CFA with binary variables

[janina steinert]

I am trying to do confirmatory factor analysis on data that is coded binary (0 no, 1 yes). UCLA suggests using tertrachoric correlation matrix, which,=
however, assumes that binary variables reflect underlying continuous variables. Is there a way to relax this assumption and use logistic procedures =
instead? I am using Stata Version 13. Could generalized SEM provide a solution?
Thank you!!
--------------

[wbuchanan]

If you look at examples 28 and 29g from the SEM manual you will see examples of 1PL and 2PL IRT models, respectively.

[janina steinert]

Thank you. This has helped me to specify my measurement model as part of the generalized SEM. My binary indicators specify the latent variable POVERTY. I wish to correlate POVERTY further with aids_illness, while controlling for household size, education status, rural/urban residence etc. Yet, when trying to correlate the latent variable with the exogeneous observed variable, Stata gives the following failure notice: "aidsillness does not identify a gaussian error or latent variable".

Also, the goodness of fit statistics that are used with SEM (such as RMSEA, CFI...) as well as the modification indices (mindices command) are not allowed for GSEM.

What could be an alternative way of fitting my model? Thank you!

[Nick Cox]

Cross-posted at http://stats.stackexchange.com/quest...using-stata-13

The Advice Guide spells out our request that you tell us about cross-postings to other forums. See Section 8 in http://www.statalist.org/forums/help#gfaq_postingadvice

[Joseph Coveney]

Is there a problem with using POVERTY as an unobserved endogenous variable predicted by aids_illness and all the others in a MIMIC-like model, and then inspecting the regression coefficient for aids_illness and its standard error in order to judge the association (correlation)?

Maybe something like

gsem (binary_indicator1 binary_indicator2 binary_indicator3 <- POVERTY, probit) ///
(POVERTY <- i.aids_illness i.household_size i.education_status i.urban_residence c.etc)

if I've got that right.

[janina steinert]

Thank you! The problem is, that I didn't want to model a directional effect given that the direction between poverty and aids_illness could be both ways and I only have cross-sectional data, i.e. a correlation is more plausible. Same with the covariates, all of them are likely to be associated with both poverty and aids_illness)...