Machine Errors and Undervotes in Florida 2006 Revisited

Transcription

1 Willim & Mry Bill of Rights Journl Volume 17 Issue 2 Article 5 Mchine Errors nd Undervotes in Florid 26 Revisited Wlter R. Mebne Repository Cittion Wlter R. Mebne, Mchine Errors nd Undervotes in Florid 26 Revisited, 17 Wm. & Mry Bill Rts. J. 375 (28), Copyright c 28 by the uthors. This rticle is brought to you by the Willim & Mry Lw School Scholrship Repository.

2 MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED* Wlter R. Mebne, Jr." ABSTRACT The 26 election for U.S. House of Representtives District 13 in Srsot County, Florid, ttrcted extensive controversy becuse n unusully high proportion of the bllots cst lcked vote for tht office, nd the unusul number of undervotes probbly chnged the election outcome. Intensive technicl studies bsed on exmining softwre nd hrdwre from the ivotronic touchscreen voting mchines used to conduct the election filed to find mechnicl flws sufficient to explin the undervotes. Studies tht exmined the bllots used in Srsot nd in some other counties concluded the high undervote rte ws cused by peculir fetures of the bllot's formt tht confused mny voters. I show tht recorded events involving power filures nd problems with the Personlized Electronic Bllots used with the mchines correlte significntly with undervote rtes in severl Florid counties. The reltionships between mchine events nd undervotes re sufficiently substntil nd vried to mke it unresonble to discount the likelihood tht mechnicl filures contributed substntilly to the high numbers of undervotes. INTRODUCTION The controversil election for U.S. House of Representtives District 13 (CD- 13) in Florid in 26 hs been extensively investigted, but the bsic question of wht hppened with the electronic voting equipment used there remins unresolved. An unusully lrge number of bllots cst in Srsot County did not include vote for tht office: 18,412 of the 238,249 bllots cst on ivotronic touchscreen mchines in the county were undervotes for tht rce.' Some of the reserch regrding the high * Prepred for the symposium, "How We Vote," Institute of Bill of Rights Lw, Willim & Mry School of Lw, Willimsburg, Virgini, Mrch 14, thnk Dvid Dill, Murice Tmmn nd Kitty Grber for dt, nd I thnk Herbie Ziskend for ssistnce. ** Professor of Politicl Science nd Professor of Sttistics, University of Michign Hven Hll, Ann Arbor, Ml Phone: Emil: wmebne@umich.edu. Voting Imge nd Log Files from Dvid Dill, Professor, Stnford University, Jennifer J. Edwrds, Collier County Supervisor of Elections, Kitty Grber, Reserch Director, Florid Fir Elections Center, Murice Tmmn, Stff Writer, Srsot Herld Tribune (26) [hereinfter Voting Imge nd Log Files], vilble t WRMreplic 13nov28.zip.

3 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 undervote rte hs occurred in the context of litigtion, 2 some hs been done by independent scholrs, 3 some ws sponsored by the Stte of Florid, 4 nd some ws conducted by the federl government. 5 The unusully high undervote rte probbly chnged the election outcome. 6 Tems of computer scientists exmined severl fetures of the softwre nd hrdwre used to conduct the election but found nothing they considered sufficient to wrrnt ttributing the lost votes to defects in the equipment's opertions. 7 Even though the dequcy of these technicl exmintions hs been seriously questioned,' mny red the technicl reports s lrgely exculpting the mchines. 9 Hence, some rgue tht the high undervote rte ws cused by peculir fetures of the bllot's formt tht confused mny voters.' Frisin et l., in prticulr, use dt from severl counties in Florid to show tht the undervote rte for the ttorney generl rce ws lso unusully high in severl counties when the choices for ttorney generl-for which, like CD- 13, there were only two cndidtes running-were plced on the sme screen s the governor's rce, which hd mny cndidtes." The correltion they demonstrte between the fetures 2 See id. t 5, 8-9; Declrtion of Chrles Stewrt III on Excess Undervotes Cst in Srsot County, Florid for the 13th Congressionl District Rce, Jennings v. Elections Cnvssing Comm'n of the Stte of Fl., No. 26-CA-2973, 26 WL (Fl. Cir. Ct. Nov. 21, 26). See Arlene Ash & John Lmperti, Florid 26: Cn Sttistics Tell Us Who Won Congressionl District-13?, CHANCE, Spring 28, t 18; Lurin Frisin et l., Bllot Formts, Touchscreens, nd Undervotes: A Study of the 26 Midterm Elections in Florid, 7 ELECTION L.J. 25 (28); Wlter R. Mebne, Jr. & Dvid L. Dill, Fctors Associted with the Excessive CD-13 Undervote in the 26 Generl Election in Srsot County, Florid (Jn. 23, 27) (unpublished mnuscript), vilble t FloridFolder/smchines 1.pdf. 4 See, e.g., ALEC YASINSAC El AL., SOFrwAREREvIEW AND SECURITY ANALYSIS OF THE ES&S IVOTRoNIc VOTING MACHINE FIRMWARE (27), vilble t 5 U.S. GOV'TACCOUNTABILTY OFFICE, GAO-8-97T, FURTHER TESTING COULD PROVIDE INCREASED BUT NOT ABSOLUTE ASSURANCE THAT VOTING SYSTEMS DID NOT CAUSE UNDER- VOTES IN FLORIDA'S 13TH CONGRESSIONAL DISTRICT (27), vilble t U.S. GOV'TACCOUNTABILrrY OFFICE, GAO-8-425T, RESULTS OF GAO'S TESTING OF VOTING SYSTEMS USED IN SARASOTA COUNTY IN FLORIDA'S 13TH CONGRESSIONAL DISTRICT (28) [hereinfter RESULTS OF GAO'S TESTING], vilble t 6 See Ash & Lmperti, supr note 3, t See RESULTS OF GAO'S TESTING, supr note 5, t 3-4; YASINSAC ET AL., supr note 4. 8 See DAVID L. DILL& DAN S. WALLACH, STONES UNTURNED: GAPS IN THE INVESTIGATION OFSARASOTA'S DSPUTEDCONGRESSIONALELECTION (27), vilble t ' See, e.g., Doug Sword, Mchines Clered in 26 Vote Flp: GAO Sys Touch-Screens Didn't Cuse Undervote, SARASOTA HERALD-TRIB. (Fl.), Feb. 7, 28, t B 1. 'o See Frisin et l., supr note 3, t " Id. t 26, 32.

