📊 statistics

Bayesian Effect Selection for Additive Quantile Regression with an Application to Air Pollution Thresholds

This paper proposes a Bayesian effect selection method for additive quantile regression using Demmler-Reinsch basis expansions to separately identify linear and nonlinear covariate effects, demonstrating its superior performance in modeling air pollution thresholds and informing regulatory decision-making through an application to NO $_2$ data in Madrid.

Original authors: Nadja Klein, Aaron Wei Qi Lee, Jorge Mateu

Published 2026-06-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nadja Klein, Aaron Wei Qi Lee, Jorge Mateu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Why We Need a Better Weather Forecast for Pollution

Imagine you are trying to predict how bad the air pollution will be in a city like Madrid. Usually, scientists look at the "average" pollution level. But for public health, the average doesn't tell the whole story. What really matters is the extreme days—the days when pollution spikes so high that it triggers health alarms and forces schools to close.

The authors of this paper argue that looking at the "average" is like checking the average temperature of the ocean to see if you'll get a sunburn. You need to know about the specific, extreme heatwaves.

To solve this, they developed a new statistical tool called QDReSS. Think of it as a "smart detective" for air pollution data that does three specific jobs:

It looks at the extremes: Instead of just the average, it focuses on the top 10%, 20%, or 30% of pollution days (the "worst-case scenarios").
It sorts the suspects: It figures out which factors (like traffic, wind, or rain) actually cause pollution.
It checks the shape of the relationship: It decides if a factor affects pollution in a straight line (simple) or a curve (complicated).

The Problem with Old Tools

Imagine you are trying to describe how a car's speed affects its fuel consumption.

Old Method (The "Mixed Model"): This method tries to fit a single, messy curve to the data. The problem is, it can't easily tell you where the "straight line" part ends and the "curvy" part begins. It's like trying to separate a smoothie back into fruit and ice; once blended, you can't tell which part was which. This makes it hard to know if a factor is acting simply or in a complex way.
The New Method (QDReSS): The authors used a special mathematical trick called the Demmler-Reinsch (DR) basis. Think of this as a high-tech blender that doesn't just mix ingredients; it keeps them in separate, labeled jars while blending. This allows the model to clearly separate the "straight line" effects from the "curvy" effects without them getting mixed up.

How the Detective Works: The "Spike and Slab"

Once the data is separated into straight lines and curves, the model needs to decide: Does this factor actually matter, or should we ignore it?

The authors use a Bayesian technique called a "Spike and Slab" prior. Here is a metaphor for how it works:

Imagine you are a hiring manager looking at a stack of resumes (the different factors like traffic, wind, and temperature).

The Spike: This is a tiny, sharp needle. If a resume (a factor) looks like it has no value, the needle pokes it, and the resume is instantly shrunk down to zero size. It disappears from the hiring pool.
The Slab: This is a wide, flat table. If a resume looks promising, it gets placed on the table, where it is given plenty of room to show its true value.

The model uses this mechanism to automatically decide:

Linear Effect: Does traffic increase pollution in a steady, straight line? (Keep it on the table).
Nonlinear Effect: Does traffic increase pollution, but only after a certain point (like a traffic jam)? (Keep it on the table, but in the "curvy" section).
No Effect: Does rain actually change the pollution levels? (Poke it with the spike and ignore it).

What They Found in Madrid

The team tested their new detective tool on real data from Madrid, Spain, looking at Nitrogen Dioxide ( $NO_2$ ), a harmful gas mostly from cars. They looked at three different "alarm levels" (moderate, high, and extreme pollution days).

Here is what their "smart detective" found:

Traffic is tricky: Old models might have thought traffic just adds pollution in a straight line. QDReSS found that traffic actually has a complex, curvy relationship with pollution. It's not just "more cars = more pollution"; the relationship gets complicated, especially on the worst pollution days.
Ozone is a double agent: They found a clear, curvy relationship between Ozone ( $O_3$ ) and $NO_2$ . As Ozone goes up, $NO_2$ tends to go down, but not in a simple way.
Weather matters differently for different days:
- Temperature: On the worst pollution days, higher temperatures lead to significantly higher pollution.
- Wind: On normal days, wind helps clear the air (a straight line). But on extreme pollution days, the relationship changes and becomes more complex.
- Rain: Rain helps clean the air, but its effect gets stronger and more complex on the worst days.

Why This Matters

The paper concludes that by using this new method, local authorities in Madrid (and potentially other cities) can get a much clearer picture of why pollution spikes happen.

Instead of guessing, they can see that on a hot, still day with heavy traffic, the pollution isn't just a little higher; it's following a specific, dangerous curve. This helps officials make better short-term decisions, like telling people to stay indoors or reducing traffic before the pollution hits a dangerous level, rather than just reacting after the fact.

In short, QDReSS is a new, sharper lens that helps us see the hidden, complex shapes of air pollution, ensuring we are ready for the worst days, not just the average ones.

