Thread Subject: Matlab probability cut-off value

Hello,
Sometimes I ran a t-test for a super significantly different groups, the Matlab often returns a p value which equals 0. that is p=0.
What does this mean, what is the p value at which Matlab stops given a number and gives a zero? P<1e-30 or p<1e-100? Etc…?

On Jan 8, 5:20 am, miki lo <myart...@gmail.com> wrote:
> Hello,
> Sometimes I ran a t-test for a super significantly different groups, the Matlab often returns a p value which equals 0. that is p=0.
> What does this mean, what is the p value at which Matlab stops given a number and gives a zero? P<1e-30 or p<1e-100? Etc...?

I had the same question and found the answer
by trial and error. Although I don't remember
the answer, you can repeat what I did.

Hope this helps.

Greg

that would be great!!! thank you.
All I need is a number.

miki lo wrote:

> Sometimes I ran a t-test for a super significantly different groups, the Matlab often returns a p value which equals 0. that is p=0.
> What does this mean, what is the p value at which Matlab stops given a number and gives a zero? P<1e-30 or p<1e-100? Etc…?

Miko, it's certainly easy enough to figure this out empirically by
experimenting with the line of code in TTEST that computes the p-value

     p = 2 * tcdf(-abs(tval), df);


and working back from there to figure out what effect size in your data
corresponds to a particular value of the t-statistic TVAL. For example,

 >> df = 25; tval = 1e8; 2 * tcdf(-abs(tval), df)
ans =
   4.7085e-184

which corresponds to having a difference in means that 1e8 times the
estimated std dev.

I think you'll find that for a t-test, the possible p-values can be
quite small -- maybe even realmin, I forget. But you ought to ask
yourself why the difference between a p-value of 1e-100, say, and a
p-value that's exactly zero (because of underflow) matters to you. The
approximation of the test statistic's sampling distribution as a t
distribution is almost certainly not very accurate that far out in the
tails in most cases involving real data.

Hope this helps.

- Peter Perkins
   The MathWorks, Inc.