Hello everyone,
I wanted to ask if anyone is familiar with “how to appropriately do propensity score matching” while working with a huge database like national inpatient sample and keeping in mind its survey design, weight, clusters etc?
I primarily work on Stata. Any guidance would be much appreciated.
Thank you.
What about the problem makes you want to do indirect covariate adjustment instead of direct covariate adjustment? If your effective sample size (number of patients in Y is continuous, minimum frequency of outcome category if Y is binary) is > for example 10 times the number of parameters in a regression model, you can just do regression modeling.
Thank you, Dr. Harrell, for your comment. I am not specifically working on any such project but just trying to understand the following paper in JAMA that used this methodology but did not really explain how they did it. I wonder if you could provide me with some insights after reviewing their methods and results?
https://www.ahajournals.org/doi/epub/10.1161/JAHA.117.007920
The paper provided no motivation for using propensity scores in that setting, and the use of PS matching resulted in exclusion of most of the patients in the database! In my view using a method that excludes most of your data is not fully scientific.
Agree. Thank you so much for guiding.