Technical Summary: Bayesian Effect Selection for Additive Quantile Regression with an Application to Air Pollution Thresholds

Problem Statement
Air pollution regulatory limits are defined by exceedances of concentration thresholds, which correspond to the upper conditional quantiles of pollutant distributions. While existing statistical methods can model nonlinear relationships between covariates and air pollution, they often fail to simultaneously address three critical needs: (1) selecting which covariates affect the response, (2) determining the functional form of these effects (linear, nonlinear, or both), and (3) doing so specifically for conditional quantiles rather than just the mean.

Common approaches, such as the Mixed Model Representation (MMR) of penalized splines, allow for flexible nonlinear effects but fail to provide a meaningful separation between linear and nonlinear components. This lack of orthogonal decomposition leads to inconsistent estimation of these distinct effect parts. Furthermore, traditional variable selection methods (e.g., LASSO, COSSO) often operate in linear or mean-regression contexts and do not naturally extend to decomposing effects within quantile regression frameworks.

Methodology
The authors propose QDReSS (Quantile Distributional Regression with Effect Selection and Structure), a Bayesian framework for additive quantile regression that integrates effect selection with effect decomposition.

Demmler-Reinsch (DR) Basis Expansion:
To resolve the inconsistency of the MMR, the authors employ a DR basis expansion. This approach performs an $L^2$ -projection of cubic B-splines onto the space of linear functions. The resulting basis functions are empirically orthogonal to the linear component and mutually orthogonal. This allows for a clean decomposition of any additive effect $f(x)$ into a linear part ( $f_{lin}$ ) and a nonlinear part ( $f_{nonlin}$ ), ensuring that these components can be estimated consistently.
Bayesian Inference with NBPSS Priors:
The framework utilizes an Asymmetric Laplace Distribution (ALD) as a working likelihood to facilitate Bayesian quantile regression. To achieve variable and effect selection, the authors place Normal Beta Prime Spike and Slab (NBPSS) priors on the importance parameters associated with the linear and nonlinear components of each covariate.
- The hierarchical prior structure allows for group-level shrinkage.
- It distinguishes between four states for each covariate: no effect, linear effect only, nonlinear effect only, or both.
- The inclusion probabilities are learned from the data via hyperparameters, avoiding the need for quantile-specific tuning.
Posterior Estimation:
Posterior inference is conducted using an efficient Gibbs sampler. The conjugate structure of the model (facilitated by the scale-mixture representation of the ALD and the Gaussian priors) allows for closed-form full conditional distributions for all parameters, including the latent weights, variance components, and inclusion indicators.

Key Contributions

Theoretical Consistency: The paper provides a theoretical proof that the DR basis expansion yields consistent estimators for both linear and nonlinear effect components in additive quantile regression, a property not guaranteed by standard MMRs.
Integrated Selection and Decomposition: Unlike existing methods that select entire functions or operate only in linear settings, QDReSS simultaneously selects covariates and determines whether their effects are linear, nonlinear, or both, specifically at user-defined quantiles.
Computational Efficiency: The orthogonal nature of the DR basis simplifies the penalty structure (reducing roughness penalties to identity matrices), resulting in faster computation compared to MMR-based approaches.
Practical Implementation: The authors provide a complete Bayesian workflow, including a streamlined prior elicitation strategy based on the scaling properties of the beta prime distribution, and an open-source implementation.

Results and Application
The method was applied to daily air pollution data from Madrid, Spain (2016–2019), modeling Nitrogen Dioxide ( $NO_2$ ) concentrations at quantiles $\tau \in \{0.6, 0.8, 0.9\}$ .

Predictive Performance: In a 10-fold cross-validation study, QDReSS consistently achieved lower quantile loss (better out-of-sample predictive accuracy) compared to the competing MMR approach across all tested quantiles.
Effect Decomposition Findings:
- Ozone ( $O_3$ ): Exhibited a strong, consistent nonlinear negative effect on $NO_2$ across all quantiles, confirming complex chemical interactions.
- Traffic: Contrary to initial univariate linear assumptions, traffic flow was found to have a significant nonlinear effect on $NO_2$ , particularly at higher quantiles (extreme pollution events).
- Climatological Variables: The effects of temperature, wind speed, and precipitation varied by quantile. For instance, precipitation showed increasing nonlinear strength at higher quantiles, while wind speed effects shifted from linear (at lower quantiles) to nonlinear (at higher quantiles).
- Carbon Monoxide (CO): Showed a strong linear effect only at the lower quantile ( $\tau=0.6$ ), becoming negligible at higher thresholds.

Significance and Claims
The authors claim that QDReSS offers a necessary advancement in modeling air pollution by enabling data-driven decisions on the functional form of covariate effects at specific pollution thresholds. The study highlights that extreme pollution events (high quantiles) are driven differently by climatological variables and traffic structures compared to average conditions.

The paper argues that these findings underscore the need for advanced statistical models that support short-term decision-making. By accurately modeling the nonlinear drivers of threshold exceedances, local authorities can better mitigate or prevent violations of $NO_2$ concentration limits. The authors modestly note that while the study focuses on a single station to elucidate intrinsic mechanisms, the framework is generalizable to other cities and can be extended to include spatial effects or simultaneous quantile estimation to prevent quantile crossing.