Dear Community,
I am interested in reporting the relative importance of my predictors in a GLM (Gamma distribution, log link). In other words, among significant predictors, I would like to know which ones influence the response variable the most. The model includes both numeric and categorical predictors.
I have read that partial R-squared could be appropriate to assess the relative importance of predictors in a GLM. Using the R package rsq (Zhang 2023), I managed to extract these partial R-squared values from my model. Yet, what worries me is that some of these values are negative - even from predictors that were returned as significant by the model. How is that possible, as partial R-squared explain some parts of the variance?
Do you know what it means exactly? As I am not sure to have fully comprehended the meaning of partial R-squared anyway, any reminder on this topic would be very helpful as well.
Thank you so much for your help,
Antoine
It could be due to the fact that you are running the GLM using unequal observations. In other words, either the number of observations in the numeric or the categorical predictors are more than the other.
Can you please give the reference that R-squared is appropriate to assess the relative importance of predictors in a GLM? Im am just curious. Or do you mean some form of dominance analysis to assess the relative importance? Here is a package, which might help:
https://github.com/clbustos/dominanceAnalysis/tree/master
Thank you so much for your answer.
- s. Rama Gokula Krishnan I have read elsewhere that unequal observations among predictors could explain negative partial R squared. Unfortunately, it does not resolve the issue here since all the predictors had the exact same number of observations (i.e., no NA in the dataset).
- Rainer Duesing This package is what I was looking for, brilliant! What is confusing although is that the outputs from dominanceanalysis::dominanceAnalysis() and rsq::rsq.partial() are different. For instance, dominanceAnalysis() does not return negative partial R-squared. I think it could be worth looking deeper into it to understand why this discrepancy occurs. For instance, partial R-squared can be computed using different methods - but it is not so clear what packge uses what definition.
I read about using partial R squared in GLM in Stoffel et al. (2021 - Article partR2 : partitioning R 2 in generalized linear mixed models
), even though this paper argues for semi-partial R-squared instead of partial R-squared.I'll go for dominanceAnalysis() for the moment but any additionnal comment on this topic would be very welcome.
Cheers!
- Badges
- Science topic
- Similar topics
- Biological Science
- Molecular Cell Biology
- Culture Cells
- Cell Line