Folks, time to stop attacking Nate Silver. The criticism I have seen, elsewhere and here, has the same intellectual depth exhibited by global warming denials. Nate does not manipulate statistics to mortify us. He has a results-validated model that he uses to make election predictions. Once the model is established, you feed it poll numbers and the model gives you the probability that Benito or Hillary may win the election. And it is methodologically incorrect and intellectually dishonest to change the model halfway through the election cycle because you don't like what the model is telling. Like any serious scientist, I am sure that Nate looks at his model after every election and considers changes that might have improved his predictions while still reproducing the predictive successes of elections past. This is how you do science and this is how you do serious statistical work.
Nate has explained the reasons why his probabilities for Hillary 2016 are lower than those for Obama 2012 even though Hillary's predicted national vote share is higher: a more "inefficient" geographical distribution of voters that increases Hillary's numbers in states that she is going to win or lose anyway, but decreases her numbers in Midwest swing states. And his model crucially incorporates correlations. If Hillary were predicted to win in every states by a 1% margin and the state polls had a 5%, error, the electoral college probabilities would be very different depending on correlation. If each state is completely independent of the other, the probability that she would lose the electoral college in spite of her 1% lead in each state would be almost zero. But if all states move in concert, then she would be close to 50/50. Empirically, there is a strong correlation, not as high as making all states move in concert, but certainly high enough to make it necessary to include correlations. Nate's model incorporates state-state correlation matrices that have been determined from past election results. He is not faking or skewing data.
Still, I believe that Nate's predictions are too pessimistic as of today because his model does not include things that we know by now, such as early voting turnout rates by party, demographic composition of early voting, etc. The case of Nevada is just one example. The models can be refined to incorporate this information. See for example the Upshot's study of North Carolina. This is perhaps the wave of the future. I personally think that when you incorporate all that info you get not only a Hillary win but perhaps even a landslide. But we cannot ask Nate's model to tell us something that wasn't built into the model to begin with. And if the Hillary landslide does materialize, it will not automatically mean that Nate's model was incorrect. More likely, it would mean that the polls fed into the model were bad. And this would not be surprising. This is a very special election. Watch Benito's closing ad and see if you would expect that from a "normal" Republican. New coalitions are being made. It is very, very difficult for pollsters.
