Let's just call this a series. In what I am now retconning to be Part I in the "Nate Silver is full of shit" series, I pointed out how absurd it is that 538 wildly overestimates Trump's chances of victory. Currently, Silver's "polls only" forecast gives Trump nearly a 15% chance of winning, and in Part I, I ridiculed that by pointing out that it is the equivalent of saying that we live in a quantum mechanical multiverse in which one out of seven universes branching out from this moment have President Trumps, even though plausible scenarios leading to his victory are vanishingly few. In what I'll now call Part II, I pointed out that if you overestimate his probability of victory by 70%, then even if the percentage point difference is small, then your betting strategy will cost you a lot of money. In Part III, I focused on the electoral college, and the statistical problems of focusing on state-level versus national level effects.
Here, let's really explore than in an historical context. Remember, Hillary Clinton is ahead. Her lead in the RealClearPolitics average is 5.5 points as of this morning. There are two ways that Trump can win: take a national lead such that the popular vote and electoral college come into alignment, or create a popular vote-electoral college split. The former is becoming absurdly unlikely. Not impossible, but absurdly unlikely. It would take one of two things: either all of the polls would have to be wrong, in which case Silver's simulation method is bullshit anyway because of the garbage in-garbage-out problem, or an intervening event would have to occur. Silver isn't really addressing either. No, Silver's method is really all about calculating a Trump victory probability based on the potential for a popular vote-electoral college split in an elaborate combinatorics game. And 15% should be a great, big, red warning flag that he's doing something wrong.
How common are popular vote-electoral college splits? We have had four. Sort of. Three real ones, one fake one. There's 1824, but that isn't informative at all. It was a multicandidate race, so nobody got a majority in the electoral college, and it went to the House of Representatives, according to the Constitution. There's 2000, but as I pointed out yesterday, that was a) an absurdly close election, which this is not, and b) thrown off by the butterfly ballot in Florida.
Gore won the nationwide popular vote, but sort of lost Florida by 537 votes. However, Palm Beach County in Florida used a weirdly structured ballot. Candidate names ran down the left and right hand sides of the ballots, with holes down the middle. Voters had to punch out a hole that lined up with the name of the candidate for whom they wanted to vote. The second name on the left was Gore's name, but the second hole on the ballot wasn't for Gore. It was for Pat Buchanan, whose name was on the right. It was a grotesquely stupid design that cost Gore the state of Florida, and the Presidency because a lot of people voted for Buchanan thinking they were voting for Gore. Also, the ballot was designed by a Democrat named Theresa LePore.
So, the 2000 election gets a big, fat asterisk next to it as a popular vote-electoral college split. And it was absurdly close anyway. Gore and Bush were separated by half a point at the national level. This one is not close. Clinton is ahead by 5.5 points in the RCP average, and that's factoring in the crappy LA Times poll, which is being thrown off by one African-American teenager in Illinois who says he's voting for Trump.
So, how far back do we have to go in history for another popular vote/electoral college split? 1888, with another in 1876, and those are the only other two. And both were close elections.
There are national forces at work here, and state-level simulations have a really hard time capturing them. That's where Silver goes wrong with his overly optimistic projections for Trump. He is too focused on the electoral college, ignoring how absurdly rare true electoral college-popular vote splits are. Even if you count 2000, we've had four in our entire history. So, focus on the national level, and nothing there gives Trump much hope.
PredictWise currently puts Trump's chances at 9%. That still seems high to me, but more realistic than Silver.
What will it take for me to change my mind? Changes in the national numbers. The latest ABC News poll had Clinton's lead down to only 4 points. That pulls her average down. How much do I make of that? At this point, not much, but I factor it in, via the RCP average. What I don't do is play Silver's combinatorics game until the national numbers get close enough for a popular vote-electoral college split to be plausible. We aren't there and are unlikely to get there.