OHDSI Home | Forums | Wiki | Github

Adding custom covariates to PLE study


Is is possible to edit a package produced in ATLAS to add custom cohort-based covariates? In this case, for a study looking at treatment effects of COVID treatments I want to add conditions such as hypertension or obesity as covariates. I think it should be possible by creating cohorts for the covariates in ATLAS and then editing the package, but I’m not sure how.

I saw this article:(Adding custom covariates based on ATLAS cohorts • SkeletonPredictionStudy) but those code lines were not present in the estimation study package.

In general I highly recommend using the default set of covariates, so all conditions, drugs, procedures, etc. We’ve shown in many papers that this approach is optimal for addressing measured and even unmeasured confounding.

Is there a reason for wanting manually-crafted covariates?


Thank you for your response. In this study, we want to measure treatment effects for covid treatments, so the idea was to include other conditions that may affect treatment allocation and outcome. However, we are still considering not including these as covariates if they do not impact treatment allocation.

In your opinion, would using the default settings of all conditions be appropriate here? (In that scenario, would every single recorded condition, ie vaccination status/hypertension/obesity be included in the propensity score?)

Yes, I would always use the default settings. However, you should not forget to explicitly exclude covariates that represent the exposures of interest, as described here in the Book of OHDSI

Thank you! I’m looking for more information about the validity of using all possible covariates, and I can’t seem to find papers about it. Do you mind pointing me towards the papers you referenced which show it reduces measured and unmeasured confounding?

Is there a concern that this method would wrongfully incorporate covariates that are only predictors of treatment but not of the outcome?

Incorporating so-called ‘instrumental variables’ can lead to some loss of power, but I would argue in observational research our concern for bias far outweighs our concern for power. Some would argue it could amplify existing bias, but that is really unlikely when you’re correcting for so many covariates.

Here are some papers:

I agree with @schuemie. On instruments, remember also that OHDSI has diagnostics to make sure there are no strong instruments. There is the univariate correlation with the treatment, which makes sure there is nothing too strong, and there is the check for equipoise (preference score) that makes sure there is no combination of variables in aggregate that serve as a strong instrument. And instruments do not amplify the confounding. They can make the bias go up in the sense that the confounding is unmasked by removing some of the good effect of the treatment.

I’m second to @hripcsa and @schuemie

The prevalence of instrumental variables in medical data is controversial. The identification of true instrumental variables is even hardly possible in RWD via univariate association analysis in hdPS. In the simulation study, Myers et al., the group that invented hdPS, found that inevitably including an instrumental variable in covariates does not amplify the bias much in many scenarios (74) . Hence, they stated that “investigators should err on the side of inclusion rather than exclusion of potential confounders” (75,76) .

  1. Myers JA, Rassen JA, Gagne JJ, Huybrechts KF, Schneeweiss S, Rothman KJ, et al. Effects of Adjusting for Instrumental Variables on Bias and Precision of Effect Estimates. Am J Epidemiol. 2011 Dec 1;174(11):1213–22.

  2. Myers JA, Rassen JA, Gagne JJ, Huybrechts KF, Schneeweiss S, Rothman KJ, et al. Myers et al. Respond to “Understanding Bias Amplification.” Am J Epidemiol. 2011 Dec 1;174(11):1228–9.

Thank you very much for these informative responses. As of now, we have decided to do as suggested and use the default set of covariates.