missing imputation, FIML, Pscore matching and AIPW

[Chiara Tasselli]

I recently submitted a research paper to a journal based on a subsample due to missing data (in my case, MAR). A reviewer suggested that I should impute all missing data using the FIML module in Stata in order to include the whole sample in the analysis. I'm currently exploring this option, but it seems to work primarily with SEM , which presents a few challenges for my analysis:

1. My regression model includes interactions, which I can address by manually creating interaction variables beforehand.
2. My main concern is that my analyses involve not only simple regressions but also estimations using "pscore" (propensity score) and "aipw" (augmented inverse probability weighting).

Below, I’ve included the relevant commands for reference:

teffects psmatch ( S_belonging) (Trattato  $covPSMATCH ) if MYsampleFINAL == 1, atet caliper(0.09)
teffects aipw (S_belonging $covPSMATCH) (Trattato Female LicSC LicClas i.VotoMicroClass MathSkill, probit) if MYsampleFINAL == 1

Another idea I had was to use a more basic approach to estimate the missing values outside of the model, for example, using simple Multiple Imputation (MI) and store them.

I would appreciate any advice or insights on how to best approach this, especially regarding the use of FIML with the specific estimations I’m working with.

Thank you in advance!

[Erik Ruzek]

There is no reason why you must use FIML to recover missing data. It is but one of many options. It just happens to be really easy to use if you can formulate your model within a SEM framework (note that FIML is only available in Stata's sem, not gsem). Perhaps this is why the reviewer mentioned it.

That said, I do not see a plausible way for you to run this analysis in sem. You have a first stage in which you estimate and predict a propensity score using a probit model. You cannot estimate such a model in sem as it only works with continuous outcomes. You would instead need to use gsem for this first stage (or probit), and as mentioned, gsem does not utilize FIML (nor does probit).

Your best bet is to use Stata's mi suite of commands. It is possible to use them with teffects, according to a response in this thread by Yulia Marchenko (StataCorp). It is a little tricky, but as long as you make sure to match the appropriate mi xeq command prefix to teffects to the style of your mi dataset. You can also bypass teffects and do the estimation by hand in the mi framework, but that is more personal preference than anything else.

[Chiara Tasselli]

Dear Erik Ruzek ,
many thanks for your suggestions which have largely confirmed my uncertainties. I also found the reference you provided very interesting. However, considering the various issues related to the imputation methodology with teffects, I believe I will opt for a simpler (MI) imputation method.

Thank you again for your time and knoledge sharing!