Demographics and the Electorate

[Erc]

So, here's one question I've had for a while.

It's clear Trump is a terrible candidate among huge swathes of the electorate.  Historically bad among Latinos, somehow doing worse than Romney among Black voters, losing (for the first time in recent memory) college-educated whites.  He seems to be doing better among non-college-educated white voters than Romney did, but surely not enough to make up for everything else.

For example, let's look at the ABC-Washington Post poll released yesterday.  Obviously, margins of error among subsamples are larger, but the numbers seem roughly in line with other samples (or subsamples) from other polls.

Clinton has an 87-point lead among Black voters.
Clinton has a 49-point lead among Hispanic voters.
Clinton has a 9-point lead among whites with a college degree.
Trump has a 32-point lead among whites without a college degree.

Plugging these into 538's demographic calculator, and assuming (perhaps poorly) that turnout levels among each demographic (and the vote share among people of other races) are unchanged from 2012, you get Clinton winning by 11.5% with 363 EVs (2012+NC+GA), and with Arizona & South Carolina just out of reach.

Making some (bad) assumptions about third-party votes, here's what the map looks like (ignoring ME-02 and NE-02 for simplicity):

So, why is Clinton only leading in the low single digits and with a narrow EV margin?  Here are some theories that get bandied about:

Geographic Polarization:

Perhaps the gains Clinton is making are in the solid-D states in the Northeast and in California, which doesn't help her EV map (or, alternatively, in solid-R states, though I can't say I buy this too much as an explanation outside of Deseret, as losses in solidly Republican areas of swing states would also hurt him.)  If this is true, it can help explain her narrow EV lead, and this is already evident in the fact that she's only winning 363 EVs in an 11.5-point blowout, above.  If it's true, it also should help us feel better about Pennsylvania, as the Philly suburbs fall squarely into the sorts of places where Clinton is gaining, according to this theory.

But, more importantly, this does nothing to explain why Clinton's national PV margin in polls is so narrow; votes are votes, no matter where they are.

Third Party Effects:

Probably still not large enough to make a huge difference, and I tried to account for it in the map above by using the margin between the candidates when making the map, rather than the % of the two-way vote each got.  Probably, this actually hurts Clinton; as undecided voters make up their mind and some third-party voters come home, we expect the margin between the candidates to widen in each group.  As Clinton's already ahead in the larger groups, this should help her more, if anything.

Weird geographic effects (e.g. Johnson in NM) could have particular effects in certain states, but again this shouldn't have any impact on the overall PV.

Differential Turnout:

This is the big one.  I was assuming that each group would turn out in the same rate as they did in 2012, which is a huge assumption.  What could change?

* Will black turnout go down with Obama off the ballot? Some assume it will, but I don't think there's any real basis for that.  
* Will Hispanic turnout go up with Trump on the ballot?  I'm less convinced of that of late, but if anything that would help Clinton.  
* Are college-educated whites so disillusioned by the choices that they aren't turning out?
* Is Trump really bringing huge swathes of non-college-educated white new voters to the polls?

More generally, are the Likely Voter screens really doing their job this year?  This is not the same sort of race we've had for the last few cycles.

Pollster Herding

All the polls are showing a close race because all the other polls are showing a close race.  Whether consciously or not, perhaps the polls are being massaged to show a closer race when in fact it is a 10+-point blowout.

Stop Looking at Subsamples, You Doofus!

Is this really just an issue of what happens when you look at subsamples for too long?  I don't think so, since the two subsamples that really matter here (the two subsets of whites) are both pretty large.

Did I happen to cherry-pick the ABC/Wapo poll because it happens to fit my narrative?  I don't think I did, but if other polls are telling different stories I'd love to hear them.

Am I just being as dumb as all the "unskew-the-polls" folks from 2012?  I hope not.

[Erc]

Looking at the 538 calculator, even if black turnout dropped to 10% with these numbers, Clinton would still win. So I don't think that this is something which she needs to be worried about, per se. Of course, this isn't to say that she should forget about black voters entirely, or that they should be written off as "safe" as they so often are these days...

A few similar scenarios:

If I swap college-educated and non-college educated white turnout from their 2012 levels (a 20% shift in each), and drop black turnout 9 points for good measure, she still wins with the same margin and map Obama did in 2012.

