Why Nate Silver’s Forecast (Probably) Overstates Trump’s Chances

By
Nate Silver.Photo: Paul Zimmerman/Getty Images

If you spend your weekends bonding with family, hanging out with friends, or pursuing worthwhile hobbies, then you probably missed the most epic Twitter battle — about the propriety of “trend line adjustments” — of the entire 2016 campaign.

Right now, FiveThirtyEight — the site founded by celebrated data wizard Nate Silver — gives Trump a nearly one in three shot of winning the presidency Tuesday night. The Huffington Post, by contrast, puts his chances at roughly 2 percent.

One reason why these forecasts diverge so radically is that they make different judgments about the predictive power of older polls. The Huffington Post’s model is not as sensitive to temporary fluctuations in polls, presuming the consensus of old surveys will generally hold up, once the dust kicked up by the latest news story has settled.

Silver’s model isn’t as sticky. It presumes that more recent polling data is more accurate data, even if those surveys depart drastically from the past norm. Or, at least, it makes that presumption more readily than the Huffington Post’s does.

And one way Silver fits his model to that presumption is by adjusting past state-level poll results to reflect the new national trend. This method helps compensate for the scarcity of polling in some key states. As Silver explains, “if Trump led in a North Carolina poll by 1 percentage point in June, but the trend line shows him having gained 3 percentage points nationally since then, the model will treat the poll as showing him up by 4 percentage points.”

On Saturday, the Huffington Post’s Ryan Grim penned a takedown of Silver’s method, likening his “trend-line adjustments” to the unscientific “unskewing” that some conservatives pursued when the polls in 2012 predicted a result they didn’t like. Grim’s central point was that, by manipulating past poll data, Silver had transgressed the bounds of data journalism and entered the fallen realm of “punditry.”

Silver … disagreed.

Grim’s piece may not have been “fucking idiotic,” but it was more than a little unfair. Silver’s model isn’t “punditry” — for better or worse, its conclusions are more dryly empirical than the average blogger’s bloviations. His model’s trend-line adjustments aren’t made on a whim, based on his reading of tea leaves, but rather proceed automatically from the assumptions that undergird his forecast.

The problem with FiveThirtyEight’s model isn’t that Silver is “putting his thumb on the scale” to generate a closer race (and thus higher traffic to his site), but rather, that the model is built on flawed assumptions that (probably) overestimate Trump’s chances of victory.

The biggest difference between FiveThirtyEight’s forecast and the Huffington Post’s isn’t actually that “trend-line” thing. More significant is the weight each model gives to uncertainty.

Silver’s model looks at the race’s volatility over the past three months — and the relatively large numbers of third-party and undecided voters — and discounts the security of Clinton’s longtime polling advantage. The Huffington Post doesn’t sweat that uncertainty.

After the campaign we’ve all just witnessed, it’s hard to fault a forecaster for biasing his model toward the unforeseen. But there are a lot of reasons to think that, if the polls prove to be wrong, they’ll be wrong in Clinton’s favor. Which is to say: The uncertainty and volatility of the race are less likely to produce a Clinton loss than a larger-than-anticipated Clinton win.

Here’s the quick rundown of why I think this is true:

Polls don’t account for ground game. And Clinton has a better ground game.

This one’s pretty straightforward. Donald Trump has invested fewer resources into turnout operations than any modern major-party nominee. Hillary Clinton, by contrast, has the best GOTV operation that George Soros and Tom Steyer can buy. All polling models assume parity in ground operations. But that’s a profoundly wrong assumption this year, and one that Trump benefits from.

Early voting suggests that ground game has already put a critical state into Clinton’s column.

If Hillary Clinton wins Nevada, even Nate Silver thinks she has a greater than 90 percent chance of winning the White House. And a spike in Latino turnout has given Democrats a bigger pre–Election Day advantage in the state than the one they enjoyed in 2012 — when Barack Obama won Nevada by nearly seven points.

Latino turnout also appears to have given the Democrats a (slimmer) advantage in Florida. Nate Silver’s model is completely ignorant of these early voting totals. Instead, its relies on public polling data, which has historically underestimated Latino turnout.

Third-party voters are disproportionately young. And young voters are disproportionately anti-Trump.

This point comes courtesy of Vox’s Matt Yglesias, who argues that supporters of both Libertarian Gary Johnson and Green Party nominee Jill Stein are highly unlikely to break en masse for Trump:

For Stein voters this is a pretty clear-cut matter of ideology. But Johnson’s own vice presidential candidate, former Massachusetts Gov. Bill Weld, has also made his preference for Clinton clear. Johnson’s voters also skew very young, which is a terrible demographic for Republicans in general and for Trump in particular.

In other words, of the four possible things a Johnson/Stein voter could do on Election Day — stick to his guns, defect to Clinton, stay home, or defect to Trump — the fourth option is a lot less likely than the other three. If third-party voters are fuzzing up the polls, in other words, they are probably doing so by understating Clinton’s true level of support — not overstating it.

Finally, on the somewhat separate question of trend-line adjustments, there’s one reason to doubt Silver’s method: The polls probably exaggerate the malleability of voters’ preferences. When a news story produces a giant polling swing among a deeply polarized electorate, chances are what we’re seeing isn’t a sea change in public opinion, but rather a change in which a party’s base is less likely to hang up when a pollster calls.

All that said, national polling averages put Clinton’s lead between three and four points. And Trump could lose the popular vote and still win the Electoral College, due to the abundance of white, non-college-educated voters in swing states. So, ya know, anything’s possible.

Just not as possible as Nate Silver’s model would have you believe.