Nugget from 538: 99.6% Chance the Map Will be Different than 2012

[100% pro-life no matter what]

They give both candidates roughly an 80% chance of flipping a state

[This account no longer in use.]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

Exactly. This is just like how people tried to defend their Michigan prediction by saying "They said >99% Clinton, not 100 percent!"

[muon2]

An identical map requires vote percentages within about 0.5% of the previous total, subgroup voting behavior (turnout and preference) within a percent or two, and sufficiently consistent state demographics. The combination is extremely unlikely, and has no precedent for happening. Demographics will shift from 2012 and the map will change. If nothing else I'll bet the vote percentages shift by more than 0.5%.

[Mallow]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

Glad to know statistics are meaningless to you.

[Ebsy]

Does that mean in 99.6% of their simulations that map was different than 2012?

[Xing]

Keep in mind the the chances of "peripheral" battlegrounds like Minnesota, Michigan, Georgia, Indiana, and Missouri flipping are much higher now than they would be (assuming the same polling numbers) in October. If it were October, the odds of the map being different would almost certainly be lower (given the same polling numbers), since only a handful of states would have a greater than 10% chance of flipping.

[Alcon]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

Exactly. This is just like how people tried to defend their Michigan prediction by saying "They said >99% Clinton, not 100 percent!"

I'm not saying it's unreasonable to suggest they underestimated the import of "unknown unknowns"...but you seem to be mocking the idea that something of 1% probability would ever happen?  Or what?

[Mr. Morden]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

So by your logic, probabilities mean nothing unless they're either 100% or 0%?

[Iosif]

We've been graced with mathematical geniuses on this thread.

[This account no longer in use.]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

Exactly. This is just like how people tried to defend their Michigan prediction by saying "They said >99% Clinton, not 100 percent!"

I'm not saying it's unreasonable to suggest they underestimated the import of "unknown unknowns"...but you seem to be mocking the idea that something of 1% probability would ever happen?  Or what?

I'm not mocking that idea. I'm talking about the people who tried to defend 538 by claiming that we were in the <1 percent, despite the fact that all of 538's predictions would be correct by that logic.

[emailking]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

Exactly. This is just like how people tried to defend their Michigan prediction by saying "They said >99% Clinton, not 100 percent!"

I'm not saying it's unreasonable to suggest they underestimated the import of "unknown unknowns"...but you seem to be mocking the idea that something of 1% probability would ever happen?  Or what?

I'm not mocking that idea. I'm talking about the people who tried to defend 538 by claiming that we were in the <1 percent, despite the fact that all of 538's predictions would be correct by that logic.

Of course they would be. You're trying to ascribe importance to the numbers that isn't there. The probabilities aren't calls. They're just probabilities.

I think those of us who like the site are more impressed by the rigorous technical analysis  and model building rather than its past success (or lack thereof) in successfully calling races (by illogically assigning prob > 50% as a call for the winner).

[emailking]

Does that mean in 99.6% of their simulations that map was different than 2012?

Yes. Although it appears to be 99.8% now in the most recent sim.

[Hatman 🍁]

And yet this is the aggregate prediction here:

[Beefalow and the Consumer]

Couple of things:

1. Consider the following outcomes:
2012, 2012+NC, 2012+NC+AZ, 2012+AZ, 2012+NC+GA, 2012+NC+AZ+GA, 2012-OH, 2012+NC-OH, 2012+NC-OH-CO, 2012-OH, 2012-OH-CO, 2012-OH-PA, 2012-OH-CO-IA, 2012-CO, 2012-CO-IA, 2012-IA, 2012-OH-VA, 2012-OH-VA-CO, 2012-OH-VA-CO-OH, 2012-OH-PA-CO, 2012-OH-PA-VA, 2012-OH-PA-VA-IA, 2012-OH-PA-VA-IA-CO, 2012-FL, 2012-FL-OH, 2012-OH-FL, 2012-OH-CO-FL, 2012-OH-PA-FL, 2012-OH-CO-IA-FL, 2012-CO-FL, 2012-CO-IA-FL, 2012-IA-FL, 2012-OH-VA-FL, 2012-OH-VA-CO-FL, 2012-OH-VA-CO-OH-FL, 2012-OH-PA-CO-FL, 2012-OH-PA-VA-FL, 2012-OH-PA-VA-IA-FL, 2012-OH-PA-VA-IA-CO-FL, 2012-FL, 2012-FL-OH.  (+17 more for all to take NV away along with CO).

