NYC readers, tonight I’ll be at the American Museum of Natural History, I’ll have a discussion with The Upshot’s Josh Katz. Our topic is The Science of Predicting an Election, with an added special focus on his approach to visual data representations using maps. It’s part of the SciCafe series and is free of charge. Doors open at 6:30pm, and we start at 7:00pm. See you under the dinosaur!
Update: Josh Katz is sick! So it’s just me. Sorry, Upshot fans!
Update2: Great crowd, full house of nearly 600 people gathered under the titanosaur. I talked about polls, polarization, and the uncanny ability of Google Correlate to predict Trump support on a state-by-state basis (read more about that here). Here are the slides, which will be better with narration, once it’s available. Thanks to AMNH for a great event!
Sounds fascinating. Will the discussion be video streamed for folks unable to attend in person?
It will be archived…I don’t know how long the delay will be before it’s available.
One of the reasons I sometimes miss NYC is getting to attend such events. The other events in the SciCafe series also look great.
Are there any statistical concepts in particular that us laymen should brush up on prior to the talk?
Cool! That’s 7 blocks from my apartment… I might be able to pull this off :)
live stream available?
Sadly no! Only available some weeks from now…long after the election!
The meta-margin is (as the Dylan songbook says) sinking like a stone. I don’t doubt the change. But, because the model runs highest of all five other models that The Upshot reports, I do wonder whether it’s getting the level right. By Election Eve, the margin may very well be half (Clinton +1.5%) what it is now. Would that be enough?
The meta-margin is the nationwide shift it would take to make the election a coin flip. If it is Clinton +0.1% on election eve and turns out to be accurate then Hillary will win. The higher it is, the more room there is for polling errors in Trump’s favor, but any positive number reflects a prediction that Clinton is currently ahead.
So if you just project out the current trend–which I would advise against–you’d land somewhere around a Meta-Margin of 2 and a median EV estimator of 300. I believe the final 95% confidence interval is going to be around +/- 10 EVs, so . . . that would seem to be enough.
I think Sam said he would eat a bug if Hillary is under 298. It looks like it might be close, unfortunately.
_Is_ it sinking like a stone?
If I recall correctly (low probability ;-) ), it was 3.2% when I went to bed last night, 3.0% this morning, and now, at 5pm it’s 3.1%.
We can actually look at recent numbers and try to anticipate whether the MM will shrink. If I’m correct, today is the first day we’re seeing polling observations (those from Nov. 1) that will be included in the final prediction (November 1-7 being the 7 day trailing average before election day).
Looking only at polls ending on Nov. 1, and only including swing states with at least 2 such polls, we find the following medians:
FL: D+2 , n=3
OH: R+4, n=2
NC: D+5, n=2
AZ: R+1, n=3
NV: R+0.5, n=4
PA: D+4, n=5
CO: D+5, n=2
We can probably expect 10-15 polls in each state in final week, so it’s still early, but safe to say there’s no evidence of a tightening that has much likelihood of flipping the Presidential race.
If anything, MM could tick upwards, with Clinton at least +4 in all of her sufficient states based on early November numbers.
@Slartibartfast, Nicely said.
Hmm: PEC’s Trump +2 map outright changes 3 states’ hue –Nevada and Iowa at 6 points each I believe, with Florida holding the EV’s lion’s share. In the Sunshine State this week, 28% of surveyed GOP-registered mail-in voters report voting for Hillary. So if state polls say anything, looks like Sam won’t be eating a bug.
@Koos van Blerk: You wrote,”I think Sam said he would eat a bug if Hillary is under 298. It looks like it might be close, unfortunately.”
Close does not a Bug Dinner make. Let’s be precise. This past Oct 19, PEC’s commenter “A” wrote here,
“Sam Wang via Twitter: It is totally over. If Trump wins more than 240 electoral votes, I will eat a bug.”
Trump might say,”Look a Chgina, look at Florida!” I’m looking at Florida, and joining in.
I hereby pledge to match Sam in bug parts per million. Ew.
While who will control the senate after the election is the real unknown, I am still very interested in subtleties of the recent MM drop. For example the change in the MM from it’s +5.1% high mark on 10/20 and now +3.1% in terms of battleground states seems to be more on the margins. From the follow data what do you think about the differences?
The Bayesian and random drift is the same for both 10/20 and today 99% and 97%
The state comparison look like this:
Date…NH…CO…NC…NV…WI…MI…FL…IA…AZ
10/20..+6.5..+11…+2…+2…+7.5..+9..+4..-2…-4
11/2…..+4….+3….+3….T…..+5…+4.5..+2..T…-2
Date…NM…PA….VA….OH
10/20..+11..+6.5..+11….T
11/2….+7…+4.5…+6…-1.5
On election day 2012 the eight battleground states looked like this: NH +3, CO+2, NC -1, NV+2.5, FL tie, IA+2, VA+2 and OH +3. The final MM +2.76.
Wow! I am so there. Maybe I’ll break my arm, get a cast put on, and ask Sam to sign it!
Will Ben Stiller be there? I might skip the World Series if he is. (Just joking.)
Sam, I’ll be coming by tonight. Where exactly is the discussion being held? In the museum or planetarium?
Unfortunately, it looks like Josh Katz won’t be there, per the AMNH SciCafe web page. Presumably someone else will step in? Regardless, a live stream of the event would be great!