If I decrease (and increase) (non-)college-educated white turnout by 10%, Trump would still need to win 76% of white non-college-educated votes (men AND women) to win the EC (and he'd still be down a point in the EV), keeping the numbers the same for everyone else.

I don't see that happening.  There doesn't seem to be a plausible way for Trump to win the popular vote, if the race remains like it is.  538 gives him a 36.1% chance of doing so at the moment.  Of course, some of this is reversion to the "mean", or the chance that there's a large shakeup in the race  that swings all white voters towards Trump, but it still seems too high.

[Erc]

I think with the tightening we are seeing that the typically-Republican college-educated white voters are more prone to fall back in line than the typically-Democratic non-college-educated working class white voters.  When Clinton was ahead, she had some real bleed in the WCW vote, but the college-educated white vote that Trump bled has all started to fall back into place.  (See the Georgie polls where this phenomenon occurred.)

What I think the issue here is that WCW are not in as good of a shape as college-educated whites; and thus they are more willing to vote for a change from their normal voting patterns than college-educated whites who have an easier time falling back in line and voting as they normally do.

I think this is the big issue we are seeing and that ultimately it is what could be putting Trump over the finish line.

Except the polls really aren't backing this up.  Clinton is maintaining, if not expanding her lead among white college-educated voters.

Comparing the late September ABC-WaPo poll to the early September poll (insert usual caveats about crosstabs):

You may be right, to some extent, about white college-educated males, who went from +1 Clinton to +11 Trump.

But white college-educated women went from +10 Clinton to +25 Clinton, more than making up for the fickle men.

[Erc]

In terms of the electorate, we can tease out what ABC/WaPo late September thinks via the cross tabs.  It breaks down as:

44% Non-College White
28% College White
27% Non-White

[This breakdown is largely unchanged from the early September poll, apart from a 2% increase in the non-white vote at the expense of the college white vote.]
----
63% Non-College (all races)
37% College (all races)



Note that that's a huge difference from the 538 calculator, which suggests (if turnout remains unchanged from 2012 up to demographic shifts):

33% Non-College White
37% College White
29% Non-White

or from the 2012 Exit Polls:

53% Non-College (all races)
47% College (all races)

28% Non-White (all education levels)

Non-white turnout isn't the real story here (just a 1-2 percentage point difference), it's the college vs. non-college white vote, which has huge differences!

In terms of the 538 calculator, that corresponds to (in addition to a pretty steep drop in black turnout):

58% turnout for white college-educated voters
75% turnout for white non-college-educated voters.

That's very close to one of my "absurd" scenarios from the first page.

Is this really what's going on?  Is Trump really getting 30% more white non-college-educated voters, while 25% of white college educated voters are so disillusioned they aren't turning up?

It's possible, I suppose, but I really can't believe that the share of white college educated registered voters who are voting for a major candidate this year (rather than staying home or voting 3rd party) is larger than the share of white non-college-educated registered voters who are voting for a major candidate this year.

[Erc]

I just came to this thread and didn't go through the whole thing, but my biggest takeaway to the original post is that perhaps some of the assumptions are erroneous. It is not written in stone that each group will vote exactly as the OP has suggested. Perhaps, Trump has gained a bit in each category, especially the Hispanic vote, but possibly also the the highly educated or even the non college-grads, and even the black vote - a few points in each category would of course be a few points overall. I'm not sure this would completely eliminate the analysis and I agree it's a great post and thread, but I do think you have to realize all of the original numbers in this post are fluid as well.

Oh, of course, both the turnout and vote share numbers within each group are fluid.

But the basic story to me seems to be (after a few pages):

1) Clinton is (even after the last few weeks) leading with white college-educated voters, in addition to her still huge leads with non-white voters.

2) Trump obviously has a large lead with white non-college-educated voters (up from Romney 2012, and up from August), but if turnout were to stay at 2012 levels within each group, it wouldn't nearly be enough.

3) In order to get the national poll results we're seeing (narrow, as opposed to huge, Clinton leads), you have to assume that turnout is way up among white non-college-educated voters and way down among white college-educated voters, to the point where the more education you have, the less likely you are to vote, period (and for a major party in particular), among white registered voters.

This last point seems, to me, to go against pretty much everything we know about elections.  Of course, 2016 is a weird year with weird candidates.  It's possible that this is indeed the case, and presumably that is what Likely Voter screens are showing.  But I've got a pretty large Bayesian prior against that, and personally I'm more willing to believe that the Likely Voter screens are wrong than that the bolded point above is true.

