Propensity Score Matching using MI data

[John Brent]

Hello Everyone –

As always, I hope all is well – please note that I am relatively new to STATA, so any/all advice is greatly appreciated.

Data Structure:
Panel data including 4 waves of survey data

Issue:
Looking to run propensity score (PS) matching/models to isolate “treatment” effects while matching and e-matching across included covariates. However, across waves, there is subject attrition and missing data I have recovered via MI methods (i.e., mvn, mice).

Goal:
Need advice with STATA syntax that enables me to perform PS using the mi data. I have read a bit of literature outlining methods to do so (example: running PS across imputed data sets and reporting the average(s)). While I have learned to run PS on non-mi data (using pscore / psmatch2 / kmatch with common support, kernel methods…), I am still trying to find the proper syntax/how tos to accomplish this with mi data.

I know this may be an elementary issue, so I appreciate any guidance. Thank you.

Best -
John

[John Brent]

Refreshing the above: Looking for initial syntax and guidance that will enable propensity score matching using mi data. - Thank you

[Rich Goldstein]

Your question is not entirely clear and the lit contains more than one way of doing this; below, I give some citations but to actually give code and guidance you need to decide on a strategy and tell us what that is; some cites (I use "mi" rather than spelling out multiple imputation in the cites):

Leyrat, C, et al. (2019), "Propensity score analysis with partially observed covariates: how should MI be used?", Statistical methods in medical research, 28(1): 3-19

Granger, E et al. (2018), "Avoiding pitfalls when combining MI and propensity scores," Statistics in Medicine, 38: 5120-5132

Cham, H and West, SG (2016), "Propensity score analysis with missing data", Psychological Methods, 21(3): 427-445

Ling, A, et al. (2020), "How to apply MI in propensity score matching with partially observed confounders", Journal of modern applied statistical methods, 19(1), eP3439. https://doi.org/10.22237/jmasm/1608552120

Segolas, C, et al. (2023), "Propensity score matching after MI when a confounder has missing data", Statistics in medicine, 42: 1082-1095

[John Brent]

Good morning, Rich:

First, thank you so much for your response, the literature, and your initial guidance. During my studies on this, I came across and read most of the articles provided above - great sources of info! and am aware of the methodological approaches that have been proposed to use mi data within PSM along with the

[John Brent]

Sorry - prematurely posted on accident:

Good Morning, Rich:

First, thank you so much for your response, the literature, and your initial guidance. During my studies on this, I came across and read most of the articles provided above - great sources of info! With that, I am aware of the methodological approaches that have been proposed to use mi data while preforming PSM along with their respective issues and pitfalls. The two approach that seems to be the most common and fitting is "within" and "between" approach - whereas the former calculated the treatment effect in each dataset and pools estimates while the “between” method averages the individual PS estimated in each complete dataset and performs inference using the pooled PS values. It would be good to learn how to perform these in STATA and see which best fits the needs and structure of the data/analysis I am doing. Does this help - not too sure - again, really new to the STATA scene!

Thank you again!

Best,
John.