Sam, in The Upshot’s large table, why is your model’s forecast so different than that of the other models on average for Maine 2, Nebraska 2, Nevada, and Iowa? Also, I notice that your model gives Clinton fewer EVs than average among the 10 states and 2 CDs that The Upshot calls competitive, and yet the highest probability of a Clinton win. Why is that?
Sam, I am surprised that it isn’t talked about very much that the surge for Trump in many places is really due to soft Johnson voters moving toward the GOP. Since there already was great affinity there between GOP and Libertarian-leaning voters, the reality is that those were always a safety deposit for Trump. The key I guess would be estimating what Johnson’s floor is.
Earlier in the year there were some surveys that seemed to indicate that Johnson’s supporters were closely divided between Trump and Clinton as their second choice. If that still holds (it may not), there may be a limit to how much of Johnson’s support is a bank of potential Trump votes.
If anyone can refer me to published analysis of this matter, I’d appreciate it. If people who answer Libertarian in polls wind up voting Trump — even if the smaller percentage who answer Green wind up voting Clinton — then NH, CO, and NC are the states that Trump would win despite a current Clinton lead or tie in the polls, in which case he’d probably get 277 EVs. (This also assumes, among other things, that the undecideds don’t vote, or vote like the decideds.)
Question for you both. With so much early voting this year, how will exit polls by the media on Nov 8th work? Do you think we will have to wait until the early votes are counted? Or do exit polls by the networks/cable already take this into account somehow? Thanks
Here is a nice article on how that works:
http://www.pewresearch.org/fact-tank/2016/11/02/just-how-does-the-general-election-exit-poll-work-anyway/
Executive summary: they are doing a very large (16000) phone survey of early and absentee voters to complement their in-person survey.
Sam, wish I could be there!
Sam. Any chance that you could make the slides from tonight available? I’m a professional geographer and love this stuff. Thanks.
Good presentation, Sam. It felt as though I was attending a live version/thread discussion of PEC. The crowd was unbelievable, appreciated the string theorist origin/bet story; it was the largest crowd ever for SciCafe, apparently. Though it was unfortunate Josh had the flu, it was, nevertheless, a great evening all around! Thanks!
What’s the string theorist/bet story? My circle of string theory acquaintances overlaps with Sam’s in a couple of instances.
The funny thing is that Sam’s field of research and mine are both about equally far in energy scale from where string theorists like to dwell.
In summer 2004, I got into a lively discussion with Steven Gubser about whether Kerry had a chance at winning. I said that whoever won two out of three states – Ohio, Florida, and Pennsylvania – would be elected. Kerry was ahead in Pennsylvania and the other two were coin tosses, so it seemed to be 3-1 in favor of Kerry. This was true at the time, given polls. He thought about it a moment and agreed. I was quite chuffed. That motivated me to write the first MATLAB script to handle all the close states.
In reference to the relative weight one should place on early votes vs. polling data, Jon Ralston dismissed one particularly bad recent CNN poll for Clinton. I concur with our shared perspective to rely on the median of multiple polls, not single likely outlier polls. But, I’m curious about your take, Dr. Wang, about his bigger point suggesting that it’s more predictive to rely on early voting patterns than polling results close to election day. Ralston tweets, “Final note on CNN, all NV polls from now on: Rely on the early vote numbers, not polls. Real votes tell you more.” I’ve often contended that PEC’s approach is the most parsimonious, but I am intrigued by Ralston’s logic. If the goal of polling is to predict voter intent and by extension, voter ballot behavior, has Ralston indeed identified a more direct measure of voter ballot behavior provided that a state offers early voting and freely shares data (e.g., party affiliation) that’s highly correlated with who voters will vote for in a general election?
See article by Jim Messina in today’s NYT. Very dismissive of public polls, especially national polls.
http://www.nytimes.com/2016/11/03/opinion/campaign-stops/the-election-polls-that-matter.html?ref=opinion
Great Talk Dr Wang!!!!
Cubs win for the first time in 108 years, and Bayesian just clicked over to 100%.
Quite a night!
100%? That seems rounded. How about >99%?
100% sounds about right to me.
Nope, there’s a glitch.
He just posted a note headilned,”Not 100%.”
The banner should read >99%
I earnestly hope that 100% is posted late Tuesday evening….
Well, it would have to be rounding, since you never do quite touch that asymptote, but you have to admit it was fun while it lasted! ;)
visualization of polarization
https://twitter.com/alec_h_tyson/status/793872116735180801
from the 2014 Pew Report
http://www.people-press.org/2014/06/12/political-polarization-in-the-american-public/
For the record: I would far rather see Sam eat a bug because Clinton wins Texas, than because Trump wins the Electoral College!
Many thanks for providing the slides from your talk, and even more thanks for (eventually) adding the actual stream! As someone living on the left coast this may well be the only way I get to see this stuff in action, even though I do come here repeatedly for my fix.
Also, big big mahaloz for putting your code out there for those of us who find pleasure in tinkering with these things! A peek into the clockwork is always a treat.
Sam, can you elaborate on the google search term correlation? Or is there a post on the site I can look at? I find the Degrassi-Trump correlation hilarious.
Google Correlate and Trump: start here.
Thanks! Apologies for missing it and having to ask.
For those who missed the talk and can’t wait for the video, here is my write-up on it. Thanks for the lecture!
https://www.facebook.com/bruce.epstein/posts/10209724927694068