I am trying out a new rounding rule on the Clinton win probability. Now rounding it to nearest 1%, not 5%. Gives you something to look at – along with the big jump in the Meta-Margin.
For the generic Congressional preference, the displayed threshold for flipping the House is now set to Democrats +6% to +8%. That estimate is fairly rough, as detailed below. I am also experimenting with the graph format. It displays every Nth day of the median. Before, N=7; now I am trying N=3.
It’s what passes for drama here!
That is what I really like here, and that is why I respect Dr Wang so much. Boring is good. Just look at Now-Cast in 538 to appreciate good solid poll consolidation.
Wow, that’s quite a jump in the MM!
This is only one poll but it’s bad news for Trump:
http://www.nbcnews.com/politics/2016-election/polls-clinton-running-table-key-battlegrounds-n629136
Technically several polls, but still…really bad for Trump.
What makes me even happier is that it also shows Senator Burr starting to fall behind and more importantly for local politics (here in NC), the governor is down 7. I sure hope it’s a good poll..getting McCrory out of office would be huge for this state.
Biggest news is the shift in the GOP senate firewall. Would like to see the (D) column shift to a mode of 51 at least.
Is there a way to see what the main driver of the big MM jump was? A particular new poll? An old, Trump-favorable one aging out?
The NBC/WSJ/Marist polls that Lee linked to in another message are what caused the MM jump. The one that had the most effect was the move in NC from Clinton +2 to Clinton +6. (The other moves were FL going from Clinton +1 to Clinton +2.5, and VA and CO going to double digits or something like it, from Clinton +5 and Clinton +7, respectively.)
If anything the MM is understating Clinton’s lead at this point, since MN is at Clinton +5 and OR is at Clinton +3, due to a lack of recent polling in both states. If you were to set those two states as solidly Clinton (say, Clinton +9), the MM would probably jump at least another 1%.
And by the way, polls don’t age out on their own. They only fall out of consideration when new polls are released for that state.
Froggy points out a missing-data problem in Minnesota and Oregon. It might be a good time to use Google Correlate to impute the less-polled states – and see where new polls would be most informative. If any readers are interested, I am all ears!
Imputing missing numbers! I’m in! (Email me.)
But I suspect states with low polling are low for a reason I.e. they are not in contention.
Using Upshot to compare against, PEC’s boldest estimates are for Oregon (4 polls) and South Carolina (3). I’d bet they are not as close as the few polls say they are.
Thanks for asking this because it is an entre to my follow-up question … I think the Marist polls drove the bump, but if I recall Sam’s explanation about how he does things, he looks at medians, not averages, so (if I understand things correctly) Marist could have been anywhere from a little bit better than other polls all the way to 100 to 0 for HRC and it wouldn’t have mattered. Because medians, not means….. Did I get that correct?
My wish is for +/- uncertainty estimates to be given on all top-line numbers.
A +/1 uncertainty on a probability seems rather redundant, doesn’t it?
See the gray shaded region on the time-series graphs of EV and Senate seats. The Meta-Margin’s error bar is effectively <=0.5%.
Is this perhaps the first big sign that state-level polls are starting to come into broad alignment with national polling?
Also, do we have a good idea at this point of how many voters are still undecided? Early in the summer IIRC the number was fairly high, something like 14% +/- 2. If that number has come down significantly, it’s hard to see how Trump has much room for error at this point.
Up to now, several key states didn’t have enough polls for the PEC model’s averaging scheme to fully reflect the post-DNC period.
The fraction of undecideds in the national polls is all over the place; recent polls have it ranging from 1% to 18%. I think they push leaners to state a choice to different degrees.
I believe the undecided number is pretty low if you look at four-way polls.
Looking at a three-way poll, the sum of undecided + Johnson has stayed constant around 15-16% since May. (On average! The individual poll numbers are all over the place as Matt says.)
http://elections.huffingtonpost.com/pollster/2016-general-election-trump-vs-clinton-vs-johnson
I think that turnout (who is least disaffected?) will matter more than the late-deciders. But haven’t seen data one-way or the other.
How does this recent analysis, including yesterday’s post talking about polling stability, fit into the post from mid-May that talked about s-dev over time?
After Labor Day I will use this year’s MM statistics to set the prior. Note that as the election approaches, the random-drift part of the calculation becomes more dominant and the prior matters less.
