Probability in a Contested Election Author(s): Tom Downs, Dennis C. Gilliland, Leo Katz Source: The American Statistician, Vol.

32, No. 4 (Nov., 1978), pp. 122-125 Published by: American Statistical Association Stable URL: http://www.jstor.org/stable/2682937 . Accessed: 06/01/2011 00:15
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at . http://www.jstor.org/action/showPublisher?publisherCode=astata. . Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to The American Statistician.

http://www.jstor.org

in Probability a ContestedElection
TOM DOWNS, DENNIS C. GILLILAND, and LEO KATZ*

models can sometimesbe used to analyze Elementaryprobability have occurred. In the results of an election where irregularities example by consideringthe 1975 this articlewe give an illustrative mayoral election in the city of Flint, Michigan. True vote is estimated with an estimatorwhich is used in randomized response models. KEY WORDS: Election challenge; True vote estimation;Randomized response model.

the 2. The Facts Concerning FlintElection The city of Flint, Michigan, held an election on November 4, 1975 forthe officeof Mayor, whereina timeundera new mayorwas to be elected forthe first recountis shown The vote formayorafter citycharter. in Table 1. The breakdownfor Precincts51 and 52 is given because these vote totalswere disputeddue to a mixupin ballot assemblies in these precincts.The vote totals from the other 143 precincts and absentee voters were not disputed. The recorded vote shows the winner with a marginof 206 votes. Rutherford However, due to the errors in ballot assemblies in the two precinctsand the closeness of the vote, the decision by the Board of Canvassers to declare James the Rutherford winnerover Floyd McCree was challenged in the courts. devices Those who voted in person used punch-card in the followingmanner. The individual voter was to given a punch-card insertin a ballot assembly. The device and is a book ballotassemblyis partofthevoting on withthenames of candidatesprinted the pages with a hole beside the name of each candidate. The individual voted for the candidate of his choice by punching throughthe hole with a stylus. This removed a square fromthe punch card. The computer used to tabulate the vote was programmedto know betweensquares on thepunchcard thecorrespondence and votes forspecificcandidates. The ballot assemblies were arranged among precandidate would always be listed cincts so thatneither firstor second. The computerwas programmedand the individualcards coordinated to take the rotation into account. Precincts51 and 52 had fiveand fourvotingdevices, respectively.The ballot assemblies were to be rotated between the two precincts so that the candidate whose name was firstin Precinct 51 was second in Precinct52. placed one ballot By mistake,the election officials assembly in Precinct 51 that should have been in Precinct 52, and one ballot assembly in Precinct 52 that should have been in Precinct51. The resultwas thata voterusingthe device withthe wrongassembly in eitherprecinctwould have his vote recordedforthe othercandidate. The errorin ballot assemblies was discovered after the polls had closed and the votingdevices were disof assembled. The commingling all punch-cardsfrom votingdevices withina precinctmade it impossibleto whichor how manyvotes were improperly distinguish recorded (reversed) withinPrecincts51 and 52.

1. Introduction The result of a close election may be challenged and the outcome may on the basis of irregularities ultimatelybe decided by the courts. A given court may uphold the result, reverse the result, invalidate the entire election, or offer any of a number of reviewofelectionlaw remedies.(For a comprehensive including postelectionremedies,see "Developments in the Law-Elections" (1975).) In general, an election will not be overturned the basis of a mere matheon maticalpossibilitythatthe resultswould be reversed Being reluctant to in the absence of irregularities. the valid electors, the courts unjustlydisenfranchise have sometimesrequiredthata challengerestablish a probabilitythat the result would be reversed in the absence of irregularities. Sometimes the evidence consists of proof that a certainnumberof persons voted who were not qualified,withno evidence as to how these persons voted. Finkelstein and Robbins(1973) computetheprobability ofreversalunderrandomremovalofa numberof votes equal to the numberof illegal votes. They point out thatthisis a neutraland propermeasure of probability of reversal in cases where there is no evidence to from indicatethat one candidate or anotherbenefited the illegal votes. In this articlewe give the facts associated withthe contestedmayoralelection in the cityof Flint, Michiin gan, on November 4, 1975. The irregularity this election is not a common one; however, the problem it created provided a situationwhere an elementary probability model could be used to quantify the of probability reversal.
Tom Downs is the senior partnerof Downs and Pirich, Attorneysat Law, 603 Capitol Savings and Loan Building,Lansing, MI 48933. Dennis C. Gilliland is Professor of Statistics and Probability,Michigan State University,East Lansing, MI 48824. ProfessorLeo Katz died unexpectedlyduringthe early stages of this work. His coauthors are indebtedto him for his contribution to this effort and to statisticsin general. The authors would like to thankthe refereeand the associate editorfor theirconstructive comments.
_

