Sunday, November 6, 2016

Sorry, Donald (and Nate): Any Polling Error Will Favor Clinton

For much of this last Sunday before election day, I grappled with the challenge of how to write -- and then stagger -- two separate blog columns. (#whitepeopleproblems.) The first of these two hypothetical columns was supposed to be a summary of where things stand with respect to early voting; the second one was supposed to be a friend-of-the-court brief regarding the geek war between the editors of the Huffington Post, and Nate Silver. And it's no small testimony to the sloppy file structures inside my head that it took until late this afternoon for me to realize that these had actually been the same column, all along.

Unless you've been vacationing on Neptune, you probably know by now that most of the leading polling firms have shown a marked tightening of the electoral horse race between Trump and Clinton -- presumably in response to the October cudgel delivered to Trump campaign HQ on a silver platter by the Director of the FBI. Timed to inflict maximum political damage, Director Comey's letter to Congress about his re-opening of the Clinton/email investigation was released at the precise moment that early voting began in many of the most crucial battleground states.

The effect has been to dramatically alter both the complexion and the media narrative of the race: from one in which Clinton was as inevitable as any Presidential candidate since Nixon's reelection bid in '72, to one in which literally anything could happen and probably would. But a funny thing happened along the way to Comey's role in upending the fate of the world: In terms of the actual dynamics of the race, there seems to have been much less impact than the polling data might otherwise suggest. The polls are, with apologies to Mitt Romney's supporters from 2012, genuinely and provably missing the drift of this thing, this time -- and Secretary Clinton is the direct and ubiquitous beneficiary of their collective mistake. This is an unusually long column even by my standards, so I hope you'll bear with me.

First off, let's be clear: Cool-headed rationality about election forecasts -- particularly cool-headed rationality posited by persons from the left -- should demand an extremely high standard of credible proof for any allegations that entire swathes of polling data are missing something both operative and systematic. Today's column simply cannot be another of those, "Oh, you'll see: John Kerry will win anyway because young people don't answer the phone" sorts of affairs. The data supporting any assertion of widespread polling error is going to need a real, verifiable chain of evidence.

Fortunately for Mrs. Clinton, we have one. In the form of the early-vote data that has already been compiled and analyzed by persons both better trained and better experienced, than I.

I take you first to Nevada, where seasoned veteran gumshoe journalist Jon Ralston has been tracking the party affiliation data of early votes cast, and compiling the results in a live blog as the new information trickles in. It's a fun read for progressives -- so much so that I hesitate to gloss it here -- but the upshot is that Mr. Trump is getting positively clobbered in the early vote, particularly in Clark County, and most especially among the Hispanic population. At the time of this writing, Ralston estimates that Secretary Clinton has an approximately 75,000 vote lead, significant both because it surpasses President Obama's leads in 2008 and 2012, and because Ralston considers it all but literally insurmountable for Mr. Trump.

In Florida the picture seems at first to be decidedly more mixed, with higher-than-expected turnout rates among some cohorts being offset by lower-than-expected participation by others. But as Steve Schale's competent reporting attests, the deeper crosstab information is nothing short of gob-smacking. Mrs. Clinton seems at this hour to have banked something like a 500,000 vote lead in the big-three counties of Palm Beach, Broward and Dade, and has enjoyed Obama-plus support levels in key areas of the I-4 corridor as well. For team Death Star to close such margins, Schale estimates that Trump would have to carry 90% of the remaining registered Republicans, as well as making serious inroads into non-affiliated women and minorities. He hasn't quite stuck a fork in Florida in the same way that Ralston has in the Silver State -- but the writing is on the wall, if not actually in the blog. According to CNN, meanwhile, Secretary Clinton's early-vote totals are far surpassing expectations among key demographics in North Carolina and Georgia as well. And Clinton wasn't even expected to compete for Georgia.

The obvious question is why this is happening. Not, "Why is Mrs. Clinton beating Mr. Trump," of course, but rather, "Why does the polling data, and why do at least some of the prediction models, seem to paint a picture of this election so radically different from the early voting performance?" And it is this question, and its answer, which ultimately tie together my two topics for the day. As I hope to convince you, the explanations for this year's apparent divergence between polls and votes are as straightforward as they are difficult to incorporate into models.

