What Chance Of Innocence Can Be Tolerated For Conviction?

[Siegmund]

The instructions from the court tend to be something a bit vague, and my answer will float a bit based on the severity of the situation.

For a traffic-ticket appeal, I am happy with 5/100, and could probably tolerate ten.

For a misdemeanor something less than 1/100 is ok - if I had to cite a specific one-size-fits-all number I'd probably pick something close to 1/100. For something like a murder trial I would be very nervous if there was as much as 1 in 1000 chance of being wrong.

There's a tradeoff between how many guilty people are let off the hook and how many innocent ones are jailed. People, especially classical statisticians, are prone to always fixing the risk of Type I error.

A better approach might be to compare the relative severities of being wrong in each direction, and find the spot on a power curve that matches that. In the case of the traffic fine, both types of error are more or less trivial; in the case of a murder, I would judge condemning an innocent man to be something like 100 times worse than freeing a guilty one--- and my equilibrium point would be something like accepting a 10% chance of freeing a guilty man in exchange for a 1/1000 chance of condeming an innocent one.