Missing data, measurement inv., ordinal

[Sara Esposito]

I am conducting a test to determine the invariance of a simple model with one latent variable and six observed items (measured on a 5-point Likert scale) across four groups. The nature of the Likert scale suggests that the Weighted Least Squares Means (WLSM) estimator would be appropriate for this analysis. However, according to Chen et al. (2020), WLSM is not recommended for ordinal data with missing values, and they suggest using the full information likelihood method with probit or logit links instead.

Chen, P. Y., Wu, W., Garnier-Villarreal, M., Kite, B. A., & Jia, F. (2020). Testing measurement invariance with ordinal missing data: A comparison of estimators and missing data techniques. Multivariate behavioral research, 55(1), 87-101.  https://doi.org/10.1080/00273171.2019.1608799

1) I was unable to locate the probit or logit options for the estimator and missing options in the ?lavOptions. Could you please provide the syntax for using either the probit or logit options?

2) If these are not available in Lavaan, could you please suggest me what other available options I could use for my data (5-point Likert scale + missing values).

Here is how missing data were handled using the default estimator (ML):
mod.cat<- '
clom=~ focus + sleep + attention + fun + space + mood
'
syntax.config <- measEq.syntax(configural.model = mod.cat,

                 data = df2,

                 ID.fac = "std.lv", 

                 ID.cat = "Wu.Estabrook.2016",
                 group = "country")
fit.config <- cfa(as.character(syntax.config), 

              data = df2, group = "country")
summary(fit.config)


Number of observations per group:               Used       Total
Country 1                                                             1043        1086
Country 2                                                              1059        1086
Country 3                                                              1044        1086
Country 4                                                               1051       1087

Thank you.

[Yves Rosseel]

On 12/16/22 14:36, Sara Esposito wrote:
> suggest using the full information likelihood method with /probit/ or
> /logit/ links instead./

Unfortunately, that is not available in lavaan.

> 2) If these are not available in Lavaan, could you please suggest me
> what other available options I could use for my data (5-point Likert
> scale + missing values).

For categorical data, the only alternative is missing = "pairwise",
which is far from perfect, but is better than nothing if you loose a lot
of data otherwise.

> Here is how missing data were handled using the default estimator (ML):

It would seem that the 'loss' of observations is minimal, so just using
listwise deletion may be ok (unless those cases with missing values are
somehow special).

If you 'pretend' the data is (semi-)continuous, you can use estimator =
"MLR" in combination with missing = "ml".

Yves.