To begin with, we are seeing a historically high turnout among Hispanics. In Florida, for example, the number of early-voting Hispanics to this point in the election cycle exceeds the total number of Hispanic voters in 2012 by well over 100%, from 260,263 in 2012, to a whopping 596,146 so far this year -- not including this, last day of early voting in the Sunshine State. Hispanic over-performance is similarly noteworthy in North Carolina, Georgia, and especially Nevada, where it is believed that Mrs. Clinton's apparent 75,000-vote margin is actually artificially low, given the abnormally high proportion of Hispanic voters who decline to affiliate themselves with a major political party.

This is a vexing issue for polling firms and their predictions-pages, since in order to be considered scientific, a poll has to screen its own data for an appropriate demographic breakdown that reflects the likely mixture of the larger electorate. You wouldn't, presumably, put much stake in a poll that showed Trump leading, but which had arrived at this conclusion by restricting the question to members of your local chapter of the VFW. So what you look for is a poll whose respondents look representative of the larger population. Ordinarily this is no problem -- indeed it's essential -- but the demographic template that gets used for this test is almost always disproportionately informed by the demographics of the most recent election. ...So that if something genuinely dis-continuous comes along, like a Presidential candidate who thinks everyone from Mexico is a rapist, the polling firms are more-or-less helpless to incorporate an anecdotal anticipation of a higher turnout rate among, say Hispanics.

But it gets worse for polling firms, from there. This is because over and above the discontinuity in likely Hispanic participation this time around, Hispanic populations are also notoriously difficult to poll. They are by far the demographic most likely to have neither a land-line telephone, nor home internet access, thus eliminating them from all but the most expensive, highest-quality polling efforts -- and even then, they also have a disproportionate proclivity to ignore incoming cell-phone calls from numbers they don't recognize. All of which adds up to a brewing statistical embarrassment, quite possibly on the sort of Dewey-Truman scale that some of us were pinning our hopes on in 2004. The short version is that Mr. Trump is losing badly, both in turnout and in preference, by a whole swathe of the electorate that is effectively invisible to the Nate Silvers of the commentariat. 

Further proof of the gravity of this issue comes from none other than Lindsey Graham, who earlier today was quoted in the act of bemoaning Trump's poor strategy with respect to diversity issues. "The story of this election may be the mobilization of the Hispanic vote," sayeth he. "So Trump deserves the award for Hispanic turnout. He did more to get them out than any Democrat has ever done.” This, you will have noted, is a quote that exhibits neither the tonality nor the open-endedness of a surrogate who thinks this thing is really tied with two days to go. Because it isn't.

There is also the problem of all those cross-pressured suburban professional women. And the problem here is that polling populations are adjusted for demographics, and for party affiliation, but they aren't adjusted for the intersection of the two. If, for example, a polling firm obtains responses from 51% women and 49% men, and if the responses are also split 40-40-20 on party affiliation, the poll is considered reflective, regardless of whether it has captured a sufficient number of women who also happen to be Republicans.

As with the Hispanic cohort, this issue is generally a non-issue because generally a Republican candidate appeals more-or-less equally across the demographic sub-groups of his or her own party. Not this year. Instead, as we've seen over and over, Mr. Trump is badly under-performing other Republican Presidential candidates, specifically among Republican women. In statistical circles, this kind of problem is called "auto-correlation," and there's little chance that our available polling data is being corrected for it, since doing so is both expensive and time-consuming. No firm that is trying to scoop the field on the latest survey out of Iowa is going to take the extra day and a half to run Durbin-Watson tests on every cohort in its crosstabs. There just isn't time.

Worse still, the problems posed for Mr. Trump by this auto-correlation are themselves positively correlated -- since a large swathe of the early vote is both white, and historically unlikely to have voted for him. Consider that, across the country as a whole, Secretary Clinton's overall lead would seem to be on the order of about four million votes, leaving Mr. Trump in the unenviable position of having to secure over 55% of the expected remaining vote, just to tie the thing.

This would be daunting enough (not to say more-or-less impossible) if those who had early-voted were disproportionately minority voters, but in fact the opposite is true. In a normal election, whites comprise about 66% of the voting electorate as a whole, and so far during this election the early voters have been a staggering 80% white. What this means for Mr. Trump is that he would need to secure an even larger portion of the remaining white vote, far larger, or else make highly improbable inroads into the minority support for Clinton, to recover. Oh, and 56% of those who have already voted, are women.

