Bayesian nonparametric quantile process regression and estimation of marginal quantile effects.

Area of Science:

• Statistics
• Bayesian Nonparametrics
• Machine Learning

Background:

• Estimating multiple conditional quantiles is crucial for understanding variable effects across the entire response distribution.
• Existing methods often focus on a limited set of quantiles or struggle with high-dimensional data and estimation uncertainty.

Purpose of the Study:

• To develop a flexible Bayesian nonparametric method for simultaneously estimating noncrossing, nonlinear conditional quantile curves.
• To enable estimation of the entire conditional distribution, not just specific quantiles, and handle high-dimensional covariates.

Main Methods:

• Utilizes I-spline basis functions to expand the conditional distribution function.
• Models covariate-dependent coefficients using neural networks, combining spline flexibility with neural network power.
• Incorporates Bayesian inference to account for estimation uncertainty.

Main Results:

• The proposed model can approximate any continuous quantile function and estimates all quantiles, unlike finite subset methods.
• Demonstrates superior performance in recovering response distribution quantiles with sparse data compared to existing models.
• Provides interpretable marginal quantile effects through local effect plots and variable importance measures.

Conclusions:

• The Bayesian nonparametric approach offers a powerful and flexible tool for comprehensive quantile estimation.
• The method effectively handles complex relationships and high-dimensional data, with applications in fields like birth weight analysis.
• It provides a robust framework for understanding variable impacts across the full spectrum of outcomes.