2016 Postmortem
Related: About this forumA C+ poll of 500 LV out of Utah just moved 538's presidential forecast 1.3% all by itself
This is why I am losing respect for 538. 1 poll showing T up 5 in a very red state should have next to no effect on the forecast. Instead it dropped Clinton's chances 1.3%. How is that based on anything real? If anything, the fact that Clinton is only 5 points down in Utah should be a plus for her chances.
getagrip_already
(14,647 posts)How do you know the single new poll is what moved it?
538 drops old polls in their weighting factor. that is automatic, and percentages sometimes change even with no new polls. All that happened was a good poll got old and its weighting factor fell, so it counted less.
With tightening polls, those older polls are the better ones, so as they fall in weight the perceentages change downward.
It's too complicated. I've found that sometimes, the more data you try to stuff into a model, the worse it does.
Ace Rothstein
(3,144 posts)vdogg
(1,384 posts)One showed CO tied with Trump winning 18-35. Another had him with 22% black support and 47% Hispanic. These polls are put out to fuck with the RCP averages and 538 to give the illusion of a Trump surge. Ignore them.
bigdarryl
(13,190 posts)Daily Kos radio this morning Amundo ,Greg Dorkin and others were saying how ridiculous Nate Silver was by putting all.these partisan polling companies in his data that changes his percentages daily
budkin
(6,699 posts)Positive Trump polls are being given more weight right now because "unpredictability."
http://projects.fivethirtyeight.com/2016-election-forecast/updates/
Dem2
(8,166 posts)I had a feeling he would turn the gears behind the scene there to make sure that it didn't look like it was bouncing back.
WillyBrandt
(3,892 posts)My best guess is that his model tries to optimize and track a number of different factors:
(a) how is any given poll or state TRENDING - versus its absolute level
(b) how might a change in state A be revelatory about state B with similar demography (KY and TV, for example)
(c) how much uncertainty is implied by the polls (e.g. by # of polled who don't answer)
Individually each of these sort of seem reasonable, but I think his aggregate model has just gotten kind of weird and spits out illogical results by a sort of statistical rules-collision. So you get that crazy Utah change you saw.
I think Sam Wang's prediction is a bit too rosy, and NYT Upshot seems most reasonable to me. But Sam Wang is an exceedingly good explainer and very smart.
Dem2
(8,166 posts)His program has gone rogue!
uponit7771
(90,304 posts)Dem2
(8,166 posts)Persondem
(1,936 posts)And whoa!! A +18 poll for VA dropped Clinton's chances!?!?!? WTF Nate!?!?
Dem2
(8,166 posts)This is the 3rd time a Utah poll with the expected result has hurt Hillary - it makes no sense unless the program can't deal with the missing vote in Utah and assumes there's a very high undecided vote there or something? He needs to fix this, he's looking like a fool lately.
Buckeye_Democrat
(14,852 posts)It doesn't help, Nate! It's more discouraging, if anything!
I don't really believe that.
I think the problem is that he keeps including too many CRAP polls -- e.g., Remington -- assuming that the bias will all "balance out" in the end.
0rganism
(23,931 posts)for some time, Utah was in question for Mr. Trump, and not winning Utah would drop his overall chance to win considerably (probably 5%+, since he needs every formerly cherry-red state he can get to even have a chance of victory)
so today a poll comes out that shows he's very likely to win Utah, which re-opens a whole bunch of possible combinations to reach 270, even if its effect is discounted to reflect the unproven status of the pollster. this, in turn, improves his chances across the board.
i think it's legit in the context of Nate's model, and i'd still rather be on this side of the odds.
Dem2
(8,166 posts)Trumps had it in the bag for a long time.
0rganism
(23,931 posts)yeah i guess that idea's dead