OHDSI Home | Forums | Wiki | Github

Adding continuous outcome support to CohortMethod

Developers
#1

Hi,

I’m interested in using the OHDSI software to answer some comparative effect estimation questions. The main limitation I am running into is that many of the packages and in particular CohortMethod don’t seem to support continuous outcomes (I have a lab result as a primary outcome). For full reference, I’m primarily interested in using propensity matching with a linear regression outcome model.

As far as I can tell, Cohort Method does not support continuous outcomes or linear regression outcome models. I looked at the code a bit and it seems like it should be possible to write a pull request to add that functionality to CohortMethod.

Am I missing some feature of CohortMethod for working with continuous outcomes? Would people be interested in a pull request for adding continuous outcome support? Is there anything I should be a aware of before writing a PR?

#2

Hi @Ethan_Steinberg! Yes, you’re right, CohortMethod currently does not support continuous outcomes, but it shouldn’t be too difficult to add that in your analysis. If you’re comfortable with R, then you could just use your own outcome model instead of calling the fitOutcomeModel in the package.

Integrating the functionality into the package would be great, although I guess that will be non-trivial. Specifically I’m thinking of how the outcome measure needs to extracted from the database and perpetuated all the way to the outcome model. Currently the package assumes outcomes are cohorts, but in this case the outcome would be closer to what we call a ‘cohort attribute’. If you really want to go down that path we should maybe brainstorm a bit about how to do that.

#3

Thanks for the quick response! For now I think we’ll just follow your suggestion and fit our own outcome model separately (probably borrowing components from the fitOutcomeModel implementation for the sake of ease).

#4

!!! I have been waiting for a long time for this moment to come, @schuemie. You made my day. :slight_smile:

#5

Just so you know we already use the cohort_attribute table as the preferred method to construct custom covariates!

This is actually the second time this week someone brought up continuous outcome models. I’ve added an issue to CohortMethod, but I don’t have a timeline for implementing it.

#6

thank you very much for your answer @schuemie. I also have a question about estimation. if I set the outcome cohort like the picture, can I use the atlas to do the estimation analysis? thank you very much

image
image824×97 3.56 KB

#7

dear all，
I am a researcher in the CDM database because I don’t have the knowledge of computers, I can’t understand the computer process. when I use the atlas to do estimation, and I set the outcome cohort like the picture, and the outcome model of analysis setting I choose is the logistic regression. My question is if the outcome cohort is like the picture, is the outcome model（logistics regression） is OK? and why？
thank you very much

image
image824×97 3.56 KB

#8

@pandamiao In this case, your outcome is binary, whether the patient will have glucose higher than 6.1 or not (so 0 or 1). So you can use logistic regression :slight_smile:

#9

thank you very much@SCYou, your answer is very important for me. At the same time, I have another question, when I use R to run the zip file, there always is an error like the picture, do you have any questions ? thank you very much!

image

#10

@pandamiao,

You are missing some of the necessary arguments required to run the createPs function.

I suggest you run help(createPs) and see which ones you’ve left out.

#11

thank you very much @gyeol99 ， I will give it a try

#12

ILLCONDITIONED means there’s a problem fitting the propensity model. This could for example be due to small sample size, or near-perfect separability.

How big are your target and comparator cohorts? Have you tried characterizing them (in ATLAS) to see if they’re (somewhat) comparable?

t