I use gamlss method from library(gamlss) on my full models with interaction terms and try to reduce them with stepGAIC. There are 3 things I want to ask.
- Do I have to specify a link for the model?
- stepGAIC give multiple timesthe following message "Model with term cov1:cov2 has failed" and doesn't continues with reducing my full model. What does this error message mean? Does it make sense to continue the calculation
- In many posts I read that forward/backward method of the stepAIC method is not good-
are there any "good" alternatives to do that?
Thanks for the answer! I'm very new to GAMLSS Theory. I actually wanted to fit a LM on my continuous variables, but I think that there is heteroscedasticity present, so I switched to GAMLSS to try out other distributions than normal distribution. I'm not really sure whether a link function is needed. Does gamlss this automatically if I specify a distribution from it's selection?
The reason I use stepwise reduction is that I have around 20 other variables to look at and I don't have much time. Strange thing is that if I make forward selection from the Null model with the step function for the LM model (with lm function) I get a model with maineffects and several interactions. Everything is ok. If I do the same procedure with the gamlss and specify the distribution as NO (Normal distribution) my best model is the Null model! Maybe I did something wrong, when I specified the backward selection:
dist<-c("NO","GU","EXP")
j=1
formula<-var~1
object<-gamlss(formula,k=k,family=dist[j],data=measurements)
red.model <- stepGAIC(object=object, scope~(cov1+cov2+cov3+cov4+cov5+
cov6+cov7+cov8+cov9+cov10)^2,
direction="backward",k=k,trace=TRUE)
Is it normal that stepGAIC is extremely slow in calculating the coefficients? It took 24 hours to calculate 4 steps in stepGAIC.