A Bayesian Semi-Parametric Scalar-On-Function Quantile Regression with Measurement Error using the GAL

TL;DR

This paper introduces a Bayesian semi-parametric scalar-on-function quantile regression method that accounts for measurement errors in covariates using the flexible GAL distribution, demonstrated through simulations and NHANES data analysis.

Contribution

It extends Bayesian quantile regression to handle complex measurement errors in scalar and functional covariates with a novel GAL-based approach.

Findings

GAL distribution offers greater flexibility than AL in Bayesian quantile regression.

The proposed method effectively adjusts for measurement errors in simulations.

Application to NHANES data reveals insights into BMI related to physical activity.

Abstract

Quantile regression provides a consistent approach to investigating the association between covariates and various aspects of the distribution of the response beyond the mean. When the regression covariates are measured with errors, measurement error (ME) adjustment steps are needed for valid inference. This is true for both scalar and functional covariates. Here, we propose extending the Bayesian measurement error and Bayesian quantile regression literature to allow for available covariates prone to potential complex measurement errors. Our approach uses the Generalized Asymmetric Laplace (GAL) distribution as a working likelihood. The family of GAL distribution has recently emerged as a more flexible distribution family in the Bayesian quantile regression modeling compared to their Asymmetric Laplace (AL) counterpart. We then compared and contrasted two approaches in our ME-adjusted steps through a battery of simulation scenarios. Finally, we apply our approach to the analysis of an NHANES dataset 2013-2014 to model quantiles of Body mass index (BMI) as a function of minute-level device-based physical activity in a cohort of an adult 50 years and above.