Multivariate normality with missing data

[C]

Hi all,

I'm wondering about the assessment of multivariate normality when there is missing data across multiple waves in longitudinal datasets. Most R packages I have come across appear to delete listwise (e.g., MVN) when providing Mardia skew and kurtosis statistics with missing data. Similar with outlier detection. However, this means that I'm not assessing this for the complete sample. Not sure what is best practice in this situation? Assessing marginal distributions at the univariate level doesn't seem sufficient for testing this assumption?

Many thanks.

[C]

My apologies, just realised this isn't strictly a lavaan question. (But appreciate any advice other lavaan users have on this!)

[Edward Rigdon]

     You need to refine the question that you are trying to answer...

     "Are the missing observations normal?" No, they are missing.

     "Are all complete cases normal?" You already have that answer.

     You could apply your tests to subsets of the variables, which would give you statistics on different subsets of cases. At some point, you will have to concede the unlikelihood of getting a different answer.

     You could use multiple imputation and apply your tests to that. Multiple imputation performs pretty well, and if you are going to use the imputed data for analysis, then "Is the imputation-result data normal?" seems to be the relevant question, assuming that normality is relevant to your analysis.

--
You received this message because you are subscribed to the Google Groups "lavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lavaan+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lavaan/b291d042-fce0-47b1-8111-96b39eedba25n%40googlegroups.com.

[Yago Luksevicius de Moraes]

In addiction to Edward's reply, it's true that lavaan's default option is listwise deletion, but you can change it to pairwise, by setting missing="pairwise" and, if you are using one of the ML estimators, you can also use full-information maximum likelihood (missing="fiml").