P value output of 0.0e+00

[Donald Lyall]

Hello,

I have worked on a problem relating to P-value output recently, and despite my research have not been able to find a solution.

I run:

"logistic Disease Variable Confounder, robust"

And recieve Variable P 0.000, however I would like this output in scientific notation. Therefore I run:

"logistic Disease Variable Confounder, robust pformat (%5.3e)"

And recieve Variable P 0.0e+00, which is not very informative. The P value for confounder is 3.9e-72, which is fine.

*****
My question is whether there is another way to format P-values to make 0.0e+00 more informative. I have been told that the smallest P-value that Stata can give is around 1x10-300, and I'm not sure how to report 0.0e+00.

Many thanks for reading.

[Joseph Coveney]

Maybe you could

test Variable
return list

Why is P = 3.9e-72 fine? Why isn't P = 0.0e+00 sufficiently informative?

[Marcos Almeida]

I'm quite curious about the aim of providing p-values under such degree of precision. By reading the commands to perform the logistic regression in #1, it seems the field is related to health sciences. Being this so, I fear most - if not all - scientific journals on health sciences demand to describe tiny p-values as just "below 0.0001", for p-values, we know, don't convey the magnitude of the effect. Maybe I'm losing the point, but I fail to envisage why going towards "nanomolecular" p-values (in this case, by the way, under "robust" calculations) would be necessary to provide trustful estimations.

[Nick Cox]

To add a little to Marcos's excellent post: Often the P-value is the most fragile result of all, so treating it as the focus of the analysis can be dangerous as well as distracting.

[Steve Samuels]

The American Statistical Association has recently issued a statement about p-values, available here. The statement itself begins on page 8. In the years when I quoted p-values, my practice was to present values like Donald's as "p< 0.01".

[Clyde Schechter]

While I agree strongly with the views expressed in #2 through #5, it is possible that Donald is researching variables relating to multiple gene alleles. In polygenic studies, it is now conventional to adopt significance levels like p < 1e-6, or p < 1e-10, etc in order to control the false discovery rate. So it may be that to fit in with the expectations of that sub-specialty, he needs a p-value that is more explicit than 0.

On the other hand, I don't know at what point Stata reports 0.e00 for a p-value rather than a more explicit extremely small number. It may be that this only happens when the test statistic falls beyond the domain where Stata can calculate the value of its distribution function, so that no more explicit value is possible.

[Marcos Almeida]

As far as I know, the New England Journal of Medicine has the highest impact factor when it comes to health sciences. Below, among the instructions for authors (http://www.nejm.org/page/author-cent...ipt-submission), I share the excerpt on the presentation of p-values:

• Results should be presented with only as much precision as is of scientific value. For example, measures of association, such as odds ratios, should ordinarily be reported to two significant digits.
• Measures of uncertainty, such as confidence intervals, should be used consistently, including in figures that present aggregated results.
• Except when one-sided tests are required by study design, such as in noninferiority trials, all reported P values should be two-sided. In general, P values larger than 0.01 should be reported to two decimal places, those between 0.01 and 0.001 to three decimal places; P values smaller than 0.001 should be reported as P<0.001. Notable exceptions to this policy include P values arising in the application of stopping rules to the analysis of clinical trials and genetic-screening studies.

[Donald Lyall]

Hi folks,

Many thanks for your comments.

> While I agree strongly with the views expressed in #2 through #5, it is possible that Donald is researching variables relating to multiple gene alleles. In polygenic studies, it is now conventional to adopt significance levels like p < 1e-6, or p < 1e-10, etc in order to control the false discovery rate. So it may be that to fit in with the expectations of that sub-specialty, he needs a p-value that is more explicit than 0.

This is precisely correct. My issue is that 'P = 0.0e+00' doesn't seem like an accurate output, and that there is perhaps another way of expressing it.

[Joseph Coveney]

.ÿforvaluesÿzÿ=ÿ37.5190(0.00001)37.5191ÿ{
ÿÿ2.ÿdisplayÿinÿsmclÿasÿtextÿ`z',ÿ2ÿ*ÿnormal(-abs(`z'))
ÿÿ3.ÿ}
37.519ÿ4.51e-308
37.51901ÿ0
37.51902ÿ0
37.51903ÿ0
37.51904ÿ0
37.51905ÿ0
37.51906ÿ0
37.51907ÿ0
37.51908ÿ0
37.51909ÿ0

So, maybe P < 4.51 × 10^-308?

[Donald Lyall]

Thankyou, Joseph - that is very helpful!