SEM with ordinal data

[katja.si...@gmail.com]

Hi,

I am PhD student and I am new to SEM. I am working on the estimation of the model with ordinal indicator variables. The indicators come from a survey responses measured with 7-point Likert scales.

I have the following questions:

1) Is Chi-square a feasible model fit statistic when the DWLS/WLSMV estimator is used?

2) When I tried using MML estimator  I got an error:

Error: cannot allocate vector of size 93.9 GB

3) If  I am using 7-point scale for my variables, can I use WLS but treat my data as continuous instead of declaring it ordinal?

4) What other estimation approaches will you recommend for the ordinal data?

5) Can I retrieve any construct validity measures for the latent variables while running SEM?

6) I would want to make results reproducible. In my case I cannot use covariance matrix for ordered data. Is it sufficient to report polychronic correlation matrix only? 

7) How should I treat significant co-variances between indicators? significant/not significant thresholds?

Model:

model_v9<-'

lq8=~q8

ITPerformance=~q13a +q13b

ITMetrics =~ q20a+q20b+q22a+q20d+q22b

BMetrics=~q20f+q20g+q20h+q20k+q20e+q20f+q20j

Trust =~ q24a+q24b + q24c +q24d

Shunder=~q25a + q25b+ q25c +q25d

ITSC=~q14b+q14c+q14d+q14e+q14f+q14g+q14h+q14i+q14j+q14k

ITPerformance~ITMetrics+FTECat+q9a+lq8+ProfitNo+q11

Trust ~ITPerformance+FTECat+q9a+lq8+ProfitNo+q11

#altern model1 (cannot be estimated with WLSMV and DWLS)

#ITPerformance~ITMetrics+FTECat+q9a+lq8+ProfitNo+q11

#Trust ~ITPerformance+ITMetrics+FTECat+q9a+lq8+ProfitNo+q11

Shunder~Trust +BMetrics+FTECat+q9a+lq8+ProfitNo+q11

ITSC~Shunder+FTECat+q9a+lq8+ProfitNo+q11

#altern model2

#ITPerformance~~ITMetrics

#Trust ~ITPerformance+ITMetrics+FTECat+q9a+lq8+ProfitNo+q11

BMetrics~~ITMetrics

q20a~~q20b

q20a~~q22a

q20b~~q22b

q20g~~q20f

q20g~~q20h

q20h~~q20f

q20e~~q20f

q20j~~q20k

q20j~~q20f

q14f~~q14j

q14f~~q14i

q14f~~q14d

q14d~~q14j

q14k~~q14j

q14d~~q14b

q25a~~q25b

q25b~~q25c

q24a~~q24b

q24c~~q24b

'

fit <- sem(model_v9, data=data,estimator="WLSMV", 

           ordered=c("ProdServ","FTECat","ProfitNo","q13a","q13b",

                     "q20a","q20b","q22a",

                     "q20f","q20g","q20k",

                     "q24a","q24b","q24c","q24d",

                     "q14d","q14f","q14i","q14j","q14k",

                     "q25a","q25b","q25c","q25d",

                     "q9a","q13c") )

Thank you very much,

Katja

[Jeremy Miles]

On Tue, 21 Aug 2018 at 17:54, <katja.si...@gmail.com> wrote:

Hi,

I am PhD student and I am new to SEM. I am working on the estimation of the model with ordinal indicator variables. The indicators come from a survey responses measured with 7-point Likert scales.

I have the following questions:

1) Is Chi-square a feasible model fit statistic when the DWLS/WLSMV estimator is used?

Yes. 

 

2) When I tried using MML estimator  I got an error:

Error: cannot allocate vector of size 93.9 GB

MML uses a lot of memory, and is very slow.  You could buy that much RAM for your computer, or you could rent a cloud instance with enough ram, but it will take a longtime. 

 

3) If  I am using 7-point scale for my variables, can I use WLS but treat my data as continuous instead of declaring it ordinal?

You can. Whether you should depends on how normal the data are - if they're not horribly non-normal, I'd treat them as continuous. 

 

4) What other estimation approaches will you recommend for the ordinal data?

MLR :)

 

5) Can I retrieve any construct validity measures for the latent variables while running SEM?

 

6) I would want to make results reproducible. In my case I cannot use covariance matrix for ordered data. Is it sufficient to report polychronic correlation matrix only? 

Yes.  

 

7) How should I treat significant co-variances between indicators? significant/not significant thresholds?

Not sure what you mean here. 

Jeremy