Many people reported difficulty voting on electronic voting machines (DREs), in particular, that attempts to vote for one candidate initially registered as votes for another. The Election Incident Reporting System (EIRS), connected to the "OUR-VOTE" telephone hotline, recorded close to 100 such incidents. TruthIsAll has asserted that 86 out of 88 reports of electronic vote-flipping favored Bush. He cites the odds of this imbalance as 1 in 79,010,724,999,066,700,000,000. Another observer reports that actually, 87 out of 94 EIRS reports favored Bush. The odds of this imbalance are, of course, still prohibitive -- on the order of 500 trillion to 1 against.
Is this compelling evidence of a stolen election? No, it really isn't at all -- because the actual election returns from DREs on crucial states don't support election theft concentrated upon DREs. In Florida, some analysts who looked for evidence of vote-switching on DREs concluded that the optical-scan results were actually more suspicious. (Of course, it is possible that both could be hacked.) In Ohio, most Ohioans voted on punch cards. Walter Mebane and Michael Herron, in an analysis of the Ohio returns for the Democratic National Committee's Voting Rights Institute, found no statistical evidence implicating either DREs or optical scanners in vote-switching. As noted earlier, the exit polls also fail to support the hypothesis of massive vote-switching on DREs -- unless one is prepared to stipulate even more massive vote-switching on lever machines.
One also wonders about the inherent plausibility of visible vote-switching as a means of widespread election fraud. If someone had the insider access to force visible vote-switching, wouldn't s/he also have the ability to engineer invisible vote-switching? This question may have a compelling answer, but I haven't seen one yet. (On the other hand, the mere fact that a fraud mechanism seems suboptimal doesn't prove that it wasn't used.)
While TIA's reported odds are breathtaking, they do not interpret themselves, for two reasons. First of all, the odds calculation depends on the underlying assumption that Bush-to-Kerry and Kerry-to-Bush switches were equally likely to be reported to EIRS. Actually, the EIRS hotline was widely publicized in liberal-leaning media sources, and the election protection volunteers who promoted the hotline were concentrated in heavily Democratic precincts. Given these facts, as well as pre-election concerns about Diebold and DREs, one would expect Kerry-to-Bush switches to be reported at a higher rate. But how much higher? There is simply no way of knowing. Certainly the imbalance of reports is suspicious.
The second problem is that these 88 or 94 reports are not a sample, but rather the universe of all EIRS vote-switching reports. (Presumably some additional vote-switching reports are not logged in the EIRS.) Moreover, most of these reports assert that the voter was ultimately able successfully to vote for the correct candidate. So, what is the likely frequency of vote-switching implied by the data? People who leap from fewer than 100 problem reports to multi-million-vote conclusions obviously are not following the data alone. Are these reports the tip of the proverbial iceberg, representing huge numbers of vote switches not detected or not reported? Again, there is simply no way of knowing. However, again, statistical analysis of the election returns generally indicates, probably not. Surely there were more problems than reported in the EIRS, and some of the problems may reflect tampering, but they seem to have had little impact on the outcome.
There is solid evidence of excessive "undervotes" on pushbutton DREs in New Mexico, disproportionately concentrated in Democratic precincts and presumably costing Kerry many net votes. The magnitude of this effect is unclear, but it possibly may have been large enough to reverse the outcome in New Mexico. The problem has been attributed in part to the implementation of the straight-ticket option on these machines: many voters may have inadvertently voided their presidential votes by initially selecting a straight ticket, then voting for another party's candidate elsewhere on the ballot.
Yes. For instance, the initial national House tabulation -- posted a bit after 7 PM Eastern time on election night -- indicates that Democratic candidates had a net margin of about 11.3 points over Republican candidates. The actual margin was probably about 7 points, depending on how uncontested races are handled.
[1/6/07: TIA asks what is my justification for the 7-percent figure. It's an estimate based on preliminary vote returns, similar to the estimate provided by Simon and O'Dell in their own argument that the 2006 exit polls evince to fraud (they peg the margin at 7.6 points). I will update it presently. But I assume that TIA's real objection is to the word "actual." I assume that it's clear in context that I was referring to the vote count -- the phrase "official returns" might have tipped him off -- but TIA misses few chances to complain that his critics assume that the vote counts are correct, although we don't.]
