Artikel
Guidance for a Causal Comparative Effectiveness Analysis Using “Big” Real World Data: the Case of When to Start Statin Treatment
Suche in Medline nach
Autoren
Veröffentlicht: | 6. März 2018 |
---|
Gliederung
Text
Purpose: Within DEXHELPP, a large FFG-funded project on decision support tools for health-policy planning, causal relations are estimated from routine data as source to gain information outside the artificial setting of clinical trials. Common challenges occurring with real world evidence (RWE) are confounding, missing/misclassified data, no explicit treatment assignment, multiple dynamic treatment regimens, and switching. Graphical, structural and statistical techniques exist to estimate causal relations using RWE. The aim of this project is to describe a causal (counterfactual) approach for analyzing such datasets using the case of when to start statin treatment to prevent cardiovascular disease (CVD).
Methods: We determined three comparative strategies, starting statin treatment when the ESC-SCORE exceeds specific thresholds, e.g. 1%, 5%, 10%. We adjust for potential time-independent and time-dependent confounding and protect against selection bias using directed acyclic graphs (DAGs) representing assumptions. We generate a study protocol following the “target trial” approach, describe data structures needed for the causal assessment, and provide potential solutions where necessity and availability of data deviate.
Results: Individuals between 40 and 75 years of age with no history of diagnosed stroke or myocardial infarction (MI) within the last month enter the target trial study at the time they first exceed the risk-threshold of 1% and are followed up for 5 years. Counterfactual “replicates” of all patients are assigned to each treatment strategy. A causal per protocol analysis is applied: individuals not following the assigned treatment protocol are censored at time of protocol violation. As censoring is informative and time-dependent confounding is present, inverse probability of censoring weighting is used to derive unbiased causal estimates. The Austrian GAP-DRG database contains ICD9/ICD10 codes from 2006-2013. As the ESC-SCORE requires continuously measured values that do not exist for all variables, rules are designed to estimate the risk score.
Conclusions: DAGs and a protocol following the “target trial” approach are important tools to guide the database structure, data assessment, and the choice of the analytic strategy in deriving causal effects from big data. Causal analytic methods are needed to provide valid estimates. Techniques using replicates can overcome the problem of missing explicit treatment assignment in secondary RWE data sets.