Tuesday, November 11, 2008

Al Franken is a slight favorite to win the Minnesota recount

The latest count shows the in the Minnesota race for Senate the incumbent Coleman leads by 206 votes. To call that razor-thin is to insult razors. It's a dead heat.

The excellent website Fivethirtyeight.com has an analysis of the undervote.
There were also about 25,000 cases in which the voter voted for the presidency but undervoted the senate race (consistent with the AP's reportined finding last week). Let's assume that in most of these cases, the voter intentionally skipped the senate race, but that in one-third of cases he did not. This equals another 8,277 votes, or a total of 15,001 cases in which the voter intended to vote for the senate race, but his vote was not recorded.

In not all of these 15,001 cases, however, will the voter's intention be clear. Let's assume that one-quarter of these ballots will be unresolvable, even upon a hand recount. This means that 11,251 ballots will actually be reclassified during the recount, or about 0.4% of the total cast.

Bitwise notes, however, that Franken did in fact perform better -- really, quite a bit better -- in precincts with more undervotes. If undervotes follow the pattern of the recorded votes, then Franken would win 52.5% of recounted ballots (excluding any ballots cast for third parties). This is a significant finding, as these are the first numbers I have seen to break the undervote down to the precinct level.
Most undervotes represent an error by the machine in reading the ballot. Upon visual inspection the intent of the voter is quite clear. (Our optical scan reader in Tarrant County kicks such a ballot back without accepting it. I think that is a setting the election administrator chooses. Does Michigan not use that setting? I called the Michigan Secretary of State to ask (1 888 767 6424), but they are closed for Veteran's Day.) [*]

The question is whether or not Al Franken wins enough of the undervotes to make up what Coleman is ahead on the latest count. If the number of ballots that included undervotes are statistically similar to the rest of the ballots in each precinct, and if the number of precincts with undervotes tended to go strongly for Obama, then the likelihood that a visual recount will shift the overall vote from Coleman to Franken is very high.

How high? Fivethirtyeight.come does a sensitivity analysis. A sensitivity analysis is an estimate of how likely there will be a change in the final conclusion if there are changes in the assumptions in the calculation. The two assumptions evaluated for sensitivity are the percentage of precincts that went for Franken and how many of the questionable ballots themselves are likely to change.
Firstly, let's assume that my estimate of 11,251 recounted ballots is correct and hold this number constant, but vary the share of such ballots that go to Franken. Here are his win percentages under various such scenarios:

11,251 recounted ballots (0.4% correctable error rate)
======================================================
Recounted Ballots
Resolved for Franken Franken Win %
50.0% 1.85%
50.1% 2.93%
50.5% 13.39%
51.0% 44.82%
51.5% 80.18%
52.0% 96.61%
52.5% 99.75%
53.0% 99.99%

Alternatively, let's assume that Bitwise's estimate of 52.5% of recounted ballots being resolved for Franken is correct, but vary the number of qualified ballots:

Franken Wins 52.5% of Recounted Ballots
======================================================
Number of Recounted Ballots Franken Win %
2,500 1.68%
5,000 54.60%
5,623 68.93%
7,500 92.49%
10,000 99.15%
11,251 99.75%
15,000 99.99%
20,000 100.00%

The long story short is as follows: if Al Franken in fact wins anywhere near 52.5% of the undercounted ballots, it is quite likely that he will prevail, even given what I would consider to be fairly pessimistic assumptions about the number of correctable errors. You could halve my estimate of the number of recounted ballots, for instance (to 5,623) and Franken still projects to prevail around 69% of the time. If, on the other hand, Franken only wins say 51% of the undercount, then the precise number of correctable errors is more important.
If there are no other significant factors to consider, then the only reason to disagree that Al Franken is a slight favorite to win is to quibble about the term "slight."

In this case it means "I am reasonably sure Franken has won the recount based on my analysis, but I want to leave room for the possibility that I have overlooked a significant factor or gotten the assumptions wrong somehow." That's what the sensitivity analysis is all about. If the number of precincts that went for Franken differs from the number of ballots the recount gives him in each precinct, that is a key difference. So is the number of qualified ballots. If they break towards Coleman rather than Franken, it changes the outcome. But the odds are on Franken's side. That's why Coleman went to court in his failed attempt to stop the recount.

My bet is that the Democrats now have 58 Senators.


[*] When I called, I got an automatic telephone response "The Secretary of State's office is closed today for Veteran's day.] Realizing I could not get through, I said "Curses!"

Without any hesitation the automatic telephone verbal response system responded to me "That response is not recognized."

It seemed almost human.

No comments: