How pollsters missed Bernie Sanders’s surge in Michigan
FOR Hillary Clinton, losing Michigan’s Democratic primary to Bernie Sanders was a nuisance. For pollsters, in contrast, the election was Armageddon. Eight different firms had surveyed the state in the previous month, and every one gave Mrs Clinton a double-digit lead. According FiveThirtyEight, a data-journalism website that had put Mrs Clinton’s chances of victory at greater than 99%, the 23-percentage-point gap between her average lead in the polls and the final result was the biggest error since Gary Hart’s victory in the 1984 New Hampshire Democratic primary.
What went wrong? To produce a misfire of this magnitude, just about everything. Pollsters both underestimated how favourable the electorate would be to Mr Sanders—according to the exit poll, it was younger, more male and included more independents than expected—and how big a share of each group he would capture. He lost black voters by merely a 2.4-to-one margin, half as big as his shortfalls among African-Americans in the South. And he limited his deficit among voters who earn more than $100,000 to three percentage points (it was 36 in Virginia).
Projecting the composition of the Michigan electorate was always going to be tough, since it was the first state in the rustbelt to vote this year, and it has not had a competitive Democratic primary since 1992. No live-interview surveys were conducted after a TV debate, preventing any measurement of its impact. Unseasonably warm weather could have boosted turnout and increased the share of less-reliable, Sanders-loving young voters. And complacent supporters of Mrs Clinton, expecting an easy victory, may have switched sides to meddle with their general-election opponents instead: 7% of voters in the Republican primary were self-identified Democrats. Now that the pollsters—who had previously performed fairly well this year by volatile primary standards (see chart), and did get Michigan’s Republican vote spot-on—have some egg on their faces, they are likely to pay extra attention to these factors before the upcoming votes in nearby Illinois and Ohio.
On the other hand, some red flags could have been spotted in advance. One-third of the polling average published by RealClearPolitics, and one-quarter of FiveThirtyEight’s, came from automated, fixed-line-phone surveys conducted by Mitchell, a Michigan firm. The sample’s demographic sub-groups were re-weighted to match targets provided by Mark Grebner, a consultant whose main business is selling voter lists in local elections. Mr Grebner foresaw an electorate resembling a retirement community, with 88% of voters aged 40 or older and 76% at least 50. Since elderly Democrats adore Mrs Clinton, the two Mitchell polls taken shortly before the vote duly estimated her lead at 27 and 37 percentage points, boosting her average.
However, the assumptions underlying the Mitchell numbers were inconsistent with exit-polls from northern states. Around half of Democratic voters were under 50 in each of Massachusetts, New Hampshire and Vermont. (The final number in the Michigan exit poll was 53%.) Had Mitchell simply copied the New England ratios, its forecast lead for Mrs Clinton’s would have shrunk to the mid-teens, the same range calculated by its competitors like Marist, YouGov and Monmouth. Such figures still would have left Mr Sanders a big underdog, but not a hopeless one.
This article appeared in the United States section of the print edition under the headline "Mich-fire"