There's 57 not-unreasonable outcomes.   And I left out some less likely ones, like Trump winning PA without OH, NV without CO, and VA without NC.  Some more likely, some less likely, but nailing exactly one is a low-probability even to begin with. 

Right now, Trump's odds of specifically picking up OH, FL, VA, or CO is relatively low, but the aggregate odds of picking up any is fairly good.  If he has only a 25% chance on each one, he has roughly a 68% chance to pick up any of them (provided they are independent events, which they aren't, but at the same time he has a better than 25% chance in all of those states).  Likewise, Clinton has a good chance of picking up NC, and lower but not insignificant chances of picking up AZ, MO, or GA.

Another thing to consider is that when two states have a roughly 50-50 probability, election outcomes come down to things like weather events driving down turnout.  WI in 2000 and 2004 was basically a tie.  Run those elections 10,000 times, and they'll split evenly, depending on when a butterfly flaps its wings in Venezuela in September.

Finally, the 0.04% probability is considering the current state of the polling, which has Clinton ahead almost 5 points nationally, which is well off the 3.86% margin in 2012.  That drives down the probability of that specific map considerably.

[Hatman 🍁]

Also relevant, the current 538 projection is just one state off from the 2012 map:

[Sorenroy]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

Exactly. This is just like how people tried to defend their Michigan prediction by saying "They said >99% Clinton, not 100 percent!"

But that's a dumb comparison.

For one, general election polls are much easier to follow and estimate. Like people on Atlas like to say, 45% of people will always vote for one side no matter what. Even if it was Jesus vs Hitler, the outcome would be 55-45 with a fairly recognizable map (I'm trying to make a point, not saying 45% of people are actually Nazi's). If you compair that to a primary where in any given year there are between 2 and 10 or more major candidates running, you get very different results.

Secondly, their forcast was a model. Weather you like Nate Silver's #analysis or not (Trump will never win the Republican primary), his forcast models are solely in the realm of numbers and statistics. Michigan was the largest upset between polling and the actual result in more than 50 years. In fact, the most accurate poll out of the state was done by an unknown polling entity (Michigan State) that did one poll over the entire cycle as a side project to something else, and did that over the course of more than a month with only 262 likely voters in that period of time (and even then they were more than six points off).

So he is not leaving a sliver of a chance that the other side will win because he is basing everything off his own beliefs and needs some way to get away from the incorrect prediction once it happens, he is leaving that sliver because, as someone who works in statistics, he knows that, no matter what, there is a chance that the unbelievable will occur.

[Mallow]

Also relevant, the current 538 projection is just one state off from the 2012 map:

And their polls-plus model is exactly the same. But that doesn't change the idea that there is only a 0.4% chance of it ACTUALLY being the same. Take the probabilities of the current projection being right in each state and multiply them, and the result is a very small number.

[Slander and/or Libel]

Also relevant, the current 538 projection is just one state off from the 2012 map:

And their polls-plus model is exactly the same. But that doesn't change the idea that there is only a 0.4% chance of it ACTUALLY being the same. Take the probabilities of the current projection being right in each state and multiply them, and the result is a very small number.

This assumes those probabilities are independent, which is completely false.

[Mallow]

Also relevant, the current 538 projection is just one state off from the 2012 map:

And their polls-plus model is exactly the same. But that doesn't change the idea that there is only a 0.4% chance of it ACTUALLY being the same. Take the probabilities of the current projection being right in each state and multiply them, and the result is a very small number.

This assumes those probabilities are independent, which is completely false.

The 0.4% certainly does NOT assume that (the 538 simulations take into account the covariance between states). My quick-and-dirty "multiply the values" was, and I should have been more clear about that. The point is (as has been noted several times earlier in this thread by other posters), while there are certainly correlations between states such as OH and PA, they're nowhere near 100%, even for the most highly-correlated swing-state pairs. Then you consider pairs like FL and CO, or PA and NV, and there's enough variability such that getting the same map is STILL extremely unlikely.

[Slander and/or Libel]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

So by your logic, probabilities mean nothing unless they're either 100% or 0%?

Perhaps not nothing, but I find little utility in this type of statistic.

Right, that's what I've been struggling with. What is this supposed to mean? If it's just a snapshot of where the race stands, cool. But it purports to be more. It purports to be a prediction, and it purports to tell us that if we ran this election 5 times, then on average, Trump would win. Is that true? What does it mean for it to be true or not? What value does it have in a predictive sense? If the map is indeed the same as 2012, is there any way to go back and say, "Yeah, but in mid-July, it really was only a 0.4% chance that it would be the same"?

