lunes, 26 de septiembre de 2016

Election Update: Polls Show A Close Race (So Our Model Does Too)

Well, folks, this is getting tight. Donald Trump is in his strongest-ever position in MARCAPOLITICA’S polls-plus forecast, which gives him a 46 percent chance of winning the election. Trump’s chances are about the same, 45 percent, according our polls-only forecast, his best standing since it showed him with a 50 percent chance in the midst of his convention bounce.

Our models have been on the move toward Trump for roughly six weeks. But with dozens of polls coming out over the past few days, he’s no longer much of an underdog at all. Hillary Clinton leads narrowly — by 1.5 percentage points — in our projection of the popular vote. But polling weakness in states that Clinton probably needs to win, particularly Colorado and Pennsylvania, makes the Electoral College almost even.

I’m aware that there’s a lot of consternation and/or excitement out there about our forecast. But there’s nothing particularly deep going on here — our numbers are just reflecting what the recent polls are saying. First, here’s a list of the 10 national polls that we’ve added to our database since Saturday. I’ve shown both the current result and — since this is how our model’s trend-line adjustment works — how it compares to the average of other polls conducted by that pollster throughout the year:

ABC News/Washington Post Clinton +2 Clinton +6 Trump +4
CVOTER International Clinton +1 Clinton +1 —
Monmouth University Clinton +4 Clinton +7 Trump +3
Morning Consult Trump +1 Clinton +4 Trump +5
Quinnipiac University Clinton +1 Clinton +4 Trump +3
RKM Research Clinton +2 Clinton +2 —
Selzer & Company Trump +2 Clinton +11 Trump +13
SurveyMonkey Clinton +5 Clinton +4 Clinton +1
USC Dornsife/LA Times Trump +4 Trump +1 Trump +3
YouGov Clinton +3 Clinton +3 — National polls added since Sept. 24

On average, Clinton is ahead by only 1.3 percentage points in these polls — right where our forecast shows the race. And the trend lines are mostly negative for her, with Clinton polling an average of 2.6 points below the previous editions of the same polls.

Meanwhile, here are the state polls we’ve added since Saturday. The list excludes the latest editions of the Ipsos/Reuters and CVOTER International 50-state tracking polls, which our model uses but assigns a relatively low weight.

Ariz. Data Orbital Trump +2

Colo. CNN Trump +1

Colo. Gravis Marketing Trump +4 Clinton +1 Trump +5
Colo. YouGov Clinton +1 Clinton +1 —
Fla. Cherry Comm. Clinton +2 Trump +4 Clinton +6
Ga. JMC Enterprises Trump +6 Clinton +7 Trump +13
Ga. Landmark Comm. Trump +4 Trump +1 Trump +3
Iowa Loras College Trump +1 Clinton +13 Trump +14
La. JMC Enterprises Trump +10 Trump +16 Clinton +6
Maine U. of New Hampshire Clinton +4 Clinton +7 Trump +3
Mass. YouGov Clinton +13

Minn. SurveyUSA Clinton +7

Minn. Gravis Marketing Tie

Mo. YouGov Trump +9

N.H. Amer. Research Group Clinton +4 Clinton +5 Trump +1
N.Y. Marist College Clinton +21 Clinton +29 Trump +8
N.C. High Point University Clinton +1

N.C. Gravis Marketing Clinton +1 Trump +1 Clinton +2
Ohio Gravis Marketing Trump +1 Clinton +2 Trump +3
Ohio TargetSmart/Wm. & Mary Clinton +3

Pa. CNN Clinton +1

Pa. Harper Polling Clinton +2 Clinton +5 Trump +3
Pa. Gravis Marketing Clinton +3 Clinton +2 Clinton +1
Pa. Mercyhurst University Clinton +1 Clinton +8 Trump +7
Pa. Muhlenberg College Clinton +2 Clinton +7 Trump +5
Utah Dan Jones & Associates Trump +9 Trump +7 Trump +2
Va. Christopher Newport U. Clinton +6 Clinton +9 Trump +3
Va. YouGov Clinton +8 Clinton +12 Trump +4
W. Va. Just Win Strategies Trump +27

State polls added since Sept. 24

These tell pretty much the same story. On average among this weekend’s polls in what we consider swing states,1 Clinton leads by only 1.2 percentage points. And the trend has moved in Trump’s direction by an average of 2.9 percentage points. Again, that’s right in line with what our forecast shows.