4 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 377 of the bllot nd the ttorney generl undervote rte is cler, but the explntion entirely in terms of voter confusion is specultive. There is no direct supporting evidence bout the voters' experiences, lthough Selker reports some suggestive results bsed on experiments conducted using subjects from Boston, Msschusetts.1 2 Frisin et l. point out, however, tht their results set the br for ny competing explntion firly high: "[A]ny explntion for the CD 13 undervote in Srsot must be cpble of explining the ttorney generl undervotes in Chrlotte nd Lee Counties nd the lck of n ttorney generl undervote elsewhere (not to mention the lck of high undervotes in other rces)."' 3 Recent reports exmining the use nd performnce of touchscreen voting mchines cross Florid in the 26 election open the door to meeting the stndrd Frisin et l. propose. 4 Pynchon nd Grber document extensive problems with the ivotronic mchines tht go well beyond the scope of the officilly sponsored technicl studies.'" They show tht undervote rtes for mny rces were higher in mny counties where ivotronic touchscreen mchines were used, regrdless of the bllot formt.' 6 Among the physicl problems they identify with the equipment, prticulrly interesting is the suggestion tht voting mchine power filures occurred frequently with strong dverse effects. 7 Grber remrks tht "power supply filures were experienced by voters nd poll workers s screen problems... [Plower problems cn ffect the responsiveness of the screen. In fct, one of the first symptoms of power problem is diminished responsiveness."'" Grber demonstrtes tht n error messge tht indictes tht voting mchine hd power filure correltes strongly with high undervote rtes in Chrlotte County. 9 2 Ted Selker, Study Shows Bllot Design nd Voter Preprtion Could Hve Eliminted Srsot Florid Voting Errors (Cltech/MIT Voting Tech. Project, VTP Working Pper No. 61, 28). 13 See Frisin et l., supr note 3, t Id. t 4. "5 See SUSAN PYNCHON & KrTY GARBER, SARASOTA'S VANISHED VOTES: AN INVES- TIGATION INTO THE CAUSE OF UNCOUNTED VOTES IN THE 26 CONGRESSIONAL DISTRICT 13 RACE IN SARASOTA COUNTY, FLORIDA (28), vilble t 16 Id. t See Krrry GARBER, LOST VOTES IN FLORIDA'S 26 GENERAL ELECTION: A LOOK AT EXTRAORDNARY UNDERVOTE RATES ON THE ES&S IVOTRONIC (27), vilble t l.pdf; KIrY GARBER, LOST VOTES IN FLORIDA'S 26 GENERAL ELECTION: A LOOK AT EXTRAORDINARY UNDERVOTE RATES ON THEES&S IVOTRONIC (28) [hereinfter GARBER, LOST VOTES PART 21, vilble t PYNCHON & GARBER, supr note 15, t GARBER, LOST VOTES PART 2, supr note 17, t 9, 32. '9 See id. t