122

? The American Statistician,November 1978, Vol. 32, No. 4

3. EstimatedTrue Vote forPrecincts51 and 52 To estimatethe true vote for a precinctwith a reversed assembly, consider the following model. Let k denotethe numberof votingdevices in theprecinct, of whichexactly one has a reversedballot assembly,and let p = l/k. Suppose that0 votes are cast forMcCree and n - 0 votes are cast for Rutherford the prein cinct. The numberm recordedforMcCree is given by m =x +y, (3.1) wherex is the numberof votes cast forMcCree on the k - 1 properassemblies and y is the numberof votes cast forRutherford the reversedassembly. Under on uniform random selectionof votingdevice by voters, one would expect k - 1 out of everyk votes cast for McCree to be recordedforMcCree and 1 out of every k votes cast forRutherford be recordedforMcCree. to Thus, the expected value of x is E(x) = 0(1 - p), the expected value of y is E(y) = (n - O)p, and the expected value of m is E(m)
=
Precinct

2. Estimates of True Vote, 1975

Candidate
McCree Rutherford

Totals
n_

51 52 Totals

185

(14.2)

270

(14.2)

455 291 746

202.5 (14.7) 387.5 (20.5)

88.5 (14.7) 358.5 (20.5)

eters0 and 1 - p, denotedbyx - B(O, 1 - p). Similarly, of y - B(n - 0, p) independently x, and the probaof bilitydistribution m = x + y is the convolutionof i.e., these two binomialdistributions; m - B(0, 1- p)*B(n - 0, p), 0 E E, (3.4) where 0 = {0,1, . . . ,n}. The mean of a binomialdisis tribution the numberof trialstimes the probability of success on a single trial,so (3.2) holds. Using the formulaforthe variance of a binomial,we findm has variance
02(m) =
0(l

0(1 - p) + (n

0)p.

(3.2)

I-p)p

+(n

- )p(I -p) =np(I- p). (3.5)

Using the methodof momentsto derive an estimator of 0, we equate m and the righthand side of (3.2). This equation can be solved for 0, providedp 4 1/2, to yield the unbiased estimator m -np 1 - 2p (33)

Thus, underour assumption,0 of (3.3) is an unbiased for estimator 0 withvariance

2(0) =

np(l-

p)

(3.6)

The estimator is a function the randomvariable 0 of m, which is the sum of the randomvariablesx and y. The probability distribution bis important assessof in ing the accuracy of 0. It can be determinedunder a varietyof assumptionson theway votersselect voting devices. We consider one such assumption. a votersindeAssumption:Within disputedprecinct, select the votingdevices on whichtheycast pendently theirballots; foreach voter,the probability p 4 1/2 is that the voting device with the reversed ballot assemblyis selected. Under thisassumption,the numberof votes cast for McCree and recorded for McCree, x, can be thought of as the number of heads in 0 independenttosses of a coin withprobability1 - p of a head on a single toss. Its probability distribution binomialwithparamis

1. Recorded Vote, 1975

Candidate Precinct 51 52 Totals (51 and 52) Other Totals McCree 202 174 376 20,099 20,475 Rutherford 253 117 370
20,s311

