Analysis: "Ballot Design" Explanation for FL 13 Doesn't Hold Up

William Steve Lang's "Notes Regarding "Ballot Formats, Touchscreens, and Undervotes"

Notes regarding "Ballot Formats, Touchscreens, and Undervotes:
A Study of the 2006 Midterm Elections in Florida"
by Frisina, Herron, Honaker, and Lewis.

[ To download the study critiqued here, click here ]

Steve Lang writes:

I put my margin notes from reading this into an email. If you're
interested in the Florida Ballot article, here's my critical review
as I read it....
Taken from my margin scribbles with quotes from the article. You
might want to get a copy of the article in front you to follow the
notes:
------------------------------------------------------------------------
Page 1: "reports of election-alterning ballot machine failures have
thus far been relatively few." What? In the first paragraph, the
authors make a claim too early in the process to know and
inconsistent with popular information, so there is a political agenda
to this article.....hmmm....what is it?

Page 3: "the legal definition of malfunction" does not discuss a
fraud or hack to flip or discard votes. Is a machine that is
hackable (not secure) in a reasonable way designed poorly or
malfunctioning? This seems to avoid a reasonable review of an
important issue.

Page 4: "our conclusion that the CD13 undervote problem resulted
from the ballot layout used in Sarasota County" is premature. There
are some logic and grammar problems in the article, but I'm not going
to note them all here. This one is a big leap where evidence might
indicate a ballot issue, but that is not the only possible
conclusion. The authors backtrack a little, "it remains possible
that a programming or design flaw in...touchscreen machines caused
low vote counts"...but "it is our belief that..." indicates a priori
belief that taints the conclusion. I believe in Santa Claus...

Page 5: Consistent with most correlation/regression analyses,
there's no way to avoid that projections based on existing data give
a linear trend for missing votes that would elect Jennings and
reverse the election. I don't like the use of "disparate impact"
here as it confuses the term with discrimination (EEOC) studies that
have a legal basis and definition. I think they misuse the term to
extend to voting machinery. A minor issue.

Page 6: BIG PROBLEM, "The fact that some CD 13 ballots do not
contain a vote for a given office does not imply a failure of the
machinery..." In fact, as many as 9 types of ballots were used in
Sarasota alone. Adding or subtracting a "roll-off" is EXACTLY what
might have cause a misalignment of the touch point or page error in
the ballot or allow a hacker an entry point or cause the published
practice ballot to differ from the actual screen...etc., etc. This
is a very poor assertion to me and seems to protect the machines at
all cost. I'm getting the idea that the authors represent DRE
manufacturers...hmmm....

Page 7: The authors avoid citing that election machinery affects
valid vote rates and undervotes? Why? There are more than three
explanations for the undervote. The exclusion of programming issues
and/or fraud makes no sense and to say they are not common
conjectures is simply WRONG! They protest too much, so they may have
a conclusion that they are unwilling to produce, so the ones they
exclude are likely the ones to investigate!

Page 8: There is little or no support for a protest vote to apply
here, the citation is old and weak. Should not be the first
suggestion! Why propose this? Dumb idea to start with...Ballot
formating has ONE famous incident, the butterfly, but it's not
typical and everyone knows about it now. There are sample ballots,
voter training, etc. Also, poll workers look for this problem now,
and there is not evidence this is a problem on touchscreen machines?
What a leap of logic?!?!

Page 9: Ah, ha! The plot thickens: "It has been conjectured...:"
Politely, that's bull****! Little or no one has worried about this
and it is not part of any certification process. Most statute, if
any applies, would only require that some races or names come in a
certain order: by party or alphabetical. In California, there are
cases of several pages of candidates for a single race that didn't
"confuse" anyone. IF you can read and have ANY experience with ATM's
etc., paging and scrolling is so common that it annoys, but doesn't
confuse most people! Would a voter NOT expect to look for a name on
a list that might not be first? Dumb, dumb, dumb...."beyond the
scope of the paper", that was your opening premise! Very bad logic.
Why show placement of a ballot that is NOT the ballot that is your
issue? In figure 1 (page 10), how would two races on a page as
shown confuse someone? Machine or engineering failure is a possible,
if not likely reason for undervotes. Needs anecdotal evidence?

Page 10 / 11 (figure 2 - optical scan). This example did NOT have a
large undervote. It seems much more confusing than the screen shot
of the touchscreen. How is that consistent with your "candidate
placement" hypothesis?

Page 11: "This paper is purely a statistical exercise..."
Unnecessary and very bad statement after 10 pages of introduction
that don't even mention statistics or statistical logic or methods.
IF you can cite and discuss "ballots", then why not cite and discuss
"engineering"? Weird!

Page 12: "Thus, if there were a generic iVotronic effect in the CD
13 race, it would have affected both Charlotte and Sarasota undervote
rates." Why would it? Different programmers, roll-offs, and
election supervisors? Different ballot designers? There is NO
support for this conclusion...it is arbitrary! These authors must be
working for iVotronics!