If it turns out I'm wrong, we'll find out pretty early on election night.  And I'll be pretty disappointed in my white college-educated brethren for not turning up; we don't have an excuse for not knowing better.

[Erc]

This last point seems, to me, to go against pretty much everything we know about elections.  Of course, 2016 is a weird year with weird candidates.  It's possible that this is indeed the case, and presumably that is what Likely Voter screens are showing.  But I've got a pretty large Bayesian prior against that, and personally I'm more willing to believe that the Likely Voter screens are wrong than that the bolded point above is true.

OK, but that goes back to my earlier question: Why are they wrong?  What are the LV screens selecting on that makes them pick up more uneducateds this time?

And also, going back to my last post, aren't there some poll releases that actually tell you what % of the electorate they have for college-educated whites, non-college-educated whites, etc.?  Can we look at them to see what they're showing, and if the fraction of the electorate for each group is indeed ridiculous?

Here's an example from the latest Bloomberg / Selzer poll.  Likely Voters are those who say they will "definitely vote" or have already voted.

The latest Bloomberg / Selzer poll had 62 non-college - 38 college in their latest poll, very similarly to ABC/WaPo. Again, this is a huge difference from the 2012 exits (53 - 47).  

Though it should be noted that it's less of a departure from previous years (56 - 44 in 2008 and 58-42 in 2004, and I believe 58-42 in 2000).

Of course, census data shows different results (see e.g. this US Census report on the 2000 election), though I'm more inclined to believe exit polling (where you know you got voters) rather than retrospective polling where people may say they voted when they did not.  That isn't to say exit polling doesn't have its own issues, however.

To compare apples-to-apples, it might be better to compare polls now to their equivalents in 2012.

[Erc]

Of course, census data shows different results (see e.g. this US Census report on the 2000 election), though I'm more inclined to believe exit polling (where you know you got voters) rather than retrospective polling where people may say they voted when they did not.  That isn't to say exit polling doesn't have its own issues, however.

With exit polls, you know you've got voters, but do you know that people are telling the truth about their education levels?  Isn't it pretty widely assumed that there's some lying going on when people are asked about their educational attainment in polls, inflating the number of those with college degrees above what we realistically think it must be?  I don't suppose there's any reason why people might be lying about their educational level less now than they were a few years ago?  Maybe Trump's made it cool to be "uneducated"? 

Very good point, though it's still a poll both times.  As you said, there could have been a change in how often people lie (although I would think it would be in the opposite direction, as Trump's made it cool to be a liar ), or people could lie more often in person than over the phone.

[Erc]

In all seriousness, there likely is a gap in self-reported likelihood to vote.  But in the end, I think I'd bet on the historical trend (better educated people are more likely to vote) rather than self-reporting.

(There's a danger for Clinton here if the gap is due to college-educated Republicans not voting who ultimately decide to come home to Trump, but I don't see evidence for this.)

Of course, this just explains self-reported likelihood LV screens; I'm not exactly sure what other pollsters use.  If you just use "did you vote in 2012" (for those 25 and older), you shouldn't have these sorts of issues (though that's not a great LV screen and you may have other sampling problems).

[Erc]

https://morningconsultintelligence.com/public/mc/160915_topline_Topicals_LIKELY_VOTERS_v3_AP.pdf

This is the type of crosstab I'm talking about. Quote: " All statistics are calculated
with demographic post-stratification weights applied."

In this poll, Clinton leads by 3% and if I read it correctly the white vote share is 75%.

Are the figures in the "Respondent Demographics Summary" raw figures, or the results after the weights are applied?

If the latter, the poll has 64% of its (weighted) respondents having less than a bachelor's degree.

EDIT: That seems just way too high for me (as does the white vote share, to a lesser extent); based on past elections somewhere in the 51-58% range seems reasonable.  I could be proven wrong in November, of course.

EDIT 2: A difference between 64% and 56%, say, may not seem to be that much.  But digging into the figures, since they're assuming a modest decrease in non-white turnout, this is entirely being driven by an increase in white non-college educated turnout.  Assuming nothing drastic is happening among the college-educated non-white vote (roughly 10% of the electorate), to replicate a 64% share of the electorate not having bachelors degrees means that turnout among white non-college-educated voters is roughly 20 points higher than turnout among white college-educated voters (and that's even including people voting third party, this time).

