Here’s the baseline:
Obama Romney
EV 332 206
Tell me how HRC out-performs President Obama, and Trump under-performs Romney in every demographic, and it’s anything other than HRC in a walk.
I think there’s a lot of analysts and pundits missing what’s right in front of them— including ignoring the starting point and demographic trends in voting for the last three elections.
I think we do have some data to evaluate, and it helps us gauge which estimates are better than others.
First, we can compare the poll averages against the actual results of previous elections.
This is particularly useful in seeing how polls (which are public opinion surveys based on population samples) vary from the true state of reality. Let’s start with 2012:
Obama Romney
This is a known result. We can compare this to a truly enormous listing of final polling numbers from 2012.
In short, as many ‘final’ polls had either a dead heat, or a one point Romney lead, as showed President Obama ahead. Very few had president Obama ahead by more than 1 point. That is to say— they were off the mark.
We can talk about normal sampling variation all day long, but the effect of these polls on polling aggregators is clear (all listing Obama vs. Romney):
Huff Post., Nov. 6, 2012: 48.2 46.7
RCP, Nov. 6, 2012: 48.1 48.1
538, Nov. 6, 2012: 50.8 48.3
Notice the trend?
All of these poll aggregators underestimated President Obama’s total, and two of three overestimated Romney’s.
Since each of these are relying on hundreds of polls, this suggests a systematic bias across the majority of polls, in the technical sense of the term:
sampling bias is consistent error that arises due to the sample selection. For example, a survey of high school students to measure teenage use of illegal drugs will be a biased sample because it does not include home schooled students or dropouts. A sample is also biased if certain members are underrepresented or overrepresented relative to others in the population.
This systematic bias should not be viewed as malicious, or based on political preferences or ulterior motives. Rather, I think the likely culprit is the stubborn refusal of polling firms and political scientists to look critically at their ‘likely voter’ screens.
For decades, until quite recently, political scientists and pollsters looked like the presumptive ‘likely voter’— older, white, Christian heterosexual male, moderate to conservative. And indeed, for most of our electoral history, this cohort had the easiest access to voting, and voted reliably for men who were pretty much like them.
But this has changed a lot in the past twenty years, and continues to move in a direction— irreversibly-- towards progressive politics, and Democratic candidates from diverse demographic cohorts. But the ‘likely voter’ models have not kept up. That’s my contention for the mismatch of polling and the actual outcomes on election day.
****************
The simplest prediction is to look at the last few elections, and assume voting patterns won’t change much. That is our null hypothesis. Any variation from prior elections would be assumed to be the effects of things like changing demographics (which is significant), the effects of external events (like an economic crash or terrorist attack), and candidate characteristics.
So, in 2004 Bush beat Kerry 50.7 / 48.3.
In 2008, President Obama beat McCain 52.9 / 45.7
In 2012, as noted above President Obama beat Romney 51.1 / 47.2
The GOP candidate has received about 46.5% of the vote in the past two elections, the Democrat 52%.
Do we have reason to think the election will revert to 2004 voting percentages? Based on what polling? What demographics?
Do we have evidence that Trump is under-performing Romney? Yes we do— across every group, including white voters. (See here, here and here.)
And yet we’re going to entertain the hypothesis Secretary Clinton will have a smaller vote margin than President Obama?
That’s a hypothesis that needs more than a little empirical heavy lifting to support.
That’s where I keep bringing PollyVote into the conversation. Pollyvote combines several different predictive models to arrive at their vote percentage prediction.
Today PollyVote stands at:
Clinton Trump
52.6 47.4
For the reasons I cited above, plus differences in GOTV and early vote data, I think the inclusion of polling averages, even in PollyVote, will underestimate Secretary Clinton’s total, and overestimate Trump’s, in a way that makes the final margin more like:
Clinton Trump
53% 45%
Why? Because that difference between polling results and the actual vote in the last two elections was about +2 Dem., -2 GOP.
I think my estimate is conservative.
To steal a line from Professor Sam Wang, if Trump get’s more than 46%, I’ll eat a bug. I’m not planning on eating any bugs.