South Australian Election - Fox & Hedgehog Polling Review
As counting for the 2026 South Australian election draws to its conclusion, we have taken the opportunity to review our final polling result, what it got right, and where there is scope for improvement.
Overall, we are pleased with our final election poll, which placed every party’s primary vote within 2% of the final result. A very strong result in what was a highly challenging polling environment.
Fox & Hedgehog SA Poll v. SA Election Result (as at 30 March 2026)
Labor - Poll: 38% - Result: 37.5% (Diff. 0.5%)
One Nation - Poll: 21% - Result: 22.9% (Diff. 1.9%)
Liberal - Poll: 18% - Result: 18.9% (Diff. 0.9%)
Greens - Poll: 11% - Result: 10.4% (Diff. 0.6%)
Other - Poll: 12% - Result: 10.3% (Diff. 0.7%)
To compare accuracy across pollsters, we use what is called ‘root-mean square error’ (RMSE) across the five major groupings (Labor, Liberal, One Nation, Greens, and Other).
RMSE is a single measure of polling accuracy that captures how far, on average, a poll’s results are from the actual outcome across all parties, with larger errors weighted more heavily. In a fragmented, multi-party system, this provides a good summary of overall polling accuracy.
On this measure, we’d like to congratulate YouGov, who appear to have recorded the strongest result with an RMSE of approximately 1.0. DemosAU and Fox & Hedgehog closely follow at around 1.3, with Newspoll also performing well at approximately 1.9.
By comparison, of the ten pollsters who released pre-election polls for the 2025 federal election, none recorded an RMSE below 1.5, based on calculations by Dr Kevin Bonham.
However, complacency can breed over-confidence. We treat these strong results as an opportunity to further refine our methods to ensure our work remains accurate. To that end, alongside the release of our March 2026 South Australian poll, we noted an emerging feature in the data relating to One Nation voters.
This election marked the first time since the 1990s that One Nation re-emerged as a major statewide political force. Rapid changes in party support of this kind introduce additional complexity into polling, particularly where standard weighting assumptions rely on relative stability over time.
Our poll estimated One Nation support at 21%, compared to an eventual result of 22.9%. While this 1.8% difference lies well within the poll’s 3.9% margin of error, we examined whether any systematic issues beyond random error contributed to the shortfall.
As outlined in our methodology, past vote is used as a core weighting variable to align samples with the composition of the electorate.
However, our national and South Australian polls have recently shown unusually high levels of recalled past vote for One Nation, well above its historical vote share. This represents a departure from previous elections, where One Nation support was comparatively stable and recall relatively reliable.
This pattern is consistent with a well-established phenomenon in survey research. Voters sometimes retrospectively align their reported past behaviour with their current political preference. As a party grows, some voters begin to report that they supported it previously, even when they did not.
In the context of One Nation, we refer to this dynamic as a form of “reverse shy-Toryism”. We don’t see a reluctance of One Nation voters to disclose their support, rather a tendency to retrospectively claim it for past elections.
In practice, this creates a specific challenge for standard weighting approaches.
There are two interacting effects.
Selection Effects: Individuals currently intending to vote for One Nation appear marginally more likely to participate in online surveys than the general population.
False Recall: A subset of voters, particularly those who previously supported right-leaning minor parties, now report having voted for One Nation at the last election.
When combined, these effects inflate the apparent size of the past One Nation vote within the sample. Because past vote is used as a weighting anchor, this places downward pressure on estimates of current One Nation support.
We applied a minimal adjustment for this effect prior to the election. With the benefit of the final result, we can now examine it more closely.
By examining the data in detail, we estimate that between 15% and 20% of current One Nation voters incorrectly report having voted for the party in 2025. We observe that the majority of those misreporting their past vote are likely to have been previous supporters of other right-leaning minor parties, such as Trumpet of Patriots or Family First.
The latter parties share supporter bases with similar demographics to current One Nation voters, but are now systemically under-reported in past vote recall. Given voters for all three parties share similar demographics, we assess the primary driver is likely to be false recall.
Adjusting for this likely misclassification has two effects. A modest increase in estimated One Nation support, and a corresponding reduction in the ‘Other’ category.
As a diagnostic exercise, applying this adjustment retrospectively to our final poll reduces the RMSE from approximately 1.3 to around 0.5. This is not a claim to retrospective accuracy, rather context of the scale of the observed underlying effect.
A natural question arises. If past vote introduces these complications, why use it at all?
The answer is that, despite its limitations, past vote remains essential for correcting large structural imbalances in survey responses.
In its absence, our testing indicates that polling would tend to overstate Labor support and understate Liberal support by a materially larger margin than the distortions observed for One Nation.
Removing past vote weighting completely would also produce an estimate for One Nation higher than what the South Australian election observed as the actual level of support, but still within the margin of error. However, it would come at the cost of significantly diminishing accuracy for the two major parties.
In our view, the challenge for pollsters is not whether to use past vote, but how it should be applied in an environment where voter alignment is increasingly fluid.
The South Australian election provides an early sign that traditional assumptions around recall stability may not hold in periods of rapid political change. Adapting to that reality will be central to maintaining our polling accuracy in future contests.