5 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 Indeed, in tht county, mchines in the sme precinct s mchine with power filure messge lso tend to exhibit high undervote rtes. 2 ' Grber suggests tht power-relted problems propgted mong mchines becuse "counties link their mchines together t the polling plce in wht is termed 'disy chin.' The first mchine is plugged into the wll outlet, the second is plugged into the first, the third into the second, nd so on." 21 Another possible hole in Frisin et l.'s explntion is their filure to recognize tht with ivotronic touchscreen equipment, distinctive bllot pttern is in fct distinctive softwre nd hrdwre configurtion. 22 ivotronic mchines use device clled "Personlized Electronic Bllot" (PEB) to lod the cndidte selection displys into the voting mchine for ech voter. 23 Indeed, most interctions with the ivotronic mchines involve the use of PEB. 24 When bllot hs different ppernce, the PEBs contin different progrmming. 25 None of the technicl exmintions of the equipment used in Srsot ever exmined the PEBs tht were used to conduct the election. 26 Mebne nd Dill demonstrte tht severl PEB-relted error messges correlte significntly with vritions in the Srsot CD-13 undervote rte. 27 It is possible, then, tht the dverse consequences of bllot formt tht Frisin et l. demonstrte should be explined, t lest in prt, s results of deficient mechnicl opertions connected to the PEBs. 28 Indeed, s I will show in the reminder of this Article, both power filure problems nd PEB-relted problems correlte significntly with undervote rtes in severl Florid counties. 29 The nlysis includes dt from nine Florid counties tht used ivotronic touchscreen voting mchines: Chrlotte, Collier, Lke, Lee, Mrtin, Mimi- Dde, Psco, Srsot, nd Sumter counties. 3 Three of these counties (Chrlotte, Lee, nd Sumter) used bllot formts precisely of the kind Frisin et l. flg s problemtic for the ttorney generl rce." Bllot formts used in Mimi-Dde County 2 Id. t Id. t See Frisin et l., supr note 3 (rguing tht the bllot design, not the technology, ws to blme for undervotes in Srsot County). 23 Mebne & Dill, supr note 3, t See id. (noting tht PEBs re used for opening nd closing the ivotronic mchines, for cncelling bllots, nd for voting). 25 See YASINSAC ETAL.., supr note 4, t 1 (explining tht election stff cretes unique election identifiction codes, defines contests, nd identifies the cndidtes in ech contest; ech PEB contins this informtion for the specific precinct t which it will be used). 26 PYNCHON & GARBER, supr note 15, t Mebne & Dill, supr note 3, t Frisin et l., supr note 3, t See infr p. 388 tbl.6, pp tbl.7, pp tbl.8, p. 391 tbl.9, p. 392 tbl.1, p. 393 tbl.11, p. 394 tbl See infr pp tbl.1, pp tbl See Frisin et l., supr note 3, t

6 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 379 exhibited the key feture of hving offices with differing numbers of cndidte choices on the sme screen. 32 Lke nd Psco Counties used bllot formts with offices rrnged in two columns on ech screen insted of the single-column formts used elsewhere. 33 Bllot formts used in the remining counties lck purportedly problemtic fetures.' The prevlence of observble problems with PEBs nd power vries cross the counties, nd the pprent consequences for undervote rtes re diverse. The mgnitude of the effects vries both cross counties nd cross offices, tending to be lrger for rces tht hve substntilly lrger undervote rtes. 35 The results, I suggest, brethe new life into the likelihood tht high rtes of undervoting in Florid in 26 were substntilly due to mechnicl filures. MECHANICAL EVENTS AND UNDERVOTES The foundtion for my nlysis is the set of vote imge nd event log files tht report, respectively, ech bllot cst nd every trnsction occurring on ech ivotronic voting mchine used by ech county during the 26 election. 36 Ech vote imge file contins records showing every cndidte selection mde on n individul bllot, including code to indicte the voting mchine, PEB, nd bllot style used to cpture the selections. 37 Cst bllots tht contin no cndidte selections re lso shown. 38 Ech event log file shows mny of the ctions tken on ech mchine, including timestmp for ech trnsction. 39 Tbles I nd 2 summrize the event log file contents for the counties used in the current nlysis. In the event log files, ech trnsction is described by numeric code nd brief descriptive phrse Events 2 nd 21 correspond to vote csting events. 41 For ech of these events, there is corresponding record in the vote imge file. 42 Becuse the vote records re supposedly included in the vote imge file in rndom order, to protect voter nonymity, it is not possible to mtch prticulr set of votes to prticulr event log file trnsction. 32 Michel C. Herron, Bllot Formt Effects in the 26 Midterm Elections in Florid, t slides (Dec. 2, 26), 33 Id. t slide 21. Id. t slide See infr pp tbl.4, p. 388 tbl.6, pp tbl.7, pp tbl.8, p. 391 tbl.9, p. 392 tbl.1, p. 393 tbl.11, p. 394 tbl Voting Imge nd Log Files, supr note id. 38 id. 39 id. 4 See id. 41 id. 42 Id.

7 38 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 Tble 1: Events Recorded in Event Log Files for Severl Florid Counties in 26 County Code Description Chrlotte Collier Lke Lee Mrtin 1 Terminl cler nd test 2 Terminl screen clibrte 4 Enter service menu 5 Service pssword Enter ECA menu ECA pssword 8 Dte/time chnge 9 Terminl open 1 Terminl close 12 Audit uplod 13 Print zero tpe 14 Print Precinct results 15 Modem Precinct results 17 Votes recollect 18 Invlid vote PEB 19 Invlid super PEB 2 Norml bllot cst 21 Super bllot cst 22 Super bllot cncel 25 Open with super votes 26 Terminl left open

8 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED Override 28 Override fil 31 Term. cler/test pssword fil 35 Modem Precinct results fil 36 Low bttery lockout 37 Nonmster PEB collection 49 Internl mlfunction 5 L nd A test run-votes clered 51 PEB/CF Election ID mismtch Tble 2: Events Recorded in Event Log Files for Severl Florid Counties in 26 Code Description Mimi-Dde County Psco Srsot Sumter 1 Terminl cler nd test 2 Terminl screen clibrte 4 Enter service menu 5 Service pssword fil 6 Enter ECA menu 7 ECA pssword fil 8 Dte/time chnge 9 Terminl open 1 Terminl close 12 Audit uplod 13 Print zero tpe

