(Caracas, 21 de Agosto de 2004 - NPS 081) - What is our claim? The Carter Center issued a report entitled “Report on an Analysis of the Representativeness of the Second Audit Sample, and the Correlation between Petition Signers and the Yes Vote in the Aug. 15, 2004 Presidential Recall Referendum in Venezuela” which is a response to our paper “In search of the black swan: Analysis of the Statistical Evidence of Electoral Fraud in Venezuela”. In preparing their response, the Carter Center never contacted us to ask questions about our methodology or asked to see our data in order to reproduce our results, although we offered to do so.
In our original paper we studied – among other things – whether the sample used by the Carter Center for the purpose of the audit was a random sample of the whole universe of automated voting precincts. We also presented what we believe to be evidence of fraud, but the Carter Center report does not deal with this aspect of our report.
It is useful to remind the reader of the reasons why we looked into this question. We asked ourselves whether it was possible to have massive electronic fraud take place and for the audit conducted under the observation of the Carter Center and the Organization of American States not to have found it. We argued that this was possible if the audit was conducted on a sample that was not truly random and hence representative of the whole universe. If fraud was committed in some precincts and not in others, it would be possible to direct the audit to the non-altered precincts causing the audit to fail to find any wrongdoing. To check this possibility we developed a test for randomness. We based ourselves on a well accepted principle in statistics which holds that if a relationship between certain variables holds for a universe as a whole, it should hold for any random sub-sample of that universe. This implies that if we estimate a relationship for the whole universe, we should not be able to reject the hypothesis that the relationship is statistically different for the audited sample.
We proceeded by estimating a relationship in logarithms between the number of votes on the one hand and four factors which should affect this result, on the other. The factors we considered were four and have a clear reason to be in the relationship we estimated:
The logic for including these variables in the relationship is straightforward. First, the higher the number of signers in a precinct, the higher the expected number of Yes votes, given that signers have expressed a preference for such a vote. Second, the higher the number of registered voters at the time of the recall referendum, the higher the potential number of additional yes votes as some voters may not have been able to sign or may have preferred not to do so given that signing was a public event while the vote was secret. Third, the higher the number of new voters, the higher the expected number of Yes (and No) votes as these voters have yet to express their political preferences one way or the other. Finally, the higher the number of voters who do not show up to vote, the lower the number of Yes (and No) votes.
When we estimate this relationship for the whole universe we find that all the variables are significant at the 1 percent confidence level. When we check whether this relationship holds with similar parameters for the audited sample, we can reject the hypothesis that it does, also with a very high confidence level. In particular we remark that the estimated elasticity between signatures and votes – conditional on controlling for the other three factors – is 10.5 percent higher in the audited sample than in the rest of the universe of automated precincts.
This is the essence of our proof. We called our paper “In search of the black swan” in reference to Karl Popper’s dictum that a thousand white swans do not prove the proposition that all swans are white, but one black swan does show that they are not. For the white-swan proposition that the sample used in the audit was random, we provided a black swan proof that it was not.
How does the Carter Center answer our claim? They make three propositions:
1.. They check whether the mean of the votes in the two samples are similar
3. They test the random number generator program used by the Electoral Council and find that it does generate a random draw of all the precincts. They also correctly point out that the numbers are not truly random in the sense that the same initial seed number generates the same sequence of numbers.
Similar sample means
With respect to the first point, the question that the Carte Center asks is whether the unconditional means of the two samples are similar. By unconditional we mean that they do not control for the fact that precincts are different in the four dimensions we include in our equation or in any other dimension. To see the importance of conditioning, let us imagine that there is fraud and let us suppose that the fraud is carried out in a large number of precincts but not in all of them. The question is: is it possible to choose an audit sample of non-tampered centers that has the same mean as the universe of tampered and un-tampered precincts? The answer is obviously yes. Let us give an example using a population with a varying level of income, say from US$ 4,000 per year to several million. Assume that half of them have been taxed 20 percent of their income while the other half has not. Is it possible to construct an audit sample of non-taxed individuals whose average income is similar to that of those that have been taxed? Obviously the answer is yes. However, if one controls for the level of education, the years of work experience and the positions they hold in the companies they work in, it should be possible to find that the audited individuals actually a higher net income than the non-audited group. That is the essence of what we do.
Now, lets go back to the case in point. Precincts vary from those where the Yes got more than 90 percent of the vote and those where it got less than 10 percent. This is a very large variation relative to the potential size of the fraud, say 10 or 20 percent. It is perfectly feasible to choose a sample that has the same mean as the rest of the universe.
However, the non-random nature of the sample would be revealed if we compare the means but controlling for the fact that each precinct is different. That is what we do and this is the randomness test that the audited sample failed.
Similar correlation coefficients
The second check consists of comparing the correlation between signatures and votes in the two samples, which they find to be very similar. This is clearly not a test of anything relevant to the case in point. To see this, suppose that in the audited sample there is a perfect relationship in which each signature becomes 2 votes and in the non-audited sample, because of fraud - the relationship implies that each signature becomes only 1 vote. However, the correlation coefficient in both samples is 1. This is due to the fact that the correlation coefficient is affected by whether the two variables move up and down together, but not by whether they do so in a relationship of 1-to-1, 2-to-1 or 10-to1. This procedure is certainly no proof of randomness or of the absence of fraud.
Test of the sample number generator
The final point is that the random number generator actually generates a sample that can potentially pick all the universe of precincts and that it was tested and appeared to actually generate random numbers. However, there are many ways in which this kind of analysis is weak. The most obvious one is that the program does not really generate random numbers but a predetermined set of numbers for each seed-number that initiates the sequence. By putting a known seed-number the Electoral Council would know beforehand which precincts would come up, and could thus decide which precincts to leave unaltered. It is our understanding that in the audit conducted on August 18-20, the seed number was provided by the Electoral Council and implemented in their computer. It does not matter if, as reported by the Carter Center, after 1000 draws, the likelihood of any precinct being chosen looks reasonably random. The point is that the first draw is completely pre-determined by the seed number.
Other problems involve the possibility that at the time and place in which the program was run, a logical bomb might have been active which would make the program work differently. The bomb could self-destruct leaving no trace.
The Carter Center report does not address the two main findings of our report. It completely disregards the evidence we put forth regarding the statistical evidence for the existence of fraud in the statistical record. It only addresses the issues we raise regarding the randomness of the sample used for the audit they observed on August 18-20, 2004. They show that the unconditional means between the audited sample and the rest of the universe are similar. However, this is no proof of randomness. Conditional on the characteristics of the precincts, we show them to be different and this result is not challenged or addressed by the report. The report also argues that the correlation coefficient between signatures and votes in the audited sample is similar to that in the rest of the precincts, but this is an irrelevant statistic for this discussion. Finally, the report checks the source code of the software used but leaves open wide avenues for fraudulent behavior.
We do not know what happened during the audit, as we were not present. We do know that the sample fails the randomness test we designed. The Carter Center has nothing to say about this fact. Paraphrasing Popper again, the Carter Center seems content in finding the odd white swan here and there. That does not prove the proposition that the sample was randomly chosen. We have presented a formal test of randomness and the sample fails it. That is a black swan.
¡La Democracia no se Rinde!!
Asociación Civil Súmate | Twitter: @sumate