In a two hour period on Friday, seven separate GOP PR and marketing firms issued an avalanche of polls that caused a roughly eight point swing in the probability of a Donald Trump victory on election day, per the 538 model.
These polls came form sources like:
Trafalgar Group – a GOP political consulting and direct mail group recently created to influence public opinion for conservative campaigns Gravis Marketing– a conservative PR and robocall shop and part of the Breitbart News Partner Network Harper Polling– NRCC political strategist and polling director Brock McCleary’s IVR firm for “testing political messaging” Strategic National– noted vote fraudster John Patrick Yob’s fly-by-night GOP campaign consulting firm Data Orbital– “grassroots engagement” shop run by Political Director and State Victory Director at the Arizona Republican PartyThese aren’t neutral players. Many know each other, and are deeply integrated into the Breitbart network of websites that constitute the alt-right media bubble, and can no longer be adequately differentiated from the Trump campaign itself. They market themselves as campaign consultancies, and their very objective is to manipulate public opinion and win campaigns.
Today’s two hour data dump prompted this salvo from Obama administration speechwriter Jon Favreau:
x@ForecasterEnten are you guys concerned about how many garbage vs high quality polls are getting thrown in the model this week?
— Jon Favreau (@jonfavs) November 4, 2016Harry Enten shrugged off the issue with a few tossed-off jokes, but it’s really own his brand and career at stake. At this late hour of an important national election, he’s playing willing dupe to political professionals, and pointing indifferently at the star belly sneetch machine of Nate’s model with a me-not-me grin. (Not that young pundits with a little jive do anything but fail upward in today’s influencer-centric media hellscape.)
But I don’t want to simply throw shade on all the young dudes just for trying.
For those here who have complained that Nate Silver has sold out, or that his models are increasingly over-engineered, I think those arguments are actually a tad oversold. Yes, I think Nate’s model may have gotten a bit too clever with its state-linking and cohort manipulations. But the true vulnerability of his model this cycle is in the data. 538 seems to have virtually no quality control, and no checks for data integrity or real verifiability. Inclusivity may be a fine principle, but not when it comes to unverifiable data. Any political consultancy with a barebones website can issue a PDF claiming to run an IVR poll (landline and English only, of course), based on voter registration files and reweighted to a proprietary demographic screen, and Enten & Silver will just ingest it and will-to-believe all is good.
I mean, dig it: That one John Yob poll – a proven liar they tried to hound out of the Virgin Islands – moved the 538 polling consensus by a full point. A full point!
That’s an embarrassment. That it even was passed through the model is a joke.
No one should be fooled: these dumps are strategic. They’re meant to manipulate the very public polling aggregates that drive the media talk tracks now.
I work in the field of audience modeling, targeting and measurement, for the past 15 years for two of the world’s largest tech firms. I’ve managed the development of behavioral modeling and ad targeting systems with billions of dollars running through them, worked with many agencies and marketing firms, and been privy to the conversations of many digital strategists, including many political pollsters and strategists on both sides of the aisle. In the course of my day job, I’ve been in the room with both HRC’s and Trump’s digital teams, and met plenty of adjacent players.
Without naming any names or going into details I simply cannot, I have heard earlier this year digital political strategists openly discuss what is clearly happening now: the potential efficacy of strategic poll dumps to manipulate the aggregates in hopes of juicing voter mobilization (and demobilizing the opposition). Everyone knows the prominence of RCP and 538, and knows that these site drive the media talk tracks.
I have no doubt that some of the data we are seeing is manipulated. Some of it is nearly fictional. The fact that it drops at just the right time, and often together, is not merely coincidental. But the pundit-aggregators will tend to hoover it all up nonetheless, whether the source is accredited, has a track record, or no.
This is not indefinitely repeatable. And it’s beginning to compromise the modeling itself, when fudge factors for “uncertainty” need to be baked in to compensate for the possibility that “all the polls might be wrong”. Not all of the polls are wrong – but there is a disinformation campaign being waged at multiple levels, and tactical faux polling is just the latest expression of it.
There’s a far better way to screen for potential pollster fallibility: do the effing hard work, and sniff out the shite. Screen non-accredited pollsters or those with non-existent or dubious track records. Insist upon call log verifiability and rational subtabs. Don’t ingest data from companies whose express mission is strategic public relations or political advocacy. Again: these are not neutral players, and accepting then as such is just stacking the deck.
There’s going to be a reckoning sooner rather than later.
Friday, Nov 4, 2016 · 11:16:38 PM +00:00 · MaxwellNate Silver just issued some analysis on the volatility of Clinton’s state polling. The jaw-dropping statistic in this one? In the past two days, there have been sixty-seven swing state polls added to his model. Of those 67, only FIVE were high-quality live-caller polls.
That’s more than a ten-to-one ratio of landline-bound and online opt-in muck to good, traditional live-caller polling.
And this, Silver feels, only deserves one throwaway caveat:
We’re also seeing a fair number of automated polls from Republican-leaning firms with middling pollster ratings.
62 IVR and experimental online polls, disproportionatly weighted to GOP campaign op shops. And only one line on the issue?
This should be the focus of multiple paragraphs of hand-wringing self-doubt on Silver’s part. You can’t confidently calibrate for partisan lean if the great majority of your data is non-representative. And most IVR polling misses second-language speakers, NPAs, new voters, and low-propensity voters.
Friday, Nov 4, 2016 · 11:48:38 PM +00:00 · MaxwellOTOH, on the side of an over-optimized model, one of this afternoon’s inputs to the 538 model was a Roanoke College poll for Virginia that showed Clinton ahead by 18 points. (They actually did two, with the other at +7.)
What was the result of a +18 pt Virginia poll cranked through 538?
A nearly full-point swing in Donald Trump’s direction.