Page 13, 14 and Figure 3, 4: Figure 3 is accurate, but poorly
labeled (y axis). Doesn't mention that the Absentee Voters are
sometimes on a different machine (optical), so why is that ballot
design and not machine error? Authors are shooting themselves in the
foot! "...and perhaps Sarasota absentee voters are simply more
attentive than Sarasota election day voters to undervotes..." DO
ABSENTEES USE DIFFERENT BALLOTS? Yes, they are sending in bubble
sheets and not using touchscreens. How does that preclude a
touchscreen error - seems to refute your hypothesis! Figure 4 is
unnessary, and badly designed. Bars are too small for the window
graph to show effects.

Page 15: Again, the authors suspect a possible issue and avoid it!
Fraud. IF there were fraud (hacking), it might show up as
differential undervote in precincts that had different proportions of
Demo/Ind/Rep voters! They know that...and blow it off as error
variance up front! What do they know? hmmm...."tend to contain
undervotes" Bull****

Page 16 and Figure 5: OK, a poorly formated graph, but the idea is
that Sarasota is different in undervotes for Election Day vs.
Absentee. Why not a SPLOM plot? Do authors have experience with
these graphics? Again, ballot design or counting process or what?

Page 17-19, Table 1-2: May not be the best model, but besides the
overkill analysis...If the conclusion is that "precincts with high
Senate undervote rates tend to have high Congressional undervote
rates" and .."we also see from this table that Democratically-leaning
precinct tend to produce disproportionately many undervotes..."
WELL, why is that consistent with BALLOT design and NOT FRAUD? If
it's not machine "failure" (presumably similar regardless of party
affiliation), and democrats are more harmed, then organized flips or
drops to throw the election might look like this regression wouldn't
it?!?! Citing yourself that democrats are dumb and can't understand
the ballot is the reference? There are too many covariates and model
choices here, and the authors don't do a good job of explaining all
the rationales for the choices. One footnote about nesting is typical.

In short, there needs to be more descriptives and simple correlations
before launching into this, so that the picture across counties,
brands of machines, and ballot types is explained. Assuming the
conclusion as given doesn't point to ballot design alone as of yet.

Page 20, Figure 6: Pretty clear that Sarasota is out of wack, but we
knew that...where are the Early Votes? The authors are doing a
little more exploration than planning here, but time is short
so...significance for Hillsboro, etc. is sample size more than
anything else. More interesting to me...why is Lee so low?
hmmm....they said earlier that they had more to say about that....not
outlined well...

Page 21-22: "If strong Democratic partisans simply dropped out of
the CD 14 race on account of its being a certain Republican victory,
then one might expect to see a negative lee County election data
effect as in Figure 6......" Pure speculation with no rational
evidence for this conclusion! Bull****.

They are correct, deliberate voters undervotes has no support...
[except from the election supervisor in the newspaper down here!].

"for reasons that transend this paper, voters are more likely to
undervote in a two-candidate race that appears immediately above or
conceivable immediately below a six-candidate race that includes a
write-in option, then we would expect to see a large Sarasota CD13
effect, precisely what we observe in Figure 6. We might also expect
to see a large Lee County CD 14 effect, which we do not see.
However, the CD 14 effect, we suspect, reflects partisanship
undervoting in a rather lopsided race." WHAT A CONVOLUTED BUNCH OF
JUNK! This is reaching a speculative (at best) prior conclusion and
rationalizing the argument to fit the predetermined conclusion in the
face of contrary evidence!

"Whether iVotroinc software is tailored for individual counties is
not something we know" YES YOU DO!! We know the ballots (roll-ons)
and local elections preclude programming for each county with
different election supervisors! How can the authors say this!!!
Bad, bad, bad. They KNOW this is a problem that is a window for
machine malfunction OR fraud and argue against it with "we don't
know" after speculating all over the place about everything else
under the sun....these authors have an agenda and it shows!
Kickbacks from iVotronics?

"or even voting machines within precincts." The authors cannot know
this because the precinct data is aggregated and not given by machine.

Page 23, Figure 7: Seems OK, but leaves out early voting (again).

Page 24-26, Table 3, Figure 8: They don't provide the analysis?!?!
Over 50 regression lines? Why not SPLOM? For that matter, this is a
mess. There may be a multivariate solution (discriminate), but his
whole thing falls into the weird category. There's NO REASON, even
if you assume that Charlotte and Lee have issues with undervote on
the Attorney General race that is not present in the other counties
that the CAUSE is Ballot Design. There is simply no reported
evidence in this paper (or anywhere that I know of) that voters
reported that they didn't see the AG race!?! Also, all these
counties had multiple variations of ballot pages with local races and
roll-ons. Most had a review screen. How do we know the review
screen in the AG race in Charlotte and Lee were not flawed resulting
in error. There are a thousand possibilities here, including fraud,
and it's a stretch to suggest ballot design and confusion -
especially with the poll workers watching for a "new machine"
confusion. The EARLY VOTING was on touchscreen. The authors leave
the touchscreen data out of this analysis all of a sudden! Is this a
manipulation? Did the touchscreen data from early voting confirm the
author's analysis, or was it left out on purpose! hmmmm.....I'd
check that!

