The R.O.A.D. to clinical trial emulation.

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• Randomized controlled trials (RCTs) are the gold standard for treatment effectiveness but are often impractical.
• Observational studies offer an alternative but are limited by confounding bias.
• Target trial emulation aims to mimic RCT designs in observational data but struggles with unmeasured confounding.

Purpose of the Study:

• To present a novel framework for target trial emulation that addresses confounding bias, including unmeasured confounding.
• To improve causal inference from observational data by overcoming limitations of existing emulation methods.
• To identify patient subgroups with heterogeneous treatment effects (HTE) for advancing precision medicine.

Main Methods:

• Applied target RCT eligibility criteria to real-world observational data.
• Corrected the observational cohort using optimization to match RCT covariate distribution and baseline prognosis.
• Addressed unmeasured confounding by adjusting treated group prognosis estimates using cost-sensitive counterfactual models.
• Utilized optimal decision trees to identify HTE subgroups.

Main Results:

• The framework successfully addressed both observed and unobserved confounding, a long-standing challenge in causal inference.
• External models verified the absence of confounding.
• The validity of estimated treatment effects was confirmed by the original trial team.
• Identified patient subgroups with differential treatment benefits.

Conclusions:

• The novel target trial emulation framework significantly improves causal inference from observational data.
• This approach overcomes critical limitations of previous emulation methods, particularly confounding bias.
• The framework has strong potential for advancing precision medicine by identifying optimal treatments for specific patient subgroups.