Robust Machine Learning for Treatment Effects in Multilevel Observational Studies Under Cluster-level Unmeasured...

Area of Science:

• Statistics
• Computer Science
• Social Sciences

Background:

• Machine learning (ML) methods are increasingly used for causal inference to mitigate model mis-specification concerns.
• Existing ML methods often assume all confounders are measured, limiting their application in observational studies with unmeasured variables.

Purpose of the Study:

• To propose novel ML methods for estimating treatment effects in the presence of cluster-level unmeasured confounders.
• To address a common challenge in multilevel observational studies where unmeasured factors are shared within clusters.

Main Methods:

• Developed a family of ML methods specifically designed to handle cluster-level unmeasured confounders.
• Validated the proposed methods through extensive simulation studies across various multilevel observational study designs.

Main Results:

• The proposed ML methods demonstrated robustness against biases introduced by unmeasured cluster-level confounders.
• Simulations confirmed the reliability of the methods in diverse multilevel settings.
• Applied the methods to analyze the impact of algebra courses on math achievement using the Early Childhood Longitudinal Study dataset.

Conclusions:

• The novel ML methods effectively estimate treatment effects in multilevel studies with unmeasured cluster-level confounders.
• These methods offer a robust solution for causal inference in complex observational data.
• The CURobustML R package provides open access to these advanced analytical tools.