The authors admit, "If Sarasota experienced a generic machine
malfunction of some type, then this same malfunction affected two
other neighboring counties but only in one particular race (and not
in the same race that the malfunction appeared in Sarasota). Of
course this could have happened: flaws in software code can interact
in ways that are hard to predict, and we can never rule out the
culprit here is i Votronic sofware or hardware that malfunctions when
races are grouped." DAMN, they said it themselves! Why argue
something when something else is easily possible. Also, why not a
hacker? It's not the data pattern that's a problem here, but the
conclusions stink to high heaven!

Page 28-29, Figure 9: The idea of a Democratic Voteshare is
OK....I'm not sure about the formula. With all the regressions in
the previous section, why not use the democratic, republican,
independent rations? Why not include a number of convariates?
Here's the early votes missing from the previous section...makes the
previous section see like a manipulation. Regardless, the point is
"Is this accident theory plausible?" I agree, assuming a linear
relationship, etc...the case that Jennings was harmed by what
happened in comparison with predictions from other precinct data is
compelling, and everyone who looks at the data would reach the same
conclusion - though maybe different predictions as to HOW much harm
or why it happened. With the election so close, you'd have to
overturn the results or revote.

"This effect seems to be more pronounced in the early voting than in
the election day voting, which is consistent with the story that some
poll workers were warned of the problematic Sarasota ballot format
problem and tried to pass on this warning to voters who voted on the
day of elections." OK, then why was there any ballot problem? [all
poll workers got the memo as reported in the paper]. This argues
AGAINST a ballot design issue. How were Lee and Charlotte poll
workers different than Sarasota? etc. etc. Why was the early data
(that followed the same pattern on the same machines) NOT included in
the above analysis on page 23-26?

Page 30-31, Figure 10: If a "constant fraction of all votes were
being randomely suppressed and converted to undervotes" and that
occurred to defeat the democratic candidate, why could that NOT have
fraud as a possibility instead of ballot design?

This indifference theory of "engaged voters" is more bull**** without
evidence or reason to think it is more plausible than anything else.
Pure speculation and not consistent with voter reports of problems of
review pages changing votes or the race not appearing even when the
voter looked for it...why would a hacker not simply drop a random
percent of independent and democratic voters? A sophisticated hack
would drop a percent of all voters, but a smaller percent of
republicans...you can imagine all kinds of possibilities. This still
is a stretch and the comparisons of regressions lines doesn't
indicate that someone who is informed they missed a vote would be
more inclined to fix the problem because they were "engaged" than
because of an accident. If you are enough motivated to vote, would
you then not care about the Senate race? We are in "la-la land"
with this theory...reminds me of stories of "reluctant responders" to
polls and other fictional characters! Can you imagine the authors
of this paper explaining the interaction effect of indifferent vs.
engaged voters and reluctant vs. non-reluctant poll responders! Wow.

Page 33: I don't believe it! I feel like Cassandra who can predict
the future! They are now speculating on interactions. "The true
process that occurred in Sarasota is undoubtedly of our accident and
indifference stories." The only thing to agree with here is that
they made up a story!

Page 34-36, Table 4: There are assumptions: "The relationship
between electoral races are the same across all precincts in CD 13."
that may not be true if a hack formula was based on registrations
within the precinct or other predictions. Regardless, the
prediction of the actual magnitude of votes to Jennings may be off,
but with this resampling model based on an "accidental" ballot design
theory is conservative and ok if WHATEVER HAPPENED is due to some
kind of random undervote issue. If the problem was not random,
Jennings may have won by more! It doesn't matter, since the election
could be fouled up and easily called for the wrong candidate, the
only solution is a revote.

I don't think the assertion of 90% that Jennings would gain > 369
votes without ballot issues is stated correctly. This result is
predicted IF ANY random error in those precincts affected were
similar to the non-affected similar precincts, then repeated revotes
from the same population would add this number of votes about 90% of
the time....if there even wasn't a random error, and if the other
local unaffected precincts were not similar, then the numbers are
off. Regardless, it makes you think that there's certainly a reason
for Jennings to be pissed off.

Page 37-40: The authors repeat my limitations above..."we assume
that, if elections were commonly administered across counties...."
and "There are no further substantial and systematic voting machine
or ballot format effects...." The authors are not idiots, but they
were on a mission to conclude something that excluded the machines
from the start so the conclusion would fit the data. We might call
this the butterfly ballot effect: if it happened in Florida, it must
be the ballot! Maybe they work for iVotronics...who knows? IF
there was a ballot design problem, it is not clear from this why
other problems were not possible.

"individual-level ballot data"? How can they get that?

Putting every race on a single page would take a LOT of pages! I
can't imagine that making it less confusing to vote! Oh, well the
authors would only put IMPORTANT races on separate pages? Maybe the
authors would give a personality test to voters to see if they were
"engaged" before voting...etc., etc...not good suggestions either.

Not really a good set of conclusion for political scientists: How
about a suggestion that the election be revoted? How about
comparison with similar undervote issues in other parts of the
country? etc., etc...

_______________________________________________