Proc CALIS Method= DWLS / LSDWLS ===> Using Polychoric Corr Matrix >?<

[-j-]

Hello All,

I'm running multiple CFA models stemming from output solutions of EFA models for a thesis project.  Being ordinal scaled data, polychoric matrices were used in the EFA processes.  I've seen PROC CALIS METHOD=ML process the input Polychoric matrix fine, but with ULS methods, I receive the error that _TYPE_=WGT is not found, and to use the raw data set.  

Using the raw, Likert-scaled data set, processes fine with ULS/LSDWLS, but will these solutions take into account the ordinal properties of the data?  Surely it would make a big difference.

Case in point, while ML is not recommended for ordinal data, I used it to generate RMSEA estimates for both the raw data in one data step and the polychoric matrix in another. Raw Data RMSEA = 0.0550, Polychoric Data= 0.1144.  

Any recommendation on ensuring the ULS/DWLS methods utilize the ordinal-scaled data properly would be appreciated.

[KDEdmiston]

There is considerable debate about how many items on the Likert scale are sufficient to treat the corresponding variable as continuous. With a sufficient number of items (whatever that number is), you might be ok with normal theory maximum likelihood. Otherwise, I would recommend using the option INWGT = <your covariance matrix>. If you don't specify the weight matrix, SAS will use the inverse diagonal matrix of variances from the input sample as the weight matrix. My guess is that this would be a Pearson matrix, so I would think you'd want to use INWGT = with your polychoric matrix. Since it wasn't mentioned, I'll note that you'll, of course, want to input a covariance matrix, not a correlation matrix. I think you can use a correlation matrix in PROC CALIS as long as the standard deviations are included so that a covariance matrix can be constructed. Or you can convert from corr to COV yourself. If M is your corr matrix, S is the vector of standard deviations, and D = diag(S), then COV = DxMxD. I had polyc in the NW corner, polys in the NE corner, polys(T) in the SW corner, and Pearson in the SE corner (had binary and continuous data), so inputting a corr matrix was not a good option. Also, I think FIML could be an option for you (Handbook of Structural Equation Modeling, ch. 12). The problem with FIML is the multiple integrations required. You should have some serious hardware if you want to go that route. Good luck.