JURIST >> OPINION >> Forum >> In Defense of Manual Recounts... 

Professor Mark R. Brown
Stetson University College of Law
JURIST Guest Columnist

In a recent article in the Hoover Institution痴 Policy Review, Harvard Law Professor Einer Elhauge asserts that the Bush-Gore debacle in the last Presidential election establishes three important truths for America痴 political system: (1) voting machinery is antiquated, (2) rules are good, and (3) manual recounts are bad.

One can hardly quibble with Professor Elhauge痴 charge that America痴 election machinery is dated. A July 2001 report of the Voting Technology Project (jointly sponsored by Cal Tech and MIT), for example, recommends that punch cards be replaced by optically scanned paper ballots. Nor can there be much debate over Professor Elhauge痴 assertion that antecedent, objective standards are preferable to standardless, post hoc rationalizations. That is what the rule of law is all about.

But what of Professor Elhauge痴 charge that manual recounts cannot enhance the accuracy of mechanically registered votes? The latest spate of close and tangled contests in the 2002 midterm elections - a number of which have resulted in recounts - suggests that this proposition merits serious scrutiny.

Professor Elhauge argues that using manual recounts to 田heck the accuracy of a machine count is rather like trying to recheck a machine痴 measurement of electron width using the human eye and a yardstick. More emphatically, he states that 甜a] human recount that is 4 percent inaccurate cannot improve upon a machine count that is 1 percent inaccurate. For these reasons, Professor Elhauge argues, manual recounts should be limited to those situations where machines 杜alfunction.

If humans were called on to recount every ballot in a state-wide election, and if human fallibility was shown to exceed that of machines, Professor Elhauge would have a valid point. However, it is not at all clear that machines are always better than humans when it comes to counting votes. The Technology Voting Project report cited above, for example, found that punch card technology has a higher failure rate than hand counts in Presidential elections. While this does not prove that humans are better than machines, it shows that humans are not necessarily worse.

More importantly, Professor Elhauge mischaracterizes what happened in the wake of Florida痴 presidential election. The Florida Supreme Court did not order a manual recount of all of Florida痴 votes. Rather, it ordered an additional hand count of thousands of ballots that could not be mechanically deciphered. This count was not intended to supplant the machine vote; rather, it was designed to supplement it. Even assuming that supplemental counting fails to uncover votes 80 or 90% of the time[1], a supplemental hand count of previously unregistered votes can enhance the certainty and accuracy of an election.

Examples best prove the point. For simplicity's sake, assume that only two candidates, A and B, run on a statewide ballot. Assume also that their respective constituencies are uniformly scattered across the state (that is, they receive the same percentage of votes in each county), one million votes are cast and that the state痴 voting machines routinely experience a 1% 吐ailure rate. A supplemental count means that only 10,000 ballots must be counted again, either by hand or by machine. Human failure rates, whether caused by lack of markings on ballots or human disagreement over their contents, are relevant only in the context of these 10,000 votes. Human errors in the supplemental count cannot detract from the accuracy of the mechanically counted votes. Assuming that the human failure rate is as high as 90%, 1000 ballots can still be accurately registered and added to the candidates totals. Under the conditions posed in this hypothetical, 991,000 votes will then have been accurately counted, compared with 990,000 if machines were to be used alone. The overall failure rate drops below 1%, and the electorate can be more confident that the outcome is correct.

Enhanced accuracy, moreover, is not changed by uncountable ballots. So long as human failures are spread randomly among candidates meaning that human counters are not predisposed to dismiss potential votes for one candidate as opposed to another the fact that some ballots remain uncounted does not impeach the added accuracy manual counts bring to elections. Using the hypothetical outlined above, 9000 uncountable ballots do not impugn the integrity of the 1000 votes that are accurately counted.

Even true errors counting B痴 vote for A do not detract from the value of supplemental counts. Again, human errors randomly spread among candidates tend to cancel each other out. Of course, should one candidate experience a significantly higher failure or error rate than another, the accuracy of the manual count is called into question. But one could not know before the votes are manually counted.

The question is not whether to use supplemental manual counts, but when to use them. The better view, it seems, is to limit supplemental counts to those instances where uncounted ballots have a credible chance of altering the outcome of an election. It is not enough that the number of unregistered ballots equals or exceeds the margin of defeat. Rather, there must be enough outstanding ballots to provide the losing candidate a plausible opportunity of overcoming the difference. Commonly employed analyses, like statistical z-tests[2],can be used to assess the credibility of a losing candidate痴 chances. If the number of votes needed to win falls outside the realm of accepted possibilities, then the need for a supplemental manual count would seem to be diminished.

How close must an election be to justify a supplemental count? It depends on the number of votes cast, the margin of victory and the number of outstanding ballots. If A were to win by 6000 votes (out of one million cast) with 10,000 left uncounted, for example, a supplemental count would still not be needed. The likelihood of B winning 6000 of these outstanding votes is infinitesimally small. Assuming that the failure rate is random and uniform, it is statistically unreasonable to believe B could prevail. A supplemental manual count in this situation could cause more mischief than good.

But what if A were to win by, say, only 5 votes out of one million cast? Could 10,000 outstanding votes be reasonably expected to change the outcome? The answer seems self-evident. B could easily win 5 or 6 more votes than A in a supplemental count that peruses thousands of ballots. Indeed, most statisticians would say that the outcome is a toss-up, and few gamblers would bet the farm on either candidate.

