It looks like we are seeing some poll-herding in the lead-up to 2025 Election. Since yesterday's blog post, two more polls were released in the 52 to 53 per cent range for Labor. It is clear that most polls since the beginning of April are narrowly dispersed in the 52 to 53 per cent range on the two-party preferred metric for Labor.
One outlier from the narrow 52/53 trend was a Freshwater poll on 15 April which had the race at 50/50. Freshwater Strategy polls typically have the Labor vote lower than the other polls. Two outliers come from Roy Morgan: 54.5 on 10 April and 55.5 on 17 April. In recent months, Roy Morgan has typically had Labor above the trend of the other polls. A further outlier comes from Resolve Strategic with Labor on 53.5 on 11 April. And a final outlier comes from YouGov with 53.5 on 19 April.
This narrow distribution is not what you would expect mathematically. Let's imagine a world where every poll is based on a random sample of (say) 1200 people, and the underlying voting intention was (for the sake of the argument) 52.8 per cent (the mean poll result since 1 April). We would expect to see a distribution of poll results that was normally distributed around that mean.
We can use a kernel density estimate to see what we actually have, and that is a distribution that is not normal, We actually have a leptokurtic distribution. It is too peaked around the mean, with thin tails. We can compare both curves, as both of them have an area of 1 square unit under the curve.
We can calculate the Chi-squared statistic for the sample means from each pollster, and test how likely the complete set of polling data is. For the polls since 1 April, they had about 1 in 206 probability of occurring randomly. That is to say, there is a 205 in 206 likelihood that these polls are not the result of independent random statistical processes. This suggests there us at least some poll-herding occurring.
Now to be clear. This is no where near as bad as the polls from the 2019 election. Those polls were very tightly correlated. The 2019 polls had something like a 1 in 100 million probability of being the result of independent random statistical processes. In the end, those 2019 polls were shockingly wrong.
Because the herding is not as pronounced, I am somewhat less concerned about the 2025 polls than I was before the 2019 election. Nonetheless, there is an increased possibility of a polling error this election compared with the 2022 election. Also, I would caution: Just because the polls have herded in 2025, it does not automatically mean they are wrong. They can herd to the right result for all the wrong reasons.
There are a couple of other things worth noting. Although the two-party preferred metric is herded, the major party vote-shares across all pollsters are not herded. [For some reason, these are not the main game on which a pollster's reputation hangs].
A little more troubling, particularly as the importance of minor parties to the two-party preferred results grows, is the massive over-dispersion of poll estimates for One Nation, the Trumpets of Patriots, and their fellow travellers. The distribution of these pollster results are platykurtic (flat in the middle). For One Nation since 1 April we have a minimum estimate of 6 and a maximum estimate of 10.5 per cent. This is not herding. But it suggests that pollsters have significant difficulty in getting reliable responses on voting intentions for minor parties. For the most part, the differences relate to pollster house effects.
With that out of the way, let's look at the latest aggregation (little changed from yesterday):
No comments:
Post a Comment