Collaborative targeted maximum likelihood estimation for variable importance measure: Illustration for functional...

Area of Science:

• Biostatistics
• Causal Inference
• Machine Learning in Healthcare

Background:

• Standard statistical methods for determining disease cause importance are often ad hoc and derived from machine learning.
• Causal inference and data-adaptive methods offer more tailored and assumption-free approaches.

Purpose of the Study:

• To propose and evaluate a fully automated semiparametric procedure for variable importance measure (VIM) estimation.
• To compare a novel collaborative targeted maximum likelihood estimation (cTMLE) approach against traditional targeted maximum likelihood estimators (TMLE).

Main Methods:

• Implementation of a fully automated VIM procedure using cTMLE.
• Application to a prospective observational study of traumatic brain injury (TBI) patients.
• Comparison of cTMLE and TMLE performance using parametric bootstrap, focusing on bias and confidence interval coverage.

Main Results:

• The cTMLE procedure demonstrated robust automated estimation of VIMs in high-dimensional data.
• cTMLE exhibited substantially less positivity bias and improved 95% confidence interval coverage compared to TMLE.
• Clinically important predictors of the Glasgow Outcome Scale - Extended (GOSE) were identified in TBI patients.

Conclusions:

• Automated cTMLE is a powerful tool for estimating VIMs in complex, high-dimensional datasets.
• This method offers advantages over traditional approaches by reducing bias and improving reliability.
• The findings support the use of cTMLE for targeted model selection in clinical research.