What is a correct approach to choose "Number of best models to see" under "all possible model selection (JMP Pro)"?
If I choose lots of models the Akaike weight (wi) of each variable will be lower than a group of models with less variables.
Thanks
It somewhat depends on the characteristics of the data table you are analysing and the questions you want to address with your model. Can you please describe the number of predictors and rows in your table, are your predictors correlated? Also what questions are you aiming to answer with your model?
Hi Malcom, thanks for your interest. I'm looking for the best model using (AICc) to predict bird abundance using and 11 not correlated environmental variables.
If your predictors are not correlated then you could use stepwise regression using AICc as the criterion. If you suspect higher order effects you may also want to add interaction and quadratic terms in addition to linear (main) effect terms. Then see what stepwise regression using AICc gives you. If you are using Pro version of JMP and you have 100+ rows I would also try using a validation column and use stepwise regression with model validation as that may gve you greater protection from over fitting.
The advantage of viewing more than the best model is that there might be an alternative not different from the best that is better explained by science. Since AIC and BIC are likelihood based, you can think of the units as standard deviations and anything within 2 or 2.5 units as not statistically different. How many alternatives you need to look at depends on the data and the model. If you're expecting certain variables to be in the model, I'd start with a small number, rerunning the platform with an increased number of models, until either those variables show up or the last model is different from the best. If not, I'd only look at the best model from each group.