14  Take-aways

14.1 What are the strengths

  1. Comprehensive comparison of various algorithms, going beyond the current literature.

  2. Plasmode simulation includes realistic considerations:

  • Complex function
  • Many noise variables available
  • Based on an open dataset
  1. Comprehensive list of performance measures considered:
  • Coverage
  • relative error in model SE

14.2 What could improve

  1. More proxies!
  • Others have used \(k = 10\) in the same context to compare
  1. Investigate additional double cross-fit versions:
  • Only kitchen sink used

14.3 Conclusions

High-dimensional proxy adjustment context
  1. Employ TMLE whenever possible for more unbiased and accurate results.
  • TMLE methods have competitive advantage while adjusting for too many proxies (high-dimensional; beyond low dimensions where it is now popular).
  1. Exercise caution with Super Learner, considering the trade-offs between smooth and non-smooth learners.