Not really. A "generic" poll is one that asks respondents whether they would vote for (in Gallup's words) "the Democratic Party's candidate or the Republican Party's candidate," rather than naming specific candidates. (Typically the order is rotated: half of respondents are asked about Republicans first.) Generic polls are common because of the logistical complexity of (1) trying to match randomly dialed telephone numbers to congressional districts and (2) prompting the interviewers to pronounce the names of the candidates in each of the 435 districts -- for an average of perhaps two or three interviews per district. Nevertheless, in real life, voters choose between named candidates, not generic ones, and this distinction makes a difference. Joseph Bafumi and colleagues state (PDF page 6):
We know that they [the generic polls] perform poorly as point estimates. For instance, an 18-point Democrat lead from early in an election year most likely will translate into a far smaller vote lead on Election Day. However, regression equations accounting for the vote in terms of the generic vote do predict well, as they properly discount the exaggerated sizes of the generic poll leads.What it means to "predict well" is a matter of perspective: Bafumi et al. report (PDF p. 2) that midterm House vote share predictions based on generic polls over the last 30 days have an estimated margin of error of plus-or-minus 3.7 points. Vote margin predictions therefore have a margin of error of more than 7 points.
The 2006 generic poll results varied widely. Seven polls conducted in the last week before the election agreed that the Democrats were ahead, but the estimated Democratic margin ranged from 4 to 20 points. The average margin in these polls was about 11.6 points. Using these polls in Bafumi et al.'s regression equation, we can estimate that the Democrats would win by about 8 points, with (as I mentioned) roughly a 7-point margin of error. (Bafumi et al., using generic polls through "early October 24," estimated that the Democrats would win by about 10 points based on an average margin of over 15 points in the polls to that date.)
David Moore and Lydia Saad noted in 1997 that the final generic Gallup poll has generally been more accurate than the generic-poll averages. They reported that from 1954 through 1990, the final Gallup poll had an average error of under 1.3 points on vote share (about 2.5 points on margin). (The 1994 through 2002 Gallup results were also close, on average, to the official totals, although the 1998 poll overstated the Democratic vote share by 2.5 points -- still within the margin of error.) The final Gallup poll in 2006 projected a 7-point Democratic margin.
Given statistical margins of error, these generic poll results could be consistent with a Democratic vote margin of 7 points or of 11-plus points -- although a 7-point margin is somewhat likelier than an 11-point margin. But we do not have to settle for the generic poll results. We can examine pre-election polls in particular House races. As Mark Blumenthal has demonstrated, overall these pre-election polls come very close to the actual returns. Thus, neither generic nor race-specific pre-election polls support the conjecture of widespread miscount in House races. (In Senate and gubernatorial races, Blumenthal showed earlier, Democratic candidates on average did as well as or better than in pre-election polls.)
[1/6/07: TIA has no substantive response to any of the arguments in this section. He offers no rationale for treating the generic poll results literally despite the historical record, while disregarding the race-specific poll results.]
Without getting into the specifics, the short answer is: I think that if voters had been able to cast their votes as they intended, the Democratic candidate Christine Jennings would have won the House race in Florida's 13th Congressional District (FL-13) by thousands of votes, instead of losing by under 400. I have seen no evidence that the events in FL-13 shed light on outcomes in any other Congressional race.
[1/6/07: TIA asks: "Are you implying that FL-13 was an isolated case of missing and/or switched votes?" No; I am stating that I have seen no evidence that any other case is like it. The VoteTrustUSA report to which TIA links is not on point. I say again: this is not a comprehensive election integrity FAQ; it focuses on TIA's baseless (and somewhat selective) faith in exit poll accuracy, as well as his other idiosyncratic polling arguments.]
In the 2004 exit poll, precincts using hand-counted paper ballots had a smaller mean (average) Within Precinct Error than precincts using any other voting system. As Edison/Mitofsky reported on p. 40 of their evaluation report, the mean WPE was -2.2 in the 40 paper ballot precincts, compared with -6.1 in optical scan precincts, -6.6 in punch card precincts, -7.1 in touch screen precincts, and -10.6 in mechanical voting (lever) machine precincts. Some observers have argued or insinuated that the inference is obvious: all the machines were hacked, but the hand counts could not be.
That inference isn't supported by the exit poll results, although they don't rule it out. Of course, I reject the assumption that the exit poll discrepancies are primarily due to vote miscount, but as a separate matter, this argument is weak on statistical grounds. Edison/Mitofsky noted that almost all the hand-count precincts (35) were in rural and small town areas, where every voting method experienced smaller mean WPEs than in urban areas: -0.8 for punch cards, -1.6 for paper ballots, -3.2 for lever machines, -4.4 for optical scan, and -6.0 for touch screen. Although Edison/Mitofsky did not report margins of error for these statistics, the margin of error for paper ballots in rural/small town precincts is probably about +/- 4 points or larger. (The five paper-ballot precincts in urban areas had a mean WPE of -6.0, of course with a much larger margin of error.) So, controlling for size of place, there is no statistically significant difference between the paper-ballot precincts and the other precincts. It may also be worth mentioning again that in 2004, an extensive hand recount of optical-scan ballots in New Hampshire found no significant discrepancies, despite New Hampshire's double-digit red shift.
I will not venture here into other arguments about hand-counted paper ballots. I simply observe that the 2004 exit polls don't prove anything about their merits.
Previous: (5) Comparing 2004 to 2000
Up to Table of Contents