OK #statistics lazyweb, I have a measured dataset that's expected to follow a gamma distribution, and a weird simulation process that takes as input the median, 5%, and 95% quantiles. Am I right that fitting my own gamma distribution and passing the quantiles of that will be more stable than just passing empirical sample quantiles?
Follow
@infotroph My hunch is that you are correct. Fitting a gamma distribution using empirical quantiles should in theory get rid of sampling noise.
I have a strong feeling that, especially for small empirical sample sizes, the 5th/95th quantiles will be particularly sensitive to small sample sizes (more than the median). Do you only have one sample? Maybe you could bootstrap the quantiles?
