Calculating p-values after regression with bootstrapping

[Carlos Toro]

Greetings,

I have performed a bias-corrected and accelerated non-parametric bootstrap for obtaining Prevalence Ratios (GLM Poisson, link log and robust variances)
Is it possible to obtain p-values for my variable "Fisica"?

Thanks in advance

[Carlos Toro]

Please some help

[Joro Kolev]

To calculate properly a bootstrap p-value, you need to use bootstrap data generating process which is consistent with the null you are testing. In other words you need to bootstrap under the null that you are testing. This is not easy in general. For an example how it is done in another context you can see the manual for -bootstrap-, the section
"Achieved significance level
Example 3", p. 243.

If you want to just report something like a p-value because you do not like confidence intervals, you can assume that your sampling distribution is normal, divide your estimate by the bootstrap standard error, and calculate the p-value from there:

. dis 2*normal(-.8878/.0604)
6.573e-49

this is the p-value for the null that the parameter on Fisica is 0.
What I have done in the second paragraph is a dodgy exercise, because the whole point of bootstrap is to not-rely on normal theory.

[Mike Lacy]

While I'm not an advocate of p-values, I would point to the built-in -permute- command, which gives a randomization-based p-value, i.e. one that enforces the null hypothesis, which as Joro points out, is needed for a p-value.