I'm reading the introduction to statistical learning book the following reason is one of three for why LDA maybe used over Logistic Regression. However it gives no further detail or explanation and i find it quite vague and lacking context.
- When classes are well-seperated, the parameters estimates for the logistic regression model are suprisingly unstable. Linear discriminant analysis does not suffer from this problem.
What does it mean by classes being well-seperated, could someone provide an example of this? For example would it mean 0 and 1 responses are clearly seperated by a boundary? Also what does it mean by the parameters being unstable? Finally why does LDA not suffer from this problem?
I really don't like skipping over important assumptions, i'd really appreciate a clear explanation so that i get a feel for what they're referring to. Thanks!!