On top of everything else, there is the matter of Mrs. Clinton's vaunted ground game. As I've written before, there's no scientific way for prediction models to incorporate the overwhelming superiority of the Clinton GOTV-database ground game, but I don't think it takes a six-paragraph rehash of that advantage for the point to stand. In a race that polls tied, only one side of that tie will be contacted with anything like the methodical professionalism necessary to ensure that it finds its way into the voting booth. Call it a field-goal unit, if you will, but in the face of its undeniable existence I just can't see a poll-tied state as a true tie.

All of this would render any projections predicated strictly on hypothetical polling data, suspect. But it gets even worse from there, since this year we're seeing an abnormally big and durable problem with something called non-response bias, a term for distortions caused by systematic non-responses. Suppose cohort-A wants candidate A, and cohort-B wants candidate-B. Obviously the pollsters can endeavor to call as many people as it needs to call until it gets a representative split between A and B, but there is no way the polling firm can know the preferences of the people who hung up on them along the way.

As with the other issues we've covered, under more normal circumstances this isn't a big problem. But in an election where both candidates have such high negatives, and during which each of them has endured such devastating news coverage, it is improbable that the political makeup of the non-respondents has consistently mirrored that of the respondents. And when all the non-respondents are from one cohort, the poll will show a disproportionately high percentage of crossovers from that cohort, to the other candidate, thus inflating that candidate's real standing among the larger electorate.

The result is that this year our projection specialists have suffered through a series of wild, almost certainly apocryphal swings in voter preferences, with very little recourse to cooler circumspection. When news was at its worst for Trump, we saw an enormous swing toward Mrs. Clinton; when the Comey letter was revealed, we saw an almost equally enormous swing toward Mr. Trump. Garden-variety common sense -- to say nothing of centuries of statistical sampling theory -- demand a truth that resists that sort of simplicity, and instead lies somewhere in-between. Indeed this is where the word "regression" comes from in the first place. But in 2016 the job of the pollster is to report the poll, and the job of the predictor is to take that poll and incorporate it into his prediction as soon as possible.

So how does Nate Silver fit into all of this?

The short answer is disappointingly, I'm afraid. Silver's fivethirtyeight website is still a popular destination for blurb-sized commentaries on the state of the election, to be sure. But for some time now his model for predicting the final results has appeared to some of us to be abnormally sensitive to transitory movements in polling data and, now that early voting has started, it paradoxically seems almost preternaturally robust to empirical contradiction by data we can literally already see. The combined effect of these two colorations is to suggest a race which is much more permissive to the electoral fortunes of Mr. Trump than seems empirically justified.

Thus it was that yesterday the Huffington Post's Ryan Grim took Silver to task regarding the former charge. Grim's article is well-constructed and carefully supported with all the necessary statistical expertise -- well worth your time, if you ever make it to the end of the present column, of course. The short version is that Mr. Silver employs an entirely ad hoc filter of his own independent design, which he calls a "trend-line adjustment," for the purposes of conforming any new polling data to his sense of the larger trajectory of the race. In this manner, an individual poll that shows Mrs. Clinton two points up in New Hampshire (e.g.) is adjusted in Trump's direction, on the grounds that other recent polls have showed Trump doing better.

Grim's problem with this is its subjectivity, and make no mistake. But mine is of a more boring, technical nature: By doing this, Silver unwittingly exacerbates the effect of any problems with non-response bias. A race which is already probably not as bad for Mrs. Clinton as it appears, is made by this practice to appear even worse, since polls that show reversion to the larger structural truth of the thing are being under-valued. In this fashion Mr Silver inadvertently adds a dramatic pro-cyclicality to his projections, and he is doing so at a time in the election season when this many people rarely change their minds, this much or this frequently either one.

Never one to sit still for being called out, Mr. Silver's late-night response to Grim was a profanity-laced, fourteen-episode tweetstorm, riven with unsupported claims of credibility-by-precedent and un-founded allegations of statistical malpractice by Mr. Grim. "Eighty percent of the time, when you go low, I go high," Silver raved in his final tweet. "The other twenty percent of the time, I kick you in the balls."

