shabbychef / sadists

Some Additional Distributions R package
GNU Lesser General Public License v3.0
8 stars 0 forks source link

Imprecise kprime CDF and PDF when the degrees of freedom are small #15

Open dcousin3 opened 3 years ago

dcousin3 commented 3 years ago

Hello,

I noticed that sadists's pdf and cdf estimates for the K' distribution are quite inacurate when degrees of freedom are small. This is visible, for e.g., n1=9 and n2=8 with noncentrality parameter = 1: image

Things are worse when the noncentrality parameter is increased to 9: image

I tried changing the parameter order.max, but it does not seems to be improving the results, and for order.max beyond 15, the functions returns NaN only.

Is there a possibility to increase a MAXITER parameter or a TOLerance parameter that would improve the estimates? Is it possible to have such arguments accessible in the R function? (in addition to order.max)

I tested the FORTRAN files from Poitevineau and Lecoutre (2010) and they seemingly return very accurate estimates for any sample sizes (TOL must be at least 10.e-6 and MAXITER, 10000). Sadly, they do not possess a r (random number generator) component.

Denis.