Donald Trump has gained ground over Hillary Clinton during the campaign, but the combined effect of events leaves her with a predicted five-point advantage on Election Day.
Back in 2012 I modelled the dynamics of national Presidential polling using a combination of time trends, survey methodologies, and campaign events. In this posting I will present a similar model for the 2016 campaign using the 190 polls archived at Huffington Post Pollster covering the period from June 1st through October 25th. All these polls include both minor party candidates, Gary Johnson and Jill Stein, in the list of alternatives.
As before I am using three types of explanatory factors to model polling dynamics:
- a simple linear time trend that measures the number of days remaining in the campaign until Election Day; using higher-order polynomials like quadratics or cubics does not improve explained variance;
- “dummy” variables that correspond to various features of each survey like the sample drawn (registered versus “likely” voters), the method of polling (live interviewers, automated interviewing, or via the Internet), and the identity of the polling organization;
- dummy variables to represent various events during the course of the campaign.
For the polling organizations, I included dummies only for those who had contributed at least nine polls, or five percent of the sample. Only six organizations met this criterion. For the events, I included both parties’ national conventions and the first Presidential debate on September 26th. I also included a term for the release of the “Access Hollywood” tape where Donald Trump was recorded as claiming to have engaged in sexual assault. Because the second debate followed only two days after the release of the tape on October 9th I have combined those events together into a single dummy variable. I have included a third variable which represents the period since the third debate on October 19th. All dates are measured from the midpoint of each poll’s fieldwork.
Measuring the effects of the conventions was especially difficult this year since the DNC took place in the week following the RNC. The RNC dummy is coded one starting on the close of the convention, July 21st, and extends through the following Sunday. Eight polls were conducted during this period. Rather than measure a separate effect for the Democratic convention, I have instead used a “post-convention” variable that is coded as one from the close of the DNC until the first debate. All models are estimated using “weighted least squares” with the weights proportional to the square root of each poll’s sample size.
I present three different specifications of the model. The first uses only the trend, method, and event variables. The second version includes effects for the six pollsters who met the criterion of nine or more polls. The last specification removes terms that were not statistically significant in prior specifications. (The marginally significant effect for Ipsos/Reuters disappears in a more restricted specification.)
Starting first with the time trend, the positive value indicates that Clinton held a larger lead early in the campaign season. A value of 0.07 means that Trump picks up about one percentage point on his opponent every thirteen days (=1/0.07). This is a much faster pace than in 2012. Four years ago, it took President Obama about forty-seven days to gain a single percentage point over Mitt Romney. The constant indicates the predicted margin between the candidates on Election Day when the “Days Before” variable is zero. Without any intervening events the model predicts a Trump victory by five to six percent.
Rather surprisingly none of the methodological variables have any effect in 2016. Poll watchers generally expect to see a one- to two-point tilt in the Republican direction when samples are constrained to “likely” voters. That difference reflects the generally higher propensity of Republicans to turn out since their age and social characteristics correlate with voting. This year we see no such effect. Nor is this likely to be a statistical artifact; polls of likely voters represent only 58 percent of the sample so there are sufficient numbers of each type of poll to generate reliable results.
In 2012, polls conducted on the Internet were about one percentage point more favorable to Obama than polls conducted by other means. This year we see no differences between Internet polls and those conducted by live interviewers. Two organizations, the Republican-leaning Rasmussen Reports and the Democratic-leaning Public Policy Polling, use automated calling systems where respondents are asked to enter their answers by pressing the phone’s dialpad or speaking directly to the calling robot. Because there are only two such agencies, I included dummy variables for each of them rather than a single variable denoting the method they use. The results for the two organizations are quite different. Rasmussen continues to show a significant bias in favor of the Republican candidate, while PPP shows no such bias. This difference parallels that found for 2012, where Rasmussen’s results showed a pro-Romney bias. Rasmussen’s polling in 2016 has an even greater Republican tilt of over four points, compared to two to three points in 2012.
What the model shows most clearly, though, is the powerful effect of campaign events on the margin between the candidates. Clinton’s lead fell after the Republican National Convention then rebounded after the Democrats convened in Philadelphia. The debates and the release of the Access Hollywood tape further boosted Clinton’s margin. Since the effects of these events must be measured against the overall pro-Trump trend in the polls, I have incorporated these data into a chart.
The aftermath of the conventions brought the race back to more or less the same place it was on June 1st with Clinton holding about a seven-point lead. Her advantage decayed over the weeks that followed until the combined effects of the first debate and the release of the Access Hollywood tape again brought her lead up to nearly eight points. The model predicts that her advantage will have fallen back to about five points on Election Day itself. Since the model has a standard error of about 0.5 percent, the confidence interval on the Election Day prediction is roughly four to six percent.
A few other observations from these results. First, the notion that there is a hidden vote for Donald Trump that does not appear in public polling is contradicted by the lack of any effects by polling method. Back in January I found that Trump did over four points better in polls of Republican primary voters when they were interviewed by automated methods. I attributed that result to the so-called “social desirability” effect; Trump supporters might have felt more shy about admitting their preference to a human interviewer. I see no such effect in the general election polls now that Trump has been legitimated by being the Republican nominee.
Second, though I do not show the results here, including the size of the vote for the two minor-party candidates, or the proportion of undecideds, has no systematic effect on the margin between the major-party candidates. If prospective supporters for one major candidate were disproportionately likely to defect to one of the minor candidates, or to remain undecided, we would expect to see fluctuations in the size of those groups influence the size of Clinton’s lead over Trump. Instead it appears that potential supporters of both those candidates have moved in and out of the minor-party columns or remained undecided at roughly equal rates. If so, as the minor candidates get squeezed as Election Day draws near, and the number of undecided voters dwindles, we should not expect to see those changes affect the competitive positions of Clinton or Trump.