Nate has his reasons for doing things as he does. He cannot come up with a better argument for dismissing some polls out of hand, than he can for keeping all polls in. You don’t throw out outliers just because they’re outliers. You just hope and expect that the bulk of the data will be sufficiently solid that the outliers won’t push your conclusion too far from the truth.

There’s also a case to be made for choosing either a certain set of high-quality pollsters, or a minimum quality standard for inclusion, PROVIDED you set those criteria beforehand and not on the fly because they produce the result you want.

Everyone who does this has a different model, but they’re all just models. None of them is reality. The designer of the model studies the variables and how they’ve looked in previous elections, and chooses the ones they deem important, and weights them according to their understanding of previous data. Every model designer thinks his is best, and they’re all logically defensible. But in the end, again, they’re all just models, and none of them is reality.

Nate Silver and Sam Wang ended up with pretty much the same accuracy in 2012, and both see Clinton winning. What differs is the uncertainty. And because there’s only one test of reality every 4 years, there’s no way to determine whose level of uncertainty is correct.

It’s also worth remembering that although we treat a 95% confidence interval as if every outcome within it is equally probable, that’s not really how it works. It’s a distribution.