In June of this year the Democratic National Committee (DNC) issued a report on the 2004 Presidential election in Ohio, in which their most significant (not to mention controversial) conclusion was “strong evidence against the claim that widespread fraud systematically misallocated votes from Kerry to Bush.
The study that led to this conclusion is presented in Section VI
of the report, and the conclusion is stated prominently in the Executive Summary of the report. It is based on strong correlations by precinct between the percentage of voters who voted for Kerry on the one hand, with voters who voted for Eric Fingerhut (Democratic candidate for Senate), “No” on Issue 1 (the ban against gay marriage), and the percent of African Americans in the precinct.
Shortly after this study was released I disputed the main conclusion of the DNC report on several grounds in a letter to Howard Dean
. The main problem that I had with the report was that it failed to consider the possibility of central tabulator fraud, whereby votes would be electronically and fraudulently deleted from the final precinct counts in Democratic precincts or added in Republican precincts in the same proportion as the actual votes. Nobody can question that this type of fraud would not have altered the correlations calculated in Section VI of the DNC report. However, one problem with this scenario (at least some people consider it a problem) is that this type of fraud would not have produced within precinct discrepancies between the exit polls and the official election results (referred to as WPE, or within precinct error), and yet Ohio demonstrated large discrepancies of this nature, consistent with the large documented discrepancy between the Ohio exit poll and the official Ohio result, of 6.2%.
With regard to the type of fraud that the DNC report purported to rule out (on a scale large enough to change the election results), most people accepted the DNC conclusion, and yet the report was never very clear about how it was determined that a small amount of vote switching within precincts could not have been sufficient to change the election results and yet maintain the strong correlation between the Kerry vote and the other variables that were analyzed.
This analysis tests the DNC conclusion that strong correlations between the Kerry vote and other variables rule out sufficient fraud via vote switching within precincts to have changed the results of the 2004 election in Ohio. Correlation between Kerry vote versus Fingerhut, Connally, and Gore 2000 vote, and “No” on Issue 1
First I looked at the correlations by county between the Kerry vote, regressed on the Fingerhut vote, Ellen Connally (Democratic candidate for Chief Justice of the Ohio Supreme Court) vote, the Gore 2000 vote, and “No” on Issue 1. I added the Gore 2000 vote to the DNC analysis because, being that it was a Presidential vote it seemed particularly applicable, and I added the Connally vote because of the controversy surrounding the fact that she outpolled Kerry in several Ohio counties in 2004. I deleted the percent of African Americans as a variable because I did not have that data.
The resulting correlation was indeed very strong. Two measures of the degree of correlation in a multivariate regression analysis are the F value and R squared. The F value was 809, and R squared was 0.97.
The variable that most correlated with the Kerry vote was the Gore 2000 vote. Next was “No” on Issue 1 and the Fingerhut vote, which were about equally correlated with the Kerry vote. And last by far was the Connally vote which, strangely, was negatively correlated with the Kerry vote. Effect on correlation of uniform switching of votes back to Kerry
Next, I attempted to see what would be the effect on the correlation by switching Bush votes back to Kerry. Since the DNC concluded that the switching of sufficient votes to change the results of the election would result in correlations much weaker than what was demonstrated in their analysis, then certainly switching a large number of votes from Bush to Kerry, starting with the “official” election results (which supposedly were reasonably valid), would greatly weaken the correlations.
That however was not the case. I added 2% to the Kerry vote in each county and subtracted 2% of the Bush vote from each county and re-ran the regression analysis. The end results were almost the same as with the original (i.e., official) vote count. In this case, the F value was 809, R squared was 0.98, and the relative contributions of the regression variables were very similar to the first analysis.
It is also important to note that the above vote switches narrowed the amount of the Bush victory from about 118,000 votes to about 6,000 votes. With more than 107,000 votes (mostly “under-votes”) as yet remaining uncounted, and given the breakdown of previously counted “under-votes”, the counting of these remaining votes would almost certainly give Kerry a victory in Ohio. Effect on correlation of non-uniform switching of votes back to Kerry
Some might argue that the vote switches that I tested above were unfair because they were unnaturally uniform. Therefore, I decided to do one more test, this time utilizing non-uniform vote switching. In doing this, I decided to test vote switching in those counties where it seemed the most likely that Kerry might have been defrauded of votes. Those were the counties where Kerry did worse than Gore did in 2000, plus Cuyahoga County, where numerous irregularities have been observed and two election workers indicted. For those counties I added 4% to Kerry’s total and subtracted 4% from Bush.
When I then re-ran the regression analysis, the correlation was even stronger than before: The F value was 1218, and R squared was 0.98. Again, the relative contributions of the four variables were similar to the above analyses.
With this intervention, Bush’s victory margin in Ohio was reduced from 118,000 votes to less than 9,000 votes, and again, the counting of the currently uncounted votes would almost certainly result in a Kerry victory. Conclusions and significance
If strong correlations between Kerry’s vote count and other relevant variables is an indication of absence of widespread fraud, as noted in the DNC report (and that is almost certainly the case, depending upon how strong those correlations are), then the finding in this analysis that switching votes back to Kerry from the current official figures results in even stronger correlations, would seem to suggest that the situation resulting from those vote switches (i.e., a virtually even election) is more plausible (in a fair election) than the current official count. Perhaps I should see what happens if I switch even more votes from Kerry to Bush.