By Gustavo Delfino Universidad Central de Venezuela Professor at the School of Mechanical Engineering and

Transcription

1 ANALYSIS OF THE VENEZUELAN PRESIDENTIAL RECALL REFERENDUM OF 2004 AND THE RELATIONSHIP BETWEEN THE OFFICIAL RESULTS AND THE SIGNATURES REQUESTING IT IN COMPUTERIZED CENTERS By Gutavo Delfino Univeridad Central de Venezuela Profeor at the School of Mechanical Engineering and By Guillermo Sala Univeridad Simón Bolivar Former Profeor at the Phyic Department On Augut 15th, 2004, Venezuelan had the opportunity to vote in a Preidential Recall Referendum to decide whether or not the Preident hould be removed from office. The proce wa largely computerized uing a touch creen ytem. In general the ballot were not manually counted. The ignificance of the high linear correlation (0.99) between the number of requeting ignature and the number of oppoition vote in computerized center i analyzed herein. The ame day audit wa found to be not only ineffective but a ource of upicion. Official reult were compared with the 1998 preidential election and other electoral event and ditortion were found. 1. Introduction. In Venezuela, electoral event are organized by the Conejo Nacional Electoral 1 (CNE). On the current Preident won the election with (57.79%) vote v (42.21%) vote for hi adverarie. The total number of voter in the electoral regitry (REP) at that time wa In 1999 a new contitution wa enacted which allow citizen to requet a recall referendum (RR) to decide whether the Preident hould continue or tep down from office. It can only be activated after half of the period for which the Preident ha been elected ha tranpired. In order to activate the referendum, a petition ha to be ent to the CNE igned by at leat 20% of the voter regitered in the REP. It i alo poible to requet a conultative 1 Before the new contitution it wa known a the Conejo Supremo Electoral (CSE) ee 1

2 2 GUSTAVO DELFINO & GUILLERMO SALAS non-binding referendum with ignature of 10% of the voter regitered in the REP. In a new CNE wa deignated but failed to organize election on chedule. Therefore, on , yet another CNE wa deignated. On the Preident wa reelected for a 6 year period with (59.76%) vote v (40.24%) for hi adverarie. The REP had regitered voter at that time. In 2002 ignature were collected requeting a conultative referendum which wa activated in the middle of a general national trike. The Supreme Court diabled the CNE, therefore the conultative referendum never took place. Citizen then collected new ignature for a recall referendum. Thi i the legal intrument which the government and oppoition repreented by the Coordinadora Democrática agreed to ue having the Organization of American State (OAS) and the Carter Center a guarantor[1]. Thi ended the trike. In 2003 the National Aembly wa unable to agree on a new CNE o the Supreme Court deignated a new temporary CNE on , even though thi procedure wa not contemplated in the Contitution. The new CNE rejected the ignature aying that they had been collected before half of the preidential period had tranpired. On the ignature were collected again, thi time under the uperviion of the CNE. On , an important part of the ignature had to be reverified by the CNE. Thi wa achieved, and on the Preidential Recall Referendum finally took place. 24 State + Embaie 335 Countie or Municipalitie 1133 Townhip or Parihe 4766 Computerized Voting Center 8142 Table Notebook Voting Machine Ballot Box 3628 Manual Voting Center 4209 Table 4209 Notebook Figure 1: Venezuelan electoral ytem layout.

3 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 3 2. Layout. Venezuela i politically organized into tate, countie (municipalitie), and townhip (parihe). Each county ha one or more voting center. There can be everal voting table (voting tation) per center, and each one ha one or more electoral notebook. In computerized center, one voting machine i aigned to each electoral notebook. One ballot box i aigned to each table. Therefore, the ballot from multiple machine are combined into a ingle box. Pleae ee Figure 1 for more detail. Each voting center ha a unique code which in time ha remained unchanged 2. Thi allow comparing electoral reult on a center by center bai. Although the number of manual center i large, the number of people regitered in thoe center i much maller than in computerized center. Thi can be een in the hitogram in Figure 2. Number of voting center Manual Center Computerized Center Number of people in REP Figure 2: Hitogram of ize of manual and computerized center. 3. The Referendum. Touch creen voting machine were ued for the firt time in Venezuela for the Referendum. Thee machine gave the voter a paper ballot to be depoited in a box. The boxe were never opened except for ome of thoe elected for auditing. The reult were ent from the machine to the CNE erver uing TCP/IP connection over telephone line, then the machine printed out the reult and a duplicate et of uncut paper ballot. The voting center had a continuou atellite TCP/IP connection to be ued only by the fingerprint machine which were uppoed to prevent anyone from voting twice, even in different voting center. In order to give the citizen confidence in the reult, two audit were made. The firt one wa made on the ame day of the event (hot audit). The econd one wa made three day later (cold audit). The whole electoral proce and audit were upervied and endored by the OAS and the Carter Center. They found no evidence of alteration (tampering) in the reult in their final report. 2 On 2006, all the voting center code where changed.

