There are 3 conditions which have to be met for the initial
audit:
1. at least 2% of the precincts in each state must be
selected;
2. at least 1 precinct in each county must be selected;
3. all precincts must be selected randomly.
These are not mutually exclusive if you write a procedure such
as:
First, DO NO HARM!...woops, never mind, that would mean
keeping lever machines! But I digress.
Let me try again:
First, randomly select 2% of the precincts in the state.
If there is at least one precinct in each county selected,
proceed with the initial audit.
If there is not at least one precinct in each county selected,
continue to select randomly until this criteria has been met.
(This will result in more than a 2% initial audit, which still
meets criterion #1 above as well as criterion #3.)
In this way, all 3 conditions are satisfied.
The question is, what percentage should the initial audit be,
and what how much auditing should there be in the event of a
close election where 2% is unlikely to find fraud that could
reverse the outcome of the race?
As far as confidence intervals, I'd say we should look for 99%
instead of the usual 95%, because the hypergeometric
thingamujig only finds the FIRST corrupted precinct, not any
additional ones, unless you audit more. I'd hate to miss the
tip of an iceberg by limiting the confidence interval to 95%
instead of 99%. See the hypergeo spreadsheet we've used in the
past below.
My Excel doesn't work with the hypergeometric distribution and
more than ~7000 precincts with a 2% audit. In a state, there
could be 20,000! So how about posting a table of the 2% audit
with 20,000 precincts using this spreadsheet? If you need a
copy, I'll send it to you again. Here it is with 7,000
precincts:
Total Corrupt % Audited Expect. Std. Odds of
finding
Prects. Prects. Corrupt Prects. Value Dev. 1st Corrupt
Prect.
7000 3500 50.00% 140 70.00 5.92 100.00%
7000 3150 45.00% 140 63.00 5.89 100.00%
7000 2800 40.00% 140 56.00 5.80 100.00%
7000 2450 35.00% 140 49.00 5.64 100.00%
7000 2100 30.00% 140 42.00 5.42 100.00%
7000 1750 25.00% 140 35.00 5.12 100.00%
7000 1400 20.00% 140 28.00 4.73 100.00%
7000 1050 15.00% 140 21.00 4.22 100.00%
7000 700 10.00% 140 14.00 3.55 100.00%
7000 350 5.00% 140 7.00 2.58 99.93%
7000 280 4.00% 140 5.60 2.32 99.69%
7000 210 3.00% 140 4.20 2.02 98.65%
7000 70 1.00% 140 1.40 1.18 75.86%
7000 35 0.50% 140 0.70 0.83 50.78%
7000 7 0.10% 140 0.14 0.37 13.19%
As you can see, with 7000 precincts, a 2% audit can find the
first corrupted precinct with as little as 4% total corrupted
precincts. If the vote swing (fraud) on the 4% is 25% or less,
a race with a margin as little as 1% can be audited
effectively. So if you have a LOT of precincts, only the
closest elections could evade such an audit. I'd like to see
this with 20,000 precincts if you don't mind. My Excel won't
do it.
Then we have another problem, which is that House races are
not statewide, and would therefore have fewer precincts.
But hey, at least we're talking about it!
Don't forget to check the "Check here if you want to
format your message in plain text. Use for posting code
snippets." box before posting from Excel.
Thanks!