25 years of conventional evaluation of data analysis proves worthless in practice
Nice
paper from Isaksson & Gustafsson at Uppsala which appears to demonstrate the unreliability of bootstrapping and cross-validation when the ratio of sample size to natural variation is too low. The problem is that it’s difficult to know what the natural variation is when you’ve got a small sample size. Looks like Bayesian confidence intervals may provide a sobering reassesment of many medical trials. See also the Ioannidis PLoS paper and E.T. Jaynes.
3 years ago
