Yuting,

regarding Question 1: I am not quite clear what you mean by "two items are collinear". In a common factor model, the items are supposed to be highly correlated (due to the common factor).

Dear Yuting,

1 – items belonging to the same dimension are expected to be highly correlated. So, try to perform the analysis maintaining all (collinearity problems use to pose in multiple regression analysis and no in CFA).

2 – considering this problem, try to perform a robust maximum likelihood estimation method (Satorra & Bentler) _ see my doc in attachment. If you are using AMOS, this program does not provide Satorra-Bentler Corrected Chi-square… This is the information in IBM webpage: “Amos provides bootstrapping for situations where the assumptions underlying maximum likelihood chi-squares and standard errors may be violated. In fact, for large samples, both Amos and the Satorra-Bentler scaled Chi- square have the same p-value. Some simulation results show that bootstrapping is even better than the Satorra-Bentler scaled Chi-square”.

Here are references for these simulation studies:

Fouladi, R.T. (1998) Covariance structure analysis techniques under conditions of multivariate normality and nonnormality--modified and bootstrap based test statistics. Paper presented at the annual meeting of

3 – just delete the more  extreme values and check if the fit improves

4 – the transformation is done in items. For example, Considering the high skewness coefficient (sk) and kurtosis coefficient (ku) for an item A3, it is necessary to perform a log10 transformation. However, if your sk and ku are ok, proceed with the analysis using a robust maximum likelihood estimation method.

Good luck with you research!

@ Heiko Breitsohl 

I found that  Tabachnick (2007) in his book  "Using Multivariate statistics" mentioned multicollinearity is one assumption of CFA, e.g. on page 734. This is because multicollinearity will make the covariance matrix in-determinant. 

1. In CFA variables are supposed to be correlated, so it's OK here

2. There a lot of normality tests, for example Shapiro-Wilk test, or Kolmogorov-Smironov test. They are based on different assumptions, though.

3. Apart from trimming (mere deleting outliers), It's also possible to perform winsorizing, which is setting extreme values to closest chosen quantiles

https://en.wikipedia.org/wiki/Winsorizing

https://en.wikipedia.org/wiki/Trimmed_estimator

You can employ robust maximum likelihood using LISREL or Mplus, if you fail to achieve normality .

Forget the skw and ku. In Amos you can find specific method to handle with non multivariate normality. In Amos you have to use ADF (Asymptotically Distribution-Free) method. Unfortunately ,AMOS can't calculate SB (Satorra & Bentler),but you can use Bollen-Stine Bootstrapping Method.