In short, this is based on analysis of a whole lot of data, but is there any reason to consume it uncritically without acknowledging the mountain of assumptions that must go into such an analysis, and potentially fatally compromise it?

[palandio]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

So by your logic, probabilities mean nothing unless they're either 100% or 0%?

Perhaps not nothing, but I find little utility in this type of statistic.

Right, that's what I've been struggling with. What is this supposed to mean? If it's just a snapshot of where the race stands, cool. But it purports to be more. It purports to be a prediction, and it purports to tell us that if we ran this election 5 times, then on average, Trump would win. Is that true?

That is in reality a deeply philosophical question about the nature of probability. The frequentist concept of probability is more or less "If we would run the election 1000 times, Trump would win 200 times". The concept that I personally prefer in many situations is the concept of probability introduced by the Bayesian school and De Finetti: Probability is subjective and encodes the lack of information that you have. Despite the lack of information on an experiment, you can still bet on its outcome. That is if Nate Silver was a bookmaker and he was forced to offer you fair odds on the outcome of the election, he would say:
"If you bet 1.00$ on Trump and he wins, I'll give you 5.00$. If you bet 1.00$ on Clinton and she wins, I'll give you 1.25$."
If he is right and Trump has a 20% chance of winning, and you bet 1.00$ on Trump, then his expected revenue is 0.8*(1.00$-0.00$)+0.2*(1.00$-5.00$)=0.00$. Similarly when you bet on Clinton his expected revenue would be 0.00$.
If he would offer different odds, he would give you the chance to make profit (on average).

You must be logged in to read this quote.

Well, when talking about the 0.4% we are approaching another phenomenon of probability. How likely is it that you throw your dart into the center of the dartboard? Not so likely after all. But at the same time the center of the dartboard is more likely to be hit than any other area of half a square-inch in its surroundings, at least if you don't systematically deviate in one direction. The 2012 outcome can be thought of as the center of the dartboard, and the 2016 outcome will likely be somewhere near the center of the dartboard, but probably not in it exactly.
At the same when you will have thrown your dart at the dartboard, every possible exact position will have been quite unlikely before you threw the dart.

You must be logged in to read this quote.

Of course, that's the challenge behind it.

[pbrower2a]

Similar elections are rare, and they either involve 49-state blowouts (trivial differences that say more about the losing nominee than about the winner) or those involving an incumbent seeking re-election (2008/2012 -- two states and NE-02; 2000/2004, three states; 1992/1996, 1952/1956, or 1940/1944 four states). Not since the 1904/1908 pair in which three states shifted and a new one (Oklahoma) appeared have there been any two consecutive elections with similar electoral maps.

There is no such thing as a normal Presidential election. 2012 had the oddity of the winner getting a median result -- but the only election since 1900 in which the winner got between 57% and 65% of the vote.

Somehow, Hillary Clinton has no resemblance to William Howard Taft. Comparing BarackObama to Theodore Roosevelt makes little sense, too. 

[Mr. Morden]

This means almost nothing; even if the map is same, they can just say "we're in the 0.4%".

So by your logic, probabilities mean nothing unless they're either 100% or 0%?

Perhaps not nothing, but I find little utility in this type of statistic.

Right, that's what I've been struggling with. What is this supposed to mean? If it's just a snapshot of where the race stands, cool. But it purports to be more. It purports to be a prediction, and it purports to tell us that if we ran this election 5 times, then on average, Trump would win. Is that true? What does it mean for it to be true or not? What value does it have in a predictive sense? If the map is indeed the same as 2012, is there any way to go back and say, "Yeah, but in mid-July, it really was only a 0.4% chance that it would be the same"?

Let’s say that your favorite football team is down by 14 points with 5 minutes left in the game.  So they’re underdogs at this point, but there’s still *some* chance that they’ll come back to win.  How big a chance?  Well, you can look back at previous games played by all teams over the past X years for situations that match this one, and calculate what percentage of the time the underdog goes on to win.  You can factor in additional variables, like what is their field position right now, at the 5 minute mark?  But still, you can use whatever variables you think are relevant to work out “In what %age of games with situations like this does the underdog win?”

That seems like a relatively straightforward thing to interpret.  What 538 does is not much different.  “In X% of cases where Candidate Y is trailing by 5 points, he comes back to win.”  It gets hairier with presidential elections compared to Senate or Governor’s races, because you have separate races in different states that are all correlated as opposed to a single statewide race, but the concept still holds.