Nate Silver in 2015. Photo by INSIDER IMAGES/Gary He, courtesy Internet Week New York.
If you are a voting-age American and you are engaged enough to be reading this, you undoubtedly know that the first televised debate between presidential candidates Hillary Clinton and Donald Trump will take place Monday, and that it could tip the balance in a close race.
Yes, by all accounts the race is close – but how close? And how reliably will we able to know whether Trump has significantly improved his chances when the debate is over, or whether Clinton has effectively beaten back his challenge to her once-commanding position?
When I sat down to write this post on Tuesday morning, my take was that although Trump’s position has clearly improved during September, his was still more of a long-shot candidacy than either camp or many pundits would have us believe. (Why do both candidates want us to think this is a cliffhanger? Because Trump’s mostly small-dollar backers lack money to throw away on a lost cause, while Clinton’s team has to fight against complacency in both her donor and voter base.)
My biggest go-to source for presidential election oddsmaking is FiveThirtyEight, the poll-aggregation website run by statistics guru Nate Silver and several competent colleagues. My starting point for analyzing where things stand right now is the do-it-yourself Electoral College vote-counting map, 270 to Win. (There are 538 votes in the Electoral College that ultimately chooses the president, and hence a majority requires 270 electoral votes. Get it?)
One of the things I like best about Silver’s shop is that although it does not shy from using decimal points, it also never pretends to know more than it does or to predict outcomes. Instead, FiveThirtyEight works like a sophisticated bookmaker: It recognizes that the outcome of any given contest can hinge on unpredictable and even random factors, but that the probability of a given outcome – the odds, in other words, that one side or the other will win – can be calculated with reasonable accuracy. Bet with the odds and you will win more often than not, but that does not mean you will win every time. You won’t.
Moreover, Silver and his colleagues recognize that polls have inherent margins of error as well as some built-in biases, that the aggregation of many polls will ultimately be more accurate than relying on any single poll and that polls become increasingly predictive as the election gets closer. They also recognize that presidents aren’t selected by majority vote, and that in fluky situations you can even end up in an Electoral College tie.
So to build their maps, the folks at FiveThirtyEight run 10,000 computer simulations each time they have numbers to crunch – multiple times every day, at this point in the election cycle. They run the simulations on a state-by-state basis, because that is how nearly all the Electoral College votes are decided. (Only Nebraska and Maine split their votes by congressional district.) Then they calculate the probability of an outcome three different ways: first, based on how the polls anticipate people will vote on Nov. 8; second, based on a proprietary “polls-plus” model that considers the economy and other external factors based on historical results; and third, counter-intuitively to all of us who are not statistics experts, based on a “now-cast” of how the election would turn out if it were held today, rather than in November. Those of us with less understanding of polling would generally presume that the now-cast would always match the polls-only projection for Nov. 8, but it doesn’t.
Silver and his staff provide a handy map, colored in different shades of pink or blue, showing the probability of an outcome in each state, and their estimated Electoral College vote under each of their three formulas. When I was writing this on Tuesday morning, their polls-only number gave Clinton a 58.9 percent probability of winning and a projected Electoral College total of 287.6 votes, against Trump’s 41.1 percent probability of winning and projected 250.1 electoral votes. (In the popular vote, the projections were Clinton 46.6 percent, Trump 44.5 percent, and third-party candidate Gary Johnson at 7.6 percent of the popular vote and with 0.3 electoral votes. In the real world, of course, nobody gets a fraction of an electoral vote.)
Taken at face value, the Tuesday morning projection was telling us that if the Nov. 8 election were to be run 1,000 times, the aggregated polls indicated that Clinton would emerge with a victory 589 times and Trump would win the other 411 contests.
But there is something about FiveThirtyEight’s numbers, or at least my understanding of their numbers, that bothers me when I look at the state-by-state breakdown. I think the figures overstate Trump’s chances by as much as 11 percentage points – in other words, that he would win no more than about 300 of those hypothetical matchups.
