Is it possible to conduct multiple imputation using H1 and a grouping variable? I am trying to follow the guidance of Enders 2011 during my imputation (see below). Also, are there any plans to include sample weighting as an option during imputation?
Thank you for your response.
Enders, C. K., & Gottschall, A. C. (2011). Multiple Imputation Strategies for Multiple Group Structural Equation Models. Structural Equation Modeling: A Multidisciplinary Journal, 18(1), 35–54. doi:10.1080/10705511.2011.532695
Thank you for your response Linda. Follow-up questions.
1) Will the imputation phase have the same behavior if I use the multiple-group syntax (grouping = female (0= male, 1=female) versus the type=mixture with a knownclass statement?
2) If I use either the multiple-group syntax or the knownclass syntax, do I need to specify group-specific model commands to free parameters across groups for the imputation phase? I would like to have a fully saturated model so all measurement parameters are free across groups during imputation.
Hi Linda, I have a follow-up question. Just as a reminder, I hope to achieve a fully saturated multiple-group model during the imputation phase (H1).
According to #12 in "General tips and observations" of http://www.statmodel.com/download/Imputations7.pdf, anything that goes into the MODEL statements is ignored during the imputation phase unless you are using H0 imputation. That document appears to give different guidance than your answer to my question #2 above. Would it help if I pasted my syntax with my question?
When you use a MODEL command, you are doing H0 imputation even if the H0 model you specify is a fully-saturated model. When you do not use the MODEL command and use TYPE=BASIC, you are doing H1 imputation.
This behavior seems to be documented differently at http://www.statmodel.com/download/Imputations7.pdf in tip #12. It says, "If the estimator is set to Bayes then you are performing an H0 imputation. If the estimator is not set to Bayes then you are performing an H1 imputation."
I am getting the sense that you are saying the inclusion of a MODEL statement will switch you to an H0 imputation. But in the documentation, it says the imputation phase will remain unchanged when a model command is included; rather, a model will be estimated based on the imputed data. Can you help clarify? Thank you!
Slide 184 in our Topic 9 short course handout on the website lays out the different imputation choices. It depends on the Estimator and the presence of a Model command. See also the other slides on this topic.
I have a 2-group SEM I'm working on, and wish to test invariance assumptions for the measurement model related to factor structure, item loadings, thresholds, and residual variances for the 2 groups using nested models. My data for factor indicators is ordinal (u) and skewed left. I also have missing data.
To examine configural invariance I imputed 20 data sets for each of the 2 groups separately using the following:
ANALYSIS: TYPE IS basic; ESTIMATOR IS WLSMV; PARAMETERIZATION IS THETA;
For each group of the 20 imputed data sets I conducted a CFA to assess the fit of the measurement model in each group, with good fits resulting for each group (.e.g., TLI = .99).
1. Is the multiple imputation ANALYSIS command appropriate for my data and goals?
2. How do I combine the two groups of 20 imputed data sets so as to be able to conduct further analyses involving both groups?