FOR THE third presidential election in a row Donald Trump has stumped America’s pollsters. As results came in on election night it became clear that polls had again underestimated enthusiasm for Mr Trump in many states. In Iowa, days before the election a well-regarded poll by Ann Selzer had caused a stir by showing Kamala Harris ahead by three percentage points. In the end, Mr Trump won the state by 13 points.

Overall, the polling miss was far smaller. Polls accurately captured a close contest in the national popular vote and correctly forecast tight races in each of the battleground states. National polls erred by less than they did in 2020, and state polls improved on their dismal performances in 2020 and 2016. Yet this will be little comfort to pollsters who have been grappling with Mr Trump’s elusive supporters for years.

The Economist’s nationwide polling average found Kamala Harris leading by 1.5 percentage points, overestimating her advantage by around three points (many votes have yet to be counted), compared with an average error of 2.7 points in past cycles. State polling averages from FiveThirtyEight, a data-journalism outfit, had an average error of 3.0, smaller than the average of 4.2 points since 1976.

But in contrast to 2016, when pollsters’ misses were concentrated in certain states, those in this cycle were nearly uniform across state and national polls. In the seven key states, polling averages underestimated Mr Trump’s margin by between 1.5 and 3.5 points (see chart). Pollsters may claim that their surveys captured the “story” of the election. But the awkward question remains: why did they underestimate Mr Trump for the third cycle in a row?

In past election cycles, pollsters have tweaked survey “weights” to make their samples of voters more representative. Although polls aim to sample the population randomly, in practice they often systematically miss certain groups. Weights are used to increase the influence of under-represented respondents. This has been especially true in recent years as response rates have plummeted.

After the 2016 election, when surveys systematically missed voters without college degrees and therefore underestimated support for Mr Trump, pollsters began accounting for respondents’ education levels. And after 2020, in an effort to ensure that Republican voters were represented, more pollsters began weighting their samples by respondents’ party registration and self-reported voting history. This caused the range of poll outcomes to narrow (weighting reduces the variance of survey results), with many pollsters finding similar results in key states and nationwide.

If there is a lesson from this year’s election, it could be that there is a limit to what weighting can solve. Although pollsters may artificially make a sample “representative” on the surface, if they do not address the root causes of differential response rates, they will not solve the underlying problem. They also introduce many subjective decisions, which can be worth almost eight points of margin in any given poll.

A pollster which gets those decisions right appears to be prophetic. But with limited transparency before the election, it is hard to know which set of assumptions each has made, and whether they are the correct ones. To their credit, the pollsters get together to conduct comprehensive post-election reviews. This year’s may be revealing. Still, without a breakthrough technology that can boost the representativeness of survey samples, weighting alone is unlikely to solve pollsters’ difficulty in getting a reliable read on what Trump voters are thinking.■