I'm working with proportion data as a responsive variable in my model. For working with this type of data, I used the Gamma distribution in the GLMM model. However, even using this distribution, the distribution of my residues is only homogeneous if I transform my response variable (I used the square root arcosene transformation - more suitable for proportion data) and remove the outliers.
Important note: When I remove outliers, my results do not change.
What is the best option for this case? Can I transform my response variable, even using a GLMM? Do I take outliers or stay with them (even though they affect my reside)?
Hi Carolina Mendes Muniz,
As far as I understood, your response variable is a counting variable, therefore a GLM with "Poisson" distribution and "log" link function is the standard.
If your response variable is ranging between 0 and 1, a binomial family with the "logit" link function should be a better approach, both approaches using the non- transformed response variable.
Start plotting your response variable and see how it looks like.
Not sure if my answer was helpful. Maybe, you will get a better return if providing a reproducible example outlining what you need.
Good luck!
- Badges
- Science topic
- Similar topics
- Computer Communications (Networks)
- Simulators