When using GAMs to model ecological count data (e.g. fish counts) using Poisson or negative-binomial distributions, the response variable is raw counts and the offset (e.g. log-transformed fishing effort) is included using the "offset()" or "offset=" functions. However, when making predictions based on model fits, the predict.gam() function doesn't explicitly incorporate the offset variable.
Questions, in no particular order:
- Why is the offset typically log-transformed? Is it a rule?
- Can the offset be included as a main effect instead of using the "offset()" or "offset=" functions in order to include it in model predictions?
- Is it acceptable to pre-calculate a standardized response variable (e.g. fish catch-per-unit-effort), multiply the number to avoid fractions (which aren't suitable for Poisson and negative-binomial models), and then avoid using the offset altogether?
In my research group, we're getting a lot of questions such as these from our professors/ committee members/ reviewers, so any insight to help us clarify/ justify our approaches would be much appreciated.
Thank you, Denise Colombano