Doctor Q
Feb 1, 2008, 02:18 PM
California will be testing a new method to determine how much hand-counting is necessary to ensure proper election outcomes based on machine tallied votes. The method is the brainchild of statistics professor Philip Stark at the University of California, Berkeley.
It seems almost obvious in hindsight, but his method is to study both the results of small hand counts and the overall automated results, then determine how much additional hand counting is necessary to achieve a given confidence level about the election result.
Current law doesn't take those factors into account, and there is often a hand count of a fixed 1% of ballots even when a race is a landslide, or when more hand counting is needed in a close race. Under the new method, they'll do initial hand counts from random precincts, then use the results to determine what further hand counts are needed.
In the U.C. Berkeley article (http://www.berkeley.edu/news/media/releases/2008/01/31_audit.shtml) about the new method, Stark makes the interesting claim thatOther experts have tackled this problem, too, but primarily from the perspective of computer science, not statistics.
That seems oversimplified to me. The difference he seems to be highlighting is that he looks for confidence levels in the automated result, while earlier methods looked for the incidences of errors. What really matters is that his method produces better decisions about when and where to do hand counting.
The article points out that, mathematically, the number of incorrectly decided election races will be less than one half of 1% when recounts are used to obtain as little as 50% confidence. By targeting a much higher level of confidence, they can use selective hand counts to reduce the chances of that to a very very small percentage.
There is a basic assumption, at least the way they handle vote reviews in California, that machine counts have a margin of error while hand counts are perfectly accurate. I often wonder why they make that assumption so easily, but I'm glad that, overall, they are using experts in statistics to help us use recounts most wisely.
It seems almost obvious in hindsight, but his method is to study both the results of small hand counts and the overall automated results, then determine how much additional hand counting is necessary to achieve a given confidence level about the election result.
Current law doesn't take those factors into account, and there is often a hand count of a fixed 1% of ballots even when a race is a landslide, or when more hand counting is needed in a close race. Under the new method, they'll do initial hand counts from random precincts, then use the results to determine what further hand counts are needed.
In the U.C. Berkeley article (http://www.berkeley.edu/news/media/releases/2008/01/31_audit.shtml) about the new method, Stark makes the interesting claim thatOther experts have tackled this problem, too, but primarily from the perspective of computer science, not statistics.
That seems oversimplified to me. The difference he seems to be highlighting is that he looks for confidence levels in the automated result, while earlier methods looked for the incidences of errors. What really matters is that his method produces better decisions about when and where to do hand counting.
The article points out that, mathematically, the number of incorrectly decided election races will be less than one half of 1% when recounts are used to obtain as little as 50% confidence. By targeting a much higher level of confidence, they can use selective hand counts to reduce the chances of that to a very very small percentage.
There is a basic assumption, at least the way they handle vote reviews in California, that machine counts have a margin of error while hand counts are perfectly accurate. I often wonder why they make that assumption so easily, but I'm glad that, overall, they are using experts in statistics to help us use recounts most wisely.
