|
Abstract:
|
‘No unmeasured confounding’ is a key assumption researchers make when producing causal inference using real world data. Given the potential for unmeasured confounding to produce bias sufficient to change directions of treatment effect estimates, a thorough assessment of bias due to unmeasured confounding is critical to produce quality RWE. However, establishing best practices for such sensitivity analyses is challenging due to the complexity, lack of broad applicability (dependence on specific information about the unmeasured confounder), and lack of comparison between methodologies. This talk will review emerging methods for unmeasured confounding sensitivity analysis and provide guidance for best practices for producing quality RWE. This will include more broadly applicable methods such as the E-value and rule out approaches that quantify the robustness through placing bounds on the potential confounding, as well as direct approaches such as Bayesian modeling which can incorporate information from external sources to provide adjusted treatment effect estimates. Guidance on best practices will be proposed to assist researchers in planning comparative real world studies.
|
