I'm not really invested in defending Nate Silver or in the results of the presidential election, I just think statistics are cool.
Being a liberal doesn't make you bad at stats. The degree to which 538's models reflect personal value judgments is almost certainly minimized, so dismissing them out of hand because they come from Nate Silver and use cutesy animal drawings doesn't really "own the libs."
"But 2016" they gave Trump more of a chance than anyone else did, and besides they've since updated their models to avoid those specific problems.
And, I'm sorry, even a really really shitty statistical model is going to be a better election predictor than your opinion on what the "mood on the street" is or whatever.
Of course, Trump's voter suppression campaign is a huge wild card in all of this.
Yes. When you dive deep you notice wild inconsistencies with the way the statistics are done and tuned. They also generalize over data sets which don't deliver the necessary information and both fail to account for past elections and for systematic reasoning about future elections.
My biggest pet peeve is that often information about spatial heterogeneous aspects gets marginalized and not explanatory homogeneous data is used in its stead. Though what at least 538 did correct is to look at individual states instead of just nation wide trends (which is pretty naive).
My understanding is that their model uses states polling. Their national estimates are not based only on national polls
I've never noticed wild inconsistencies in 538 results but admittedly I've only ever done some loose verifying of their numbers in my head (being math brained - assuming you're the same based on this comment). Gimme some ideas where to look? Im genuinely interested to see.