Saturday, October 22, 2016

Republicans are (Probably) in Bigger Trouble Than They Think

Election polling is, as ever, a good news / bad news proposition. The good news about following polling data in an election season is that there is always a great mass of data and there are always a great many stone-cold pros crunching that data and writing articles about it, as well as Nate Silver. But this volume of data and expert analysis happens also to be the bad news -- because the more stone-cold pros there are, the more likely their predictions are to diverge in meaningful ways. And never is this more true than when it comes to predicting so-called "wave elections," in which the mood of the electorate turns so decisively against one or the other of the two political parties that even the best models under-count the damage.

Worse, this kind of mistake is what statisticians refer to as a "systematic error": one in which the divergence between prediction and outcome only manifests in one tail of the probability distribution. For recent examples of this phenomenon, the United States has experienced "wave" elections in three consecutive midterms -- 2006, 2010, and 2014 -- and in all three of them the most trusted and professional pollsters and poll analysts, and Nate Silver, have all missed the victorious party's margin to the low side. But then this much is nearly tautological: Having failed to diagnose a wave, it would be impossible to then over-estimate its size, since missing a congressional margin to the high side would itself constitute the correct prediction of a wave.

So the bigger, more vexing question, then, is why pollsters and poll-analysts so often miss waves altogether -- and whether those explanations could proffer any guidance in predicting the outcome of this year's electoral horse race. Personally I believe that the explanations for un-diagnosed waves are mostly straightforward, and that they do, in fact, offer real promise that the Democrats could be positioned to do much better down-ballot than the punditry is currently predicting. Let's look at the factors that might substantiate this possibility.

Polling isn't done across the whole ticket. When looking at polling data in House- and Senate races, it's important to remember that a polling firm such as ARG or Marist or Quinnipiac doesn't generally ask a respondent who he or she plans to support in both the Presidential and down-ticket races. Most persons who participate in telephone surveying at all do so under a banner of at least mild irritation, and the pollster is therefore limited to no more than a very short handful of questions before the risk of hangup reaches critical mass. Worse, the surveyor must collect certain demographic data in order for the response to provide a full battery of so-called "cross-tabs" -- the bread and butter of insider validation, without which the entire exercise is usually disqualified. A survey that doesn't ask the right number of women, Hispanics, elderly, etc., won't be taken seriously by the beltway types or the media either one.

The issue here is that persons who respond to a down-ballot-only poll have no real way of honestly weighting the effect of a Presidential blowout on their proclivity to vote at all. If I call a Republican in New Hampshire, and I ask which candidate he or she plans to support for Senate, he or she will probably say Kelly Ayotte. But it would be a fruitless exercise for me to ask something on the order of, "Okay, from a scale of zero to ten, how likely is it that your dejection at the Trump campaign's certain humiliating defeat, will suppress your likelihood to cast a ballot?" To begin with it would be a leading question, and therefore unscientific, but even if it were phrased far more cleverly it would be impossible for the person to give me an assessment that was both scientific and reliable, since neither of us know how he or she will feel when the time comes.

Let's be clear: So-called "likely voter screens" are extremely sophisticated metrics, devised by people a good deal smarter than I am after having gone to school for a very long time to study and perfect this extremely high craft. ...But then, of course, they get it wrong when it comes to waves. And I believe the reason may be simple and unmeasurable despondency on the part of the supporters of the "losing" side. Such a possibility would explain at a stroke how a Presidential poll in New Hampshire could have Mrs. Clinton leading by double digits, while a separate and ostensibly equally scientific poll could have Republican Kelly Ayotte enjoying a small lead in her bid for reelection from that state to the US Senate -- which doesn't pass even a basic sniff-test in an election season as partisan and negative as this.