To see my logic, we have to use 270 to Win’s handy little map. I started by assuming that Trump will win every state where Silver’s aggregated polls showed him leading on Tuesday morning, including extremely close Nevada and the critical large states of Florida and Ohio, as well as North Carolina. I gave Clinton every West Coast state save Alaska, and every East Coast state from Virginia to Maine, including Pennsylvania. She got Colorado, which is fairly close right now, and New Mexico, which isn’t. She also dominates Illinois and gets credit for Minnesota, Wisconsin and Michigan; Trump carries Indiana (home state of his running mate, Mike Pence) and Iowa.
Trump could conceivably pick up an electoral vote in Maine, but I gave all of them to Clinton. She could conceivably grab one vote in Nebraska, but I gave the whole state to Trump. The split-state uncertainties cancel one another.
This starting map gives Clinton 273 electoral votes to Trump’s 265, so she wins the White House. What are the prospects Trump can change this before Nov. 8?
The closest blue states for Trump are New Hampshire, where he has about a 37 percent predicted chance of winning four electoral votes, and Colorado, where he has a 36 percent chance of winning nine. Assuming all the other states end up as projected, either of those would give the presidency to Trump. But that is an enormous assumption, considering that Clinton’s chances in Nevada are nearly 50 percent, along with 45 percent in North Carolina, 48 percent in Florida, 42 percent in Ohio and 37 percent in Iowa. If Trump’s results in nearly all-Anglo New Hampshire turn out to be noncorrelated with, say, his performance in significantly Latino Nevada, he might pick up the former but lose the latter, leaving him worse off than before.
Silver’s many fans would note that his 10,000 simulated elections account for these variations; he just tallies the results. They would say his methodology is more scientific than my finger-in-the-air speculation, and I agree. But still, results either move randomly – as would be more likely in a knife-edge state like Nevada – or they move concurrently, as they would in a mass trend toward either candidate. If Trump’s best chance of picking up a state is New Hampshire’s 37 percent, it is hard to see how his overall chance of changing the map is any better than that as you move into terrain he will find even more challenging. At this writing, the chances of him losing Nevada are much better than his odds of winning New Hampshire.
Now we turn our attention to the big industrial states that are considered in play: Pennsylvania, Michigan and Wisconsin. Silver gives Trump about a 30 percent chance in each of them. Since Trump only needs one of them to win given our starting map, his chances would be significantly better than 30 percent if they moved independently of one another.
Historically, however, they don’t. All three of those states have voted for the same presidential candidate, in lockstep, in every presidential election since 1980, with the exception of 1988, when Wisconsin went for Michael Dukakis, the Democrat, while the other two went for Republican George H.W. Bush. Since 1992 all three states have gone to the Democratic candidate in every election.
It is reasonable to believe the correlation may break down this year, but not necessarily in a way that favors Trump. The Republican will do very well in the hard-pressed coal and industrial country in the western part of Pennsylvania, but he is not popularly favored in Philadelphia and its relatively affluent and highly educated suburbs. Philly itself has become a sort of mini-Brooklyn of late, attracting large numbers of college-educated millennials, who as a group have no more affinity for Trump than does the city’s large minority community.
So while any one of those industrial states breaking for Trump could put him over the top, there seems to be little reason to believe that state would be Pennsylvania, which is getting most of the outside attention due to its 20 electoral votes. Michigan (at 16) or Wisconsin (10) appear to be likelier to break from the pack and follow Ohio into the Trump camp, if any of those states in fact end up in his column. And if I had to bet on one of those two, it would be Wisconsin, where conservative Gov. Scott Walker survived a recall and a close re-election campaign before briefly entering his year’s presidential contest himself. Oddly, however, Wisconsin’s conservative talk radio hosts – who have made no secret of their hostility to Trump – may be the factor that keeps him from grabbing the state and, thus, that could put Clinton into the Oval Office.
All of this oddsmaking could change based on next week’s debate or any other surprises in store during the remainder of the campaign. And the odds do not always determine the winner, in presidential elections any more than they do in the Super Bowl. But they are not irrelevant either. Trump has rebounded from his August lows, but the numbers and the geography tell us he still has an uphill climb to reach 1600 Pennsylvania Avenue.