0

I'm working on a regression analysis which involves eight symptom domains comprised of individual survey items. As is the procedure in my lab, the domain scores are simple averages of the corresponding items.

I started with a stepwise logistic regression procedure to determine which domains best predicted my binary outcome. If a domain was found to be statistically associated with my outcome, I tested the corresponding individual items in a similar stepwise process to determine which items were "driving" the observed associations in the domain regression.

My question is this - are folks aware of any statistical references or research that I could cite which would support this sort of drill-down model building process?

letsplayhorse
  • 1
  • 1
    Many people in this community are aware of plenty of references that suggest this may *not* be a good approach: see https://stats.stackexchange.com/search?q=stepwise+regression – whuber Dec 17 '20 at 22:05
  • thanks @whuber. understand the concerns regarding stepwise. i suppose i could have highlighted my primary question - if you find a factor/domain/latent construct is a significant predictor of an some outcome in a regression model, is there a reference that can be cited (as requested to find by my PI) that would justify regression using the individual items? – letsplayhorse Dec 18 '20 at 03:45
  • I would hope there isn't much literature on this, because--if I understand your procedure correctly--the standard methods of multiple regression (namely, F tests for groups of variables) will be more powerful and informative. – whuber Dec 18 '20 at 13:04

0 Answers0