Polling Firm Wedgewood Polls Admits Fabricating Survey Data After Being Confronted With Evidence
Wedgewood Polls admitted its survey data was fabricated after the election-analysis site FiftyPlusOne confronted the firm with evidence that raw data from a Mississippi Senate poll did not match the expected patterns of the survey platform.
A self-described polling firm whose surveys were being picked up by election forecasters admitted on Tuesday that it had been fabricating its data, after the election-analysis site FiftyPlusOne confronted it with evidence gathered during an investigation that began with a tip.
FiftyPlusOne, whose contributors include Mary Radcliffe, G. Elliott Morris and Cooper Burton, said it reached out to Wedgewood Polls on Oct. 20 for comment on the investigation. Instead of answering the questions, the firm posted a public statement on social media admitting the fabrication, according to FiftyPlusOne. The firm also sent the site direct messages that read “Fuck you,” along with the name of an individual not affiliated with FiftyPlusOne.
In the statement, Wedgewood described itself as “an experimental project to demonstrate the vulnerabilities, lack of oversight, and overall low barrier to entry in the modern political polling industry,” according to the Columbia University statistics blog that analyzed the episode. The statement went up the same day FiftyPlusOne made its request.
How the raw data gave the firm away
The investigation started with what FiftyPlusOne described as a credible tip from a trusted source: that Wedgewood might be faking its numbers. The site then asked Wedgewood for the raw data from its most recent survey, a Mississippi Senate poll fielded Oct. 12-14.
The poll had arrived just hours after the University of Virginia’s Sabato’s Crystal Ball moved the race from Safe Republican to Likely Republican. Wedgewood’s version put Sen. Cindy Hyde-Smith four points ahead of her challenger Colom.
When the raw data arrived, it looked, in FiftyPlusOne’s words, “surprisingly clean”: the unweighted sample was 67 percent white and 31 percent Black, close to a plausible statewide estimate, and the design effect (a measure of how much weighting a pollster had to do) was 1.18, which the site described as unusually low for a sample built without quotas. Wedgewood told the site it had not used quotas.
The more direct problem was in the ID columns. In real data from the online survey platform Pollfish, respondent IDs mix plain number strings with hyphen-prefixed ones, and the transaction ID tracks the time a person submitted the survey almost exactly. FiftyPlusOne confirmed those patterns with Pollfish and with data from another pollster on the same platform. In Wedgewood’s data, the IDs did not behave that way.
A dispute over who vetted whom
Wedgewood’s statement claimed that FiftyPlusOne, the New York Times and other poll aggregators had not vetted the firm before including its polls. FiftyPlusOne called that claim false, saying it had asked Wedgewood for the identities of its pollsters, its sampling and weighting methods, and its funding sources during a vetting process — and that it had independently confirmed that Jack Sims, one of the two names provided, had prior experience in the polling industry.
FiftyPlusOne also said Wedgewood had collected money from donors to run crowd-funded polls and had shared raw survey data — now confirmed as fraudulent — with the site on Oct. 15.
Reporting based on coverage by FiftyPlusOne and the Columbia University Statistical Modeling blog.
