Re: Stat function requests

["David J. Braden" <dbraden rochester rr com>]

On Wednesday, October 23, 2002, at 04:57  PM, Morten Welinder wrote:

>    (1) For distributions, return 0 for x <= lower-bound, and 1 for x >=
>    upper-bound.
>
> While XL-incompatible, this might be useful.  Not my call, though.

It *is* useful, especially in simulations when working with mixtures on 
dist. parameters.  It doesn't hurt that it is standard practice as well 
<g>.

I would argue that while it doesn't replicate Excel's behavior, it 
certainly isn't a strong incompatibility.  Indeed, your GAMMAINV is 
incompatible with Excel along the same lines: when p gets too close for 
your algorithm to work, you switch to #VALUE! instead of returning 5E+06 
(a la Excel, which BTW is inconsistent with what it does for NORMINV, 
but hey...). Your binomdist does the reasonable thing (you evidently are 
using beta-tails for the calcs) to give a far better implementation than 
Excel's BINOMDIST; no one will complain about the incompatibility there, 
will they?

<<snip>>

>    (3) Fix the discrete densities to correct what Excel does, namely,
>    truncating doubles, where integers are expected, before evaluation.
>
> We cannot do this.  All floating point arithmetic has rounding errors
> and we would get some really weird failures if we tested for intergers.
> Things might work for an explicit "3", but fail for (3*1.1-0.3).

Good point. A colleague suggests a global "fudge factor" that is 
user-accessible to take care of this.  Something as simple as abs(x - 
round(x)) < fudge.

DaveB