Scalable collaborative targeted learning for high-dimensional data.

Area of Science:

• Statistical inference
• Machine learning
• Computational statistics

Background:

• Robust inference in large semi-parametric models requires external estimators.
• Optimizing multiple estimators improves bias-variance trade-offs for parameter estimation.
• The original greedy C-TMLE algorithm has scalability issues with numerous covariates.

Purpose of the Study:

• To develop scalable C-TMLE algorithms for large semi-parametric models.
• To introduce novel pre-ordering strategies for improved computational efficiency.
• To present a data-driven approach for selecting optimal pre-ordering strategies.

Main Methods:

• Developed a novel C-TMLE instantiation with pre-ordered covariates, reducing time complexity.
• Proposed two pre-ordering strategies and a rule for developing others.
• Introduced the SL-C-TMLE algorithm for data-driven strategy selection.
• Compared algorithms via simulations and real-world electronic health database analyses.

Main Results:

• The greedy C-TMLE algorithm demonstrated unacceptably slow performance on large electronic health databases.
• Scalable C-TMLE and SL-C-TMLE algorithms showed good performance in simulations.
• The novel algorithms offer significant computational gains ($$ vs. $$).

Conclusions:

• Scalable C-TMLE and SL-C-TMLE algorithms effectively address the limitations of the original greedy approach.
• The developed methods are computationally efficient and perform well in practice.
• All C-TMLE algorithms are available in a public Julia software package.