To buy both of these results, one would have to believe that people will split their ticket: by voting for, say, Hillary Clinton and Kelly Ayotte, the Republican Senator from New Hampshire. But the support for this theory strikes me as even more suspect and indeed maybe even self-deceptive, than usual. Yes, both Mr. Trump and Mrs. Clinton are deeply divisive figures -- and this could, as has been argued, prompt disaffected Republicans to vote for Mrs. Ayotte as a means of "blocking" Mrs. Clinton. But wait: if the polarizing influence of Mrs. Clinton's candidacy isn't sufficient to motivate a Republican to vote directly against her, is it really plausible to predict that the same voter will expend the effort necessary to indulge a Machiavellian triangulation by voting for Ayotte? Perhaps in 1996, when it was Mrs. Clinton's husband, but not now that the country has had twenty more years to segregate itself politically. To my mind, the far more plausible explanation is that pollsters are under-valuing the effect of Republican disillusionment on suppressing turnout for Ms. Ayotte and others like her.

The most famous example of this phenomenon was the 1974 midterms, in which a relatively un-damaged President Ford headed the recently-disgraced party of Lincoln. In that contest the professionals badly under-estimated the size of the Democratic wave, having failed to account for the suppressing effect of a demoralized Republican base. And it could easily be imagined that the very same thing happened to Republicans in 2006, at the nadir of our national folly in Iraq. The 2010 and 2014 waves are probably better explained by other phenomena, to be covered below -- but the point stands: We have at least one solid data-point suggesting that, at least some of the time, waves are under-predicted because the polling firms can't adequately adjust for uncharacteristically high enthusiasm gaps without blowing up their likely voter screens. And that's before the eventual winning side starts leveraging its advantage. Which they do, of course:

Waves have built-in factors of self-amplification. When a race that should be D+4, like the Wisconsin Senate seat, is instead polling D+6, everyone can see that difference but only one side can react to it. Specifically, what happens in such cases is that the leading party in such a race can divert some or all of its star surrogates, turnout machine, and deployment of paid advertising and internal staff, to a race that was D-1 or D-2, and in the process move the needle in a way that has much more bearing on the size of the wave, since doing so results in the potential pickup of an additional seat.

In recent days, both Obamas have appeared in Florida to campaign -- on behalf of Democratic Senate challenger Patrick Murphy, as much as on behalf of Secretary Clinton. This may or may not have had any direct effect, but the fact of the situation is that a race that was once solidly in the Republican camp has become a statistical tie. If one imagines Murphy defeating Rubio in Florida, and if one imagines that at least part of the reason was that the Democrats decided they could breathe easier in Wisconsin, it's easy to see how the beginnings of a wave (say, in Wisconsin) could lead to the amplification of the wave (say, in Florida). And because there's no offsetting scenario in the opposite tail of the distribution, the result is a one-sided miss by pollsters, even a few days out from the election. Which brings us to the third problem in predicting waves.

As the name suggests, waves are acute and often manifest at the last minute. As we've explored several times in these pages, polls don't do well with rapid changes. A poll must be in the field for a certain number of days, then collated and data-entered, then analyzed, then written-up, and *then* published. My own unprofessional sense of these things is that the modern lag between polling data and manifestation in the narrative is somewhere on the order of ten days, depending on the severity of the swings that happen in the meantime. (We're only just now, I believe, starting to see the solidification of the Access-Hollywood fallout in the map of the electoral college.)

This is not a big deal in August and it's not a big deal in a relatively close and stable election. But waves are, by definition, none of those three things. They happen with thunderclap intensity and they happen in the ballot booth. And since many Americans don't actually decide for whom (or whether) they'll vote until literally the last minute, it's next to impossible for even the most sophisticated likely voter screens to notice -- or even predict -- the big, herd-mentality breaks in voter behavior that characterize the biggest waves.

This is particularly bad news for Republicans in 2016, since the likelihood of further bombshells by and about the Trump campaign cannot be overstated. After all, this is a man who appeared on the Howard Stern Show dozens of times in the 1990s and early 2000s, and whose private opinions regarding race are only failing to gain traction because of his even more obviously repugnant private opinions about women. My guess is that someone, somewhere, has a recording of Mr. Trump saying something characteristically exploitative and disgusting about women, but which also manages to insult one or more minorities at the same time, and that this one is being saved for the big finish. Dismiss that as idle speculation if you must, but if I'm right, there would be simply no way for a set of down-ticket polling to adequately price the effects of such a disclosure into their findings. And this is only speaking of the campaigning acuity of Mr. Trump himself, of course: The down-ticket Republicans would still have to run flawless campaigns in these last two weeks, a feat which is rendered far more difficult in a strong political headwind like this one.

