In MLR,,,the main reason for low tolerance or high VIF is the overlapping {similarity} between predictors. present to us the variables...,, or you can do bivariate correlations for your IV and see if the high correlation is the same source for the multicollinearity.
This depends on the cause of the high multicollinearity. Did you include an interaction term? Did you include highly correlating predictors, which might be redundant? Please provide some more information about your design.
In MLR,,,the main reason for low tolerance or high VIF is the overlapping {similarity} between predictors. present to us the variables...,, or you can do bivariate correlations for your IV and see if the high correlation is the same source for the multicollinearity.
Solving a problem of multicollinearity requires a qualitative judgment about which variables to remove from your model as there is no statistical way of making this choice.
For example, owning a waistcoat watch might predict mortality age but it is not as useful as measuring wealth or diet in explaining this phenomenon. There might be a high correlation between owning a waistcoat watch and wealth but wealth would be a better variable to retain.
You do not necessarily remove variables completely from your model. Another option would be to run a PCA on the highly correlating variables and to use the factor score as a predictor in your multiple regression. But again, the strategy depends on your original design and your variables as Peter pointed out.
In near future, I will upload the paper about this problem. I asked Dr. Sall about the threshold of VIF. He answered me it depended on the data. I think you are careful in the case "VIF>=20". Download my papers about CPD data from RG.
The independent variables are Body mass(kg), BMI, WC(cm), Body fat(%), FM(fat mass) (kg) and FFM (fat free mass)(kg). Some of these variables are associated with each other (weight and BMI, body fat% and FM)(r>0.95). I used multiple liner regression to show relationship between these variables and Omentin (something like hormone).( In fact, I want to determine some predictors). After analysing by SPSS, VIF was more than 10. can I remove Body mass and Body fat variables? when I removed these variables, VIF reached to 2-3. But I am not sure whether this way is true or not? (I used SPSS).
if you include variables which are created from each other (like an interaction term in moderation analysis), just like BMI, mass and height then it is not astonishing that you have high VIF values. What is the gain of including all of the variables if they have such high intercorrelations??? The information is highly redundant.
The question which variables you can omit should be theory driven. Do you want to make predictions basing on the mass or the BMI, what is the theory behind it? It is not a good idea to put everything into the model and hope for the best, i.e. somthing gets significant.
As you noticed, if you omit highly correlating variables, your VIF values decreases, but nobody can tell you which one to eleminate, except you and the theory.