Tuesday, October 13, 2020

Polling Difference From 2016

The popular notion that the polls were way off in 2016 is wrong. If a poll says that "Smith" is ahead of "Jones" 49% to 47% with a margin of error of 4 points, what that means is that the pollster is predicting that there is a 95% chance that Smith will score in the range 45% to 53% and that Jones will come in between 43% and 51%. Victory by Jones, 50% to 44%, would mean the pollster still got it right. In 2016, the national polls had Hillary Clinton winning by 3%. She won by 2.1%, which is close to perfect. The state polls weren't as good. The worst state was Wisconsin. We had Clinton ahead 46% to 41%. She indeed got 46% but Trump got 47%, so he was slightly outside the predicted range of about 37% to 45%. The final predictions for Michigan and Pennsylvania were correct in terms of the ranges predicted.

That said, pollsters are being much more careful this time. In particular, in 2016 they didn't realize how predictive educational level was of how someone (especially men) would vote. Having learned from 2016, pollsters are doing things differently this year. FiveThirtyEight contacted 21 pollsters, of whom 15 responded (an undreamed of 71% response rate) and learned what has changed this time.

To start with, just about every pollster is now weighting for education in order to have the right number of noncollege whites in the weighted sample. This is a no-brainer given how predictive education is of party preference nowadays. Actually, it always was, only the roles are reversed now. It used to be that college graduates were Republicans and high school graduates were Democrats.

Some pollsters, including Pew and Ipsos, and now weighting education within racial groups, making sure they have the right number of white college and noncollege voters and also the right number of Black college and noncollege voters, since different races have different percentages of college attendance.

Lee Miringoff, head of polling at Marist College, said that he is paying more attention to geography. He wants to be sure he has the right number of urban, suburban, and rural voters in his samples. The higher the population density, the more Democratic a location is, so he is careful not to undersample rural areas.

Recruiting respondents has also changed, with random-digit dialing less popular than it was. Pew Research now first contacts potential respondents by sending them a letter, to make them aware that the call later is legitimate. The hope is to raise the dismal response rates. Some pollsters have completely abolished random-digit dialing and are drawing a random sample from lists of registered voters instead. One advantage of this in states that register voters by party and publish the party registration is that the pollster can more accurately get whatever percentage of Democrats, Republicans, and independents it wants.

Another change is far more calls are being made to cell phones. Some statistics show that 96% of Americans own a cell phone. Suffolk University now makes 88% of its calls to cell phones. The downside of this is approach is that it raises costs. It is illegal to have a computer call a cell phone, so all these calls have to be made manually, which reduces the number of calls per hour an interviewer can make.

Some pollsters, including Cygnal, PPP, Emerson College, and SurveyUSA are conducting polls by text message. Tom Jensen, PPP's director, said that men and people in urban areas prefer answering polls by text rather than by phone, so the response rates go up when using texts. Jay Leve, SurveyUSA's CEO, said that on any given day, they have four different methodologies in use: live phone interviews, robocalls, online surveys with prerecruited respondents, and text polls. Other pollsters are also moving toward online polling with prerecruited panels. On the minus side, making sure the panels are representative is hard. On the plus side, longitudinal studies are much easier. If Jane Smith is polled online month after month, and in June, July, August, and September she was planning to vote for Donald Trump and in October she switched to Biden, the pollster has detected a real change that is not just due to a bad sample.

It seems unlikely that pollsters will make the same mistakes in 2020 that they made in 2016 (especially concerning education), but there are plenty of new mistakes they can make. Miringoff says that the obsession with education may obscure other problems, like the sudden reliance on listed telephone numbers, which misses people with unlisted numbers, something random-digit dialing is immune to.

The biggest worry that all pollsters have is the effect the pandemic will have on turnout. For example, a voter may say: "I am absolutely, 100% going to vote in person, even if have to stand in line for 12 hours." This counts as a likely voter. Then on election eve, the local news says that deaths from COVID-19 in his area have tripled in the past week and the voter stays home. How can a pollster correct for this?

Then there is another worry: The polls are perfect but the vote count is not. What happens if large numbers of absentee ballots don't arrive on time due to the mail being (possibly intentionally) slow or are rejected due to signature or witness errors or stray marks on the envelope? What happens if many voters are turned away at the polls because their student ID card is not an acceptable ID in their state and they don't have a gun permit (which is)? What happens if the lines to vote are so long that some voters give up? In short, the polls may give a perfect reflection of how people wanted to vote but not how the actual electorate turned out. Then the pollsters will get blamed for something they had no control over.

On the plus side, in 2016, most of the undecideds voted for Trump. The polls couldn't and didn't take that into account. This year the number of undecideds is under 5% and may be even smaller by Election Day, so that factor is minimized. Also, there wasn't enough polling in the Upper Midwest last time. That's not going to happen this time. We currently have 60 polls of Michigan and 53 polls of Wisconsin so we have a pretty good idea of what is going on in those states. So, maybe the polls will nail it completely this time.

No comments:

Post a Comment