Unfortunately for Clinton, her state-by-state polls are configured in a way that makes her Electoral College position relatively vulnerable. Particularly problematic for Clinton were the numbers in Colorado, where two of the three new polls this weekend had her trailing Trump. A couple of those pollsters (Gravis Marketing and CNN) have Trump-leaning house effects, but still, it’s a close race there, and Clinton leads by only 1.6 percentage points in our Colorado forecast. Without Colorado in her column, Clinton would need to win a state that she currently appears to trail in, such as North Carolina or Florida.

There were also five polls of Pennsylvania that showed Clinton ahead by only 1 to 3 points there. She leads in Pennsylvania by 2.4 percentage points in our forecast.

Not every poll was bad for Clinton: She led fairly comfortably in two new polls of Virginia, although they showed negative trend lines for her. She got relatively good polls in Florida and Ohio. And as with any long list of polls, this one contained a mix of good and not-so-good pollsters. But there was no clear pattern of better pollsters showing better numbers for Clinton, or vice versa. For instance, the single poll that hurt Clinton the most in our forecast was a national survey from Selzer & Co. on behalf of Bloomberg Politics, which showed her trailing Trump by 2 percentage points. Selzer is one of our highest-rated pollsters and had shown strong numbers for Clinton earlier in the cycle.

Recently, FiveThirtyEight has shown better better odds for Trump than other models have, for several reasons. First, our model is generally quicker to update than others, because of its use of the trend-line adjustment. That allows us to make inferences about how the polls are moving in every state, even when they haven’t been polled recently. For instance, the model correctly anticipated significant tightening in Colorado and Pennsylvania, even after we went a long stretch without many new polls there.

A good test of whether a model is too conservative, too aggressive or “just right” is whether it does a good job of matching new polls as they come out in a state. So far in this election, the FiveThirtyEight and Daily Kos Elections model — which also uses a trend-line adjustment — have done a good job of this, while other models sometimes lag behind the trend.

A good, related question is whether polls are mean-reverting. Clinton has generally led Trump by more than the 1 or 2 percentage point lead she has now. Does that mean she’s more likely to gain ground than to lose ground from this point onward?

Our polls-only model makes no assumptions about this, instead taking the polls at face value. Polls-plus does account for mean reversion, but it assumes that polls revert toward a mean established by an index of economic conditions, rather than the long-term average of polls. Because economic conditions project a very close race right now, the polls-plus forecast is about the same as polls-only.

One could argue for reverting polls toward a long-term average instead, as at least one other forecaster (Princeton Election Consortium) does. We’re not totally sold on the empirical case for this, but theoretically it’s perfectly sound: A model could have the race as a dead heat in the event of a hypothetical election held today but nonetheless have Clinton favored on Nov. 8.

FiveThirtyEight’s models also generally account for more uncertainty than other models — or at least they do in this election because the presence of a large number of undecided and third-party voters, who contribute to polling volatility. That helps Trump’s odds, since he’s (narrowly) the underdog in our forecast.

Another difference is whether one uses the version of the polls with third-party candidates included, as FiveThirtyEight’s forecasts do. Clinton’s leads are often slightly larger in two-way matchups. But those two-way matchups describe a hypothetical election — in actuality, Libertarian Gary Johnson will be on the ballot in every state, and the Green Party’s Jill Stein will be on the ballot in all but a handful of them. That’s why we prefer the version of the polls that include their names. It’s up to Clinton and Trump to earn those votes and not up to us to make assumptions about how those voters will behave.

So to summarize:
FiveThirtyEight’s models are faster to incorporate new data and identify trends than most others. For the time being, this helps Trump, since he’s been gaining in the polls.
FiveThirtyEight’s models account for more uncertainty than most others. For the time being, this helps Trump, since he’s the underdog — although it potentially also means we give Clinton a better chance of a landslide than other models do.
FiveThirtyEight’s models use the version of the polls that include third-party candidates. For the time being, this helps Trump, since he’s losing less to third-party candidates than Clinton is.

None of these will necessarily help Trump permanently, however. It hasn’t always been the case that third-party candidates so disproportionately hurt Clinton, for instance. And if Clinton gains following the debates, FiveThirtyEight’s models will probably be among the quicker ones to detect it.

For now, however, the polls show a very close race. Clinton leads in the majority of national polls, but not by much, and there are several that have Trump ahead. Likewise, she leads in the narrow majority of swing state polls, but there are many Trump leads in the swing state polls as well, and Clinton does not have clear leads in enough states to win the Electoral College. Therefore, the race is close. This ought to be clear whether you’re looking at relatively simple averages like those at RealClearPolitics or considering more complex methods like FiveThirtyEight’s.

No hay comentarios.:

Publicar un comentario