4 4 GUSTAVO DELFINO & GUILLERMO SALAS There were only two way to vote 3 : SI (ye) or NO. In order for the Preident to tep down, the number of SI vote had to be greater than and greater than the number of NO vote. The official reult were (40,64%) SI vote v (59,10%) NO vote, with regitered voter in the REP. Mot of thoe vote (87.1%) were cat at computerized center. 4. The Signature Introduction. In order to activate the Referendum, on ignature and fingerprint were collected in a four day event organized by the CNE, with witnee from all political partie. Special form, with erial number were upplied by the CNE to all political partie. There were 2676 ignature collection center (SCC), all of them in Venezuela. No ignature collection wa allowed outide Venezuela. There were two kind of form: type A and B. Type A form were ued in the SCC. Type B form were alo aigned to SCC, but they were meant to be ued for houe to houe ignature collecting (under pro government witne uperviion). There were type A form and type B form. Each form had a maximum capacity of 10 ignature. The number of ignature required to activate the Referendum wa 20% of the REP ued to elect the Preident, i.e = ignature. The law required the publihing of all ignature collected in a newpaper in order for thi to be publicly viewed. The CNE divided thee ignature into three categorie: valid, invalid and quetionable. An important number of quetionable ignature had to be recollected in order to reach the required minimum number of ignature. Oppoition group claimed to have ubmitted ignature to the CNE. Once inide the CNE, ignature were lot. An additional indeterminate number of ignature were lot before reaching the CNE. The preumed intention of the igner wa to vote YES 4. However, it wa poible that ome igner voted NO. Thi i the cae for government upporter who igned the petition becaue they believed they could ue the referendum to help olve the high tate of political confrontation. There were alo igner who changed their political preference between the time of the ignature collection and the vote. In the following ection, the reult of the referendum will be compared with the ignature collected. Thi will reveal important fact about thee 3 In manual voting center it wa alo poible to cat a null vote 4 The OAS and the Carter Certer concur with thi tatement. See [2], Section 5, econd paragraph.

5 reult. ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS SI Vote Uncertainty with Regard to Signature. Let k be the relative number of SI vote, a defined in Equation 1: (1) k = SI vote ignature Alo, let be relative number of ignature in a voting center, a defined in Equation 2: (2) = ignature SI vote + NO vote + Null vote = ignature total vote For each value of, there i a maximum poible k which i jut 1/ a hown in Equation 3: (3) k max = max(si vote) ignature = total vote ignature = total vote total vote = 1 In voting center with a large value of, it i expected to have a k around 1. Thi i becaue, each ignature tend to become a SI vote, and at the ame time k max get cloe to 1. For example, in a voting center with 1000 total vote and 900 ignature, the number of expected SI vote i between 900 and Here = 900/1000 = 0.9 and k max = 1/0.9 = Therefore, the uncertainty in the value of k i very mall, a it hould be between 1 and The value of k could be lower than 1 if, for any reaon, the number of vote wa low with repect to the electoral regitry (REP). The ituation i completely different in voting center with a mall value of. Notice that there i an eential ingularity in k at = 0 a hown in Equation 4: (4) k = SI vote/total vote Thi ingularity can produce very high value of k in the neighborhood of = 0. Hence, the level of uncertainty in k become very large. For example, in a voting center with 1000 total vote and 2 ignature, the number of expected SI vote i between 2 and Here = 2/1000 = and k max = 1/0.002 = 500. Therefore, the uncertainty in the value of k i extremely large, a it hould be between 1 and 500. The aforementioned reaon for uncertainty in k are purely mathematical. In practical term, high value of k in center with a mall happen due to the following fact:

6 6 GUSTAVO DELFINO & GUILLERMO SALAS There were only 2676 SCC compared to 8394 voting center. Therefore, people living far from a SCC could not ign the petition, even if they wanted to. Thi affected motly rural area. There were many people who did not ign the petition becaue of the fear of negative conequence from the government. On the other hand, voting wa ecret. There were SI vote, from people who could not ign becaue they were not in the REP or were outide the country at the time of ignature collection. Some SCC ran out of form. Not everyone wa able to go to a more ditant SCC. An undetermined amount of ignature got lot. There were SI vote from people who jut didn t bother to ign the petition. Notice that all thee iue with the ignature did not affect all voting center equally. Center with a mall value of are more likely to have been affected by thee iue than center with a high value of. A plot of k v. i hown in Figure 3. Notice that when i not large, all the computerized center are very far away from k max, clearly contradicting the expected non-linear behavior with repect to. On the other hand, the manual center reult are effectively ditributed in the allowed range regardle of the relative number of ignature Manual Voting Center Computerized Voting Center k (a) (b) Figure 3: Relationhip between k and for computerized and manual center. The hadowed area contain the mathematically impoible value of k. The maximum k value i 1/. The hollow dot repreent voting center located in conular office.