9 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17: Print Precinct results 15 Modem Precinct results 17 Votes recollect 18 Invlid vote PEB 19 Invlid super PEB 2 Norml bllot cst 21 Super bllot cst 22 Super bllot cncel 25 Open with super votes 26 Terminl left open 27 Override 28 Override fil 31 Term. cler/test pssword fil 35 Modem Precinct results fll 36 Low bttery lockout 37 Nonmster PEB collection 49 Internl mlfunction 5 L nd A test run-votes clered 51 PEB/CF Election ID mismtch Severl trnsction codes in the event log files indicte events tht my be relevnt for studying undervote rtes. A need to clibrte the terminl screen (event 2) my suggest mislignment between where vote choices disply on the screen nd

10 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 383 where touching the screen ctivtes ech choice. 43 Such mislignment cn cuse votes to be misrecorded or lost.' Other events indicte problems with psswords or possibly inpproprite ccess to service menus, nd suggest possible security filures (events 5, 6, 7, nd 31). Still other events indicte vrious mlfunctions (events 36 nd 49) or some problem with PEBs (events 18, 19, 37, nd 51). The record of problems contined in the event log files is not necessrily complete. There re severl documented instnces in which problems with ivotronic mchines my hve cused the event log files to be corrupted.' A bug dignosed in in prticulr cuses incomplete records in some cses where the power supplied to mchine fils.' Whether tht bug or other problems ffected the event log files vilble from the 26 election is unknown. For the current nlysis I focus on two kinds of events. Event 18 ("Invlid vote PEB") flgs PEB-relted error nd ws focus of study of dt from Srsot County by Mebne nd Dill. 47 Event 36 ("Low bttery lockout") is the event used by Grber to indicte tht voting mchine experienced power filure. 48 Becuse it is not possible to mtch prticulr events to prticulr individul vote records, nor in generl to tell whether prticulr vote occurred prior to n event of interest, I define vribles to indicte whether prticulr voting mchine hs n occurrence of event 18 (E18) or event 36 (E36). In light of Pynchon nd Grber' s observtion bout voting mchines being "disy chined," with the likelihood tht ny consequences of power problems on one mchine propgted to other mchines connected to it, 49 I lso define vrible to indicte whether voting mchines were in the sme loction s nother mchine tht hd power filure. No dt exist to show which mchines were ctully connected to one nother in ech polling plce. Grber used the fct of being in the sme precinct to mesure such connections on election dy in Chrlotte County. 5 ' I use n lterntive proxy: mchines re considered to be relted ccording to E36 if the sme PEB ws ever used on both of the mchines. If mchine A is relted to mchine B by PEB, nd mchine B is relted to mchine C by nother PEB, then ll three mchines re considered to be prt of the sme PEB cluster. All mchines in the sme PEB cluster re considered to be t the sme loction. The power-loction vrible E36L is set on for ll mchines in cluster whenever ny mchine in the cluster hs n occurrence of E36. The rtionle for this mesure is tht mchines tht were used with the sme PEB probbly were locted close to one nother in the polling plce, nd mchines tht were locted close to one nother were more likely to hve been disy chined. 13 See Dill & Wllch, supr note 8, t 4. 4Id. 4' See PYNCHON & GARBER, supr note 15, t Id. t '4 See Mebne & Dill, supr note 3, t GARBER, LOST VOTES PART 2, supr note 17, t PYNCHON & GARBER, supr note 15, t 4. 5 GARBER, LOST VOTES PART 2, supr note 17, t

11 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 Tble 3 shows the distribution of the three event-mesuring vribles in terms of the number nd proportion of bllots cst on the ffected mchines in ech county. E18 occurs much more frequently during erly voting thn on election dy. 5 Typiclly bout hlf of the erly voting bllots re on mchines tht hve n occurrence of E18, while on election dy the frequency rnges from low of 6.8% in Sumter County up to bout qurter of the bllots in Lee, Mrtin, nd Mimi-Dde Counties. The frequency of E36 is reltively low on election dy, rnging from 1.2% in Sumter County to 3.9% in Lke County. Tking PEB clusters into ccount in most counties gretly increses the number of bllots potentilly ffected. Sumter County remins reltively low, with 3.1% of bllots hving positive vlue for E36L, but the percentge for Lke County jumps to 26.6, nd Mimi-Dde County tops the percentges with vlue of Only four counties (Chrlotte, Collier, Lke, nd Lee) show substntil proportions of bllots on mchines or in PEB clusters with power filure indictors during erly voting. Tble 3: Frequency of Invlid Vote PEB nd Low Bttery Lockout Events Percentges County E18 Election Dy E36 E36L E18 Erly Voting E36 E36L Chrlotte Collier Lke Lee Mrtin Mimi Psco Srsot Sumter Counts County E18 Election Dy E36 E36L E18 Erly Voting E36 E36L Chrlotte Collier " Whether mchine ws used during erly voting or on election dy is determined from the timestmps for the events reported for the mchine. For Mimi-Dde County, none of the event records hve timestmp prior to election dy. See Voting Imge nd Log Files, supr note 1.

