Hi all,
David King wrote:
Moving discussion of this issue here, at Regina's request.
Hi Regina
The function ZTEST should depend on NORMSDIST.
Therefore NORMSDIST has to be controlled and improved first.
I'm afraid I disagree pretty whole heartedly :)
NORMSDIST may or may not be more accurate in Excel, but in Calc
it is accurate enough to use.
For example for our 2nd example the stats package R gives a p-
value of 0.02275 or 2.275% - accurate to just 4 digits - that's
good enough for practical purposes. Both Excel and Calc have more
digits.
R works internally with 64 bits internal precision if I am correct. You
can obtain more digits, but the truth is, you almost never do need them
for a p-value.
From a very old version of R:
The floating-point control setting has been changed to use 64-bit
rather than 53-bit mantissas, which will result in slightly
different but generally more accurate results.
I only added the accuracy of the SD-algorithm into the discussion,
because the naive single pass algorithm used inside the ztest routine
may give very wrong results, invalidating the whole z-test (wrong by
orders of magnitude in either direction and even negative variances,
therefore raising an error).
Sincerely,
Leonard
The point is that Calc ZTEST is *broken* and has probably always
been - there is a clear and simple bug that I've found (and even
provided a couple of lines of code to solve).
Where Excel calculates 2.275%, Calc calculates 0.00039%. The idea
of this statistic is to compare it against a preset significance
level of maybe 5%, or 1% or 0.1%.
Calc is giving the wrong answer; in a situation that demands a
simple pass/fail answer Calc gives the wrong one. That's serious
and needs action - it can't wait until someone figures out how to
improve NORMSDIST's accuracy from 7 places to 13 or whatever.
The good news is that it's very easy to fix :)
All the best
David
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