7 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 7 In ummary: It i expected that k from voting center with a mall value of will be much more variable than thoe with large value of Behavior of k with regard to the ize and characteritic of the voting center. Although the manual center tend to have fewer voter than the computerized center, thi doe not eem to be the only reaon for the different behavior in k. Thi can be een in Figure 4. k 20 (k = 104) Manual Voting Center (k = 36.5) 20 Computerized Voting Center total vote total vote Figure 4: Relationhip between k and total number of vote for computerized and manual center in the ame ize range. The hollow dot repreent voting center located in conular office. There were many mall computerized voting center in rural area. In fact, many of thoe ued mobile phone line to connect the voting machine to the CNE erver in order to tranmit the reult. Thi wa often done due to the lack of regular phone line in thee remote area. There were 586 mixed townhip which included both manual and computerized voting center. Thee mixed townhip had 5449 voting center (2538 manual and 2911 computerized). Notice in Figure 5 (top) that the behavior of k in thee mixed townhip, i very different for manual and computerized center. Appendix B how an example of a mixed townhip. Another intereting comparion i the one related to hamlet ( caerío ). A total of 2162 voting center in hamlet were found 5 (1852 manual and 310 computerized). 5 The official lit of voting center wa earched for the word CASERIO in the addre field. Thee produced the lit of 2162 voting center.

8 8 GUSTAVO DELFINO & GUILLERMO SALAS 10 Manual Voting Center 10 Computerized Voting Center Mixed Townhip k Hamlet (caerio) k Figure 5: Relationhip between k and for computerized (right) and manual center (left) for mixed townhip (top) and hamlet (bottom). Due to the reaon mentioned in Section 4.2, many hamlet mut have been far away from a SCC. For thi reaon voting center located in hamlet hould include large value of k. In Figure 5 (bottom) it can be een that thee large value are found only in manual voting center. Furthermore, Figure 5 how that the behavior of the k value in computerized voting center in hamlet, look more like the ret of the computerized center than the behavior of the 1852 manual center located in the ret of the hamlet Correlation between SI vote and requeting ignature. Let r SI be the correlation of SI vote with repect to the number of ignature. The Carter Center and the OAS aid the following in one of their report[2]: A very high correlation between the number of igner and the number of YES vote per center in the univere of automated voting machine ha been found a correlation coefficient of Thi mean that in voting center where a high igner turnout wa obtained, a high YES vote alo wa obtained. 6 What thi report doe not ay i that for manual voting center, the correlation i 0.607, a much lower value. Thi difference can be viualized in 6 Thi correlation value wa reproduced with a difference of jut which i negligible.

9 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 9 Figure 6. Notice that a traight line from the origin to each of the point ha a lope of k. The high correlation value for computerized center tranlate into imilar k value (or lope) for mot center. 600 Manual Center y = 1.81 x Computerized Center y = 1.15 x zoom ignature (a) ignature (b) Figure 6: Manual Center have a correlation of with repect to the ignature while computerized center have a correlation of A correlation of 1 would look like a traight line. In thi cae, the high correlation in computerized voting center alo implie that in voting center where a low igner turnout wa obtained, a low SI vote wa alo obtained. Thi can be een in the origin of Figure 6b. Hence, when the number of ignature tend to zero, the number of SI vote alo tend to zero. But, a oberved in Figure 6a, manual center do not exhibit thi behavior. The behavior in computerized center eem unexpected becaue the relationhip between ignature and SI vote hould not be linear, epecially when the number of ignature i mall. A explained in ection 4.2, you could expect a large number of SI vote if there were a large number of ignature, but a the number of ignature per center decreae, the level of uncertainty in the number SI vote with repect to the number of ignature increae. In Table 1 the correlation are calculated for center where igner were a minority ( 0.5) and a majority ( > 0.5). Notice that a expected, the correlation for manual center i much higher when there are many ignature (0.947) than when there are fewer ignature (0.613). Thi i the

10 10 GUSTAVO DELFINO & GUILLERMO SALAS expected behavior becaue when you have many ignature the uncertainty of k i mall, and the number of SI vote i equal to k ignature o the uncertainty in the abolute number of SI vote i alo mall. 0.5 > 0.5 All r SI # r SI # r SI # Manual Computerized Both Table 1: Correlation of SI vote with repect to the relative number of ignature per center, for manual and computerized voting center. In the cae of the 645 computerized voting center where > 0.5 the correlation wa which i very high. It tand out that in the computerized voting center where igner were a minority, the correlation i till very high at Furthermore, there i not a ingle computerized voting center with much more SI vote than ignature a een in Figure 6b. In other word, for ome reaon, computerized center do not eem to how the expected non linear relationhip between ignature and SI vote Correlation Plot. In order to further invetigate the change of uncertainty a the relative number of ignature varie, a technique imilar to a moving average i ued. The difference i that intead of calculating an average, a correlation i calculated. A window ize of 150 voting center wa ued. Thi i the ame number of center that were audited. In order to do thi, the firt tep i to ort the voting center, computerized and manual, according to their value. Then r SI i calculated for center in poition 1 to 150. Subequently r SI i calculated for center in poition 2 to 151, and o on. The reult i hown in Figure 7. For manual center, there are large variation in the correlation in the firt half of Figure 7. Thi i the reult of outlier coming in and out of the 150 center calculation window. A the outlier are real official data, they hould not be dropped. Intead, logarithm can be ued for both the number of vote and ignature. Thi way the effect of the outlier i taken into account in a better way. The reult of uing thi technique i hown in Figure 8. Regardle of whether correlation are calculated on a linear cale (Figure 7) or on a logarithmic cale (Figure 8), the important fact to point out i that the reduction in correlation a decreae i large for manual center, wherea in comparion it i negligible for computerized center.

