[R] Parameter estimation of gamma distribution

[Prof Brian Ripley]

On 15/02/2012 14:18, Yogs wrote:
> Hi,
>
> I am trying to estiamte parameters  for  gamma distribution using mle for
> below data using fitdist&  fitdistr functions which are from "fitdistrplus"
> &  "MASS"packages . I am getting errors for both functions. Can someone
> please let me know how to overcome this issue??
>
> data
> y1<-
> c(256656, 76376, 6467673, 46446, 3400, 3100, 5760, 4562, 8000, 512, 4545,
> 4562, 4645, 4521, 4624, 5110, 4678, 4782, 4587, 491100,48426, 457940, 45000,
> 4820, 48225, 5415, 482466, 185535, 174346, 46875, 4876467)
>
> Results
>> fitdist(y1,"gamma")
> Error in fitdist(y1, "gamma") :
>    the function mle failed to estimate the parameters,
>                  with the error code 100
> In addition: There were 17 warnings (use warnings() to see them)
>
>> fitdistr(y1,"gamma")
> Error in optim(x = c(256656, 76376, 6467673, 46446, 3400, 3100, 5760,  :
>    non-finite finite-difference value [1]
> In addition: Warning messages:
> 1: In dgamma(x, shape, scale, log) : NaNs produced
> 2: In dgamma(x, shape, scale, log) : NaNs produced

Rescale your values, as the help page suggests.

Note:

      Numerical optimization cannot work miracles: please note the
      comments in ‘optim’ on scaling data.  If the fitted parameters are
      far away from one, consider re-fitting specifying the control
      parameter ‘parscale’.

Using 1e-6*y1 works.


-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595