CFA with multiple groups and continous/categorical variables

[Jordane Boudesseul]

Dear all,

I have some difficulties trying to fit this model in Lavaan:

1) The thing is that I have some continous variables but also some ordinal and one binary (gender). Is it possible in Lavaan to fit different type of variables in the same model?

2) Also, does the function "group" would permit me to account for a clustering effect? (participants "nested" into location). I have many many location, can it poses a problem of statistical power?

3) Also I know there is no clear-cut answer to this, but since I have predictions for most of the relations above (full vs. dash lines), does a CFA model is realistic? I'm sure some of you may have insights on that issue.

I'm attaching my code so you can take a look: if I'm correct, there is 3 models since I have 3 IVs. Based on that, I tried to approximate a code (there is no moderators yet). Tell me if you see some huge mistakes so far.

Thanks in advance for all your help and forgive my English,

j

[Mikko Rönkkö]

Hi,

On 20 Jan 2017, at 1.10, Jordane Boudesseul <jmj.bou...@gmail.com> wrote:

Dear all,

I have some difficulties trying to fit this model in Lavaan:

1) The thing is that I have some continous variables but also some ordinal and one binary (gender). Is it possible in Lavaan to fit different type of variables in the same model?

Yes. See e.g. http://lavaan.ugent.be/tutorial/cat.html

2) Also, does the function "group" would permit me to account for a clustering effect? (participants "nested" into location). I have many many location, can it poses a problem of statistical power?

If the number of locations is small, then fitting a multi group analysis is a feasible strategy. But this depends on what exactly you think that the effects of clustering are on your analysis. Do you expect the intercepts of slopes to vary? 

Some people use survey features to calculate cluster robust standard errors

https://groups.google.com/forum/#!topic/lavaan/B4r80vSeJLc

3) Also I know there is no clear-cut answer to this, but since I have predictions for most of the relations above (full vs. dash lines), does a CFA model is realistic? I'm sure some of you may have insights on that issue.

I am not sure what you are trying to do. The path diagram that you provided contains no latent variables (circles), but it simply a path analytical model. In the code you specify DV1-DV3 as three factors of MED1-MED5. If DV1-DV3 are latent, then the model is not identified because the factors overlap perfectly.

Mikko

[Jordane Boudesseul]

Hi, thanks for your feedbacks Mikko ! I respond down below:

Hi,

On 20 Jan 2017, at 1.10, Jordane Boudesseul <jmj.bou...@gmail.com> wrote:

Dear all,

I have some difficulties trying to fit this model in Lavaan:

1) The thing is that I have some continous variables but also some ordinal and one binary (gender). Is it possible in Lavaan to fit different type of variables in the same model?

Yes. See e.g. http://lavaan.ugent.be/tutorial/cat.html

Thank you. I went through that but I was not sure I could combine several types, nice! 

2) Also, does the function "group" would permit me to account for a clustering effect? (participants "nested" into location). I have many many location, can it poses a problem of statistical power?

If the number of locations is small, then fitting a multi group analysis is a feasible strategy. But this depends on what exactly you think that the effects of clustering are on your analysis. Do you expect the intercepts of slopes to vary? 

Some people use survey features to calculate cluster robust standard errors

https://groups.google.com/forum/#!topic/lavaan/B4r80vSeJLc

Thanks for the robust SE link, really interesting. I have somehow 100s of locations so idk how much of this group function may work... 

3) Also I know there is no clear-cut answer to this, but since I have predictions for most of the relations above (full vs. dash lines), does a CFA model is realistic? I'm sure some of you may have insights on that issue.

I am not sure what you are trying to do. The path diagram that you provided contains no latent variables (circles), but it simply a path analytical model. In the code you specify DV1-DV3 as three factors of MED1-MED5. If DV1-DV3 are latent, then the model is not identified because the factors overlap perfectly.

Yes it's a path analytical model. But I have predictions for all the paths excepting the the covariates basically (and the clustering effect).

Thanks again !

j 

Mikko