Stat Cafe - Alexander Coulter
fastfrechet and distQTL: New R Packages For Distributional Regression (And Bonus Topic*)
- Time: Monday 4/7/2025 from 11:30 AM to 12:30 PM
- Location: BLOC 457
Description
Regression methods with complex-structured responses, such as empirical distributions, are becoming more prevalent as data collection technologies become more advanced. These include inferential tests like asymptotic F-tests, and variable selection methods. At the same time, fast algorithms need to be developed so these methods are accessible and feasible. I’ll discuss algorithmic developments I’ve made in the specific context of global Fréchet regression with univariate distribution responses, with applications to diabetes research and single-cell transcriptomics. New algorithms apply these methods upward of 10,000x faster than before, and thousands of times faster than competing large-scale modeling efforts like generalized mixed effects models. In particular, the optimization technique I applied for variable selection can likely be generalized to other contexts. Time permitting, I’ll walk through specific examples.