How to conduct FIML for growth curve/mixed model in Stata?

[Emma Zhu]

Hello everyone,

I am using the Stata "mixed" command to run growth curve analysis within the multilevel framework for my dissertation project. I want to conduct full information maximum likelihood (FIML) to deal with missing data and make sure that my analysis has consistent sample size across models that use different independent variables. Does anyone know how to do that? I couldn't find any resource online about how to apply FIML to a growth model with "mixed" command in Stata.

Below is my growth curve model code (without FIML):
mixed cognition i.media##c.wave ///
i.edu i.gender i.race c.basemem ///
c.age i.mar i.live i.work c.inc ///
|| hhidpn: wave, var cov (unstr)

Any help will be appreicated! Thank you very much!

[Erik Ruzek]

If by FIML you are referring to maximum likelihood that can accommodate missing x and y data, then you need to move over to the structural equation modeling (sem in Stata) paradigm. See this extremely useful UCLA resource. To get FIML, you will need to add the option method(mlmv), which is Stata's shortcut for FIML. It stands for maximum likelihood with missing values.

[Emma Zhu]

Thank you a lot, Erik! I am now learning sem and hopefully can transfer my work to sem framework efficiently. Just to clarify, do you mean that multilevel modeling is not compatible with FIML? What are some other typical ways people do to deal with missing values if they use the mixed command for growth curve models? Thank you!

[Erik Ruzek]

The key insight that you need to wrap your head around is that a mixed model can be exactly expressed as a structural equation model. This took me years to understand, so I don't expect you will get it right away. But for folks who don't want to spend their time working in sem, but have missing data, they would employ multiple imputation. Stata has its mi set of commands, but those are not ideal for multilevel data. Instead, you would want to use a purpose built implementation of multiple imputation for panel or multilevel data. Check out BLIMP for this. You can impute in BLIMP and then import the imputed data into Stata for analysis. Enders and colleagues have documentation for doing this at the BLIMP website.

[Richard Williams]

You might check out Discovering Structural Equation Modeling Using Stata, Revised Edition

https://www.stata.com/bookstore/disc...g-using-stata/

“Chapter 4 details the application of SEM to growth curve modeling. After introducing the basic linear latent growth curve model, Acock extends this to more complex cases such as the inclusion of quadratic terms, time-varying covariates, and time-invariant covariates.“

One of my longtime wish items is adding a FIML option to many commands, e.g. even regress.

[Roman Mostazir]

Besides all the good suggestions above, bear in mind that -sem- will not allow factor variable notations as you have i.edu, i.gender and i.race in your model. If you have all of them in two levels like 'gender' (I am assuming as it is in most cases) then ignoring the factor variable notation will do the job. -gsem- will accommodate the factor variable notation but you don't have the 'mlmv' option for that so it won't help in your case.

[Emma Zhu]

These are very helpful comments. Thank you all!