Should other factors also drive final model selection?
I have used a linear mixed effects model to identify factors associated with depression scores over time (measured at 3 time points). I have been advised to use the AIC/BIC to assist with model selection. I used a forward selection approach (although you could say it is more bidirectional elimination, as I added and removed variables depending on significance and the AIC/BIC).
I ended up with two models. The final model (I will call this Model A) which only included factors significant at p
Thanks Pat, yes I don't feel super comfortable with the stepwise approach but there is limited research in this particular area around what are the longer term factors/predictors. Are you suggesting I based my model(s) around a more sound scientific theory e.g. If the literature suggests trait anxiety is a strong predictor I start with this? Issue is I have well over 40 predictors and some of them have never really been tested in the area.
Model selection seems to divide opinions. For example, Zuur et al. 2009, Mixed Effect Models and Extensions in Ecology with R provides procedures to backward/forward model selection. A part of their proposed protocol (in many cases) is model selection using AIC/BIC. They also discuss the ways to spot correlation handling correlation in model selection (among other things in data).
Thank you Patrick and Petri, yes these divided opinions make it hard for the novice to know which way to go! I like your suggested approach Patrick. I will check out the Zuur et al paper as well Petri. Cheers Bonnie
If your theory proposes a model that is too complex to fit with your data, in practice you have might have no alternative to model selection. With BIC you end up with simpler model than with AIC (but you should do model validation to ensure that the model is good: e.g., assumption about the normal distribution of the response variable holds).
If you have a large number of variables, start with variable selection methods based on the variance/covariance structure of the data or using supervised variable selection methods.
Parallelization of a Modified Firefly Algorithm using GPU for Variable Selection in a Multivariate Calibration Problem
Hello, Bonnie! I suggest you to read one of my papers:
I believe it might be helpful.
I want to select best possion model base on aic and bic ? Which I should select? The highest bic or aic mean that model better than other. Am I correct or wrong?
- Badges
- Science topic