gms | German Medical Science

Background: Health technology assessment bodies increasingly ask for and use results from real-world observational studies to complement the evidence of randomized clinical trials in order to support clinical guidelines and guide health care policy decisions. However, real world evidence (RWE) poses additional and severe methodological challenges for evaluating causality, including time-fixed confounding, time-varying confounding, selection bias, missing/misclassified data, no clear treatment assignment, immortal time bias, dynamic treatment regimens, non-adherence and treatment switching. In the presence of time-varying confounders (i.e., variables that both influence the treatment and are affected by the treatment), “traditional” methods to control for confounding such as multivariate regression analysis or propensity score may fail, and more complex causal inference methods (g-methods) must be used instead. One of the g-method approaches is using a marginal structural model with inverse probability weighting based on repeated measurement data.

Methods: This presentation will address the design of a target trial based on real-world observational data, the potential biases that must be considered, and the steps of the causal analysis. These steps are illustrated with a real world example from oncology.

Results: The causal framework presented includes the following four parts:

1.

Understanding time-varying confounding: Using a directed acyclic graph (DAG), we will demonstrate why traditional regression analysis fails in the presence of time-varying confounding.

2.

Designing a target trial: The causal framework is based on the specification and emulation of a target trial, which is a hypothetical randomized trial addressing the research question of interest. This step prevents from several biases, including immortal time bias.

3.

Applying a marginal structural model with inverse probability of treatment weighting: After generating a pseudo population with treated and non-treated patients using “replicates”, a marginal structural Cox model will be applied to control for time-varying confounding an selection bias.

4.

Case example: The causal framework will be demonstrated using a study addressing the research question “Should 2nd-line therapy be initiated after successful first-line therapy and progression of the cancer?” The results of the observational study will be compared to those of a randomized clinical trial with the same research question and compared strategies.

Finally, key assumptions, limitations and pitfalls of the causal framework are discussed.

Conclusion: When using RWE data, the application of the target trial approach with replicates in combination with a causal marginal structural model with inverse probability weighting can protect against multiple severe biases and procuce results that match those of a randomised clinical trial.

The study presented on this methodological presentation was funded by an independent research grant from Eli Lilly and Company.

The authors declare that an ethics committee vote is not required.