06-25-2017 06:28 AM
I need help on choosing the cut-off p-value for interaction term in linear regression model.
It seems to be a rule of thumb for setting p-value<0.25 or <0.2 for including potential interaction terms in logistic regression model, as mentioned by this website
I ran linear regression models (outcome is continuous variable, sample size=4300) and I'm not sure that I can use the same cut-off p-value to distinguish an interaction term in my model.
I would love to hear from your experience.
06-25-2017 11:03 AM - edited 06-25-2017 11:05 AM
One person's "rule" is another person's "ridiculous", or vice versa. "Rules" that work fine in one subject matter may be total hogwash in a very different subject matter.
There is no mathematical or objective way to determine the proper p-value. You pick one that you feel comfortable with. You can be conservative, and pick small p-values, or liberal and pick large p-values, it's up to you.
Just remember that whatever rule you pick, there are consequences — you may reject hypotheses that should be accepted, or vice versa.
06-25-2017 01:24 PM
That method of analysis is also known as 'p-value' hacking to some degree these days and you should be very careful with this, especially if you're trying to publish your results.