11 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 11 1 Manual Center 1 Computerized Center correlation r SI correlation window initial point correlation window initial point Figure 7: Correlation plot uing a window of 150 voting center. 1 Manual Center 1 Computerized Center correlation r SI correlation window initial point correlation window initial point Figure 8: Correlation plot (logarithmic cale) uing a 150 voting center window.

12 12 GUSTAVO DELFINO & GUILLERMO SALAS 5. The Hypothei. What ha been preented thu far hould be enough to cat a eriou hadow of doubt with regard to the official reult in the computerized center. Baed on thi, it i natural to conider the following hypothei 7 : Hypothei: In computerized center, official reult were forced to follow a linear relationhip with repect to the number of ignature. If thi hypothei were true, becaue of the reaon explained in ection 4.2, the reult would be ditorted with repect to reality, epecially in voting center with a mall value. In place where the ignature did not correctly capture the political intention of the people, two thing would happen: 1. The number of SI vote, according to the official CNE reult, would tend to be much le than the number of real SI vote. 2. The official reult of thoe computerized voting center would be a poor repreentation of the political intention in the area. In the next ection the reult of the referendum will be compared to thoe of the 1998 preidential election in order to find out if thee ditortion are indeed preent Election Comparion. Depite the fact that more than 5 year eparate the 1998 Preidential election and the Referendum, and that the Referendum wa not an election, there are reaon to make the comparion of both event intereting: In both cae the future of the preidency wa at take. In Venezuela, ince 1958 a new Preident had been elected every 5 year. Immediate reelection wa prohibited by the 1961 contitution. Between the 1998 election and the 2004 Referendum, 5 year and 8 month had gone by. On the other hand, the preident had repeatedly claimed that he would tay in office at leat until the year Both event were open for all Venezuelan citizen in the electoral regitry. 7 The mechanic of how vote could have been altered, and by whom i not tudied here. However, the fact that the machine etablihed a TCP/IP connection to the CNE, diconnected and only then printed the reult, open many ecurity hole. Thee iue are beyond the cope of thi article.

13 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 13 Both cae involved a polarization. In 1998 the top two candidate obtained 96.17% of the valid vote. The other 3.83% of the vote were candidate which were alo politically oppoed to the winning candidate. There were 8431 voting center in 1998 and 8394 voting center for the Referendum. Both event involved 8328 common voting center. Comparing 1998 and Referendum reult give an etimation of whether the popularity of the preident increaed or decreaed in the vicinity of each voting center. Additionally, the 1998 electoral reult were ued for comparion, becaue at that time, the CNE could not be under the influence of the current government Correlation between % of oppoition in 1998 and in RR. By comparing the electoral reult (percentage of oppoition) on a townhip by townhip bai, it wa detected that ome of them had a high correlation with repect to previou reult while other had a very low correlation. Furthermore, it eemed a if townhip with higher oppoition reult with repect to 1998 had a much higher correlation than thoe where the government increaed it performance. Thi correlation will be called r 1998, and the percentage of oppoition difference will be called % RR 1998 a defined in Equation 5. (5) % RR 1998 = (% Oppoition in RR) (% Oppoition in 1998) In order to illutrate thi, the reult of two townhip are plotted in Figure 9. In the Olegario Villalobo townhip, the correlation with repect to the ignature and the 1998 percentage of oppoition i large at r SI = and r 1998 = repectively. Additionally, notice that the average i 0.639, o igner were the majority in thi townhip. Therefore, the ignature are likely to have captured the political intention of voter here. In the cae of the Vita al Sol townhip, the average i very low. Therefore, the uncertainty in the number of SI vote with repect to the ignature could be large, a wa hown in Section 4.2. In other word, the ignature are not likely to have captured the political intention of the townhip accurately. Thi uncertainty i jut not een in the official reult, a the correlation of SI vote with repect to the ignature i Furthermore, the referendum reult eem very ditorted with repect to the 1998 election, with a negative correlation of In thi townhip,

14 14 GUSTAVO DELFINO & GUILLERMO SALAS the center with the mot oppoition in 1998 ended up being the mot pro government, and vice vera 8. State: Zulia County: Maracaibo Townhip: Olegario Villalobo r 1998 = = r SI = %Oppoition %Oppoition Referendum Election Voting Center Code r 1998 = = r SI = Election Referendum Voting Center Code Figure 9: Two ample townhip. All the center hown are computerized voting center. The two townhip hown in Figure 9 behave conitently with the hypothei. Olegario Villalobo wa able to increae it percentage of oppoition becaue many ignature were collected, wherea Vita al Sol could not increae it percentage of oppoition becaue only a few ignature were collected. If thi repeat itelf in the ret of the country, then r 1998 would be large when % RR 1998 i large, and r 1998 would be mall when % RR 1998 i mall. In an untouched proce, thee two variable hould be independent. In Figure 10, it i hown that, indeed in all of the country there i a trong relationhip between % RR 1998 and r 1998 for computerized center at the townhip, county and tate level. Thi relationhip i much weaker 8 Thi center returned to being the one with the mot oppoition 77 day later in the tate governor election.

