Artikel
Assessing type and impact of biases potentially occurring when analyzing real world evidence: comparing emulating a target trial with traditional methods: the case of second line treatment for ovarian cancer
Suche in Medline nach
Autoren
-
Felicitas Kühne
- UMIT – University for Health Sciences, Medical Informatics and Technology, Hall i.T., Austria
-
Marjan Arvandi
- UMIT – University for Health Sciences, Medical Informatics and Technology, HAll.i.T., Austria
-
Lisa Hess
- Eli Lilly and Company, Indianapolis, United States
-
Douglas Faries
- Eli Lilly and Company, Indianapolis, United States
-
Raffaella Gothe
- UMIT – University for Health Sciences, Medical Informatics and Technology, Hall in Tirol, Austria
-
Holger Gothe
- UMIT – University for Health Sciences, Medical Informatics and Technology, Berlin, Austria
-
Julie Beyrer
- Eli Lilly and Company, Indianapolis, United States
-
Uwe Siebert
- UMIT – University for Health Sciences, Medical Informatics and Technology, Hall i.T., Austria
| Veröffentlicht: | 26. Februar 2021 |
|---|
Gliederung
Text
Background: Real world evidence (RWE) pose additional methodological challenges for evaluating causality, including confounding, missing/misclassified data, no clear treatment assignment, dynamic treatment regimens, and switching. We aimed to assess the type and impact of biases potentially occurring when analyzing RWE using the case of ovarian cancer treatment.
Methods: We compared overall survival (OS) with and without second-line treatment in ovarian cancer patients (n=1581) using retrospective IMS Oncology electronic medical records. We identified potential confounding and other biases using directed acyclic graphs (DAG). To assess the biases, we applied several analytic approaches starting with simple Cox regression with and without baseline variables, including time-dependent covariates to reduce immortal time bias, applying the “target trial” approach, including pseudo-populations and marginal structural models with inverse probability of censoring weighting (IPCW) (causal). We compared hazard ratios (HR) and 95% confidence intervals (95%CI) to assess the bias associated with each of these approaches compared to the causal analysis. We used a randomized controlled trial as a reference case.
Results: The crude and baseline-adjusted analyses yielded a HR for second-line versus no second-line therapy of 0.565 (95%CI 0.495-0.645) and 0.535 (95%CI 0.468-0.613), respectively. Including treatment as a time-dependent covariate to account for immortal time bias, the corresponding crude and adjusted HR increased to 1.665 (95%CI 1.459-1.901) and 1.683 (95%CI 1.407-2.014). Applying a causal (counterfactual) analysis using IPCW and replication yielded a HR of 1.067 (95%CI 1.020-1.115), which matched the results of a published randomized clinical trial.
Conclusion: When using routine RWE data, DAGs can guide the identification of potential biases and variables that need to be controlled for. In our analysis, potential biases were substantial with different directions. Only the application of the target trial and replication approach in combination with a causal analysis matched data from clinical trials.
The authors declare that they have no competing interests.
The authors declare that an ethics committee vote is not required.