Maybe Trump really is having that much of an effect on increasing (or decreasing) turnout, but that just seems preposterous to me.

Let's take two really basic assumptions:
1) Turnout among whites with bachelor's degrees will be at least as high as among whites without bachelor's degrees.
2) The white share of the electorate will be at least 70%.  (Any poll in which this isn't true isn't going to be Trump-friendly, anyway.)

Under those assumptions, the absolute highest you can crank the share of the electorate without bachelor's degrees is 58%.  Any poll that has a figure higher than that is implicitly violating one of these assumptions (most likely the first).

[Erc]

(snip)
I don't suppose anyone with a lot of time on their hands would be willing to go through all the national polls one by one to see if they have crosstab info on the %age of the sample with college degrees, and then we could look at whether there's really a systematic deviation from what we would expect?  Also could see if there's a strong correlation between the %age of the sample that has college degrees and the overall support for Clinton.

Thinking about this a bit harder, the abormally low percentage of people with college degrees in these polls is weird, but it's not going to affect the top line that much.

If I'm dialing the turnout of just white RVs with college degrees, this is a group of people that is voting pretty similar to the rest of the country (in aggregate).  Clinton has consistently had a solid lead (5-12 points or so) among these voters, which is pretty similar, in the grand scheme of things, to the polling as a whole, where Clinton has consistently had a lead of 1-6 points. You can decrease white college turnout by 50%, and not really affect the national numbers by more than about a percentage point.

What really matters, in terms of turnout, are the actually polarized groups: white voters without a bachelor's degree (favoring Trump about 2-to-1) and non-white voters (favoring Clinton by at least 4-to-1).

So, the $64,000 question is: in the final voting electorate, what's the ratio of non-college white voters to all non-white voters?

The latest ABC-WaPo poll has that ratio at 163%.
538's demographic calculator (as an extrapolation from 2012) has that ratio at 113%.

That's a big difference.  Is non-white turnout really going to decrease by more than a few points (if that)?  If not, is non-college white turnout really going up by 44%?  (That would require something like an 81% total turnout among non-college white voters, which seems insanely large.  Turnout hasn't been anywhere near such levels since the 19th century.)

Even that 163% figure, plugging it into the calculator, still gives a 4-point Clinton win even if every single white college-educated voter stays home.

[Erc]

FiveThirtyEight has updated its calculators, which has a pretty significant bearing on this discussion.

A point I alluded to briefly earlier in this thread was the large difference between exit polling and census data when it comes to what share of voters have a college degree.  The previous version of the calculator largely used exit polls (47% in 2012), while census data showed 37%.  This is a big discrepancy, and I was (perhaps unfairly) questioning the reality of polls that more closely matched the latter number.

This is not a question that's mattered that much in previous elections, as (controlling for race) education didn't matter all that much.  In this election, it's a big difference, and this is an important question.

The new version of the calculator now "blends" exit poll and census data, along with other sources of information, to break down college vs. non-college white, and the differences are pretty stark.

The old version of the calculator had (with turnout levels unchanged from 2012):
33% Non-College White
37% College White
29% Non-White

The new version has:
42% Non-College White
31% College White
27% Non-White

These are huge changes; they're definitely more in line with what pollsters are assuming about the electorate.

My guess is the polls are weighting their samples to match the census data.  If the census data is wrong (because a lot of people lie about their past voting behavior), we could be in for a whopping Clinton victory on election day, as I argued earlier in this thread.

If the converse is true, and exit polls are oversampling college educated voters (or people with associate's degrees are responding Yes to "Are you a college graduate?"), we could be really excited early in the evening and then less so as the night goes on.

The changes do certainly make it easier to get within the calculator the sorts of results one would expect from the state polls (Trump picking up ME-2 and IA despite losing NC, for example).

[Erc]

How does the census dataset work?  The census bureau asks people some demographic questions, and also includes the question "Did you vote in the last presidential election?", then releases the demographic breakdown of people who answered "yes" to that question?

It seems like, as a supplement to the November 2012 Current Population Survey, they ask people whether they are registered and whether they voted in the general election.

And yeah, they release all the appropriate crosstabs.