I appreciate the greater accuracy.
I also want to un-lurk to say thanks for all the work and explanations. PEC is my go-to place for statistical analysis of the election. Please do keep doing it future elections or at least hand it off to a capable caretaker if you choose to not pursue it in future cycles.
My goodness! Just noticed that Clinton now has enough safe states to win the presidency even if she ultimately loses every state where she currently has less than a 95% chance of winning, including Oregon.
The Republicans really picked the wrong candidate. Looks like the real drama in this election will be control of Congress. When does everyone think the GOP should essentially abandon Trump ala ’96 and try to save Congress?
I think politicians are generally wary of rebuking Trump because they don’t want to angry his strongest supporters. But make no mistake, the GOP is essentially backing away from Trump right now, but slowly and quietly.
Months ago the Koch brothers and other well-heeled conservative GOP backers decided to focus their spending on congressional races, and the results of that can be seen today on the airwaves.
The 2014 red dot looks like a clear outlier with respect to the gray polygon. But.. imagine if the polarity were reversed. Then instead of a 47%D vote with 43% House, we’d have a 53%R with a 57% House. That dot would fall neatly into the gray box. (For 2012, the same but a closer to edge.)
My small point is that the graph is subtly misleading. If the graph doubly plotted the points (normal and polarity reversed), it might be better. (If it’s true that the majority party has usually gotten more than its exact share of the house, then it’d be pretty similar to results for the electoral college.)
My larger point is that there’s large variance (height) in the gray box. Whatever the bar (7% is fine) to deliver 218 seats on average, then a spread of +/- 3% would be more consistent with the gray box.
Sam,
I see that NJ is “pinned” in the jerseyvotes list (for obvious reason). Otherwise it looks like simply the top-N according to “Power”.
Could you also either “pin” or add a second threshold, to make it pull in a few states that may be lower “Power” but are near the current balance point? e.g. as I write this GA is present, but AZ and MO are not. Maybe “anything within the Margin range R+2 .. D+2”; or specifically pin GA, AZ and maybe a couple of others.
Alternatively, can you offer a “full” jerseyvotes page with all 51 states+DC sorted by “Power”? (I realize some would be conjectural at best, with no state polls at all, and might have to be omitted.)
Drew Linzer indicates that he will be back with forecasting analysis for this cycle, in conjunction with Daily Kos. Sometime in the next few weeks. A bit late in the season, but I’ll still be interested to see what he has to offer.
Dr. Wang is the last pollster remaining that hasnt undergone regulatory capture by a cartel.
Thats why he is so good, and so essential.
I checked out votamatic. It looks like he’s a bit more bearish on Hillary than most other forecasters but he still gives her about a 75% chance as of today. Does his use of economic data to make his forecast even help him at this point, though? Seems like he’s pretty late to the game this time around.
Dr Wang has revealed himself to be in the fuzzy, snuggly pocket of Big Sweater.
https://election.princeton.edu/2016/07/29/a-local-experiment-in-crowd-wisdom/
I recall a scatter plot by Sam (ages ago) showing that as the margin of victory gets large, the poll errors get big. Linear around 50-50, they diverge once the winner is not in doubt.
While 6% – 8% margins might get measured ok, if it gets into double digits the polls might undershoot.
Errors get larger & consistently undershooting?
Both seem at odds with how sampling should work. (If 95% is truth, it seems unlikely to sample 100 people and be off by 10%.)
I’d like to see that plot!
By error I meant prediction minus result. The stat uncertainty they report sqrt(N) as ever.
I am cracking up at this “Arizona Freedom Alliance” poll showing trump up 60+ to 19 with 50k respondents. Then they say to share it on Facebook to get the word out. And the fringe folks are lapping it up.
Denial and lunacy
ah come on, Sam–that new rounding formula is just click-baiting! We know you’re in talks with ESPN, for as soon as they drop Nate 🙂
Snark, people. I go to this site daily, regardless. I think it’s cool that NYT Upshot now includes PEC, and I’ve no doubt that drives considerable traffic. However, Sam does need to launch an ‘unskewed’ model where e.g. Trump is winning PA unless voter fraud.
Would Sam or anyone care to comment on the Alan Abramowitz Time For Change Model, either the Methodology or anything else about it?
Leave a Reply