Totals 455 291 746 40,410 41,156

20,681

The family distributions of B(H, 1 - p)*B(n - 0, p), can be shown to be complete (provided p ? 1/2), so that 0 is the unique unbiased estimator of 0. The estimator0 has the nice propertythat its variance (3.6) does not depend on 0. (Since 0 takes values outside the parameterset, it would ordinarily be truncatedback to the closest value from0, which can only reduce mean squared error of estimation.) The maximumlikelihoodestimatorof 0 has not been determined. We now apply (3.3) to Precinct 51 (p = '15) and Precinct52 (p = 1/4), using the data of Table 1 to find unbiased estimatesof the true vote forMcCree in the is two precincts.The true vote for Rutherford n - 0 and is estimated by n - 0. Of course E(n - 0) = n - 0 and ((n - 0) =_(O). Standarddeviationsare indicated in parenthesesbeside each estimatein Table 2. The standard deviation found in the bottom row comes fromthe well-knownformulaforthe standard deviation of the sum of two independent random 1/ ? variables; namely, _(0 02) = (o.2() + 14.72)1/2 =20.5. Of course, here we are -(14.22 correctlytakingx and y for Precinct 51 to be independent of x and y for Precinct 52, fromwhich the 0 separate estimators for Precinct51 and 02 for Precinct52 are independent. The estimated total for McCree, 387.5, and its standarddeviation20.5 indicatethat,in all probability, McCree did not have a total 0, + 02 2 479 sufficient to overcome the undisputed margin of 212 votes Rutherfordheld outside these precincts. S3ince 0,
0 E 0,

?DThe American Statistician,November 1978, Vol. 32, No. 4

123

and 479 is normallydistributed + 62 is approximately 4.5 standard deviations from 387.5, the evidence is in overwhelming favorof this conclusion. 4. Critiqueand Comments Our assumptionmay be attacked as unrealisticbecause it says that voters select voting devices independently.During a busy period on election day, the votingdevice (booth) selected by the voter does influencethe cho9iceof voting device by other voters. was considered In the instantcase, the voter turnout relativelylight.The robustness of our calculation in the face of departurefromindependence is an interesting study in itself and is not dealt with here. Of course, the assumptionthatthe votingdevice withthe reversed ballot assembly is selected with probability p = 1Ik may also be unrealistic if it is physically located so as to be selected with less or greater likelihoodthan other devices. In the instantcase, no was accepted in court on these matters,so testimony p the specification = 1Ikremainsthe only neutraland specification. nonarbitrary Some evidence, however,can be obtained to either supportor attack the reasonableness of our assumpthe tion, by monitoring votingdevice use patternsat polling places. In precincts with electronic voting machines,for instance,each such machine automatically keeps trackof how manyvoters have used it. The problem of estimationof 0 in our model is similarto an estimationproblem firstconsidered by Warner(1965) in a randomizedresponse model. Supto pose thatn people respond truthfully either(i) Do you smoke?, or (ii) Do you not smoke? Suppose that which of the two questions a person is asked is determinedby an experimentwherein (i) is selected with probability 1 - p and (ii) is selected with p, probability and suppose thatthe selection of (i) or for individuals. Then the (ii) is done independently number m of yes responses is the sum of x, the numberof smokerswho answer (i), and y, the number who answer (ii). The Warnerestimator of nonsmokers of 0, the numberof smokers among the n people, is 0 of (3.3). In situationswhere randomized response is used, the n people usually constitute a random a sample from largerpopulationabout whichinference is to be made. For example, if X- represents the of populationproportion smokers,then -T = O/nis unbiased and consistentfor 7r 5. Resolutionof the InstantCase the The Board of Canvassers declared Rutherford winnerof the November 4, 1975 election on the basis of the 206-votemarginreportedin the bottomline of Table 1. The court case which ensued was heard in CircuitCourt of the County of Genesee on February presentedmodels with certainas20, 1976. Plaintiffs

