Hello everybody,
I suspect multicollinarity in my structural equation model because a standardized beta weight is above 1. Thanks to other questions & answers here, I learned that this might be a sign of multicollinearity in the model. Before I add any relationships within my model, I would like to test whether there actually is multicollinearity in the SEM.
In multiple regression, I have already used VIF and tolerance to inspect multicollinearity. For this, I used the R command vif(model). However, when I try this with my SEM model, it doesn't work.
Now my question is: Which test and which R command do I use in order to test for multicollinearity in my SEM model?
And if I detect multicollinearity, I read that I could add relationships to my model (e.g. correlation or causation). How many and which relationships would I add?
Thank you very much in advance for your answers!
You could examine a confirmatory factor analysis (CFA) model first, in which you specify all your factors that play a role in your SEM. In the CFA model, you can allow all factors to be correlated. The estimated correlations (to be found in the completely standardized solution) tell you which factors may be highly correlated/collinear/redundant.
Also, you can specify separate models in which each of your predictors is regressed on the remaining predictors. The R squared values from these models tell you about collinearity. The tolerance statistic is equal to 1 - R^2 and VIF is equal to 1/tolerance and thus is easy to calculate based on these additional models.
Hello Antonia,
beyond what Christian mentioned, we need (as always) more info about the model (optimally an attached path diagram) and a bit background.
Principally, standardized effects (no matter if in a regression or SEM) can be above 1 (it is not as in a correlation which is bound to 1) but I agree that most often there are some issues. I doubt however that this has something to do with multicolinearity as this would not case large effects but large SE.
But let's see that after more info.
All the best,
Holger
Calculte VIF if the value is less than 10, you are good to go. No multi-collinearity
Calculte VIF if the value is less than 10, you are good to go. No multi-collinearity
Dear Antonia Fischer
R has three diagnostics for detecting multicollinearity:
1. Examine the correlation matrix for strong predictor factors.
2. Calculate the VIF and the tolerance statistic.
3. Find Eigenvalues.
- Badges
- Science topic
- Similar topics
- Psychology