Hello Ahmad,

First, did you do Principal component analysis or something else? People usually do PCA on categorical or dichotomous variables and it's wrong. You can CAT PCA, but not the "regular" PCA.

Second, I think that collapsing variables with  different scaling into one factor  is not so good. But it's just my opinion, I don't have references on that.

So, I would suggest to enter your variables into regression model separately. The problem of multicollinearity might appear when your variables are continuous, so in your case, you have nothing to worry about.

Bests.

http://www-01.ibm.com/support/docview.wss?uid=swg21477550

I agree with Marina

Another option is to run a Latent Class Analysis... given that you want to combine categorical data. 

a good intuitive explanation is given at:

https://methodology.psu.edu/ra/lca/example

Thanks Dear Marina and Seven. Clear and logical solutions!

Hi Ahmad. I believe your main issue is that using traditional approaches to dimension reduction assume a Pearson correlation (i.e, numeric measures). In your instance you will need to use polychoric correlations, which provides appropriate correlations for your measures. I'm guessing that the CAT PCA suggested by Marina is doing just that; not sure which program.

Assuming you have an EFA that suggests a 3-item factor, then your next question is how to use this in some predictive/regression analysis. I believe your best approach is to use these items as a latent value. If you are using a program such as AMOS, then your best bet is to enter the dataset as a polychoric correlation matrix (with SD and means). Unfortunately, this does not give you a factor score for each participant - although you could use the loadings to make a calculation. Alternatively, you may consider using a program that is more suited to handling latent variable for regression (e.g., Mplus as an example, but there are others). 

Multicollinearity is not an issue if you use a single factor score (or as a latent variable with 3 indicators); a factor represents the shared covariance between the items. 

Hope this is helpful, Ahmad. Wishing you well in your research.

Thank you dear Tim for you time and effort on this matter. 

Best wishes!

My initial data was 1- 10 point Likert scale. So, I was working with ordinal logit. But now I am working on life satisfaction with just 3 categories Very satisfied, satisfied and unsatisfied. Is it possible to use Multonomial Categorical model? I just want to compare the three categories.