I would prefer to blame the NY Times article on the ignorance of the reporter rather than on the abdication of professional responsibility by the statisticians involved, but clearly some big-name statisticians need to respond to this article.
To suggest that there is no way to get a more accurate result than to recount multiple times and pick a number somewhere in the middle is profoundly ignorant of standard statistical methodology for checking data quality. Every practicing statistician routinely applies multiple checks for data quality. Many such checks could be used to detect voting irregularities. Some obvious examples of checks that should trigger investigations would include a precinct in which voter turnout was far higher or lower than average, or a candidate receiving far more votes in a precinct than there are voters registered to his or her party. Both of these checks would have detected actual mistakes in the Florida vote count (see the second article below).
Obviously, US law does not allow a statistical adjustment of voting counts, but just as obviously, the statistical analyses of the Florida votes posted on edstat-l by Robert Dawson and at http://elections.fas.harvard.edu/election.html by Jonathan N. Wand, Kenneth W. Shotts, Jasjeet S. Sekhon, Walter R. Mebane, Jr., and Michael C. Herron show that the will of the voters in Palm Beach County was indeed thwarted. One can quibble about the details of the analysis, but the results for Palm Beach County are so extreme--especially when analyzed at the precinct level--that there can hardly be any doubt that Buchanan received between 2000 and 3000 votes intended for Gore.
This is a wonderful opportunity to promote the value of sound statistical analysis, to get some rare good publicity for the statistical profession, and to improve the democratic process in the USA.
In article <200011101603.LAA11640@nightingale.hcp.med.harvard.edu>, email@example.com (Alan Zaslavsky) writes: > The following might be interest for those following press coverage of the > possible role of statistics in this dispute. (The printed version in the > edition I receive contained additional comments by David Freedman, also > downplaying the potential of statistics in this highly charged situation. > I would not follow Persi very far on the analogy to census undercount > adjustment, since anything that would be done now on the elections would > be post hoc and supported by little research ... that's another argument!) > -------------------------------------------------------------------------- > IN RESEARCH, RECOUNTS ARE NORM > http://www.nytimes.com/2000/11/10/politics/10MATH.html > > November 10, 2000 > THE SCIENCE OF COUNTING > By GINA KOLATA > > First George W. Bush led Al Gore by 1,784 votes in Florida. Then, an > unofficial count by The Associated Press suggested his lead was slashed > to less than half that. So, which number is right mathematically? > > Statisticians chuckled at the idea. > > "There's always going to be an error," said Howard Wainer, a statistician > at the Educational Testing Service in Princeton, N.J. "Every time you > count them, you're going to get a different answer." > > In research, said Diana C. Mutz, a professor of political science at Ohio > State University, scientists will repeat a process multiple times and > choose a number somewhere in the middle of their data as most likely to > reflect the truth. But, she and others said, multiple recounts are > probably not desirable in the presidential election because they would > add to the delay and uncertainty, not to say the bickering. Whoever was > losing could argue for one more recount. > > Even if it were just a research question, Professor Mutz was not sure how > many counts would be needed to make her confident the Florida vote was as > accurate as it could be. How many times she would count it "depends on > how many graduate students I have," she joked. > > Then there is the problem of Palm Beach County, Fla., where residents > said confusion over the ballot led more than 19,000 voters in a heavily > Democratic area to mark two candidates instead of one for president. > Their ballots were discarded as invalid. And, adding to the confusion, > Patrick J. Buchanan won more than 3,000 votes in Palm Beach. Some > Democrats said many people accidentally voted for Mr. Buchanan when they > meant to vote for Mr. Gore. Isn't there a way to fix that? A statistical > adjustment, perhaps? > > Sorry, say the statisticians. Any adjustment would only make matters > worse. Persi Diaconis, a statistician at Stanford University, said the > situation reminded him of attempts to adjust the census to correct for > the undercount, the people who were missed. The census recount turned out > to be a nightmare, he said, with new errors introduced and even more > squabbling. > > "The process really degenerates," Professor Diaconis said. "It's not at > all simple." > > Copyright 2000 The New York Times Company
From comp.risks > Florida vote counts > > "Peter G. Neumann" <firstname.lastname@example.org> > Fri, 10 Nov 2000 10:23:07 PST > > The recount in Florida presents another interesting lesson in risks in the > election process. > > * The recount in Palm Beach County increased the totals for Gore (+751) and > Bush (+108). > > * An entire precinct had been left uncounted. The ballots had been run > through the card reader, but the operator had pressed CLEAR instead of SET. > (The recount gave Gore +368, Bush +23.) > > * In Deland, Volusia County, a disk glitch caused 16,000 votes to be > subtracted from Gore and hundreds added to Bush in the original totals. > This was detected when 9,888 votes were noticed for the Socialist Workers > Party candidate, and a new disk was created. (The corrected results were > Gore 193, Bush 22, Harris 8.) > > * The day after the election, an election worker discovered a sack of about > 800 ballots in the back of his car that obviously had not been included > in the official results. > > * Voting cards failed to fit properly in the slots of some voting machines > in Osceola County, giving 300 votes to the Libertarian candidate (where > only 100 Libertarian voters are registered). Misaligned card machines > have long been a source of errors. > > * In Pinellas County, election workers were conducting a SECOND recount > after the first recount gave Gore more than 400 new votes. Some cards > that were thought to have been counted were not. > > [Source: Democrats tell of problems at the polls across Florida, > *The New York Times*, 10 Nov 2000, National Edition A24] > > Punched cards are inherently subject to differences on successive recounts. > Hanging chad is clearly a problem, and successive mechanical recounts > normally change the results each time. Human inspection is typically > necessary to resolve conflicts. > > Although electronic voting systems reduce the mechanical uncertainty that > sometimes makes recounts necessary in punched-card elections, they also > introduce different uncertainties in the integrity of the election process, > and particularly in the integrity of the computer systems. Certainly, > hanging chad problems, paper fatigue, and tampering with punch cards would > disappear, and recounts would be unnecessary: votes could be tabulated only > as originally entered. But many new problems are also introduced. The > opportunities for accidents and fraud are transformed into different > categories -- such as tampering with software development and operation. > And the desire for voter privacy is fundamentally in conflict with any > requirements for accountability (e.g., audit trails). > > In the Florida case, we still have to wait for the absentee ballots, and any > possible further recounts in other states.
Warren S. Sarle SAS Institute Inc. The opinions expressed here email@example.com SAS Campus Drive are mine and not necessarily (919) 677-8000 Cary, NC 27513, USA those of SAS Institute.