12 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 385 Lke Lee Mrtin Mimi Psco Srsot Sumter Note: Percentges nd counts of bllots on mchine with the indicted event in ech county. E18 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 on mchine in the bllot's PEB cluster, Erly voting dt not vilble. A bllot hs n undervote for n office if the vote imge does not include cndidte choice for the office. 2 For offices tht pper on the bllot only in specific jurisdictions, such s U.S. House elections, non-vote is n undervote only if the vote imge is for precinct nd bllot style tht indictes the voter is eligible to vote for the cndidtes in the referent jurisdiction. I focus on votes for five sttewide offices (Governor, Attorney Generl, Chief Finncil Officer (CFO), Commissioner of Agriculture, nd U.S. Sente) nd for sets in the U.S. House of Representtives. Tble 4 shows the distribution of undervotes for these vrious offices cross the nine counties. Undervote rtes re typiclly lowest for governor nd U.S. Sente, nd in mny instnces they re lso low for U.S. House sets. Undervote rtes re noticebly higher for commissioner of griculture nd CFO, nd they vry quite widely for ttorney generl. The three ttorney generl undervote rtes tht re greter thn 2%-for Chrlotte, Lee nd Sumter Counties-re for the counties where the ttorney generl rce ppered on the sme pge with the governor's rce. 3 Among the U.S. House rces, the CD-13 rce hs the second highest undervote percentge. 4 The result for the U.S. House rce with the highest percentge-cd-17 in Mimi-Dde-is lmost certinly explined by the fct tht the Democrtic cndidte in tht district did not fce mjor prty chllenger. Tble 4: Undervote Percentges by Office County Office Chrlotte Collier Lke Lee Mrtin Mimi Psco Srsot Sumter Att'y Gen Comm'r Agric See Voting Imge nd Log Files, supr note 1. 5 See PYNCHON & GARBER, supr note 15, t id.

13 386 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 CFO Governor US House US House _ US House _ US House US House US House US House US House US Housel _ US House _ - US House US House US Sente Note: Undervotes s percentge of ll bllots cst. Results for "US House" with district include only bllots in the referent district. "US House" with no district combines ll districts for ech county. Office not included in county. The stndrd proposed by Frisin et l. clls for effects of the invlid PEB nd power filure events on the undervote rte tht vry cross counties, 55 but nonetheless, it is worthwhile to begin by looking t the verge rte of undervoting for ech combintion of the events, ignoring ny differences mong counties. Tble 5 shows undervote rtes for ech office for combintions of the two kinds of events, seprtely for election dy nd erly voting bllots but pooling cross ll counties. For this disply, the undervote counts for ll U.S. House sets re treted together. Tble 5: Undervote Percentges by Occurrence of Mchine Events No E36 No E36L Election Dy Erly Voting Election Dy Erly Voting Office NoE18 E18 NoE18 E18 NoE18 E18 NoEl8 E18 Att'y Gen Comm'r Agric CFO " See Frisin et l., supr note 3 (observing nomlous undervoting t the precinct level).

14 28] MACHNE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 387 Governor US House US Sente E36 E36L Election Dy Erly Voting Election Dy Erly Voting Office No E18 E18 No El8 E18 NoEl8 E18 No E18 E18 Att'y Gen Comn'r Agric CFO Governor US House US Sente Note: Undervote percentges with nd without the indicted type of event, pooling dt from ll nine counties. E18 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 on mchine in the bllot' s PEB cluster. "US House" includes ll districts in ll counties. In the subtble showing the brekdowns mong the election dy bllots, in the bsence of power filure indiction for the prticulr mchine (No E36) or for the PEB cluster (No E36L), the undervote rte is lwys higher on mchines tht hve n invlid vote PEB event thn on mchines tht do not hve such n event. The lrgest such differences re differences of 1% for ttorney generl in the No E36 cse nd for the CFO in the No E36L cse. Other rces in these sme conditions typiclly show difference of bout hlf percent in the undervote rte cross vlues of El 8, except for the governor rce where the differences re smller. In the subtble showing the election dy brekdowns in the presence of power filures (E36 or E36L), the pttern of differences in the undervote rte cross vlues of E18 is similr except the differences re smller. Also, there is one reversl: for U.S. House with E36, the undervote percentge is smller with E18 thn it is with No El8. An invlid vote PEB event is on verge typiclly ssocited with higher undervote rte on election dy. In the subtbles showing the undervote percentge brekdowns for erly voting bllots, the differences with respect to E18 do not exhibit consistent pttern. Focusing on the differences with respect to E36 nd E36L for comprble vlues of E 18, mong the election dy bllots the undervote rte is typiclly greter for n office for the bllots tht hve E36 or E36L thn for bllots tht hve No E36 or No E36L. Power filures re on verge typiclly ssocited with higher undervote rte on election dy. Agin, the pttern of differences mong the erly voting bllots is mixed.

