On 04/01/2013 14:23, gus gassmann wrote: > If you only have 10^4 samples, then any estimate of the 10^(-7) quantile > is going to have pretty serious estimation error attached to it. Even if > you have 10^7 samples, the 10^(-7) quantile is determined by a single > observation. I would not put too much stock into that estimate.
I think the same. That is the reason for which I'm asking here. :-)
> More to the point, one is tempted to ask *why* you need this tail quantiles.
I need a version of the AD test which returns uniformly distributed p-values in the widest possible range.
> What kind of precision do you require?
Precision is a word I wouldn't use, because I calculate the confidence interval for the critical values of the AD statistic. I'm using a confidence interval of +/-5*10^-5 at 99% confidence level.