Thursday, January 5, 2017

From the Washington Post: State pollsters, pummeled by 2016, analyze what went wrong

According to the author, the national polls nailed the popular vote for president. The results were within 1% of the final poll numbers. The error was in state polls, especially the ones that mattered: Wisconsin, Michigan, Ohio, and Pennsylvania.

The article explains what happened.

- Click here for it.

. . . when the national popular vote was certified, the major national pollsters were nearly redeemed. The final Washington Post/ABC News tracking poll put Clinton three points ahead of Trump. She won the popular vote by 2.1 percentage points, about the same as Jimmy Carter's 1976 margin over Gerald Ford. No national pollster was as badly burned by 2016 as by 2012, when those projecting a tie between President Obama and Republican challenger Mitt Romney had to explain a clear Obama majority.
It was a different story in the states, where a half-dozen pollsters, seen as rock-solid for their command of the local numbers, saw a Clinton win that never materialized. In Ohio, which seemed to have slipped away from Clinton in the summer, the Columbus Dispatch's unique mail poll — praised by Silver as the country's most accurate — seemed to find late movement her way. Its final numbers, released on the Sunday before the election, found Clinton and Trump deadlocked. Two days later, Trump triumphed in Ohio by eight points, the biggest Republican victory in the state since 1988.
“I realized that our poll, which showed Clinton leading by one point the weekend before the election, was going to be 'wrong' as I was monitoring results from across Ohio showing her underperforming in urban counties and Trump rolling up extraordinary margins in more rural areas,” said Darrel Rowland, the Dispatch reporter who runs the poll. “Our presidential poll has never been that far off in my memory nor that of my predecessor, who started at the Dispatch in the early 1970s.”
In Michigan, which went red for the first time in 28 years, the final EPIC/MRA poll found Clinton clinging to a four-point lead over Trump. Unlike in Wisconsin, neither Trump's nor Clinton's campaign responded like the state was deadlocked. Nonetheless, that poll — and all but one Michigan poll conducted in 2016 — pointed to a Clinton victory that never happened.
Like Marquette Law's poll, the EPIC/MRA survey had concluded after a few bad Clinton news cycles but before the campaign's final days. “We had an 11-point lead after the debates, it dropped down to seven after WikiLeaks, then after the [FBI Director James B.] Comey letter it dropped down to four points,” said Bernie Porn, the poll's director. “The only thing I wish we'd done is take one more poll the weekend before the election. I assumed, based on the numbers we found, that Clinton would win unless she wasn't able to turn out enough blacks in Detroit and Flint. And that was what happened.”
In Pennsylvania, two reliable state pollsters — Susquehanna and Morning Call/Muhlenberg College — released final numbers that suggested Clinton was going to repeat recent Democratic history, with a narrow win. The latter poll, which suggested Clinton was ahead by six points — outside the margin of error — also surmised a seven-point Democratic registration advantage. On Election Day, said Muhlenberg College pollster Chris Borick, the advantage was only three points.

Here's a summary of the points made:

  • Key error: not picking up Clinton underperformance in urban areas and Trump over performance in rural areas. 
  • Polls did not pick up late breaking voters in favor of Trump.
  • Robocalls were more effective in measuring the silent Trump voter
  • Cuts to local papers minimized the number of polls that could be taken, a late movement towards Trump was not measured.
  • A better job needs to be done explaining what polls are actually saying. Some of what went wrong was less the fault of pollsters, than in the interpretation of the results.

For more: How to recover from the polling disaster of 2016? Look beyond polls.

No comments:

Post a Comment