to sumptions show thatMcCree was the actual winner of the November electionwitha likelihoodas greatas 1 in 5. Plaintiffs sought reliefin the formof a new citywideelection. Defendantspresenteda probability analysis similarto that discussed in this articlein an attemptto establish that Rutherford was the actual winner in all probability.They argued that if relief should be given, it should be in the formof a special to "mini-election"restricted the voters who voted in Precincts51 and 52 in November. On March 4, the Court ruled in favorof the special mini-election and in the Order indicated that "those voters are instructed cast theirballot in the same to way" (as cast in November). The letterto individual voters phrased it, "You are instructed recast your to ballot at this rerunelection so that we may ascertain how you voted on November 4, 1975." of We hoped that the mini-election April 13, 1976 would reveal the true vote of November 4, 1975, so that the Table 2 estimates could be compared to the actuals. Many factors caused this hope to be unrealized. To begin with, 763 voters voted in person at Precincts 51 and 52 in November, of whom only 746 cast votes in the mayoral election. All 763 were by necessity eligible to vote in the rerun. Not all people eligibleto vote in the rerunbotheredto or were around to participatein the April election, as Table 3 shows. We have notbeen able to discover the precinct originsof the 32 absentee votes reportedin Table 3. In additionto the 661 votes, seven ballots were cast withno vote shown foreithercandidate. The polarizationevidentin Table 3 maybe explained bybehind-the-scenes in campaigning thetwoprecincts, whichwas notexplicitly prohibited theCourtOrder. by On the basis of demographiccharacteristics and voting patterns,and ignoring the recorded vote for Precincts 51 and 52 in November 4, 1975 (Table 1), it was fairlyclear before the rerun mini-electionthat Precinct51 favoredRutherford Precinct52 favored and McCree. The shift the favorites to withinthe two precincts in excess of what was suggestedby the model and Table 2 may be attributed behind-the-scenes to Of campaigningin the mini-election. course, the discrepancybetweenTables 2 and 3 can also be explained by the conjecture that the probability of using the p reversedballotassemblywas considerably greater than 1/5 in Precinct 51 and greaterthan 14 in Precinct 52; i.e., that the reversed devices were in positions to
3. Results of April 13, 1976 Mini-Election
Candi date Precinct 51 McCree 122 Rutherford 255 Totals 377

52 Absentee Voters Total s

217 10 349

35 22 312

252 32 661

124

?) The American Statistician,November 1978, Vol. 32, No. 4

of attracta largerproportion the voters than assumed in the model. In any event, such postelection analyses became academic in this case. The net effectof the rerun "firmly"in was to establish Rutherford mini-election officewitha citywidemarginof 175 votes in a total of over 41,000 votes cast.
[Received February14, 1977. Revised June 20, 1978.]

References
"Developments in the Law-Elections" (1975), Harvard Law Review, 88, 1111-1339. Finkelstein, Michael O., and Robbins, Herbert E. (1973), "Mathematical Probabilityin Election Challenges," Columbia Law Review, 73, 241-248. Warner,StanleyL. (1965), "Randomized Response: A SurveyTechnique for EliminatingEvasive Answer Bias," Journal of the American StatisticalAssociation, 60, 63-69.

PUBLICATION POLICY FOR ARTICLES IN THE AMERICAN STATISTICIAN

of articles general to The AmericanStatisticianattempts publish These articlesshould be profession. to interest the statistical of to generality appeal to largenumbers and timely of sufficient not be ordinarily of an expository, highly should Articles readers. topicsthatwouldbe list The nature. following illustrates technical appropriate: probstatistical nationaland international current 1. Important lemsandprograms. 2. Public policy mattersof direct interestto the statistical in and of 3. The training statisticians education statistics. statisaffecting with developments discussions dealing 4. General ticalpractice. of and 5. Short paperson thehistory statistics probability. curriculum, matters, classroom including of 6. Teaching statistics, development. and program surveysof the disstate-of-the-art or 7. Nontechnical tutorial
cipline. ExpositoryPapers profession.

The Teacher's Corner "The Teacher's Corner" Department carries articles on all levels of statisticaleducation. These include new and unique approaches to curriculumdevelopment, course organization, and useful in the teachingof particular originaltechnicalcontributions statisticalconcepts. Letters to the Editor to Letters on mattersof interest the statisticalprofessionwill be considered. They should be concise; when necessary the Editorial Board may requirea shorterversion.

SUBMISSION POLICY
welcomes the submissionofmanuscripts The EditorialCommittee forpossible publication.Four copies, double-spaced, of all material (includingLetters to the Editor) should be sent to the Editor, John Neter, QuantitativeBusiness Analysis, College of Bus. Admin., Universityof Georgia, Athens, GA 30602. Manuscripts should follow the Style Guide published in the August 1977 issue of The the AmericanStatistician.Additionalcopies are available from ASA office. Manuscripts will not be returned. Articles published are page. Payment page chargeof $30 per printed subject to a voluntary of this page charge is not a prerequisitefor publicationnor is the authorexpected to pay these charges himself.

desirous of attracting The EditorialBoard is particularly to importance the on articles topicsof current expository general styled shouldbe particularly Such articles community. statistical and and forour generalreadership shouldbe authoritative well papersare plansforpossibleexpository with Authors researched. of or withan outline draft with to urged communicate theEditor article. theproposed

?DThe American Statistician,November 1978, Vol. 32, No. 4

125