Wrap-Up, Conclusions, and Questions

Key Ideas, Discussion, and Next Steps

Closing session for the short course, including key takeaways and suggested next steps.

What to Take Away from the Short Course

By the end of this short course, you should be able to:

Distinguish prediction tasks from downstream inference tasks.
Explain why treating predicted outcomes as observed data can bias estimates and understate uncertainty.
Use the ipd package to compare naive, classical, and prediction-based inference workflows.
Recognize how the same statistical issues arise across epidemiology, proteomics, and model-interpretation examples.

Across the modules, a few ideas repeat:

Predictions can be useful without being interchangeable with truth.
A small labeled set can be enough to recover valid inference when used carefully.
Predictive accuracy alone is not a sufficient criterion for scientific conclusions.
Good workflows make the assumptions behind surrogate outcomes explicit.

Use this session to revisit any of the following:

When does a surrogate or predicted outcome become too weak for downstream inference to be useful?
How large does the labeled sample need to be in a new application?
Which PB inference methods seem most appropriate for your own applied work?
What practical barriers would you face when adapting these methods to your data?

If you want to keep exploring after the short course:

Revisit the Getting Started module to compare methods on simulated data.
Work through the Additional Information page for method summaries and references.
Explore the supplemental BCR-ABL Fusion module if you want a genomics example.
Read more about the ipd package and try adapting one of the workflows to your own problem.

This short course adapts and extends Stephen Salerno’s original IPD workshop materials.

For follow-up questions, please contact Tyler H. McCormick at thmccormick@gmail.com.