Hi all,
I'm running active comparator new-user estimation studies using CohortMethod v6 on a 1M-patient OMOP CDM database in our
Parthenon platform, and I'm hitting the "High correlation between covariate(s) and treatment detected" error during
propensity score fitting. I'd appreciate guidance on the recommended approach.
Setup:
I have three comparative effectiveness studies, all using PS-matched Cox regression:
1. Simvastatin vs Atorvastatin (same-class statin comparison) — both arms are true new users with a 365-day washout
excluding all prior statins
2. Metformin vs Insulin in T2DM (cross-class but same therapeutic area) — target: first metformin + T2DM ±180d,
comparator: first insulin + T2DM, no prior metformin ever
3. Opioid vs NSAID for chronic pain (cross-class, different therapeutic area) — this one works perfectly (HR=0.913, PS
AUC=0.775, 12,867 matched pairs)
What I've tried:
- excludedCovariateConceptIds with the 4 statin ingredient-level RxNorm concept IDs (1539403, 1545958, 1510813, 1551860)
— still fails because descendant drug product covariates (e.g., "simvastatin 20 MG Oral Tablet") perfectly separate the
arms
- The opioid vs NSAID study works because the drug classes are sufficiently different that excluding the 7 ingredient
concepts is enough
My current workaround:
Setting useDrugExposureLongTerm: FALSE, useDrugExposureShortTerm: FALSE, and useDrugEraLongTerm: FALSE in the covariate
settings, relying only on demographics, conditions, procedures, measurements, and Charlson index for PS estimation. This
eliminates all drug covariates from the PS model.
My questions:
1. Is disabling all drug covariates the recommended approach for same-class comparisons, or is there a more targeted way
to exclude only the exposure-related drug covariates while keeping other drug classes (e.g., antihypertensives,
diabetes drugs) in the PS model?
2. For the metformin vs insulin study — these are different drug classes but both are glucose-lowering agents. Is there
a standard concept set of "all glucose-lowering drugs" that should be excluded, or should I exclude at the ATC-4 level?
3. Does excludedCovariateConceptIds with ingredient-level IDs automatically exclude descendant concepts (drug products,
clinical drugs) in FeatureExtraction? The OHDSI documentation suggests it should, but I'm still getting separation with
ingredient-level exclusions only.
4. Would using includedCovariateConceptIds with a curated list (conditions + procedures + measurements only) be a better
approach than trying to exclude problematic drug covariates?
Environment:
- CohortMethod v6+, FeatureExtraction (latest HADES), PostgreSQL 17
- Cohort sizes: simvastatin 225K vs atorvastatin 25K; metformin 8.2K vs insulin 1.6K
- Covariate settings: demographics (age, gender, race, index year/month), conditions (365d + 30d), drugs (365d + 30d),
procedures (365d), measurements (365d), Charlson index
- PS: 1:1 matching, caliper 0.2 SD logit, preference score trim 0.05
Any guidance from the community would be greatly appreciated. Happy to share the full analysis specifications if
helpful.
Thanks!