15 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 15 r1998 r1998 r r = AMAZONAS APURE TRUJILLO ARAGUA DTTO. CAPITAL MONAGAS YARACUY PORTUGUESA ZULIA MIRANDA BARINAS BOLIVAR SUCRE FALCON ANZOATEGUI DELTA AMACURO GUARICO r = COJEDES CARABOBO r = Manual Center VARGAS LARA MERIDA TACHIRA EMBASSIES % RR 1998 State Countie Townhip Computerized Center APURE AMAZONAS GUARICO COJEDES SUCRE VISTA AL SOL YARACUY TRUJILLO CARABOBO DELTA AMACURO LARA BARINAS MONAGAS MIRANDA VARGAS ARAGUA BOLIVAR PORTUGUESA FALCON NUEVA ESPARTA MERIDA r = TACHIRA ZULIA DTTO. CAPITAL ANZOATEGUI r = r = OLEGARIO VILLALOBOS % RR 1998 Figure 10: Relationhip between r 1998 and % RR 1998 at tate, county and townhip level for manual and computerized voting center. The correlation between % RR 1998 and r 1998 i hown a r. Auming that % RR 1998 and r 1998 are independent, the probability of eeing thoe r value i calculated in Appendix D

16 16 GUSTAVO DELFINO & GUILLERMO SALAS almot inexitant for manual voting center. Thi finding i conitent with the hypothei. 7. Variability in value of k and the correlation between percentage of oppoition and value of, for variou electoral event. In Section 4.2, it wa tated that a the value of decreae, the variability in k i expected to increae. According to Equation 4 thi variability mut alo be preent in the relation between and the percentage of oppoition. Therefore, a become mall, it hould poorly correlate with the percentage of oppoition. For thi reaon, when i mall, it hould not determine the percentage of oppoition. On the other hand, when become large, it hould correlate better with the percentage of oppoition. Let r be the correlation of the percentage of oppoition and, and let be the median of all the value of for computerized center. For the ubet of computerized center with thi correlation will be called r,, and for the remaining center where > the correlation will be called r,>. The value of r, hould be maller than r,>. Thee propertie jut defined are calculated for variou electoral event in Table 2. Table 2: Correlation r for computerized center with above and below, for different electoral event. Date Event r, r,> r,> r, Preidential Election Preidential Election Referendum Official Reult Exit Poll State Governor Election The exit poll referenced in Table 2 wa made under the uperviion of Penn, Schoen, and Berland Aociate. The State Governor election took place jut 77 day after the Referendum. By counting vote for and againt the pro government candidate, a percentage of oppoition wa calculated. During thi election, the ame voting machine were ued, but there wa an important difference: the paper ballot were manually counted for a randomly elected voting machine in each and every voting center. The reult for the correlation r for thi election are hown in Table 2. From Table 2 it i clear that only the Referendum official reult fail to exhibit a poitive correlation difference. Alo notice in Figure 11 that for

17 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 17 the Referendum official reult, there i not a ingle voting center with a mall and large percentage of oppoition. The fact that only in the official Referendum reult r, i not maller than r,> i conitent with the hypothei. 8. Hot Audit. In general, the paper ballot from the computerized center were not manually counted. The CNE aured the Venezuelan citizen that the voting machine had to accurately reflect the voter intention, becaue a ample of 192 machine (1% of them) would be randomly elected and audited the ame day of the referendum. Thi i indeed a valid way of eliminating upicion, a long a the election i a truly random ample of all the voting machine. The day of the referendum, the CNE informed the public that becaue of logitical reaon, the ample would be taken from a retricted univere of 20 countie located in urban area, leaving out of the audit more than 300 countie. At thi moment, confidence in the reult wa adverely affected to ay the leat. Computerized voting center inide and outide of the 20 countie, to which the hot audit univere wa reduced, are hown in Figure 12. It i clear that thee 20 countie are not repreentative of all the computerized voting center. See Appendix E for further detail on thi ubject. Furthermore, out of 192 center elected for hot audit, only 26 were actually audited in preence of witnee repreenting the oppoition and the international oberver. The following excerpt from the Carter Center Comprehenive Report [4] i very illutrative: Auditor, table member, and military peronnel were not properly informed that the audit would occur nor were they clear about the procedure to be followed. The intruction themelve did not clearly call for a eparate tally of the Ye and No vote, and in ome center, the auditor only counted the total number of voter. (... ) Neverthele, Carter Center oberver were able to witne ix auditing procee. In only one of the ix auditing ite oberved by The Carter Center did the paper ballot receipt counting actually occur. In thi place, the auditing wa conducted by the mea preident, and the recount of the ballot produced exactly the ame reult a the acta printed by the voting machine. In the ret of the ite oberved, the auditor appointed by the CNE did not allow the opening of the ballot box, explaining hi/her intruction did not include the counting of the Ye and No ballot from multiple machine. There were alo complaint of military denying acce to voting center where audit were being conducted. Carter Center oberver could not confirm thi claim. (... ) The CNE provided The Carter Center with copie of the audit report of 25 center. It wa clear from the form that the audit wa not carried out in many place becaue the field in the form were left empty, there were no ignature of pro government or oppoition witnee, etc. The form were poorly filled out, clearly howing inadequate training. The intruction iued by the CNE to the auditor were either incomplete or unclear. Thi i a direct conequence of iuing the audit regulation three day before