The down-ticket candidates have dis-equal challenges to persuade the undecideds. In 2014, one of the embattled Democratic Senators was a woman from North Carolina named Kay Hagen. A first-term, moderate Democrat, she'd been swept to power during the original Obama victory in 2008, having defeated none other than Elizabeth Dole (!!!), only to see the mood of her state swerve sharply and continually against the Administration over the entirety of her term. Popular in her own right, Hagen polled slightly but consistently ahead of Republican Thom Thillis through the summer and fall, but at every turn the warning signs of her imminent electoral difficulty could be found no further afield than her repeated disavowals of the head of her own party. "I, too, have disagreed with President Obama on many occasions," she was heard to say one crisp October afternoon -- and not long after this statement she lost by 45,000 votes.

Democrats love to say that candidates who play the game this way have no one else to blame, but consider for a moment the alternative strategy available to Hagen: to hug the President on live television and remind her state of the historic economic turnaround we've all just experienced. It's tempting to think this would have worked, but this is because in my experience Democrats like facts and rationality a little too much for their own electoral good: The general public, particularly in North Carolina, was simply in no mood to hear about how great things had gotten under President Obama (despite the fact that they had), and as such a campaign based on touting the accomplishments of the Administration, and Hagen's ties to it, would have rung as hollow as a shrill college lecture on the importance of doing one's homework.

This year the shoe is firmly on the other foot, with down-ballot Republicans reaping what they sowed in '10 and '14, in the form of a historically unpopular figure at the head of their party's electoral fortunes. Their choices, as we've seen, have been as follows: They can (a) disavow Mr. Trump and thus inflame his supporters, who comprise a plurality of the Republican cohort in this year's contests. They can (b) embrace Mr. Trump as the misunderstood and swashbuckling underdog who'll block Hillary Clinton if not imprison her -- thus alienating the majority of the electorate, non-Republican and Republican alike. Or, in some cases, they can (c) vacillate without plan or principle between the two.

It's easy to see why (a) and (c) are easier choices than (b), but it's equally easy to see how someone who chooses (a) or (c) -- especially (c) -- could poll ahead of his or her actual performance on election day. As we've seen elsewhere in this column, a demoralized base has much less difficulty telling a pollster that they will support a down-ballot slate, than they do with actually summoning the necessary grit to deliver themselves to a losing exercise, and to *then* cast a vote for someone who sounds like he doesn't always, or even often, support their point of view unless it's electorally expedient. Make no mistake, if Republicans hate anything more than losers, it's ideological relativism.  

The effect of these dynamics would be to galvanize an Obama coalition that has thusfar been lukewarm to Secretary Clinton, increasing her margin and pushing the down-ballot races in the same direction -- not least by sheer force of turnout advantage. Which is yet another thing about waves that makes predicting one this time so elusive.

In waves, the effect of a good turnout operation is monolithic across the entire ballot. At the same time that Democrats seem poised to force-multiply their recent advantages in turnout operations, the unprecedented discord in the GOP has left a strong impression that even rudimentary ground-game efforts are less likely to benefit Republicans, this cycle. Much has been written about the paucity of paid turnout professionals on the Trump team, the lack of coordination with the RNC, and the virtual civil war that has been transpiring between Trump people and various state parties. These things alone would probably have been enough to consolidate the likelihood of a Clinton victory on November 8th, of course -- but additionally I fail to see how any poll targeted at some other race on the ballot can possibly account for the effect of this enormous disparity in ground operations, down the ticket.

