Statisticians again reject Edison/Mitofsky's explanation of US exit poll discrepancies

international | politics / elections | news report Sunday September 11, 2005 20:15 by Dr. Coilín Oscar ÓhAiseadha - Spider Force

US Count Votes publishes latest statistical analysis of exit poll discrepancy

The US Count Votes project's most recent statistical analysis presents a thorough scientific rebuttal of Edison/Mitofsky´s proposed explanation for the discrepancy between last November's exit poll figures, which predicted a win for Kerry with a margin of 3%, and final vote tallies, which gave Bush the victory with a margin of 2.5%.

According to the statisticians who have authored this latest analysis of the exit poll and vote tally data, Edison/Mitofsky's explanation just does not hold water.

The hypothesis that the discrepancy was due to vote miscounting persists.

The exit polling firm Edison/Mitofsky (E/M) and the independent election research project US Count Votes agree that the historically unprecedented discrepancy between the exit polls and the reported vote count for the 2004 U.S. Presidential election cannot be a result of random sampling error. This leaves either exit poll error or vote miscount as the only two possible explanations for the exit poll discrepancy.

In the project's most recent statistical analysis, US Count Votes presents a thorough scientific rebuttal of Edison/Mitofsky´s proposed explanation for the discrepancy between exit poll figures, which predicted a win for Kerry with a margin of 3%, and final vote tallies, which gave Bush the victory with a margin of 2.5%.

According to the statisticians who have authored this latest analysis of the exit poll and vote tally data, Edison/Mitofsky's explanation just does not hold water. The numbers do not add up.

The hypothesis that the discrepancy was due to vote miscounting persists.

To exclude suspicions of errors in the official vote tallies, US Count Votes repeats its demand for more detailed data:
"It is a matter of the utmost national importance that detailed precinct level exit polling and election data that would allow for investigation by independent analysts, such as USCV, be publicly released."

Please take the time to read the analysis for yourself, and if you find it difficult to understand, please do ask a statistician for help.

The media are preoccupied with the flood in New Orleans, and many American activists are preoccupied with trying to impeach a president whom they believe has led the country into military and diplomatic defeat in Iraq and disabled emergency management mechanisms by sending members of the National Guard to Iraq, refusing to provide funding to raise the levees, and appointing his poorly qualified cronies to senior and middle management positions in the Federal Emergency Management Agency.

But already there is a commitment to rebuilding New Orleans, and even if Bush lasts the remaining three years of his term, the problem presented by the undeniable possibility that the wrong man was (again) appointed president through a miscount is not going to go away any time soon.

Until further notice, I am standing by the hypothesis that George W. Bush's team cheated in order to keep him in the White House in November 2004.

You may call this a conspiracy theory, but please remember my motto: "I don't believe in conspiracy theories - except those that are true." Apart from the statistical evidence presented by US Count Votes, the body of anecdotal evidence to support this one is overwhelming.

**********

US Count Votes
National Election Data Archive
---
The 2004 Presidential Election:
Exit Poll Error or Vote Miscount?
September 8, 2005

Ron Baiman, Ph.D – Institute of Government and Public Affairs, University of Illinois at Chicago
David Dodge - BSMME Engineer, Composite Developments, Inc., Technology and Innovation
Kathy Dopp -MS in mathematics - USCountVotes, President
Reviewed via USCountVotes’ email discussion list for statisticians, mathematicians and pollsters.

Press Contact:
Kathy Dopp, National Election Data Archive, President [email protected]

This paper can be found on the Internet at:
http://uscountvotes.org/ucvAnalysis/US/exit-polls/USCV_exit_poll_analysis.pdf



Introduction
The exit polling firm Edison/Mitofsky1 (E/M), and USCV, agree that the historically unprecedented discrepancy between the exit polls and the reported vote count for the 2004 U.S. Presidential election cannot be a result of random sampling error.2 This leaves either exit poll error or vote miscount as the only two possible explanations for the exit poll discrepancy.
E/M has claimed that the exit poll discrepancy is exclusively a result of “within precinct error” (WPE)3, and that the entire WPE observed in 2004 could be explained by a hypothetical exit poll completion rate of 56% among Kerry voters and 50% among Bush voters (herein referred to as “the E/M hypothetical”).4 The E/M hypothetical was widely interpreted by the media and by USCV as a claim that the 2004 exit poll discrepancy was caused by a pervasive, and on average uniform, shortfall in Bush voter exit poll response relative to Kerry voter exit poll response that was dubbed the “reluctant Bush response” (rBr) hypothesis.5
A recent clarification by E/M indicates that the “E/M hypothetical” should be interpreted as referring to hypothetical average (rather than constant average) partisan exit poll response rates.6 In this interpretation, average precinct partisan response rates may vary widely by reported precinct vote shares, yet all of the reported WPE could be explained by partisan response rates whose average across the sample is K=0.56 and B=0.50.7 This interpretation of the “E/M hypothetical” does not depend on the “rBr hypothesis” of constant average partisan response rates, which was shown by USCV to be inconsistent with the pattern of the exit poll discrepancy.8
However, our analysis below shows that even if the “E/M hypothetical” is interpreted as referring to average, rather than constant average, partisan exit poll response rates, it is inconsistent with the reported WPE data. There is no configuration of partisan response rates, however varying across precinct partisanship categories, with overall averages of K=0.56 and B=0.5, that can produce the reported values of the actual E/M exit poll data for mean WPE and median WPE.9 Thus, neither a "reluctant Bush responder" (rBr) hypothesis interpretation nor an overall average interpretation of the “E/M hypothetical” is consistent with the WPE pattern shown by the Edison/Mitofsky exit polling data.
A larger overall average partisan exit poll response gap than that proposed by E/M could account for the observed WPE, but the large differences in mean and median partisan response bias10 necessary to produce the actual observed WPE levels across partisan precinct categories would need to be explained. In particular, the very large mean and median WPE for precincts with over 80% reported Bush vote, and the almost zero mean and median WPE for precincts with over 80% reported Kerry vote, requires an explanation.
Ten months after the election, no plausible explanation of the 2004 exit poll discrepancy, based on exit polling error, has been provided by E/M. The precinct level exit polling and official vote count data that would enable independent investigators to analyze the exit poll discrepancy has not been provided to the public. Perhaps an exit poll explanation for the discrepancy does exist. However, a cloud of suspicion is cast on the 2004 presidential election results because the possibility that a “vote miscounts” explanation is required to generate the reported exit poll discrepancies is still open. It is a matter of the utmost national importance that detailed precinct level exit polling and election data that would allow for investigation by independent analysts, such as USCV, be publicly released.11

