I am looking for a way to standardize the coefficients of determination (beta coefficient) of the explanatory variables in a negative binomial function. I tried the lm.beta function in QuantPsyc package of R, but the returned coefficients are too small (seems to be divided by 100, compared to usual glm). Also, standardizing the variables to Z-scores prior to modelling also returns low values for the coefficients.
Does anyone know of a way to standardize these coefficients across different explanatory variables so I can select the variables that cause greater change in the response variable?
Thank you!
Hi Jonas,
If you are interested in variable selection, I suggest you to have a look at LASSO regression (which can be also used for glm, in your case with family being a negative binomial). You can find an answer that I have already made here : https://www.researchgate.net/post/How_do_I_reduce_the_number_of_control_variables_in_logistic_regression
Best regards,
Geoffrey
Hi Jonas,
You might try using a dominance analysis approach. Dominance analysis has been applied to logitstic regression and clearly the same logic can apply to a different glm.
Basically, if using the general dominance statistic, it would result in the percentage of the pseudo-R2 chosen attributable to that predictor. That interpretation sounds similar to what you were looking for. Though this technique is not useful for model selection.
Take a look at the attached paper and see if it applies in your case.
Alternatively, you might try MuMIn - it provides information criterion data which is better for selection (esp. the importance() function). Link to the package is pasted below.
- joe
https://cran.r-project.org/web/packages/MuMIn/MuMIn.pdf
Article Using Dominance Analysis to Determine Predictor Importance i...
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression