Finding my control group PSM, Matching or regression model and correct for confounder

Longo

New Member
#1
Hey there,

First off I'm hoping I'm posting this in the right place.

I'm doing research where there is a patient group of n=9 that underwent our new treatment. Whilst there is a group of n=140 that underwent the regular treatment.

Now my prof. wants me to find the 18 most matching patients from the regular group.
So we can ofcourse compare all our findings.

However I have no idea which method is the best and why.
Should I use Propensity Score Matching, regular matching based on a few key-variables, or use a regression model and correct for the confounders?

Hope somebody can help me out with this.

Kind regards,
Longo
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
Re: Finding my control group PSM, Matching or regression model and correct for confou

So given the treatment was not randomized, you are trying to balance baseline covariates that may confound the results. So you need to start off by figuring out what variables may be confounders and/or predictors of the outcome. Also, you are not bounded to only having 2 controls. You can typically gain a little more power given more controls, but like to stop around 2-4 before things become too unbalanced.


Is there a reasonably large treatment effect? Given your very small sample sizes power may be an issue regardless. You may be just fine doing this by hand (e.g., not automating it), but you would want to randomize your lists to make sure you are not biasing your selections. I believe some people dislike PS matching, perhaps watch Gary King's (of Harvard) Youtube video on propensity score matching to give you ideas about its potential weaknesses.
 

Longo

New Member
#3
Re: Finding my control group PSM, Matching or regression model and correct for confou

Thank you for responding!

The treatment was given based on availability (it's about treated bad quality organs vs untreated regular organs transplantation). I guess that's randomized?
So I use a logistic regression model where I drop off a variable and see the effect, right?

The hypothesis is that these treated bad quality organs will perform as good or better than untreated regular ones. So that's a reasonably large treatment effect, right?

Currently (based on other articles) I have chosen my control group based on time. I would look at the time of each treated transplant and find the two nearest untreated transplants, while also considering same surgery method, cause of transplantation and some others. The groups that I got with that were not significantly different from eachother on any variable, (age, sex, etc.)
Is this then a totally invalid method? Or would this be defendable when asked upon? (I find it strange that others have successfully applied this, but I can't find why it would be wrong).

What do you mean by doing this by hand? Just randomly picking out my n=18?

Again, thank you for all the information and tips.
 

hlsmith

Less is more. Stay pure. Stay poor.
#4
Re: Finding my control group PSM, Matching or regression model and correct for confou

It almost seem like you described a non-inferiority test, as good or better!