high dimensional propensity scores

#1
Hi all,
I have used high dimensional propensity scores in a research paper. One of the reviewers asks: if the computation of this score involves estimates that may be biased in any way, discuss this putative bias.

What do you think?
Thank you,
Best,
PM
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
The phrase high dimensional PS, does not mean anything to me. If you had a bunch of covariates I'd wonder about the positivity assumption. I'd also want to see the overlap of the scores. Love plots are also nice but not totally necessary.

Please describe what you did by copying over the exact text you used in the paper.

Thanks!
 
#3
Hi !

I actually worked on a very large healthcare dataset and I investigated the association between hypertension and cardiovascular mortality. I calculated a high-dimensional propensity score (hd-PS) for every subject in my dataset. This hd-PS represents the predicted probability of having hypertension versus no hypertension, conditional on all covariates in the dataset. Then I grouped my patients into hd-PS based deciles and used logistic regression to estimate the risk of cardiovascular mortality.

This is the reference of the main paper on hd-PS.

High-dimensional propensity score adjustment in studies of
treatment effects using health care claims data
Sebastian Schneeweiss, Jeremy A. Rassen, Robert J. Glynn, Jerry Avorn, Helen Mogun,
and M. Alan Brookhart

One of the reviewers asks: if the computation of this score involves estimates that may be biased in any way, discuss this putative bias. I don't really know what is expected here... any idea?

Thank you,
Best,
PM
 

hlsmith

Less is more. Stay pure. Stay poor.
#4
So you conducted PS stratification. With PS you can do stratification, adjust for scores in model, use scores as weights, or match based on weights. I typically don't use stratification, but it likely has the same issues.

You need to ensure balance between treatment scores, you can also look at covariate balance via love plots, absolute standardized mean differences, and variance ratios. You also want to make sure the positivity assumption is met.

However given you stated that you all conditioned on all covariates in the set, this sets off red flags. Issues would be if you included instrumental variables (causes of confounders) or common effects of confounders and outcome in the model. Such actions could biases and cause a lack of precision in the estimate.
 

hlsmith

Less is more. Stay pure. Stay poor.
#6
PS, the weights are only as good as your data. So if you are missing an important variable, you are at that limitation.