Uncertainty Quantification and Calibration in Stellar Parameter Estimation

Practitioners often take datasets they are given and run downstream analyses "plugging in" any estimates and/or uncertainties provided, regardless of origin (eg, constructing a histogram of the estimated values, selecting a subset based on the estimates and their uncertainties). This behaviour can lead to various biases, especially when dealing with estimates derived from Bayesian or machine learning-driven approaches (where the priors can have outsized impact, especially when data are out of distribution) and when underlying estimators are multimodal. To address some of these challenges, I will present ongoing work on hybrid "Frasian" (Frequentist-Bayesian) inference approaches that can recalibrate Bayesian predictions under various marginal/conditional coverage settings using a separate calibration set. An application of these strategies to data on stars from the Gaia and APOGEE surveys will also be provided.

This work was done in collaboration with James Carzon, Luca Masserano, Joshua D. Ingram, Alex Shen, Antonio Carlos Herling Ribeiro Junior, Tommaso Dorigo, Michele Doro, Rafael Izbicki, and Ann B. Lee, as well as Ricardo Baptista.