18 18 GUSTAVO DELFINO & GUILLERMO SALAS Election %Oppoition Exit Poll %Oppoition Referendum %Oppoition Election %Oppoition Election %Oppoition r = r = r = r = r = median r = r = r = r = r = correlation comparion bar plot Figure 11: Correlation between percentage of oppoition and for the lower two ( ) and upper two quartile ( > ) for computerized center. The correlation for the lower two quartile i expected to be maller than the correlation in the upper two quartile. Thi expected difference i not een in the Referendum official reult.

19 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 19 Computerized Center Subject to Hot Audit 100 Computerized Center Left Out of Hot Audit RR %SI RR %SI (a) 1998 %Oppoition (b) 1998 %Oppoition Figure 12: Center inide (a) and outide (b) of the 20 countie to where the hot audit drawing wa retricted. the election. The final reult wa that the CNE quandered a crucial opportunity to build confidence and trut in the electoral ytem and outcome of the recall referendum. Auditing only 26 center out of 192 elected center, i baically a cancellation of the auditing proce. But, i there anything pecial about thee 26 center? If thi dratic reduction in audit ize wa becaue it wa poorly executed, and poor execution i independent of the value of, then the value of of thee 26 center would behave a a random ample within the value of the 192 elected center. From Figure 13, it i clear that the 26 center that were actually audited eem to have a much higher value of than the 192 center from where they come from. The average for the 192 elected center i elected = while for the audited one it i audited = Additionally, the ditribution of the 192 elected center i poitively kewed while the ditribution of the 26 audited center i negatively kewed. Can thi be jut a coincidence? A Monte Carlo imulation wa done, electing 26 random center out of the 192 elected for auditing. The reult of thi imulation i that the probability of having a audited = i 1 in ; and thi doe not take into account the difference in kewne. Thi reult i conitent with the hypothei, becaue center with a mall value of are the one mot uceptible to ditortion. It ha been hown that the audited center are not repreentative of neither the univere of all computerized center, nor the retricted univere

20 20 GUSTAVO DELFINO & GUILLERMO SALAS 26 Center Audited v. 192 Selected Military Hopital Center audited Center elected but not audited 192 # Center Center Audited Selected = γ 1 = = γ 1 = mean of γ 1 kewne Leaf Stem Leaf Figure 13: Comparion between the value of the 192 elected center and the 26 audited center. TOP: The elected center are ordered according to the value of and plotted. BOTTOM: Back to back tem and leaf plot howing the ame value of a in the top figure. To obtain value, multiply tem by 0.1 and leave by 0.01.

21 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 21 ued to elect them. The audited center are not repreentative of the univere of computerized voting center (ee Figure 14) becaue: 1. In the audited center, the SI won by 63.47% to 40.91%. 2. % RR 1998 i very different. 3. The value of i much larger. Additionally, the townhip, countie and tate where center were audited are not repreentative of the other townhip, countie and tate. They are not repreentative with regard to their % RR 1998 and the correlation with repect to the 1998 election r Thi can be een in Figure 15. % RR Audited Center Average Not Audited Center Average Audited and Not Audited Computerized Center Figure 14: In thi figure, the 26 computerized audited center are compared with the univere of computerized center. The average and % RR 1998 are indicated with line. 9. Cold Audit. Given the fact that the hot audit failed to erve it purpoe, another audit wa made three (3) day after the referendum. Thi audit cannot validate the official reult mainly becaue of two reaon: The audited entity itelf cannot elect the center to be audited. According to the OAS/Carter report [3] The ample wa generated by CNE taff on it own computer uing it own oftware. The control mechanim that had been implemented to certify that the ample were unaltered were not ued.

22 22 GUSTAVO DELFINO & GUILLERMO SALAS State: VARGAS MIRANDA ARAGUA DTTO. CAPITAL r1998 r1998 r Countie: Townhip: MERIDA ZULIA TACHIRA Unaudited % RR 1998 Include 1 or more audited center Figure 15: Townhip, countie and tate where the 26 audited center are located. The vertical axi i the correlation with repect to the percentage of oppoition in 1998 (r 1998 ). The horizontal axi i the difference in percentage of oppoition with repect to 1998 ( % RR 1998 ).

