Update: There is some additional analysis (plus a couple of new graphs) here.
Here come more graphs of the undervote in the Minnesota race for US Senate.
The big change here is that I figured out a way to extract the precinct ballot count* from the Minnesota SecState website, so I have a much more accurate count on the true number of undervotes, without having my data polluted by the presidential undervote. To my knowledge, I'm the first to extract the true undervote count-- everyone else seems to still be using the "pres minus senate" method, which is off by about 30%.
The freshest data, pulled from the state website minutes ago, shows Franken down by 206 votes. The total presidential undervote is 10086. The total senate undervote is 34916. If the senate undervote is allocated to Coleman and Franken along their fraction of the Coleman+Franken vote in that precinct, Coleman would receive 16573 new votes, Franken 18342, for a Franken gain of +1769 (see update below). This means about 12%, or 1 in 8 undervotes would need to turn into a newly counted vote for Franken to tie the race. But of course, take that number with a huge grain of salt because there's no reason to believe that things will break down that neatly.
Figure 1 shows the distribution of the presidential undervote, as a function of the Franken part of the Franken+Coleman** vote:
Figure 2 shows the distribution of the senate undervote. Note the extra "bulge" on the right side; that should give Franken fans some hope:
Figure 3 shows the distribution of the total ballots cast:
Figure 4 shows the senate undervote rate. There seems to be a gentle upward slope from left to right:
I should spend a little bit of time cautioning people that it's almost certain that votes newly counted during the recount won't split along the lines I've suggested. It's impossible to tell which undervotes are intentional and which are not, which are due to machine error, and which are due to voter error, where the voter errors are likely to happen, which precincts used ES&S and which used Diebold machines, or any of a million other things that could dramatically change this race. 206 votes is a tiny margin, and the smallest issues could easily drown out any effects I've shown here. End result: this race is still a toss-up. I also recommend Nate Silver's analysis for a look at the undervote from the demographic angle.
Late update: It's probably unfair to allocate undervotes along the F+C breakdown, so if I allocate them taking third-party votes into consideration, the breakdown is Franken +15619 (44.7%), Coleman +13956 (40.0%), third-party 5341 (15.3%), for a Franken gain of +1663.
Extra bonus figure: this is what the undervote distribution looks like (red=Coleman, blue=Franken, green=other):
The graph seems pretty clear: Al Franken is likely to pick up a big boost from those precincts where he dominated by a large margin.
Looks like Nate Silver has taken this even further.
* The data is slighly inconsistent, with a few precincts showing higher vote totals than ballots cast. I assume this is because there are absentee ballots that did not arrive at the polling place before the polls were closed. For example, county 73 (Stearns), precinct 0450 (Zion Twp P-1), is showing a ballot count of 50, timestamped 11/4 10:52 PM, but there are 107 ballots counted (and Coleman beat Franken here 2:1, so I doubt they're going to call any attention to
this discrepancy). In these cases, I simply take the higher number as the true ballot count. This raises the final ballot count by 238 votes.
** Because there was a significant third-party vote in this race, I categorized the data based on Franken's percent of the total votes cast for Franken or Coleman. That is, the 50% mark means ballots were cast in even numbers for Franken and Coleman, regardless of the number of ballots for Barkley in those precincts.