Iterative Causal Forest: A Novel Algorithm for Subgroup Identification.

Area of Science:

• Biostatistics
• Machine Learning
• Real-World Evidence

Background:

• Identifying subgroups with heterogeneous treatment effects (HTEs) in real-world evidence (RWE) is crucial but challenging.
• Existing methods may not efficiently pinpoint these subgroups based on key variables.

Purpose of the Study:

• To develop and validate an iterative causal forest (iCF) algorithm for precise HTE subgroup identification.
• To assess the performance of iCF compared to other machine learning methods in simulation studies.
• To apply iCF to identify HTE subgroups for specific diabetes medications in Medicare beneficiaries.

Main Methods:

• Developed an iterative causal forest (iCF) algorithm.
• iCF iteratively grows causal forests (CFs) with varying depths and uses plurality votes for subgroup decisions.
• Cross-validation selects the optimal subgroup decision that best predicts treatment effects.

Main Results:

• Simulations across 12 scenarios demonstrated iCF's superior performance over other methods for subgroup identification.
• Application to Medicare data identified a subpopulation of heart failure patients benefiting from sodium-glucose cotransporter-2 inhibitors.
• iCF identified HTEs and additive interactions consistent with existing clinical knowledge.

Conclusions:

• The iterative causal forest (iCF) is a promising method for identifying HTE subgroups in RWE.
• iCF can effectively pinpoint patient subpopulations with differential treatment responses.
• This method is valuable in RWE studies where unmeasured confounding can be mitigated by study design.