...


Footnotes:
1 "Evaluation of the Edison/Mitofsky Election System 2004" January 19, 2005
http://www.exit-poll.net/election-night/EvaluationJan192005.pdf
2 See “Analysis of the 2004 Presidential Election Exit Poll Discrepancies,” USCV, updated April 12, 2005. E/M state that the average “within precinct error” (WPE) of the exit polls in 2005 of -6.5% was the largest since 1988 and was 30% higher than the next largest mean WPE of -5.0% in 1992, op. cit. January 19 E/M report, p. 34.
3 WPE is defined by E/M is "an average of the difference between the percentage margin between the leading candidates in the exit poll and the actual vote for all sample precincts in a state."
4 Op. cit., January 19 E/M report, p. 28 and 31.
5 It should be noted that “rBr” does not necessarily imply that Bush voters were “psychologically” more adverse than Kerry voters to completing exit polls. The partisan exit poll response gap could for example be linked to the characteristics and methods of the exit pollsters, or it could be a function of the external circumstances in which exit polling was conducted. However, the initial interpretation of “rBr” as stipulating a pervasive and, on average, uniform, bias in exit poll response, seemed to indicate that Bush voters as a group (regardless of the characteristics of the exit pollsters and any other possible factors) had a lower completion rate than Kerry voters. This seemed to suggest that the explanation had to do with, on average, differential behavior of Bush and Kerry partisans that was unrelated to any other factor, i.e. a “psychological” or “group behavioral” explanation.
6 In a (5/26/2005) communication to Ron Baiman, Warren Mitofsky states that: “There is no constant mean bias conjecture on our part. This is wholly USCV's invention.” Mitofsky's statement would appear to vindicate USCV’s position that the “constant mean bias” or rBr hypothesis cannot explain the exit poll discrepancy.
7 Average K and B levels will be equal to weighted average K and B for partisan precinct categories, where the weights are the relative sample sizes of the precinct categories - see Table 6 in Appendix F.
8 See USCV April 12 report, op. cit. Elizabeth Liddle recently published (http://www.geocities.com/lizzielid/WPEpaper.pdf ) a simulation-based analysis suggesting that a new “unconfounded” index shows that a constant mean response bias is consistent with Edison/Mitofsky’s reported exit poll discrepancies.
However, as Liddle’s analysis is based on the same variables investigated earlier by USCV (see Appendix A, Liddle’s index equals LN(K/B)), its conclusions cannot logically be upheld, if USCV’s analysis is correct. The recent statement by Mitfosky (op. cit.) appears to support USCV’s analysis.
9 See Appendix A for derivations of partisan exit poll response rates K and B. See Appendices B and Table 6 in Appendix F for a proof of this statement. "Precinct partisanship categories" refers to precincts grouped together by the percentage of official votes that Kerry and Bush received in them.
10 “Bias” equals K/B and will sometimes also be labeled “Alpha” in this paper, following Liddle’s notation, op. cit.
11 The claim that E/M cannot release this data because of concerns over “respondent confidentiality” is belied by the fact that a “blurred” version of this data for the state of Ohio has been released to the Election Science Institute. See discussion below of apparent misstatements and inconsistencies in the Kyle, Samuleson, Scheuren, and Vicinanza report on the Ohio exit polls based on this data (see: http://www.votewatch.us/reports/view_reports). This report appears to support a vote miscount hypothesis rather than refute it as is claimed by the authors. In any case it is of limited value as it does not include information on the factors that influence WPE necessary for a substantive statistical analysis of the exit poll discrepancies. It should be noted that since this additional data on (anonymous) exit pollsters and polling conditions has no relation to exit poll respondents, its release would not compromise confidentiality in any way.



Read the full analysis here:
http://tinyurl.com/8hh3p