In the interest of rigour alone I support Edward J. Felten's call for rigourous analysis of the 'effect of electronic voting'
http://www.freedom-to-tinker.com/archives/000714.html

This data gives a some very troubling empirical support for that call.

I can't find any evidence that there was funny business in Florida.

I constructed a quick dataset for each county in Florida. The dataset included number of votes cast for each candidate, the number of registered voters for each candidate, the technology used in each county, the vendor supplying the technology, the votes cast in the 2000 election, and a lot of census bureau data for the county.

Attached is a log file (http://www.halplotkin.com/Nickerson ) for a quick analysis I ran (flat text file -- best viewed in courier or courier new).

Pay attention to the coefficient on "touchscreen." Initially, it appears that counties that used touchscreens were more likely to vote against Bush by 6 percentage points (and it is almost statistically significant, but with only 69 counties -- one shouldn't expect statistical significance).

However, once you control for Gore's two-party vote total in 2000, that number falls to 1.3 percentage points. A much smaller number,
but enough to swing the election.

Alternatively, you can drop the advantage of optical scan systems (i.e., the disadvantage of touchscreens), by controlling for the
percent of a county that is classified as rural. Again, 2.4 percentage points is sizable.

Controlling for both Gore's vote share and the rural composition reduces the technology difference to 0.2 percentage points. The difference is no longer meaningful.

Substituting in the difference in the Republican and Democratic registration in the county ("demgopregdiff") actually flips the sign
on the touchscreen variable. When a sign flips, the chance of a variable measuring something meaningful is very small.

Adding controls for the percentage of black residents returns the touchscreen coefficient to the negative realm, but it is now less than 0.1 percentage points.

My conclusion is that it is EXTREMELY unlikely that funny business occured with the Florida election. Gore's vote total is the single
best predictor of a county's vote (in fact, it explains 94% of the overall variance alone) and technology washes out very quickly as a factor. I've tried a number of different ways of measuring and operationalizing -- there is no stable finding for touchscreen v. optical scanning.

I'm not happy with the result of the election, but I don't think the Republicans stole it.

Thanks, David.

Two questions:

Does your analysis get at the question of whether their was a difference in the final vote vs. the exit polls (pre post election adjustment) in counties that used electronic systems vs. counties that used paper ballots? (or similarly, in the states that used paper ballots vs. those that used voting machines, per the chart at: http://www.plotkin.com/ChargesofElectronicCoup.htm

and;

Does it make sense to rely on Gore's 2000 vote total per county as a variable given the questions that have been raised about the tens of thousands of voters who were allegedly unable to cast their votes or have them counted for Gore? Wouldn't those two errors kind of cancel each other out?

No, it doesn't look at exit poll results. The only reason it does not is that I couldn't find county-by-county data for exit polls. If you have the data available and could pass it on, I would happily look at the difference in exit polling between electronic and optical scanning technology. From looking at Professor Duster's graphs, it looks like he has only state level data.

I should be able to track that down easily enough. Maybe I can get to that in the next few days (but no promises).

Does it make sense to rely on Gore's 2000 vote as a variable?

I think it does for a couple of reasons:

1) The correlation between vote in 2000 and vote in 2004 for counties is EXTREMELY tight nationwide. (see
http://polysigh.blogspot.com/2004/11/county-level-regression.html ). Most of the country didn't exhibit the problems that plagued Florida in 2000.

2) The two biggest errors I heard of that hurt Gore were:

a) Over voting (7000+ voters punched Al Gore and wrote in Al Gore, thereby invalidating their ballot);

b) Palm Beach's butterfly ballot (costing him another 8000 or so);

c) General intimidation and long lines;

d) Felon's list.

"A" is widespread and I don't think would be concentrated in any particular county. "B" is located in one particular county -- so we might expect Palm Beach to be an outlier, but that doesn't invalidate the analysis. "C" occured this year as well, so it is just as much a factor. And "D" was minor in the grand scheme of millions of votes. A personal tragedy and troublesome and it should be fixed, but in a state with millions of voters -- it isn't going to bias estimates.

