Hi! Is there a good example (reference, chapter, article) of how to use propensity score matching (PSM) or propensity score estimation (PSE) with structural equation modeling in MPLUS? Thank you so much, Clara
Here is an answer from a colleague in the causal inference field:
The short answer is that no, I donít know of a good reference for propensity scores with SEM.
The longer answer is that I think in part I donít know of a good reference for this because the goals of the two methods are somewhat different, at least from what I understand about SEM. From my understanding SEM is designed for looking at relationships between a large number of variables (and not necessarily always in a causal way). Propensity scores are designed when you really just have one variable of interest (the ďtreatmentĒ) and you want to look at the treatmentís effects on some outcome(s), basically treating all of the other variables as nuisance confounders. So the methods donít necessarily generally fit together very well. As one specific example, there has been very limited work using propensity scores in the context of post-treatment variables and mediation, but that is still basically a very simple type of SEM (if one even wants to think about it in that way at all).
I should add that there is at least one example of research that uses propensity scores as an explanatory variable in an SEM setting. You can email me if you want me to give you more information on that.
Can you send it to me? Your email is not available. Thanks so much Clara
Emil Coman posted on Wednesday, August 08, 2012 - 7:06 am
Don't want to boast, but I will: I presented recently this: Coman E., Yanovitzky I., Coman M., Weeks M. R. (2012). Understanding propensity score matching through a more flexible causal modeling alternative. Mixture and multilevel causal modeling of true effects. 2012 Modern Modeling Methods (M3) Conference. http://www.modeling.uconn.edu/ They have not posted my PPT yet, so I can email it.
Hello, I'm attempting to use Mplus to estimate propensity scores using MLR to handle missingness on predictor variables. The model is a simple logit with a binary outcome. When using listwise deletion, I'm able to save the propensity scores.
When including the list of predictor variables in the model statement to address missing data, I'm also required to use montecarlo integration. Doing this leads to the following error:
The SAVE=PROPENSITY setting is only available for analysis with regressions. Request for SAVE=PROPENSITY will be ignored.
I'm unsure why the model is not considered a regression. Is it possible to create propensity scores in Mplus using MLR to address missing data on covariates?
Your findings are correct. The propensity score is computed using true covariates only (and no missing values). Because you have missing values you are essentially converting the predictors to endogenous variables and the program is telling you that there are no real covariates left.
Here is what you can do. Take a look at User's Guide example 11.5 to see how to impute the missing values. Make sure you specify categorical variables as such. If you have more than 25% missing data I would use 100 imputations. If less, you can probably use 10 or 20. For each imputed data set compute the propensity scores. Average the propensity scores over all the imputations.
Hi Tihomir, Thank you for the guidance. Will averaging the propensity scores require merging the imputed data sets together then calculating manually? I tried to create the pooled pscore file using imputed data and received the following warnings:
*** WARNING in SAVEDATA command The FILE option is not available for TYPE=MONTECARLO or TYPE=IMPUTATION. The FILE option will be ignored. *** WARNING in SAVEDATA command The SAVE option is not available for TYPE=MONTECARLO or TYPE=IMPUTATION. The SAVE option will be ignored.
Correct. You will have to run each imputed data set separately/manually and then extract the propensity score column from the savedata file using excel and then average all those columns also in excel. You might find the MplusAutomatio utility useful https://www.statmodel.com/usingmplusviar.shtml or you might setup a batch mode https://en.wikipedia.org/wiki/Batch_file to automate the process but if you have 20 imputations I would just do it manually.