15 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 In terms of gross verges, then, ech of the two types of voting mchine events is typiclly ssocited with tendency to hve higher undervote rte, t lest mong election dy bllots. In view of the lrge number of bllots being considered cross ll nine counties, it is perhps not surprising tht most of the differences pprent in Tble 5 re sttisticlly significnt if ll of the votes for ech office re treted s sttisticlly independent of one nother. Nonetheless, the differences cross conditions in Tble 5 re smll s percentges of bllots cst, nd of course the overll verges do not ddress the need to demonstrte tht the differences vry cross counties. Tble 6 presents the first set of results tht ber on the question of diverse effects. The tble shows the simple percentge difference in the undervote rte for ech office in single county when the condition mesured by ech of the event vribles is, respectively, present versus bsent. I ssess the sttisticl significnce of ech difference by testing for independence between ech event vrible nd vrible mesuring whether ech bllot hs n undervote for the respective office. The test level is the conventionl vlue.5, but symbols in the tble lso ddress the fct tht cross the whole nlysis we re looking t dozens of seprte tests. One symbol indictes tht n ssocition is significnt even when we djust for the number of tests being done for n event in the prticulr county, nd one indictes significnce even when we djust for the number of tests being done in ll nine counties. 56 Tble 6: Undervote Percentge Differences by Occurrence of Mchine Events I Election Dy Srsot Erly Voting Office E18 E36 E36L E18 E36 E36L Att'y Gen Comm'r Agric CFO Governor..3.3* US House 13.9* 2.3*.8* -2.4t US Sente Note: Differences between undervote percentges with nd without the indicted type of event. El8 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 on mchine in the bllot's PEB cluster. *: Two-wy ssocition significnt t level.5; :: t level.5 djusting for tests cross ll counties use the Bonferroni djustment. Letting mjdenote the number of offices being nlyzed in countyj nd m the number summing cross ll nine counties the single test level is.5, the within county test level is.5/mj, nd the cross-county test level is.5/m. Person chi-squred is the test sttistic.

16 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 389 The results in Tble 6 show undervote rte differences for Srsot County. The substntil nd significnt differences occur minly for the CD- 13 rce. There re significnt differences for E18 mong both the election dy nd the erly voting bllots. Oddly, the direction of the difference vries between the two election periods. The undervote rte is lower on mchines with n invlid vote PEB event during erly voting, but the rte is higher on such mchines on election dy. Also, mong the election dy bllots, the undervote percentge is higher in the presence of power filure event. The effect is smller for E36L thn for E36, but it is lrge enough to give some support to the ide tht the consequences of power filure propgted mong nerby mchines. 5 7 The results for Chrlotte nd Sumter Counties likewise show substntil differences in the undervote rte occurring lrgely for the office tht ws subject to the peculir bllot feture. 5 " In both counties there re significnt reltionships between power filure events nd election dy ttorney generl undervote rtes. Tble 7 shows the significnt differences for Chrlotte County re of comprble mgnitude whether mchine-specific events or PEB cluster events re considered. There re lso significnt power-relted increses in the undervote rte for few of the other offices, but these increses re smller in mgnitude thn the increses for the ttorney generl rce. Tble 7 shows tht in Sumter County power filures re ssocited with significnt decreses in the election dy ttorney generl undervote rte. In Sumter on election dy, there is lso significnt increse in the CD-5 undervote rte ssocited with mchine-specific power filure events. Neither county exhibits significnt undervote rte differences ssocited with invlid vote PEB events, except for smll decline in the erly voting governor undervote rte tht chieves single-test significnce. Tble 7: Undervote Percentge Differences by Occurrence of Mchine Events H Chrlotte Election Dy Erly Voting Office E18 E36 E36L E18 E36 E36L Att'y Gen t 4.6f: Comm'r Agric * 1.8:.4.7. CFO.1. 1.* * Governor US House.2.6.7* " Tble 3 shows tht nine times s mny bllots hve positive vlue for E36L s hve positive vlue for E36, but the percentge difference in Tble 6 is not correspondingly nine times smller. Compre supr pp tbl.3, with supr p. 388 tbl.6. 5 See infr pp tbl.7.

17 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 US House 13 US House 14 US House 16 US Sente _ b b * Office Att'y Gen. Comm'r Agric. CFO Governor US House 5 US Sente E Election Dy E36-6.6* * -.1 E36L -5.5* Sumter E Erly Voting E36 b b b b E36L b b b b b b Note: Differences between undervote percentges with nd without the indicted type of event. E18 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 on mchine in the bllot's PEB cluster. *: Two-wy ssocition significnt t level.5; t: t level.5 djusting for ll tests in this county; t: t level.5 djusting for tests cross ll counties. Results for "US House" with district number include only bllots in the referent district. "US House" with no district number combines ll districts for ech county. b No events. Lee County provides the principl exception to the generl finding tht undervote rtes for the office tht ppered with the peculir bllot feture re significntly ssocited with invlid vote PEB or power filure events. Tble 8 shows no significnt differences for the ttorney generl rce. A significnt increse in the election dy commissioner of griculture undervote rte is ssocited with invlid vote PEB events. Tble 8: Undervote Percentge Differences by Occurrence of Mchine Events EI Lee Election D) Office E18 E36 E36L E18 Erly Voting E36 E36L Att'y Gen. Comm'r Agric. CFO

