Markos Moulitsas fired Research 2000 as the pollster retained by Daily Kos a few weeks ago after R2K fared poorly in "pollster ratings" compiled by FiveThirtyEight's Nate Silver. At the time I wondered whether Markos reacted a bit harshly, since Silver himself admitted, "The absolute difference in the pollster ratings is not very great." In addition, some polling experts had raised questions about Silver's rating system (see also here).
Today Markos published a remarkable analysis of "problems in plain sight" with Research 2000's polling. Three researchers uncovered "extreme anomalies" in certain results and concluded, "We do not know exactly how the weekly R2K results were created, but we are confident they could not accurately describe random polls." You should click over and read the whole thing, but here are the anomalies in question:
1. A large set of number pairs which should be independent of each other in detail, yet almost always are either both even or both odd.
2. A set of polls on separate groups which track each other far too closely, given the statistical uncertainties.
3. The collection of week-to-week changes, in which one particular small change (zero) occurs far too rarely. This test is particularly valuable because the reports exhibit a property known to show up when people try to make up random sequences.
Markos has renounced "any post we've written based exclusively on Research 2000 polling" and asked polling sites to "remove any Research 2000 polls commissioned by us from their databases."
Based on the report of the statisticians, it's clear that we did not get what we paid for. We were defrauded by Research 2000, and while we don't know if some or all of the data was fabricated or manipulated beyond recognition, we know we can't trust it. Meanwhile, Research 2000 has refused to offer any explanation.
This analysis only covered R2K's weekly national tracking polls for Daily Kos, but based on the findings I no longer have confidence in R2K's state polling either, including various Iowa polls I've discussed at Bleeding Heartland. Some of those were commissioned by Daily Kos, and others were commissioned by KCCI-TV, the CBS affiliate in Des Moines.
Last year the Strategic Vision polling firm was brought down by convincing allegations that at least some of its polling results had been fabricated. Research 2000 had a much better reputation than Strategic Vision, though. Markos listed some of the news organizations that have commissioned R2K polls. I am seeking comment from KCCI News Director Dave Busiek about the company's future plans regarding polls, and I'll update this post when I hear back from him.
Share any relevant thoughts in this thread.
WEDNESDAY UPDATE: Mark Blumenthal contacted a forensic data guru for his take on the statistical anomalies. Excerpt:
[Walter] Mebane says he finds the evidence presented "convincing," though whether the polls are "fradulent" as Kos claims "is unclear...Could be some kind of smoothing algorithm is being used, either smoothing over time or toward some prior distribution."
When I asked about the specific patterns reported by Grebner, et. al., he replied:
None of these imply that no new data informed the numbers reported for each poll, but if there were new data for each poll the data seems to have been combined with some other information---which is not necessarily bad practice depending on the goal of the polling---and then jittered.
In other words, again, the strange patterns in the Research 2000 data suggest they were produced by some sort of weighting or statistical process, though it is unclear exactly what that process was.
JULY 4 UPDATE: Mark Blumenthal reviews what we know so far about this "troubling" story at Pollster.com.