@Habtu Birhan Co-varariance based SEM (in AMOS) requires the normality assumption and works well with large dataset (preferably > 400) while variance based SEM (in Smart-PLS) doesn't require the normality assumption and works well even with low data sample. If your data is not normally distributed and highly skewed (beyond the prescribed range of -1 and +1) you are advised to use Smart-PLS.

The multivariate normality assumption is made when using maximum likelihood (ML) estimation in SEM or CFA. When this assumption is violated, parameter standard errors and test statistics (e.g., the chi-square test of model fit) may be incorrect, leading to incorrect statistical inference. The parameter estimates themselves are usually not as strongly affected by non-normality under ML estimation.

In cases of non-normality, you can use robust ML estimation methods that provide the same parameter estimates but use corrected standard errors and test statistics such as for example the Satorra-Bentler correction or the Bollen-Stine bootstrap. Another option is to not use ML estimation at all and instead use an estimation method that does not require normality such as weighted least squares (WLS). However, WLS estimation requires very large samples to work properly. So the best option is usually to employ robust ML estimation methods as mentioned above.

Thanks @Dr.Imran Anwar and Dr.Christian Geiser , I learn a lot.