|
|
Christine wrote:
I'm trying to fit my data to a log gamma distribution, i.e. fitting the logs
> of my data to a gamma distribution. How and when do I take the exponential
to get statistics on my original data set?
Hi Christine -
I know next to nothing about the log-gamma distribution, but I'll take a shot:
The reference I have, Johnson&Kotz's Continuous Univariate Distributions,
describes the log-gamma as the distribution of y when -log(y) has a gamma
distribution, where 0<y<1. They also parameterize the LG in the same way as the
LN is, that is, the same parameters as the standard gamma dist'n. If you have
the Statistics Toolbox, you can just log and negate your data, then use GAMFIT
to estimate parameters. You can also use GAMCDF to computed estimated quantiles
of the fitted gamma distribution, and then negate and exponentiate those to your
data scale. What you can't do is simply exponentiate the mean or variance of
the estimated gamma. Hopefully you have a reference on the LG that gives the
first few moments of the dist'n in terms of the parameters. Johnson&Kotz seems
to think that the mean is (b/(b+1))^a and the variance is (b/(b+2))^a -
(b/(b+1))^(2*a). You should check this on your own.
Hope this helps.
- Peter Perkins
The MathWorks, Inc.
|
|