The exit polls are correct; it's the machines that are not

[Goeres Ross P Contr AFOTEC/AS]

        Here's some background info on an interesting project for
mathemagicians.

        I was the Presiding Judge for Precinct 214 in ABQ--a heavily
Democrat/Green population of registered voters.  At the end of the night
when the paper tapes were printed and removed for anomaly reconciliation and
certification.  One of the Republican Judges was anxious to see the results
so he could see who would win nationwide.  Kerry votes outnumbered Bush
votes, of course, but he made the announcement that Bush would win and
predicted 300 electoral votes.  When asked how he was so sure, he said it
was due to experience in that precinct with past elections.  He said one
could expect about a quarter of the Democrats to vote for Bush--an entirely
unexpected conclusion.

    In the voter registers, the party affiliation of the actual voters are
there.  He said he counted the number of registered Republicans,
non-Democrats and Democrats that were issued permission slips to vote on the
machines.  On machine A, he said roughly 1 in 5 Democrats voted for Bush.
On machine B, about 1 in 3 of the Democrats must have voted for Bush.  In
aggregate, it was indeed roughly 1 in 4 Democrats must have voted for Bush.
I was so shocked that I didn't have the presence of mind to reproduce his
calculations until after I had delivered the results to the County Clerk and
Secretary of State's office.

    We had a slight anomaly in the counters because we had issued one more
permission slip than votes tallied.  We compared the voter roster line
numbers from the permission slips to the annotated paper registers and found
the person (Democrat, I think) who was issued a slip but did not vote for
some reason.  

    We shared the room with Precinct 221; the Presiding Judge conducted his
precinct less strictly than I (we had some words regarding jurisdiction--I
had to point out that election laws overrode his authority to ignore
possible infractions).  On their machine B, it was reported that pushing the
Kerry button illuminated the light next to another candidate.  I could not
observe this personally but it was certainly a possibility on these old
machines.  I tried to make sure everyone in the lines was instructed as to
what the lights meant, how to change their selections and to make absolutely
sure the steady lights are next to the intended selections before pushing
the green "VOTE" button in the lower right-hand corner.  These notifications
seemed to irk the Presiding Judge because they were not in my precinct.

    I find it to be extremely unlikely that, assuming all Republicans voted
for Bush (a faulty assumption as I am a registered Republican but I did not
vote for Bush--at least I hope the machine tallied it according to my
wishes) that 1 in 4 of the Democrats would wait in line on a cold, windy day
just to vote for Bush and more of the same--only worse.  This doesn't pass
any sort of "smell" or "common sense" test that I can think of, but that's
impossible to quantify.  Or is it?

    If one has access to a copy of the rosters used by the election
officials and the tally sheets for all of the precincts, one can estimate
the percentage of Democrats who must have voted for Bush for the tally sheet
numbers to jibe (e.g., those precincts where there were more votes for Bush
than the number of non-Democrats who actually voted).

    There were only 209 permission slips issued in my precinct and the
roster has the (theoretically) current mailing address.  If the Democrats
who voted were contacted and asked if they voted for Bush, what would be the
sampling parameters necessary to get a significant chi-squared statistic
that indicates whether the samples (votes) from the machines were from the
same population of the Democrats who would answer the question (the null
hypothesis, H0)?


http://www.blackboxvoting.org/ has documented some interesting events that
indicate more reasons why the exit polls and the voter tallies were so far
out of whack.

    The 3-hour gap in the activity logs would seem to indicate that the
tampering was not done quickly by scripts but probably manually (or might
have to cover the effects from the Central, Mountain and Pacific time zones
for a "scheduled" event that needed to be erased).  Humans are notortiously
bad  random number generators.  Could Zipf's Law or Benford's Law be used
for anomaly detection in suspected tampering situations?  How big would a
"spike" have to be in the distribution to be statistically significant?

    The RNC's Voter Vault (their mission-critical privacy-invasion and
voter-targeting system) was out-sourced to an Indian firm that uses nothing
but Microsoft development tools.  If the attacking machines were using the
faulty random number generators (that started from the same seed each time),
could artifacts be extracted to predict a likely "random" sequence for
tampering at a pre-selected range of rates of "defecting Democrats" (e.g.,
from hidden markov model analysis)?

    The Republican Judge was suspiciously cocksure about the 1 in 4
ratio--any agent-based models that might be useful in testing the a priori
assumption of 25% Defecting Democrats?

Anyone up for some number crunching?

--Ross