On 7/20/2007, "Eike Rathke" <> wrote:
>Hi Andreas,
>
>On Thursday, 2007-07-19 14:31:07 -0600, Andreas J. Guelzow wrote:
>
>> In the current draft functions such as TINV require the degree of
>> freedom to be an integer that is at least 1. Our don't see a
>> mathematical reason for the "integer" requirement.
>
>Correct me if I'm wrong, but isn't degrees of freedom the number of
>parameters that can vary?
The "degrees of freedom" is a parameter for the Student T Distribution.
In many places where the Student T Distribution is used this happens to
be an integer. Occasionally an non-integer is being approximated by the
next larger integer since caclulating the Student T Distribution is hard
and so tables were used.
> How much sense would one and a half parameter
>make?
I'll try to remember to dig out an example where non-integer df is used.
(These are not your common text book exercises since they assume
tables...)
>
>> Is there some other reason?
>
>The spreadsheet applications I checked (Calc, Gnumeric, Excel; Kspread
>1.5.0 seems not to have it) deliver identical results for fractional
>values as if truncated to an integer.
True for Gnumeric 1.7.10 and before, not true for current cvs. (We had a
bug report with respect to large degrees of freedom which caused this to
become an issue.)
If you give a mathematical program such as Mathematica a try you will see
that they give you different values for non-integer df, in fact the
Student T Distribution is continuous in df.
I think that this is a situation where we have to decide whether to go
with what spreadsheets do or with what is mathematically sensible.
Andreas
