (LNP) Coalition, despite the fact that every single one of the 16 polls published since the election was called predicted the LNP Coalition losing with 48 percent or 49 percent of the two-party vote, while putting  Labor ahead with 51 percent or 52 percent.

In fact, the results were reversed with LNP shocking the country by taking 51 percent to Labor’s 49 percent.

Some claimed there’s nothing to see here. After all, the polls were only off by two or three points on the vote for each party.

That perspective seriously understates the severity of the problem.

It’s not just a single poll that was off by 2 or 3 points —16 different polls all had nearly identical results.

All polls have sampling error and that should result in a spread of results.

If you flip a coin 10 times you can expect to get five heads frequently, but you’ll only get that exact result about 25 percent of the time. Professor Brian Schmidt, Australian Nobel laureate in physics, calculated the odds against 16 polls coming in with the same small spread of answers as greater than 100,000 to 1.

In other words, about as unlikely as things get.

This bizarrely narrow dispersion suggests herding. No media pollster wants to be wrong, so everyone makes their poll come out like the others.


In truth, if polls all come out the same, something unnatural is happening. The laws of probability say they should be different.

If the true vote percentage were 48.5 percent fewer, than half the polls should come up with 48 percent or 49 percent. A quarter should show 50 percent or higher and a quarter would be 47 percent or below.

Even if we assume the best about Australian pollsters and believe their violations of the laws of statistics were not willful herding, there are lessons worth taking away.

One is a point I made after a previous British election: report undecideds. In Australia, they don’t.

If the poll results had been 46 percent Labor, 44 percent LNP, with 10 percent undecided, people would have been less surprised by a 51 percent-49 percent LNP victory, and those following the polls would have had a clearer picture of both the state of the race and the uncertainly about the final outcome.

Second, leadership counts. Like many other countries, Australia is a parliamentary system. Voters are not casting ballots for prime minster, but the votes for local representative determine who’ll occupy that position.

In addition to asking for which party respondents would vote, some of these polls asked which party leader they would rather have as prime minister. The LNP candidate led Labor’s leader each time, by 7 points to 11 points. In most (but not all) parliamentary elections in Australia, and elsewhere, the party with the most popular leader (and putative prime minister) wins.

Finally, it appears that Australian pollsters made a huge leap in calculating this two-party vote. (I can’t confirm this, because none of them are transparent about their methodology.)

There are actually several other parties competing in Australian elections and garnering meaningful shares of the vote, even if they win few seats.

In this initial vote choice, the LNP was leading Labor by 2 points on average (recall that was their eventual margin). It was only after “calculating” the two-party vote that Labor pulled ahead.

What happened in the black box of that calculation only the pollsters know for sure, but the key to the error may be locked away inside it, along with the truth about the crocodile hunters’ death.

Mellman is president of The Mellman Group and has helped elect 30 U.S. senators, 12 governors and dozens of House members. Mellman served as pollster to Senate Democratic leaders for over 20 years and as president of the American Association of Political Consultants.