This paper introduces a framework for condition-number diagnostics in Interactive Double Machine Learning (IDML) models, providing early warning indicators for numerical instability and theoretical bounds on the condition number for IDML estimators.
Key findings
Proposes a novel framework for condition-number diagnostics in IDML models.
Develops theoretical bounds on the condition number for IDML estimators.
Establishes connections between conditioning and the validity of Neyman-orthogonal moments.
Demonstrates through simulation studies that condition-number diagnostics can identify scenarios where standard DML inference breaks down.
Limitations & open questions
The study focuses on IDML models and may not be directly applicable to other causal inference frameworks.
The practical implementation of the proposed diagnostics requires further validation in real-world datasets.