In 2001, the National Opinion Research Center at the University of Chicago undertook a comprehensive review of almost all ballots rejected in initial counts in the state of Florida in 2000, a process known as the Florida Ballot Project. NORC provides a wealth of data for public consumption, some of which is especially helpful to our purposes here.
Specifically, I examined records for a set of 6,902 undervotes in a series of 37 Florida counties that used optical scanning technology (as Minnesota does) in 2000. This excludes data from Volusia County, which NORC recommends against using because of ambiguities surrounding which ballots were and weren't included in its initial count.
NORC researchers went through each of these ballots, attempting to determine whether voter intent was discernible as a result of any kind of marking on the ballot. Among the 6,902 undervotes:
-- In 4,505 cases (65.3%), there were no clear markings on the ballot and voter intent was not discernible. In most or many of these cases, the voter probably skipped the presidential race intentionally.
-- In 179 cases (2.6%), there were markings beside more than one candidate. Although these votes should probably have been classified as overvotes rather than undervotes, in either eventuality they are not very useful to us.
-- In the remaining 2,217 cases (32.1%), there was a marking next to one candidate only -or- there was an affirmative marking next to one candidate and a negative marking (such as the candidate's name being crossed out) next to one or more candidates (such negative markings were rare and represented just a small fraction of the total).
There were 34,916 unrecorded votes in Minnesota's senate race. If the voter's intent can in fact be discerned in 32.1% of these cases, that would total 11,208 reclassified ballots. This would likely be more than enough to give Franken a victory given that his voters are probably more likely to have cast a discarded ballot.
However, I believe this may overestimate the number of correctable errors for two reasons. Firstly, the proportion of intentional undervotes (i.e. where the voter skipped the race on purpose) is liable to be higher the further one goes down the ballot; a top-of-the-ticket race like the presidential election in Florida does not make for an apples-to-apples comparison with a second-on-the-ticket race like the senate election in Minnesota . Secondly, I have only examined the data for undervotes and not overvotes, and overvotes are usually harder to correct during a recount.
Still, let's dig a little deeper and see what we might find. The NORC data set also provides a brief description of the nature of the voter's mark in cases when a ballot is reclassified upon a hand recount. Among our 2,217 reclassified ballots, the description of the marks were as follows:
746 Arrow/oval filledIn 746 of 2,217 cases -- almost exactly one-third -- the voter appeared to have filled in the oval or arrow properly, but the machine did not record the vote. These votes can be thought of as cases of machine error (although, in some subset of cases, they may consist of instances where a voter used a nonreadable marking device such as a dull pencil). The other two-thirds of cases are various forms of voter error, such as a voter circling a candidate's name rather than filling in his oval.
442 Arrow/oval marked other than fill
(circled, X, ?, checked, scribble)
418 Other mark near arrow/oval
263 Other mark on or near party name
227 Other mark on or near candidate name
89 Circled party name
32 Circled candidate name
This distinction between machine error and voter error happens to be particularly helpful to us because the state of Minnesota periodically conducts an audit of its optical scanning systems. This audit is designed to detect the incidence of machine error and machine error only; cases of voter error are explicitly disregarded by the audit.
In Minnesota's 2006 senate race, the audit detected just 53 discrepancies out of 94,073 ballots tested, or an error rate of 0.056%. However, these are the cases of machine error only, whereas the state has a liberal voter intent law to cover cases of voter error as well during the process of an actual recount.
From our Florida data set, we believe that machine error represents approximately one-third of the total number of correctable errors. That would imply that about 0.169% of ballots -- roughly 1 ballot out of every 600 cast -- will be reclassified in Minnesota. Given the total number of ballots cast in Minnesota's senate race, this translates to 4,835 ballots that will in fact be reclassified during the hand recount.
Would this number be sufficient to provide Al Franken with a victory? It is very, very close. Using the Daily Kos estimate that 52.5% of recounted ballots will go to Franken (after dropping votes for third parties), we estimate a net gain of 206 votes for him, which is almost exactly the margin by which he presently trails Norm Coleman. (The margin is in fact exactly 206 votes as of this writing).
Another piece of information from the Florida data set, however, contains heartening news for Franken. Among the 2,217 reclassified ballots, 1,129 (50.9%) went to Al Gore, and 1,013 (45.7%) went to George W. Bush, with the balance going to third party candidates. This doesn't seem like that big a discrepancy. However, the counties that were using the optical scanning system in 2000 tended to be wealthy, suburban and quite red. When we look at the results of the officially-tallied ballots in those counties (weighted based on the number of recounted ballots) we find that 56.2% of those votes went to Bush and just 42.2% to Gore. Instead of losing to Bush by 15 points, however, as he had among regular ballots in these counties, Gore beat Bush by 5 points among the reclassified undervotes.
(As an aside, Gore and Bush split roughly evenly those ballots that were rejected due to apparent machine error, while Gore beat Bush about 3:2 among undervotes due to voter error).
The Daily Kos estimate of 52.5% reclassified ballots going to Franken is based on an examination of precinct-level data (specifically, the weighted average of already-counted ballots in those precincts which were scored for Franken), and therefore may be closer to the mark than the county-level data I examined in Florida. Even so, however, there is a certain amount of heterogeneity within a given precinct, and it is probable that those voters within a precinct that mismarked their ballots are more likely to be disposed toward Franken. If anything, therefore, the 52.5% estimate may be conservative.
The more that I examine this data, the more I'm beginning to believe that the number of reclassifiable ballots may be relatively low, but that the proportion of such ballots that are resolved in Franken's favor may be relatively high. How these two factors will ultimately reconcile themselves, I don't know.