Guest lecture by Wixted (of the power law Forgetting curve).
The central point here was that people thought that in criminal trials, eyewitness’s confidence ratings should be ignored, since they don’t correlate with accuracy (i.e. using a correlation coefficient). But with binary values, a correlation coefficient is a poor estimator. It’s better to use signal detection approaches… i.e. “When they give a confidence rating of 5/5, what’s their accuracy?” With this approach, we find that eyewitness memory is actually quite reliable: notorious cases of innocent convictions are almost always those with low confidence; high confidence is generally quite accurate.