Missing values and CFA
Dear JASP,
I intend to conduct a CFA and I have missing data. I have three questions concerning JASP and one concerning statistics (hopefully, you can provide answers to all three!):
- Is there a way in JASP to test whether the data are missing at random?
- How does CFA in JASP deal with missing data?
- Can I choose how CFA deals with missing data, e.g., by using the SEM lavaan option (I've seen this option in another answer in this forum, but then for another topic)?
- Would it make sense to conduct the CFA with all the cases and then with the cases that have no missing values as a sort of sensitivity test?
Many thanks!
Georgios.
Comments
I've forwarded this to our experts
E.J.
Dear gvleioras,
Is there a way in JASP to test whether the data are missing at random?
Would it make sense to conduct the CFA with all the cases and then with the cases that have no missing values as a sort of sensitivity test?
It depends on the exact question.
It is possible to get an indication whether data is missing completely at random (MCAR) or missing at random (MAR) by creating dummy variables for whether a variable is missing: 1 = missing 0 = observed. Run t-tests (continuous) and chi-square (categorical) tests between this dummy and other variables to see if the missingness is related to the values of other variables. Tests which return a finding of significance indicate MAR rather than MCAR.
However, it is not that easy to distinguish MAR (missingness depends on observed variables) from MNAR (missingness depends on unobserved variables) and to my knowledge, the following is not yet possible in JASP (although we will probably eploit the MICE package with such functionality in the near future). A method to perform a sensitivity analysis by imputing data under various delta adjustments is clearly described here: https://stefvanbuuren.name/fimd/sec-sensitivity.html.
How does CFA in JASP deal with missing data?
Can I choose how CFA deals with missing data, e.g., by using the SEM lavaan option (I've seen this option in another answer in this forum, but then for another topic)?
When all observed variables are continuous, full information maximum likelihood (FIML) is available for handling missing data, for both CFA and SEM in JASP 0.17 (https://jasp-stats.org/download/). For CFA, this option is located in the 'Advanced' tab. This method (FIML) would be valid under the the MAR assumption.
I hope this will help you out! If something is unclear or if you have another question, feel free to ask!
LLindeloo
It looks like JASP v0.17 uses FIML by default for both CFA and SEM, and you can change this option. If you prefer to use the SEM analysis in JASP instead of the CFA analysis you can specify the syntax for the measurement model manually (e.g. f1 =~ item1 + item1 + item3). lavaan (the library JASP uses in the background for both CFA and SEM) is super good at "guessing" it's a CFA from the SEM syntax and estimating the right model (e.g. the intercepts will be fixed to 0, the factors will be allowed to covary, etc.).
@LLindeloo,
Thank you for your thorough explanation. When I create this dummy variable, what variables do I use for the chi-square or t-tests? All other variables of my dataset? Or the demographics? Or something else?
Also, does it make sense to exclude from the beginning cases with many missing answers (e.g., more than 10% of the items), or should I include all cases?
@patc3
Thanks for the tip! I would like to also start playing a bit with lavaan, and maybe this is a good place to start...