What Trump needs to win: A polling error much bigger than 2016’s

If the polls are right, Joe Biden could post the most decisive victory in a presidential election in 3 1/2 decades, surpassing Bill Clinton’s win in 1996.

That’s a big “if.”

The indelible memory of 2016’s polling misfire, when Donald Trump trailed in virtually every pre-election poll and yet swept the battleground states and won the Electoral College, has hovered over the 2020 campaign. Biden’s unusually persistent lead has done little to dispel questions about whether the polls could be off again.

But while Trump’s surprising victory has imbued him with an aura of political invincibility, the polls today put him in a far bigger predicament than the one he faced heading into Election Day in 2016. The polls show Biden with a far more significant lead than the one held by Hillary Clinton, and many of the likeliest explanations for the polling misfire do not appear to be in play today.

Of course, it’s possible the polls could be off by even more than they were four years ago. But to win, that’s exactly what Trump needs. He would need polls to be even worse than they were in the Northern battleground states four years ago. Crucially, he would also need polls to be off to a far greater extent at the national level as well as in the Sun Belt — and those polls have been relatively accurate in recent contests.

Another way to think of it: Pollsters would have far fewer excuses than they did for missing the mark four years ago. Trump’s upset victory was undoubtedly a surprise, but pollsters argued, with credibility, that the polling wasn’t quite as bad as it seemed. Clinton did win the national vote, as polls suggested she would, and even the state polls weren’t so bad outside of a handful of mostly white working-class states where there were relatively few high-quality polls late in the election.

In postelection post-mortems, pollsters arrived at a series of valid explanations for what went wrong. None of those would hold up if Trump won this time.

Here are the many ways the polls are different today than they were in 2016.

The national polls show a decisive Biden win. Four years ago, the national polls showed Clinton with a lead of around 4 percentage points, quite close to her eventual 2.1-point margin in the national vote. This year, the national polls show Biden up by 8.5 percentage points, according to our average. The higher-quality national surveys generally show him ahead by even more.

Unlike in 2016, the national polls do not foreshadow the gains that Trump made in the Northern battleground states.

Four years ago, national polls showed Trump making huge gains among white voters without a college degree. It hinted that he was within striking distance of winning in the Electoral College, with possible victories in relatively white working-class states like Wisconsin, even though the state polls still showed Clinton ahead.

This year, the national polls have consistently shown Biden making big gains among white voters and particularly among white voters without a degree. In this respect, the national polls are quite similar to state polls showing Biden running well in relatively white Northern battleground states like Wisconsin and Michigan. The national pollsters won’t be able to sidestep blame while pointing fingers at the state pollsters.

There are far fewer undecided or minor-party voters. Four years ago, polls showed a large number of voters who were either undecided or backing a minor-party candidate, and it was always an open question how these voters would break at the end.

Overall, Clinton led Trump, 45.7% to 41.8%, in the FiveThirtyEight average, and 12.5% of voters were either undecided or supporting a minor-party candidate like Gary Johnson or Jill Stein.

There’s significant evidence that undecided and minor-party voters shifted to Trump in 2016. The exit polls found that late deciders broke toward him, 45% to 42% — but by even higher margins in the states where the polling error was worst, like Wisconsin, where late deciders broke toward him, 59% to 30%, in the last week. Post-election surveys, which sought to re-contact voters reached in pre-election polls, found voters drifting to Trump. And all of this was foreshadowed by pre-election polls, which showed the race tightening after the third debate and the James Comey letter. It doesn’t explain the whole polling error four years ago, but it probably does explain part of it.

This year, just 4.6% are undecided or backing a minor-party candidate, according to the FiveThirtyEight average. Even if these voters broke unanimously to Trump, he would be far short of victory across the battleground states and nationwide.

Some pollsters — including The New York Times/Siena poll — do show more undecided voters, voters backing a minor-party candidate or voters who simply refuse to state whom they’ll back for president. Yet there’s little evidence that they’re poised to break unanimously for the president.

In the final Times/Siena polls of the six battleground states likeliest to decide the election, the 8% of likely voters who didn’t back either Trump or Biden were slightly likelier than average to be young, nonwhite, less educated and male. They were slightly likelier than average to be registered Democrats. They disapproved of the president’s performance by the same modest margin as voters overall and didn’t have a favorable view of either Biden or Trump. They were far less likely to have voted in a recent election. One wonders whether many of these voters will ultimately turn out at all, even though they say they will.

— <strong>Many more state pollsters now properly represent voters without a college degree.</strong> The failure of many state pollsters to do so four years ago is probably one of the biggest reasons the polls underestimated Trump. It’s not 100% solved in 2020, but it’s a lot better.

The issue is simple: Voters without a college degree are less likely to respond to telephone surveys. To compensate, pollsters need to weight by education, which means giving more weight to certain respondents to ensure that less-educated voters represent the appropriate share of a survey.

This has been true for decades, but Democrats and Republicans used to fare about the same among white voters in both groups, so many political pollsters glossed over whether their samples had too many college graduates. That changed in 2016: Trump fared far better among white voters without a degree, and suddenly polls that had been accurate for years were woefully biased against Trump.

By Upshot estimates, failing to weight by education would have biased a national survey by 4 points against Trump in 2016. It would have had no effect at all in 2012.

Importantly, most national surveys in recent cycles weighted by education. There’s an arcane reason: They mainly sample all adults, and adjust their samples to match census demographic variables — like educational attainment. Many state polls, in contrast, called voters from lists of registered voters and adjusted their samples to match variables that voters provided when they registered to vote, like their party registration or age — but not their educational attainment.

Fortunately, most state pollsters now weight by education. There are a couple of exceptions, but they’re generally not polls that get talked about too much anyway. Virtually all of the polling you’re looking at shows white voters without a degree as a very large share of the electorate. They’re just supporting Biden in far greater numbers than four years ago.

No guaranteed improvement. There’s no reason to assume the polls will be very accurate this year. There’s not even reason to be sure that the polls will be better than they were in 2016, which wasn’t exactly the worst polling error of all time. In fact, the polls were even worse in 2014 and quite bad in 2012 — although few cared since they erred in understating the winner’s eventual margin of victory. The polls could easily be worse than last time.

Even if the polls do fare better than they did in 2016, they might still be off in ways that matter. In the 2018 midterms, the polls were far more accurate than they were in 2016, but the geographic distribution of the polling error was still highly reminiscent of the error in the presidential election.

Today, polls show Biden faring best in many of the same states where the polls were off by the most four years ago. Take Wisconsin. It was the highest-profile miss of 2016; now, it’s a battleground state that Biden seems to have put away.

We won’t know until Election Day whether that simply reflects real strength among white voters, as shown repeatedly in national polls, or whether it’s an artifact of an underlying bias in polls of states. Four years ago, undecided voters broke to Trump at the end, leading to an error in his direction; today, perhaps they’ve swung back to Biden.

The survey research industry faces real challenges. Response rates to telephone polls are in decline. More and more polls are conducted online, and it’s still hard to collect a representative sample from the internet. Polling has always depended on whether a pollster can design a survey that yields an unbiased sample, but now it increasingly depends on whether a pollster can identify and control for a source of bias.

Nonetheless, pollsters emerged from the 2016 election mostly if not completely convinced that the underestimation of Trump was either circumstantial — like the late movement among a large number of undecided voters — or could be fixed if pollsters adhered to traditional survey research standards like weighting by education. If Trump wins this time, they will be in for a whole new round of self-examination. This time, they might not find a satisfactory answer.