23 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 23 The drawing where the center for audit were elected wa broadcat live on the official TV tation, but the reult were not hown. Uually, the whole idea of tranmitting a drawing on TV, i to let the public know the reult a they are being generated. When the ballot boxe were cloed and ealed, and the electoral center cloed, the eal wa igned by witnee. The boxe were then taken in the cutody of the military. The following excerpt from the OAS/Carter Center report [3] explain the mechanim ued to certify that the boxe were unaltered: Each box wa phyically checked to ee whether: 1. The material ued to eal the box wa intact or whether there were ign that it had been taken off and then replaced. 2. There were crack or hole through which vote might have been extracted or inerted. If a box wa defective in regard to ealing, crack, or hole, all the boxe of that polling tation were excluded from the audit and a note to that effect recorded in the minute. However, the witnee who had igned the boxe were not called to certify the authenticity of the box. When thi audit wa quetioned, the Carter Center and OAS repone wa that: Furthermore, the correlation between the igner and the YES vote i almot identical in the univere and in the ample. The difference between the correlation i le than 1 percent: Correlation Coefficient Univere Sample Thi certainly can be ued to argue that the boxe opened were repreentative of the official reult, but doe not indicate anything in regard to validating the official reult. Interetingly, the drawing for the hot and cold audit included ixteen (16) common center. Thee ixteen (16) center were uccefully cold audited, but none of them were allowed to be hot audited. 10. Concluion. Regarding the computerized voting proce, taking into account that: 1. In general, the paper ballot were not counted. 2. The voting machine were connected to a central CNE erver before printing the reult.

24 24 GUSTAVO DELFINO & GUILLERMO SALAS 3. Neither of the audit could remove the doubt on the official reult, let alone certify the reult. 4. That the computerized center in the 20 countie to which the hot audit wa retricted are not repreentative of the univere of all computerized center (ee Figure 12). 5. That the hot audited center are not repreentative of the other computerized center (ee Figure 14). 6. That the townhip, countie and tate where computerized center were hot audited are not repreentative of the other townhip, countie and tate (ee Figure 15). 7. The official SI reult eem to behave in an exceively linear way with repect to the requeting ignature (Section 4). 8. The problem with the hot audit eem to be more evere than jut poor execution becaue the center actually audited belong to a ditribution with a much greater value than the ditribution of the 192 elected center (Figure 13). Preciely where the high linearity i more unexpected (mall value) i where mot audit were upended. 9. The official SI reult eem to be ditorted with repect to the 1998 preidential election (Section 6). 10. The percentage of oppoition eem to be exceively tied to particularly in center with mall (Section 7) 11. That if the hypothei i true, all the aforementioned item would be expected. It mut be concluded that there are undiputed reaonable doubt a to the truthfulne of the computerized center official reult. The OAS and the Carter Center are encouraged to review the ignificance of having uch a high linear correlation between the SI vote and the ignature. Their current interpretation i found in the following excerpt from the OAS/Carter Center report [3]: There i a high correlation between the number of YES vote per voting center and the number of igner of the preidential recall requet per voting center; the place where more ignature were collected alo are the place where more YES vote were cat. There i no anomaly in the characteritic of the YES vote when compared to the preumed intention of the igner to recall the preident. APPENDIX A: DATA PROCESSING METHODOLOGY All the data wa tored on a MySQL databae. Calculation were made uing Mathematica 5.2 which wa connected to MySQL uing the DatabaeLink package. Referendum official reult were downloaded from the official CNE webite uing a Perl cript. Electoral reult from the 1998 preidential elecimart-t ver. 2005/02/28 file: delfino-ala.tex date: November 2, 2007

25 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 25 tion were obtained on an original CNE CD-ROM, and the data wa converted from MS Acce to MySQL. The REP from July 2004 wa alo converted from MS Acce to MySQL. The CNE ignature data wa obtained on a CD from Súmate, and i the ame verion given to the OAS and the Carter Center. Thi data wa upplied in a ingle text file. By matching people ID number (cédula number) from the ignature and REP data, it wa poible to find the number of ignature per voting center. In order to claify voting center into manual and computerized, the following ource of information were ued: Súmate lit of computerized and manual voting center. On CNE web ite, computerized center how reult down to the voting machine level, wherea manual voting center how reult down to the table level. The lit of computerized and manual center obtained uing the aforementioned ource wa compared on a townhip by townhip bai with the CNE infratructure document[5]. The lit of center effectively audited on the day of the Referendum wa obtained from a document given by the Coordinadora Democrática to the Carter Center and OAS. A copy of thi document can be found at: gdelfino/rr/hot-audit.xyz The coordinate of the voting center hown in Appendix B were provided by Delta Electoral. The data needed to reproduce thi tudy i available at: gdelfino/rr/. The imulation wa done uing a deck of card huffling algorithm. The random number generator ued by thi algorithm wa the Wolfram rule 30 cellular automaton generator for integer, which i provided by Mathematica. APPENDIX B: A MIXED TOWNSHIP EXAMPLE. Miguel Peña i a townhip in Valencia county, Carabobo tate. It i one of the townhip with more population in the country. It had 32 voting center, 28 computerized and 4 manual. In Figure 16, a partial aerial view of thi townhip i hown. In it, notice that manual and computerized voting center are in the ame urban neighborhood. Depite thi, the value of k are much higher for the manual center than for the urrounding computerized center, regardle of the total number of vote.

