www.cmm.bristol.ac.uk/forum

Hi,

I fit multilevel nominal models using mcmc by running runmlwin in Stata. In their results tables, I found a pattern that the p values of coefficients tend to be smaller than expected. For example, there is one coefficient like this:
Mean S.D. EES P 95% cred. Interval
-.2965969 .1812569 548 0.049 -.6563954 .0521605

Based on the Mean and sd, the p value should not have been below 0.05. The CI (-.6563954, .0521605) also suggests that zero can not be ruled out.
Why the p value is 0.049?

Does anyone know this? Thank you very much!

Best wishes,

Rodrigo

Dear Rodrigo,

The default presented p-value is a Bayesian one-sided p-value (the proportion of the MCMC chain which is the opposite sign to the point estimate).

You report a p-value of 0.049 so a conventional two-sided p-value would be more like 0.100

You can actually request the 'properly calculated' classical frequentist two-sided p-value to be displayed in the runmlwin output when fitting models by MCMC by specifying the zratio option.

Best wishes

George

Thank you, George. Your reply is very helpful!
best wishes,
Rodrigo