In short, I don't see why adding Gore's percentage would bias results. It goes a long way to explaining Kerry's vote total in '04. It offers the single best baseline by which to predict this year's election results.

Here are links to two other posts that seem to contradict David Nickerson's calculations regarding the vote in Florida.

The first was compiled by Elizabeth Little. It can be found at: http://ustogether.org/election04/Little_Analysis.html

and the second was done by Jeff Chambers at Tulane University. It can be found it: http://synapse.princeton.edu/~sam/Bias%20in%20Florida%20opscan%20vote.pdf

I have to say I am somewhat mystified by the synapse analysis. Yes, the technology is correlated with the size of the county. Why?
Because most of the problems with voting technology in 2000 came in urban areas, so those are the areas that were updated to touchscreen. Most of the rural areas had received optical scan technology prior to 2000 (which was the cause of complaints from groups like the NAACP -- optical scan systems tend to have low error rates compared to punch cards).

The second synapse analysis controls for party registration. That is a step in the right direction, but it completely ignores the history of the two parties in the south. there are still counties where the population is conservative and votes for Republican Presidential candidates, but everyone is still a Democrat. These counties voted overwhelmingly for Bush in both 2000 and 2004. Once you take this into account, the difference more or less disappears.

Is it possible that the county vote tallies were rigged?

I doubt it because they were the same in 2000 and they hand counted all the ballots in 2000. The recount was performed by members of both parties -- someone would have spoken out. Furthermore, when 75% of a county is Democratic and 85% vote for a GOP candidate, surely SOME local person
would have spoken out. Obviously, the local people think the result makes sense. So I think the synapse people are missing the boat. I stand by my results.

I am further mystified by the insistance of the synapse people to use the total vote tallies. What they are really picking up on is the degree to which urban areas used touch screens and voted for Kerry. To give an example of how weak this is, attached is a text file (see: http://www.halplotkin.com/NickersonDataSet2.log )
with more regression results. The dependent variable is the margin of victory for Bush over Kerry in each county. Initially, the touchscreen sure seems to favor Kerry. But once you control for the number of registered voters in the county (regression #3) or the total population in the county (regression #4) the result becomes statistically insignificant and flips sign.

Furthermore, look at the R-squared term -- very little of the overall variance is being explained. [Note: One shouldn't fetishize the
R-Sq term, but it does suggest there is a lot going on in the analysis that is not being explained. The fact that 2000 vote tallies explain almost all the variance is extremely telling -- see the following set of regressions: http://www.halplotkin.com/Florida_Quick_Instability_Demonstration.log ]

As for the exit polling data, the NEP hasn't released the raw numbers yet. Eventually it will probably go into the Roper Collection. Until then, all we have are the midday numbers (4pm) and the final weighted numbers (which are adjusted to correspond with actual turnout). It really is impossible to know what is going on without the raw unweighted numbers. If Democrats showed up to the polls earlier than Republicans, then the afternoon polls may still be overweighted with Democrats. [Note: if Republicans were less likely to take the survey at all time periods, then every exit poll would understate the number of Republicans. It drives me crazy that refusal rates aren't reported for exit polls.]

The Mystery Pollster has a good post on this issue today: http://www.mysterypollster.com/main/2004/11/vote_fraud_.html

If I get suitably inspired, maybe I will look at the state level exit polls. But the state level analysis doesn't allow for differences in technology. Cheating in areas where one can follow the paper trail and recount ballot is extremely risky. A scandal of that proportion would make Watergate look like a college prank.

Since every major newspaper has access to the raw exit polling numbers, I have to imagine that someone would pick up the story and would look at the data carefully. Talk about a way to secure your place in history.

I hope this helps to clear things up.