The question of whether ESPN wishes to continue its contractual arrangement with someone capable of threatening a casual critic's testicles, is not for me to consider either way. And, full disclosure, I've always noted this sort of behavior from Mr. Silver, together with the arresting frequency of his mistaken predictions, even when it was far less socially fashionable to do so.

In 2010 he confidently predicted a uniform swing in the UK federal elections instead of a hung parliament, shortly after having posited that U.S. republicans would not under-count themselves in the census -- the latter of which he supported by weighting the Republican states equally, regardless of population, so that under-responses in Texas counted the same, regardless of total, as over-responses in Wyoming. His entire self-mythology, it seems to me, is predicated on correctly calling two blowout Presidential elections -- something rather more than a few of the rest of us have also managed to do without threatening anyone else's reproductive anatomy, at least in print.

But the more alarming question, in my mind, is how Silver can continue to claim the more refined expertise, even implicitly, in the face of his model's continued insistence that the early vote simply hasn't taken place. Or so it would seem.

At the time of this writing (with a mere forty-eight hours before polls start closing around the country), Mr. Silver's website gives Mr. Trump a 50.8% chance of winning Nevada, a 51.2% chance of winning North Carolina, a 51.6% chance of winning Florida, and a whopping 66.5% chance of winning Ohio -- the last of these despite the fact that the Columbus Dispatch shows Mrs. Clinton with a small but durable lead in the Buckeye state. At most, one of these four things is actually going to happen, which would make Mr. Silver 1-3 in the four biggest battlegrounds left on the map this time around. (That is, if he doesn't hurriedly change his "predictions" at the last possible second, like he did in both of his supposedly wizardly successes in the past.) I mean, just look at these screen captures, taking special note of the volatility of Silver's predictions about each of these states (graphs, lower right):

There simply is not this kind of volatility in the American electorate in the October and November of an election year. Any October and November of any election year, much less an October and November at the end of a historically polarizing and divisive campaign like this one. Indeed if we widen our focus to Silver's national map, we can see that his confidence in Trump's late-autumn surge leaves him even more vulnerable to omelet wearing than these four states suggest in isolation. Consider his projected electoral-vote map -- with the caveat that the darker shades of red indicate progressively higher "certainty," in Mr. Silver's mind, at least, that Mr. Trump will win that state.

According to this map, Silver believes that Mr. Trump is a near-mortal lock in Iowa, Arizona, and Georgia -- and I put it to you that he's wrong in at least one of these instances as well, perhaps as many as all three of them. Indeed if we can agree on nothing else, I hope we can agree that these projections are shockingly, one might say willfully tone-deaf to the massively pro-Clinton numbers we've been seeing for early voting in every state listed in this discussion.

At the end of the day, the beauty of statistics is of course its own trap. Or, more accurately, its own series of traps. The more reflective of past precedent one's model is, the less accommodating it will be of structural discontinuities in historical trends. The more responsive to new data, the less loyal to its own best-fit line. These are the sorts of dilemmas for which the big boys get paid the big bucks.

But with his pro-cyclical decision to adjust polls in accordance with recent swings, and with his apparent tone-deafness to the hard early-voting evidence we have in hand, Mr. Silver seems to be committing himself to the worst of both of those tradeoffs: a model that is at once insufficiently respectful of its own larger trends, and insufficiently receptive to the truest source of data that might suggest a movement away from them.

The fact that Mr. Trump's electoral prospects come out looking better for both of these reasons could easily be an accident of providence for the Orange One and his followers. But it does seem, at least to this commentator, that the page-views and associated ad revenue for fivethirtyeight can't possibly be hurt by Mr. Silver's depiction of a race ostensibly closer than it really is. As a friend of mine and equally well-trained statistician suggested late yesterday evening, Silver's prediction at the moment is enjoying the opportunity to have it both ways: If Mr. Trump wins, Silver will be the one who came closest to predicting it; if Clinton wins, he still said that too -- albeit at a lower level of probability than literally any of the other projections widely available on the worldwide web. And this possibility would raise the very unsavory dichotomy that Silver himself is either guilty of statistical malpractice, or disingenuous self-promotion.

Or, even worse, that it's some of both.

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

No comments: