Hi All.
I am building a churn model and using a host of models - logistic, Decision Tree & Random Forrest.
Currently i am in the initial phase of my first model - Logistic Regression.
There are multiple variables of which only 3-4 are numeric. If the outcome variable would have been numeric [Linear regression] we could have used VIF & tolerance to study which one to keep or remove, whereas the outcome in the case is dichotomous [ 1/0]. There is a bit of confusion as in some of the research articles people are still recommending using VIF for binary outcome variables while in some cases they are not.
Please advice.
For multicollinearity problem you can use correlation mxtrix for the independent or explanatory variables of the model. Correlation coefficient higher than 0.8 indicate multicolinearity.
Hair et al 2010 recommend percentage of correctly clasified cases at least 75% so the result of the logistic model to be considered.
Regards
Alma
I'm not sure I understand your question fully, but if your VIF and tolerance levels are your outcome variables of interest, and they are not binary (i.e., are continuous variables), then you can classify them as binary by using a cutoff value. For example if your values range from 0 to 1, you may to choose 0.5 as the cutoff---higher than 0.5 is "yes" (or 1), lower than 0.5 is "no" (or 0). Similarly, if they are predictor variables, you can change them from continuous to binary by using a cutoff. The cutoff value should be based on their optimal predicting power, and not necessarily at 0.
- Badges
- Science topic
- Similar topics
- Mathematics