Decision trees with binary split-offs can be modeled using nested and sequential logistic regressions. I feel rather confused about the difference between the two, although I do have some thoughts.
Nested logit seems to be appropriate in situations where
- Every subjects walks through the entire decision tree
- There are random effects across the subjects at each level
- We are interested in within-level variation
- Covariates are at level-specific
Sequential logit seems to be appropriate in situations where:
- At every split-off, some subjects are lost whereas others continue
- We are interested in between-level variation
- Covariates are subject-specific
In Stata, there are the corresponding commands
seqlogit. Any comments would be much appreciated.