Propensity Score Matching

#1
Hello,

I have retrospective non-randomised data and wish to investigate the relationship between blood transfusions (2 treatment groups : transfused or not transfused) and infections (binary outcome - true/false) after major heart surgery in children. To do this I want to use a propensity score (propensity to receive a blood transfusion) to match patients from both treatment groups.

I've been using various methods in R to do the propensity score matching (GenMatch() and Match() with default settings) but consistently have the problem that AFTER MATCHING the means for 2 covariates are significantly different (p << 0.0001) between the two groups.

The 2 covariates are body weight and blood loss. My question is a general one - is it legitimate for me to report that "the two treatment groups cannot be matched for all covariates because of the strength of association between the treatment group and certain covariates" OR should it be possible to adequately match two treatment groups (i.e. I just need to iteratively adjust the settings / weightings in the propensity score algorithms until I have achieved adequate matching).

I would really appreciate any advice from members with experience in propensity score matching. With my thanks in advance,

Martin
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
How big is your dataset? I would look at doing some of the matching manually to try and really understand the problem. Can you look at BMI instead of weight? I also think that before you through in the towel, you may try to just match one of the two variables (if that would be feasible).
 
#3
How big is your dataset? I would look at doing some of the matching manually to try and really understand the problem. Can you look at BMI instead of weight? I also think that before you through in the towel, you may try to just match one of the two variables (if that would be feasible).
Thank you for your reply,

The dataset contains 800 children with 24 possible covariates. Yes I can and will look at BMI.

There are compelling medical reasons why smaller children and children who lose more blood will require more transfusion during heart operations. Is it not dangerous to leave one of these variables out?

From a medical point of view it would be interesting to say that the relationship between transfusion and volume of blood loss and body weight is so strong, that assignment to treatment is not ignorable in these types of non-randomised studies.

I suppose my question is, is it justifiable to say that : "I have performed state of the art iterative matching methods and am unable to adequately match these patients"?

Thanks again for your reply
 

hlsmith

Less is more. Stay pure. Stay poor.
#4
As you may already know, when you match on a variable you have controlled for its effects and can no longer examine its relationship with the dependent variable. If you cannot match on it, you would then just include it in the model. When running the multivariate procedure you can then hold it constant and interpret the effect of transfusion on infection. If you have a large enough sample then you don't need to have as many concerns about matching on the variable because you can examine its effect or partial effects later on. Matching is more imperative when you have smaller samples that cannot support many independent variables and become over-parameterized. Its seems like if you incorporate these two variables into the final model they may explain quite a bit of outcome.