The numbers, of course, don't match up perfectly with exit polls or with the actual turnout results as reported by the state election divisions.  This mismatch with actual turnout is pretty bad in midterm years, but isn't so bad for Presidential years.

In 2012:
Actual Turnout: 129.2 m
CPS Turnout: 132.9 m

In 2008:
Actual Turnout: 131.4 m
CPS Turnout: 131.1 m

This suggests that (barring weird cancelling differences among different demographics) there isn't a huge prevalence of people lying about whether they voted in Presidential years.

So it comes down to whether you expect larger systematic issues with sampling in exit polls or in the CPS.  Honestly, since lying about voting doesn't seem to be a large effect on net, I'd be more inclined to trust the CPS rather than the exit polls.  It's very easy to see how exit polls might oversample college-educated voters (for a variety of reasons); less so with the CPS.

The education question for the exit poll asks whether you have:
1) No high school diploma
2) High school graduate
3) Some college / assoc. degree
4) College graduate
5) Postgraduate study

Since associate degrees are explicitly mentioned, it's unlikely that the reason for the discrepancy is people with associate degrees claiming they are college graduates.

[Erc]

In 2012:
Actual Turnout: 129.2 m
CPS Turnout: 132.9 m

That's only a 3% difference, which is actually much less of a mismatch than I would have guessed.  It seems odd to me that a 3% discrepancy in turnout #s would translate into a 10% discrepancy in the fraction of voters with a college degree.  The best explanation I can think of is that people are more likely to lie to exit pollsters about their educational status than they are to lie to the Census Bureau.

What alternative explanation is there?  Did either the Census Bureau or the exit pollsters do a poor job of selecting a representative sample of voters?

My personal guess is that exit polls are just oversampling college educated voters.  College educated voters are more enthusiastic about voting / civics in the first place (as evidenced by their higher turnout, by any measure), and perhaps are more interested in spending time with an exit pollster.

There are of course also the usual geographic sampling concerns with exit polls, but it's hard to say a priori what sort of effect that would have.

[Erc]

And then there's the question...even if the exit poll #s are biased in a way that the Census #s are not, does that really mean that we should be weighting the current round of telephone polls to match the Census #s?  The telephone polls could have the same bias that the exit polls have.

If it's a response bias that's an issue in the exit polls, then we should certainly be weighting them to match the census; that's why pollsters do demographic weights.

If the bias is due to people lying about their education, then they shouldn't.

I definitely think it's more of the former than the latter, and thus the pollsters may very well be doing it right.  Of course, how this actually interacts with their likely voter screens is another question; these levels are not fixed in stone.

[Erc]

And then there's the question...even if the exit poll #s are biased in a way that the Census #s are not, does that really mean that we should be weighting the current round of telephone polls to match the Census #s?  The telephone polls could have the same bias that the exit polls have.

If it's a response bias that's an issue in the exit polls, then we should certainly be weighting them to match the census; that's why pollsters do demographic weights.

If the bias is due to people lying about their education, then they shouldn't.

Good point.  If you're right that it's a response bias in the exit polls, then I think that also means that the exit polls from past elections have also gotten it wrong wrt what % of the electorate with a college degree was voting Democratic vs. what % were voting Republican....since they are weighting the toplines to match the election outcome.

I guess one thing to do if someone has a lot of free time on their hands is to go back to pre-election polls from 2012, and look at what they were estimating *before* the election both for the fraction of the electorate with a college degree, and the fraction of each educational group supporting Obama or Romney.  Do those #s line up more with the exit polls or the census #s?

Here's an example from the last YouGov poll before the 2012 election (sadly, they don't have education crosstabs this cycle).

Topline: Obama 49 - Romney 47.

College Graduates (34%): Obama 53 - Romney 43
Non-College Graduates (66%): Obama 48 - Romney 47

Compare this to the exit polls:

Topline (adjusted to match the result): Obama 51 - Romney 47

College Graduates (47%): Obama 50 - Romney 48
Non-College Graduates (53%): Obama 51 - Romney 47

Note that the shares are quite different, and that the polarization is essentially nonexistent in the exit polling (and is in fact slightly in the opposite direction).

YouGov is a lot closer to the census data; this is not a surprise, as they heavily, heavily used the 2010 CPS to weight their responses (at least when it comes to registered voters; likely voters is a different question).