SEM with ordinal data What are the prerequisites for SEM with ordinal data? I have read that there should be no missing data, no multicolinearity and the data should follow a normal distribution (does this also apply to ordinal data). I have also read that tests for assumptions of normality are not relevant to ordinal data...I have found 2 variables that violate skew and kurtosis norms and do not know what to do with them. Can I use them without any transformations? 
Steps I have taken so far to prepare for SEM are:
 - grouping similar variables
 - multiple imputation
 - QQ plots 
 - checking for correlations 
 - checking skew and kurtosis 
I am beginning to doubt everything, as I cannot find concrete information on how to prepare grouped ordinal data for SEM. 
Additional information: My independant variables are school results and students motivation, all my dependant variables are related to parental involvement (homework help, expectations, contact with school, cultural capital etc.) I am trying to determing how parental involvement influences motivation and results AND to create a single latent variable using CFA that represents parental invovlement. 
 A: SEM can be applied to ordinal data. In fact, it is a "cheap trick" that allows one to ignore issues of non-normality (or skewness and kurtosis).
Let me briefly explain, and I will follow-up if requested.
First, let’s not think of this as SEM, but as categorical confirmatory factor analysis (CCFA). The assumption that variables are normally distributed still applies, but now all of the variables are latent (unobserved) variables.  Let's use this context:  X is measured by (influences) y1, y2 & y3 (lower case is important here). But, X does not influence y's directly...X influences an intermediary set of latent variables, Y1, Y2 & Y3. The lower case (observed) values are obtained from the upper case (latent) values.  Of course, we can't "see" X, so no harm in assuming it is normally distributed. But, we also need to assume that the Ys are normally distributed. But, if the observed y's (lower case indicating ordinal variables) are not "normal", this does not imply the preliminary Y (upper case indicating predicted latent variables) are not normally distributed. Again, if we can't actually "see" the variable, there is no way to assess if it is normal or not. And this is where the connection between Y and y makes all our "normality" problems go away. CCFA estimates the thresholds at which the underlying normal distribution would be broken into order categories. Because the placing of these thresholds (tau's) can be estimated anywhere, it is possible to generate very non-normal distributions amongst the observed variables.
Again, happy to share more if this seems useful.
A: Skewness and Kurtosis are not typically calculated for ordinal variables. You may want to talk about symmetry or asymmetry of your ordinal responses. Are all the variables in your data set ordinal? How many response categories do they have? As pointed out by Jeremy Miles, what do you mean by 'grouped ordinal data'? \
Also, it seems like you have mixed up your dependent and independent variables. As you want to determine 'how parental involvement influences motivation and results', your independent variable are parental involvement characteristics and the dependent variables are motivation and results.  \
I would refrain from drawing causal inferences based on the analysis if the design and methods don't support it. You can examine the association between parental involvement and motivation but not the influence of parental involvement on motivation. 
I can answer your question better if you provide more details about your design, variables in the study, and methods tried so far.  
