Phantom poll changes

From Benjamin Lauderdale and Doug Rivers writing at YouGov:

We believe that most of the bounces seen in surveys this year represent sampling noise that can be reduced or eliminated by adopting by better statistical methodology. We risk a repetition of 2012 where polling swings were largely statistical mirages. The convention and first debate bounces in 2012 were mostly the consequence of transitory variations in response rates. Fewer voters were changing their minds than were changing their inclination to respond to surveys.

Most telephone polls use independent samples, so the respondents in one week’s poll are different from those in another week’s. This makes it impossible to distinguish change in individual vote intentions from changes in sample composition from week to week. It is possible that five percent of the electorate switched from Clinton to Trump over the past week (decreasing Clinton’s lead by 10 points). But it’s also possible that nobody switched and apparent swings are due to differences in sample composition.

YouGov draws its samples from a large panel of respondents. In most of our polls, there is little overlap from one sample to another. However, sometimes the same respondents are recontacted to see whether their opinions have changed. For example, after the first presidential debate in September, we reinterviewed 2,132 people who had told us their vote intentions a month before. 95 percent of the September Clinton supporters said they intended to vote for her. None of them said they intended to vote for Donald Trump, but five percent said they were now undecided, would vote for a third party candidate, or would not vote. Of the Trump supporters, only 91 percent said they were still planning on voting for Trump. Five percent moved to undecided, one percent to Clinton, and the rest to third party candidates or not voting. The net effect was to increase Clinton’s lead by almost four points. That was real change, though significantly less that the ten point change to Clinton’s lead seen in some polls.

Other events, however, have not had any detectable impact on voting intentions. We did not see any shifts after the release of the Access Hollywood video, the second or third presidential debates, or the reopening of the FBI investigation into Clinton’s emails. When the same people were reinterviewed, almost all said they were supporting the same candidate they had told us they were supporting in prior interviews. The small number who did change their voting intentions shifted about evenly toward Clinton and Trump so the net real change was close to zero.

Although we didn’t find much vote switching, we did notice a different type of change: the willingness of Clinton and Trump supporters to participate in our polls varied by a significant amount depending upon what was happening at the time of the poll: when things are going badly for a candidate, their supporters tend to stop participating in polls. For example, after the release of the Access Hollywood video, Trump supporters were four percent less likely than Clinton supporters to participate in our poll. The same phenomenon occurred this weekend for Clinton supporters after the announcement of the FBI investigation: Clinton supporters responded at a three percent lower rate than Trump supporters (who could finally take a survey about a subject they liked).

The bolded italics are theirs.