18 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 391 Governor US House *.1 US Sente Note: Differences between undervote percentges with nd without the indicted type of event. E18 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 on mchine in the bllot's PEB cluster. *: Two-wy ssocition significnt t level.5; t: t level.5 djusting for ll tests in this county. The results for Collier County, reported in Tble 9, show no significnt election dy undervote rte differences. Among the erly voting bllots one significnt increse in undervote rtes (for the governor rce) is ssocited with invlid vote PEB events, one is ssocited with mchine-level power filures (for the U.S. Sente rce) nd one is ssocited with PEB cluster-level power filures (for the combined U.S. House rces, but principlly for the CD- 14 rce). Frisin et l. plce Collier County in contrst to Chrlotte nd Lee Counties becuse the bllot formt used in Collier did not hve the purportedly confusing feture of rces with very different numbers of cndidte choices plced together on the sme screen. 59 Frisin et l. do not consider undervote rtes in the erly voting dt.' Tble 9: Undervote Percentge Differences by Occurrence of Mchine Events IV Collier Election Dy Erly Voting Office E18 E36 E36L E18 E36 E36L Att'y Gen Comm'r Agric CFO Governor * US House * US House US House US Sente *.1 Note: Differences between undervote percentges with nd without the indicted type of event. E18 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 '9 See Frisin et l., supr note 3, t See id.

19 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 on mchine in the bllot's PEB cluster. *: Two-wy ssocition significnt t level.5. Results for "US House" with district number include only bllots in the referent district. "US House" with no district number combines ll districts for ech county. The bllot formt used in Mrtin County exhibited no remrkble fetures, yet Tble 1 shows significnt ssocition between mchine events nd undervote rtes for couple of offices. For CFO there is n erly voting ssocition with invlid vote PEB events nd n election dy ssocition with mchine-level power filure events. There is lso significnt negtive ssocition between election dy ttorney generl undervotes nd PEB cluster-level power filure events. Grber observes tht the voting mchines in Mrtin County hd mintennce histories tht differed in severl respects from wht ws done in other Florid counties nd tht officils in Mrtin County exerted specil efforts to reclibrte the screens of the voting mchines. 6 Tble 1: Undervote Percentge Differences by Occurrence of Mchine Events V Election Dy Mrtin Erly Voting Office E18 E36 E36L E18 E36 E36L Att'y Gen *.2 -b -b Comm'r Agric b CFO.6 1.9*.1 1.ot -b -b Governor b -b US House b -b US Sente b -b Note: Differences between undervote percentges with nd without the indicted type of event. E18 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 on mchine in the bllot's PEB cluster. *: Two-wy ssocition significnt t level.5; t: t level.5 djusting for ll tests in this county. b No events. The bllot formt used in Mimi-Dde County differed in significnt respects from the bllot formt used in the other counties, feturing in prticulr the property of hving offices with very different numbers of cndidte choices together on the sme screen. 62 As Tble 11 shows, the county lso exhibits lrge number of significnt ssocitions between election dy undervote rtes nd mchine events. Every office except CD-21 hs significnt ssocition between the undervote rte nd 61 See GARBER, LOST VOTES PART 2, supr note 17, t See Herron, supr note 32, t slides 23, 24.

20 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 393 either invlid vote PEB events or power filure events, nd severl offices show significnt ssocitions with both kinds of events. Tble 11: Undervote Percentge Differences by Occurrence of Mchine Events VI Election Dy Mimi-Dde Erly Voting Office E18 E36 E36L E18 E36 E36L Att'y Gen..3*.7t Comm'r Agric..1.7* CFO.6t Governor.2It US House.8t t US House L.ot - - US House 18.6* 1.1* -.1 US House US House t US Sente.1.3.2* - - Note: Differences between undervote percentges with nd without the indicted type of event. E18 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 on mchine in the bllot's PEB cluster. *: Two-wy ssocition significnt t level.5; t: t level.5 djusting for ll tests in this county; t: t level.5 djusting for tests cross ll counties. Results for "US House" with district number include only bllots in the referent district. "US House" with no district number combines ll districts for ech county. Erly voting dt not vilble. Lke nd Psco Counties differ from the other ivotronic counties becuse the bllot formts used in the two counties presented choices on the voting mchine screens in two columns. 63 Tble 12 shows tht between the two counties the ssocitions between undervote rtes nd mchine events differ considerbly. In Lke County, undervotes for lmost every office re significntly ssocited with one or the other type of mchine event, nd sometimes with both types. Only the governor rce nd perhps the rce for CD-8 re exempt. In Psco County, only two offices exhibit significnt ssocition between undervote rtes nd one of the mchine events: PEB cluster-level power filures re negtively ssocited with election dy 63 See Herron, supr note 32, t slide 21; see lso E-mil from Kitty Grber, Reserch Director, Florid Fir Elections Center, to Wlter R. Mebne, Jr., Professor of Politicl Science nd Sttistics, University of Michign (Mr. 8, 28, 2:2:55 EST) (on file with uthor).

