How to choose appropriate linear model? (ANOVA)

I'm wondering how to choose an appropriate linear model for a given
problem. I have been reading Applied Linear Regression Models by John
Neter, Michael H Kutner, William Wasserman and Christopher J.
Nachtsheim. I'm still not clear how to choose an appropriate linear
model.

For multi-factor ANOVA, shall I start with all the interaction terms
and do an F-test to see with interaction terms are not significant,
then do a linear regression on a model without the non-significant
iteration term?

Could somebody point me some good book or chapters on this topic?