Date: Nov 12, 2000 6:29 PM
Author: Rich Ulrich
Subject: manual recount - of punched ballots

posted also to sci.stat.edu  where more discussion has taken place.

On Sun, 12 Nov 2000 11:36:54 -0500, Thomas Kragh
<tkraghNOSPAM@NOSPAMeecs.umich.edu> wrote:
re: "How do you feel about manual recounts in Florida?"
>
> About the same as I feel about the use of an "Actual Enumeration" during
> the census. A manual recount is obviously less accurate then a machine
> tally, but for some bizaare reason the luddites & politicians in our
> society believe it to be better in some sense...
>


You have apparently missed the point. It has been in the newspapers,
on TV, and in sci.stat.edu:

*In particular* with punched ballots, it is possible for the punched
"chad" to fail to fall out. As a result of that problem (and others),
Maine (for one instance) abandoned/outlawed the use of that style of
ballot.

The manual recount is thus more *complete* and accurate. Its main
function is to correct for that error; Bush -- this proves that the
gods are toying with us -- signed a law for his state of Texas, to
provide for *manual recounts*.

With 10,000 no-punches where only half that many no-votes should be
expected (in Palm Beach County), they re-counted a 1% sample and came
up with 47 additional votes -- about half of the 100 or so that were
possible, and consistent with the number of no-votes that typically
are seen. There was no report of how many no-punches had existed.
Gore gained, as he was expected to, because Gore carried the county by
almost 2-to-1.

Gore gained 19, which the official extrapolated to 1900.
My extrapolation would be a smaller pickup -- but I haven't nailed
down how to place limits.

At one extreme, we could extrapolate 47, as 1%, to "4700" as the total
vote, and then I would prefer to apportion 4700 by the observed
breakdown between candidates. That would give a point estimate of
1400 or so. If I place 95% confidence limits on '47' as Poisson
variation, that adds in about 30% error, around the 1400.

There must be other ways to estimate. I don't know how I justify
ignoring the 19 vote marginal gain (33 vs 14) that was observed.

Anyway, ignoring the 19,000 double-punched ballots which were mostly
intended for Gore, and ignoring the 3000 or so intended-for-Gore votes
that Buchanan seems to have picked up elsewhere, it does appear that
Gore will carry the state without going beyond this most obvious step
of correction. The lead as reported by AP was 325, and the absentee
difference favoring Bush is expected to be perhaps that size
(apparently, 54%-46% on 2300 ballots in 1996: about 200 votes).

I don't think that ANY re-vote is pragmatically feasible.

Speaking of "actual enumeration" as with the census -- using
statistical estimation and adjustment is the best way to actually
enumerate, that is, "count," for changes in population < I don't know
which political point TK was trying to make >; but there is not a
reliable starting point for voters. And we don't want to encourage
anyone to play games by intentionally messing up ballots.

--
Rich Ulrich, wpilib@pitt.edu
http://www.pitt.edu/~wpilib/index.html