Consider a Hispanic woman who has never voted before in her life, and who was inspired by a well-oiled and well-targeted GOTV effort on the part of the Clinton campaign. Such a person is not going to walk into the booth, bubble the circle for Clinton, and go home. And she's not going to split her ticket, either. The question, then, is whether a team of intentionally sterile-minded scientists could possibly model such survey contamination from a Presidential turnout operation, as it would affect their own, down-ballot race. It would look like confirmation bias, and the findings would suffer reduced credibility as a result. But the analogous effect on the other side's voters is even more chilling....

Collapse is hegemonic. Nobody likes to say that they backed a loser, and this seems particularly true of an eventual loser who spent his entire public career vilifying losers. The result is what engineers call a "self-exciting vibration," whereby the more Mr. Trump sinks in the polls, the more he appears to be a loser, and thus the further he sinks. We've seen this dynamic with other, similar (?) candidacies in recent elections, notably Herman Cain: a man whose crash became the reason for his crash, and whose base of support evaporated with a geometric swiftness that few people remember for the spectacle it was.

And yes, it's true that in this particular election the opposition to Hillary Clinton has been, and continues to be, hegemonic as well. Just ask Bernie Sanders. But the principal difference between the stubborn support of Sanders' supporters, and the base following Trump, is that the empty machismo of greatness and winning had little or nothing to do with the political identity of the Sanders movement, except perhaps in a sort of rumpled, lovable negative. Tell a Trump supporter, by contrast, that his candidate is in serious danger of failing to hold Utah, Texas, Georgia, South Carolina, *and* Mississippi, and at some point the narrative of a rigged election seems unlikely to withstand the weight of the evidence. Thus will the Trump supporter in question be left with the Hobson's choice of casting his ballot for a loser, or staying home. And this is something that even a sophisticated statistical model is unlikely to detect on-the-fly, precisely because of how statistical models behave, and how they are interpreted....

Statistical models, and their analysts, abhor their own tails. Extreme outcomes are so anathema to professional statisticians that they've engendered an entire sub-lingo designed to identify them and root them out. Indeed the term for establishing a best-fit line through a plot of noisy data -- regression -- owes its origins to this precept among statisticians. Fluky-looking data points will quickly be subsumed by the overall data's tendency to *regress* to its more centralized outcomes, and accordingly statisticians tend to discount the flukes as empty noise.

Ordinarily this is absolutely the right thing for the statistician to do. If he didn't, there would be no such thing as statistics in the first place. But when it comes to election-polling data there's a crucial and wave-averse problem with this thinking, in that it's entirely too easy for the analyst to be distracted by the apparent mean outcome of an election, which is to say a tie, and to therefore discount any data that strays from fifty-fifty to the far side of a race's *true* mean condition.

Display to the commentariat, especially a statistically knowledgeable commentariat, a set of three recent polls showing Secretary Clinton with leads of 7, 10, and 12 points nationally, and the pundits will paradoxically be more likely to under-value the 12-point survey than they will the 7-point one, even though the 7-point result is further from the true average of the three, because the 12-point result is further from fifty-fifty -- and is therefore seen, wrongly, as the outlier.

Conclusion. So is it possible that the predicted volume of apparent split-ticket voting will in fact occur? Is it possible that both Hillary Clinton and John McCain are about to win Arizona by five points? Yes, it is of course possible: We won't know that it wasn't until November 9th, and if things do indeed turn out this way, the lesson for Republicans will be the one that many of them expected to have to learn all along -- that a primary system designed to satisfy a polarized base isn't worth the eventual nomination of such a toxic and self-evidently un-electable candidate for President.

To my semi-professional eye, however, it seems far more likely that a Democratic wave is about to defy all efforts to predict it. Indeed I would be willing to wager that at least some of the apparent split-ticket polling we've seen (notably in New Hampshire and Arizona) is illusory for some or all of the reasons I've elucidated here today: The Democrats will recapture the Senate and, I believe, come extremely close to recapturing the House of Representatives.

Or maybe -- just maybe -- they'll take that, too.

Dave O'Gorman
Associate Professor of Economics
Santa Fe College
Gainesville, Florida


warren johnston said...

bull shit

Dave O'Gorman said...