Regression Discussion > Multicollinearity - regression
Great questions Gavrilovici!
There is not a specific correlation level that two predictors cannot exceed to be in a regression together, but multicollinearity is certainly something to consider as you conduct your analysis. One way to do that is to examine diagnostic statistics that will let you know if multicollinearity is a problem in your model. In SPSS, you can examine the "Variance Inflation Factor" (VIF) and/or the "Tolerance" to see if there may be a problem. For the VIF, you start to worry about tolerance if the VIF is above 2.0. For Tolerance, there may be a problem if it gets below .3. HOWEVER, these are only general guidelines and I would be very careful about following them to rigidly. Multicollinearity, as a assumption problem, should be considered on a spectrum, not as either "there" or "not".
There is a really great example and explanation of how to diagnose and deal with this issue here:
Jeremy Taylor
Firstly, I want to excuse my grammar mistakes, as I am a student from Romania and english is not my native language.
I am currently working on the statistics part of my thesis. My objective is to discover which factors predict the appearance of muscle dysmorphia simptoms in male weight lifters. Basically, I plan to test an etiological model of the disturbance that was created a few years ago.
The model has 'muscle dysmorphia symptoms' as an outcome (dependent variable) and, among others, perfectionism, body dissatisfaction, self-esteem and negative affect as factors with a possible influence on the development of muscle dysmorphia. I know that the best way to test a conceptual model is by using a path analysis or SEM tehnique, but my coordinator and I have agreed that these statistics procedures are way above my level of knowledge (Bachelor level). So, my coordinator suggested using a regression analysis, but I can't figure out which method of regression to choose, as my predictors interact with each other in a complex manner and I've found medium correlation coeficients between them. So, my question is: if I want to use the forced-entry method, is there any maximum level of multicolinearity admitted between the predictors?
If so, which other method of multilinear regression do you suggest using?
Thank you very much for your help