26 26 GUSTAVO DELFINO & GUILLERMO SALAS In Figure 17 notice that in thi townhip, the lowet k value of the 4 manual center i greater than the maximum k value of the 28 computerized voting center. Ecuela Ruiz Pineda TV = TV = TV = 6715 Preecolar Ricardo Urrieta Ecuela Moneñor Lui Eduardo Henriquez Ec. Ba. Bella Vita II TV = 2661 Ecuela Báica El Socorro TV = 743 Ec. Ba. Joé Regino Peña 4.38 Unidad Ecolar Negro Primero 1.00 Unidad Ecolar Renny Otolina TV = Ecuela Etadal Joé Antonio Paez Ec. Felix Manuel Luce 1.78 TV = 940 TV = 4844 Unidad Ecolar Dr. Enrique Tejera TV = TV = Unidad Educativa Fundación Valencia TV = TV = 4274 Ec. Antonio Ricaurte TV = 3006 Manual Voting Center 1.00 Computerized Voting Center 1.00 k k Col. Miguel A. Pérez II TV = 2652 Figure 16: Partial aerial view of Miguel Pen a townhip (taken from Google Earth). Manual v. computerized voting center are compared in regard to their k value and total number of vote (TV). The image i centered at Latitude N and Longitude W. APPENDIX C: ADDITIONAL NON LINEARITY PLOTS According to the exit poll made under the uperviion of Penn, Schoen, and Berland Aociate, the oppoition won the Referendum by a wide margin. By changing the numerator of Equation 4 from percentage of SI vote to percentage of SI from exit poll, a value of kexit poll can be calculated. The reult, for computerized center only, i plotted on Figure 18. Similarly, k1998 can be calculated by uing the percentage of oppoition in the 1998 preidential election in the numerator of Equation 4. The reult, for computerized center only, i hown on Figure 19 imart-t ver. 2005/02/28 file: delfino-ala.tex date: November 2, 2007

27 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 27 k State: Carabobo County: Valencia Townhip: Miguel Pena Manual Center Computerized Center total vote Figure 17: Behavior of k v. total vote in all of Miguel Peña voting center. Manual and computerized center are hown. APPENDIX D: MONTE CARLO SIMULATIONS FOR CORRELATION BETWEEN % RR 1998 AND r Auming that % RR 1998 and r 1998 are independent, regardle of being calculated at tate, county or townhip level, then the correlation between them r mut be caual. In order to find the probability that the oberved r i caual, it i poible to reorder the value of r 1998 with repect to % RR Thi reordering wa made time and the correlation wa calculated each time. In all cae, the reulting ditribution wa found to be normal. The etimated probabilitie for manual and computerized center at tate, county or townhip level are hown in Figure 20. APPENDIX E: DIFFERENCES IN CHARACTERISTICS, OFFICIAL RESULTS AND REP VARIATION OF THE 20 COUNTIES SUBJECT TO HOT AUDIT DRAWING IN COMPARISON TO THE OTHER COUNTIES. When the CNE decided to retrict the audit to 20 urban countie, it created two group of computerized center: 2040 computerized center inide the 20 countie and therefore ubject to be elected in the drawing. Variable referring to thee center will ue a 20 a a ubindex ( 20 ).

28 28 GUSTAVO DELFINO & GUILLERMO SALAS 10 Exit Poll on Referendum Computerized Center 8 6 k exit poll Figure 18: Exit Poll at computerized center Reult on Referendum Computerized Center 8 6 k Figure 19: 1998 Preidential Election at computerized center.

29 ANALYSIS OF THE 2004 VENEZUELAN REFERENDUM RESULTS 29 Manual Center p = State Computerized Center p = tandard deviation Official Reult r tandard deviation Official Reult r p = tandard deviation Official Reult Countie tandard deviation p = Official Reult r r p = Townhip p = Official Reult 1.2 tandard deviation tandard deviation Official Reult r r Figure 20: Comparion of official reult correlation r v. expected value ditribution found after imulation for manual and computerized center at tate, county or townhip level. The imulation reult follow a normal ditribution, which i hown a a dotted line. The probability of the official r happening by chance i indicated a p.

30 30 GUSTAVO DELFINO & GUILLERMO SALAS 2553 computerized center not ubject to hot audit at all. Variable referring to thee center will ue a a a ubindex ( ). In Figure 12 it i hown that the behavior in computerized center in the 20 countie i very different from that of the ret of the country. E.1. Difference in Characteritic. When the CNE et up the ignature collection event, it etablihed the number of ignature collection center (SCC) directly in proportion to the number of people in the electoral regitry (REP) for each county. A lot of people live in urban countie, therefore, a lot of SCC were aigned to thee countie. So the acce from where the people lived to where they had to ign wa much eaier in thee 20 countie. On the other hand, voting center are more numerou and better ditributed throughout the national territory. For example, a county like Chacao in the Miranda tate ha 27 km 2 of area and 11 SCC. In Chacao there were 24 voting center, all of them computerized. On the other hand, the much larger Macanao Peninula in Margarita Iland ha an area of km 2 and only had 3 SCC. There were 8 voting center in Macanao, all of them computerized. In Figure 21, it can clearly be een that the 20 countie have higher value which i conitent with the idea jut explained. pdf All Computerized Center cdf pdf Computerized Center ubject to Exit Poll cdf 20 countie in the hot audit drawing 302 countie exluded from the hot audit drawing Figure 21: Comparion of probability denity function (pdf) and cumulative denity function (cmf) for computerized center inide the 20 countie of the hot audit and in the 302 excluded countie. There were many computerized center in rural area where it wa much