Of course, not all of the 10,000 outstanding votes will be countable. Most probably are not. Uncounted votes, however, do not by themselves impugn the integrity of a supplemental count. Whatever result, it is more certain with the additional votes than without them. Could a supplemental count痴 failure rate affect B痴 chances of catching A? Sure. If only 1000 votes are countable, as opposed to 10,000, B痴 likelihood of garnering an extra 5 votes is reduced. It may be that the number of outstanding countable ballots is too small to afford the loser a statistically credible chance of success. The problem is that no one knows until the outstanding votes are counted. Arguing that a likelihood of not being able to count votes precludes counting them puts the cart before the horse.

What does all this mean in terms of the 2000 Presidential Election? Should Gore have been awarded a supplemental count? Did he have a statistically credible chance of overtaking Bush? With hindsight as guide, the answer is 惣es. After the second machine count, of the six million votes cast in the state of Florida, Bush痴 total exceeded Gore痴 by only 537 votes. Over 175,000 votes, including undervotes (those that failed to register a mechanically verifiable mark) and overvotes (those that contained more than one mark), were found by NORC in its year-long review of the election to have escaped mechanical counting. Assuming that these uncounted votes occurred randomly and uniformly across the state, Gore痴 likelihood of success roughly one in nine was credible enough to warrant a supplemental count.

The Florida Supreme Court, of course, did not order a supplemental count of all uncounted votes. It ordered a supplemental count of only undervotes. Although no one at the time knew the exact number, NORC uncovered over 60,000 undervotes. Assuming this number was known beforehand, and further assuming that failure and error rates were uniform across the state, the Florida Supreme Court probably ought not have ordered a supplemental count. Gore痴 likelihood of success under these facts about one in 70 was not credible. However, it was reasonably clear that mechanical failure rates were not uniform. Nor were the candidates constituents. For example, voters in two relatively small counties, Bay and Collier, preferred Bush two-to-one over Gore. Voters in populated Broward and Palm Beach Counties opted for Gore two-to-one. Palm Beach County experienced a higher failure rate (6.4%) than either Bay (1.11%) or Collier (3.34%) Counties. Because disparities like these undermine statistical predictions, the Florida Supreme Court痴 order is hardly surprising. True, it should have included overvotes. But no one imagined that the number of overvotes would be so large. Contrary to Professor Elhauge痴 argument, the numbers do not suggest any impropriety on behalf of the Florida Supreme Court.

I am not blind to the possibility of fraud or collusion. It may be that Democrats were conspiring to 都teal Bush痴 election. Potential malfeasance, however, is a constant. If the mere possibility of bias justifies abandoning votes, America would have to stop holding elections. Some elections officials probably hoped a manual count would favor Gore. But this is a far cry from saying that Florida痴 manual count could only be understood in those terms. Like it or not, supplemental manual counts offer a useful tool for uncovering electoral truth.

[1] The National Opinion Research Center (NORC) found that only 14% (roughly 25,000 of 175,000) of Florida痴 uncounted votes could be clearly interpreted. A Cal Tech/MCI study, by way of contrast, estimates that 75% of those whose votes were not counted intended to vote for President.

[2] At the risk of losing sight of the electoral forest for statistical trees, let me digress just a bit into the science behind z-tests. Take a coin from your pocket and flip it. Record the result heads or tails. Do it again. Do it 9998 more times and record the results. You have now produced a single 都ample of 10,000 coin flips. Most readers with the stamina to perform this experiment will record somewhere around 5000 heads in their samples. Some will record more than 5000 heads, and some less, but most samples will closely approach 5000. The reason is simple the odds of flipping a head are 1 in 2 (50%). If one were to count the respective numbers of readers who recorded no heads, one head, 1000 heads, 4999 heads, 5000 heads, 5001 heads, 5002 heads, 5059 heads, etcetera, and plot these totals on a bar graph, one would discover a 澱ell curve, with the number of readers reporting 5000 heads at its apex. Less common results, like 4500 heads and 5500 heads, would fall under the tails of this curve. Statisticians are quite familiar with this curve and can use it to make confident predictions about the potential outcomes of future samples. For example, using a z-test, one can confidently say that flipping 8000 heads in 10,000 attempts is infinitesimally improbable. In the language of statisticians, the difference between 8000 heads and 5000 heads the expected result is 都tatistically significant. If a sample痴 likelihood exceeds one in twenty, on the other hand, most statisticians would say that it is not statistically improbable i.e., though not expected, it could occur by chance. Flipping 5050 heads, for example, should occur in 1 out of 6 samples. This likelihood is not so improbable that it ought to be dismissed. Hence, were an election to produce a margin of victory of only 50 votes, with 10,000 ballots outstanding, a supplemental count would prove quite useful.

Mark R. Brown is a professor of law at Stetson University College of Law, where he teaches Constitutional Law, Administrative Law, and Civil Rights Litigation.

November 18, 2002


JURIST Guest Columnist Mark R. Brown is Professor of Law at Stetson University College of Law, where he teaches Constitutional Law, Administrative Law, and Civil Rights Litigation. Before coming to Stetson, Professor Brown clerked for the Honorable Harry W. Wellford of the United States Court of Appeals for the Sixth Circuit. Professor Brown also has taught at the University of Illinois College of Law (1991-1992, 1998-1999), The Ohio State University College of Law (Fall 1999) and served as a Judicial Fellow at the Supreme Court of the United States (1993-1994). He graduated as the valedictorian from the University of Louisville School of Law where he served on the Law Review's editorial board.