pbr6cornell
commented
8 years ago Correct, the features include the index date as part of their time at risk, and you do need to exclude the corvariates for drugs that you use to define your target/comparator group.
A potential future feature to add to the package would be to have the ability to further refine the 'time at risk' for covariates, so that users could choose whether they want to include the index date or not.
On Fri, Nov 4, 2016 at 10:29 AM, Christophe Lambert < notifications@github.com> wrote:
I used createCovariateSettings() within Atlas to generate a cohort study. The desired behavior is to not compute covariates for propensity scoring on the index date. I assumed setting useCovariateDrugEraOverlap = FALSE would do that. However, when I run propensity scoring code, it gives errors about a high correlation between covariates(s) and treatment, listing the drugs that are my treatment and comparator drugs as the culprits. I specified a 365-day pre-treatment observation period, limiting events to the earliest event per person. My assumption is that I would have 365 days of no treatment, followed by treatment on the index date (perhaps this is wrong?). It would seem to me that the problem is that the treatment on the index date is what is picked up as the high correlation.
If I list my treatment and comparator drugs in the excluded concepts, the high correlations go away, but I'm concerned other variables such as Conditions and Observations on the index date are also getting into the propensity model.
Here is the documentation for useCovariateDrugEraOverlap:
useCovariateDrugEraOverlap: A boolean value (TRUE/FALSE) to determine if covariates will be created and used in models that look for presence/absence of drug era that overlaps the cohort index date. Only applicable if useCovariateDrugEra = TRUE.
If this is in error, I would question if useCovariateConditionEraOverlap works. Also, for symmetry, it does not appear that there is a way to exclude Observations, Procedures, or Measurements via analogous "Overlap" parameters for those covariates.
Below is the JSON code defining one of my cohorts:
{ "ConceptSets": [ { "id": 0, "name": "Bipolar Disorder", "expression": { "items": [ { "concept": { "CONCEPT_ID": 436665, "CONCEPT_NAME": "Bipolar disorder", "STANDARD_CONCEPT": "S", "INVALID_REASON": "V", "CONCEPT_CODE": "13746004", "DOMAIN_ID": "Condition", "VOCABULARY_ID": "SNOMED", "CONCEPT_CLASS_ID": "Clinical Finding", "STANDARD_CONCEPT_CAPTION": "Standard", "INVALID_REASON_CAPTION": "Valid" }, "includeDescendants": true } ] } }, { "id": 1, "name": "Lithium", "expression": { "items": [ { "concept": { "CONCEPT_ID": 751246, "CONCEPT_NAME": "Lithium Carbonate", "STANDARD_CONCEPT": "S", "INVALID_REASON": "V", "CONCEPT_CODE": "42351", "DOMAIN_ID": "Drug", "VOCABULARY_ID": "RxNorm", "CONCEPT_CLASS_ID": "Ingredient", "STANDARD_CONCEPT_CAPTION": "Standard", "INVALID_REASON_CAPTION": "Valid" }, "includeDescendants": true } ] } }, { "id": 2, "name": "Drug Lithium", "expression": { "items": [ { "concept": { "CONCEPT_ID": 751246, "CONCEPT_NAME": "Lithium Carbonate", "STANDARD_CONCEPT": "S", "INVALID_REASON": "V", "CONCEPT_CODE": "42351", "DOMAIN_ID": "Drug", "VOCABULARY_ID": "RxNorm", "CONCEPT_CLASS_ID": "Ingredient", "STANDARD_CONCEPT_CAPTION": "Standard", "INVALID_REASON_CAPTION": "Valid" }, "includeDescendants": true } ] } } ], "PrimaryCriteria": { "CriteriaList": [ { "DrugEra": { "CodesetId": 2 } } ], "ObservationWindow": { "PriorDays": "365", "PostDays": 0 }, "PrimaryCriteriaLimit": { "Type": "First" } }, "AdditionalCriteria": { "Type": "ANY", "CriteriaList": [ { "Criteria": { "ConditionOccurrence": { "CodesetId": 0 } }, "StartWindow": { "Start": { "Coeff": -1 }, "End": { "Coeff": 1 } }, "Occurrence": { "Type": 2, "Count": 2 } } ], "DemographicCriteriaList": [], "Groups": [] }, "QualifiedLimit": { "Type": "All" }, "ExpressionLimit": { "Type": "All" }, "InclusionRules": [], "EndStrategy": { "CustomEra": { "DrugCodesetId": 2, "GapDays": "30", "Offset": 0 } } }
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/OHDSI/FeatureExtraction/issues/8, or mute the thread https://github.com/notifications/unsubscribe-auth/AAsrGvo2bDrVIaBORkcG87GFlhEJhPNyks5q60FXgaJpZM4KpmTp .
ChristopheLambert
schuemie
I used createCovariateSettings() within Atlas to generate a cohort study. The desired behavior is to not compute covariates for propensity scoring on the index date. I assumed setting useCovariateDrugEraOverlap = FALSE would do that. However, when I run propensity scoring code, it gives errors about a high correlation between covariates(s) and treatment, listing the drugs that are my treatment and comparator drugs as the culprits. I specified a 365-day pre-treatment observation period, limiting events to the earliest event per person. My assumption is that I would have 365 days of no treatment, followed by treatment on the index date (perhaps this is wrong?). It would seem to me that the problem is that the treatment on the index date is what is picked up as the high correlation.
If I list my treatment and comparator drugs in the excluded concepts, the high correlations go away, but I'm concerned other variables such as Conditions and Observations on the index date are also getting into the propensity model.
Here is the documentation for useCovariateDrugEraOverlap:
useCovariateDrugEraOverlap: A boolean value (TRUE/FALSE) to determine if covariates will be created and used in models that look for presence/absence of drug era that overlaps the cohort index date. Only applicable if useCovariateDrugEra = TRUE.If this is in error, I would question if useCovariateConditionEraOverlap works. Also, for symmetry, it does not appear that there is a way to exclude Observations, Procedures, or Measurements via analogous "Overlap" parameters for those covariates.
Below is the JSON code defining one of my cohorts: