I am analyzing data of a mathematics test consisting of 200 categorical items. The 200 items were distributed across 10 booklets. About 200 pupils worked on each booklet. The booklets were randomly assigned to the pupils (thus data for several items are missing by design). The booklets are linked pairwise by 5 items, e.g. 5 items appear in booklet A and B, 5 different items appear in booklet B and C, etc.
As I learned from the Mplus users´ guide and previous discussions on this list neither the pattern option of Mplus2 nor the "classical" multigroup approach to missing data (Muthen, Kaplan, & Hollis, 1987) will work.
I also learned that Mplus3 may deal with missing by design and categorical data. But unfortunately I cannot wait until the release of Mplus3.
Is there any possibility to establish measurement invariance across the booklets?
I am thinking of constraining thresholds and factor loadings of the linking items (e.g. for booklet B) to the values for those linking items that I optained when analyzing the item set of booklet A.
Any help is highly appreciated.
Muthen, B., Kaplan, D., & Hollis, M. (1987). On structural equation modeling with data that are not missing completely at random. Psychometrika, 52(3), 431-462.
To deal with the designed missingness, you could treat this as a multiple group analysis where each test booklet is a group. If you have missing data within test booklet, you would have listwise deletion for each test booklet. You would then place appropriate equalities on the anchor items.
Hi, is the modeling feature for LGM with missing data a pairwise deletion method? I am writing a methods section and am not sure how I should describe the missing data method for categorical dependent variables.
WLS missing uses a three stage (just like WLS without missing) estimation that is pairwise deletion based. The method guarantees more than MCAR consistency but less than the full MAR consistency. The exact condition is called MAR-covariates, the estimates are consistent even if covariates influence the missing data patterns. For the full MAR consistency use the ML estimator.
I am analyzing a model with latent constructs composed of some categorical items (which are indicator variables). Therefore, I am using TYPE=COMPLEX and the estimator is WLSMV. Is Mplus 6 conducting FIML to account for missing data when I run this type of model? If not, how is the model dealing with missing data?