21 394 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 ttorney generl undervotes nd positively relted to U.S. House undervotes. The ltter ssocition mnifestly does not describe the seprte congressionl districts, so on the whole the evidence for substntil mchine event ssocitions with undervote rtes in Psco County should be viewed s wek. It is not the ppernce of the bllots tht shrply distinguishes the undervote experience in these two counties. Rther, it is the ssocition with voting mchine error conditions tht differs. Tble 12: Undervote Percentge Differences by Occurrence of Mchine Events VII Election Dy Lke Erly Voting Office E18 E36 E36L E18 E36 E36L Att'y Gen * Comm'r Agric * CFO.2 1.2* Governor US House.7* 1.5*.7t t US House * US House * US House US Sente.1.7t *.9.5* Election Dy Psco Erly Voting Office E18 E36 E36L E18 E36 E36L Att'y Gen *.1 -b -b Comm'r Agric b CFO b -b Governor b -b US House.4.6.5*.3 - -b US House b -b US House b -b US Sente b -b Note: Differences between undervote percentges with nd without the indicted type of event. E18 nd E36: events 18 nd 36 on the bllot's mchine. E36L: event 36 on mchine in the bllot's PEB cluster. *: Two-wy ssocition significnt t level

22 28] MACHINE ERRORS AND UNDERVOTES IN FLORIDA 26 REVISITED 395.5; t: t level.5 djusting for ll tests in this county;f: t level.5 djusting for tests cross ll counties. Results for "US House" with district include only bllots in the referent district. "US House" with no district combines ll districts for ech county. bno events. CONCLUSION Plinly the two kinds of PEB nd mchine events exmined here re not sufficient to explin the pttern of undervotes, neither in the CD- 13 rce in Srsot County nor in the other Florid counties included here. The mesured events re not lwys ssocited with high undervote rtes (e.g., Lee County), nd even where there re strong ssocitions, the mgnitude of the pprent effects is usully too smll to fully ccount for high proportion of the undervotes tht occurred. In few instnces, the invlid vote PEB or power filure events re ssocited with declines, not increses, in the percentge of undervotes. Citing the kinds of events studied here hrdly settles the question of wht cused the excessive numbers of undervotes. The point of the current exercise is not to supply n dequte explntion for the undervotes, but merely to demonstrte tht the reltionships between mchine events nd undervotes re sufficiently substntil nd vried to mke it unresonble to discount the likelihood tht mechnicl filures contributed substntilly to the high numbers of undervotes. In three of the four cses considered by Frisin et l.' nd Selker, 65 where bllot formt purportedly confused voters nd consequently produced drmtic increses in undervoting-the CD-13 rce in Srsot County nd the ttorney generl rce in Chrlotte, Lee nd Sumter Counties-I find significnt ssocitions between invlid vote PEB events or power filure events nd substntil vritions in undervoting, prticulrly for the offices of interest. Even if one focuses nrrowly on those four rces in those plces, there is every reson not to rule out mechnicl effects. Undervotes lso pper significntly relted to events for other offices in other counties. Of the nine counties exmined here, only Collier nd Psco Counties emerge reltively unscthed in terms of significnt ssocitions between events nd election dy undervotes. When erly voting is included, no county escpes unblemished. When thinking bout the smll mgnitude of the percentge differences ssocited with the invlid vote PEB nd power filure events, it is importnt to remember tht the event log file entries tht re the bsis for mesuring these conditions re t best symptoms of whtever ws wrong with the PEBs, mchines, or polling environments where the problems occurred. There is reson to believe the event logs do not include every occurrence even of the precise kinds of flws in focus here. Recll, for instnce, the softwre bug dignosed in tht sometimes cused events not to be 64 See Frisin et l., supr note See Selker, supr note 12.

23 WILLIAM & MARY BILL OF RIGHTS JOURNAL [Vol. 17:375 reported in cses of power filure. 66 And it is merely hypothesis tht the two kinds of events highlighted here re especilly importnt mrkers for high undervotes. As we sw in Tbles 1 nd 2, severl other kinds of events might well lso indicte conditions tht produce excessive undervotes. The current nlysis most likely understtes how much mechnicl defects contributed to the problem of excessive undervotes in the 26 election. A definitive explntion for the undervotes requires technicl exmintions of the hrdwre, softwre, nd prctices used in the election tht go considerbly beyond the officilly sponsored reviews tht hve occurred. Detiled dministrtive nd mintennce records of the kind prtilly collected nd reviewed by Pynchon nd Grber 67 re needed to supplement the kinds of tests described by Dill nd Wllch 68 nd others. Unfortuntely, it is likely tht we will never hve sufficient explntion for the pttern of undervotes in the 26 election. Key equipment, such s the PEBs ctully used during the election, hs not been preserved. 69 Experts hve not been llowed sufficient ltitude to fully test even the equipment, both hrdwre nd softwre, tht ws preserved. 7 From some counties it is impossible to obtin the necessry kinds of dministrtive records, nd for others to do so is infesible. Nonetheless, simply becuse sufficient explntion is not forthcoming, there is no reson to tret prtil nd incomplete explntion s if it were dequte. While it is tempting to fll bck on n explntion tht blmes the voters-were voters confused?-there is not sufficient evidence to support doing tht. While undoubtedly voter confusion prompted by unfortunte bllot formts hd hnd in incresing the number of undervotes for some offices, the mgnitude of this effect is uncler, nd clerly tht is not ll of wht hppened. 66 See PYNCHON & GARBER, supr note 15, t See PYNCHON & GARBER, supr note See Dill & Wllch, supr note 8. 9 See PYNCHON & GARBER, supr note 15, t 29 (suggesting tht it my be too lte to exmine the PEBs ctully used in the election). 7 See GARBER, LOST VOTES PART 2, supr note 17, t 48.