Nonparametric Hierarchical Bayesian Quantile Modeling

• The paper presents nonparametric hierarchical Bayesian quantile models that enable robust estimation by leveraging flexible priors on quantile functions and adaptive shrinkage across groups.
• It details methods like quantile pyramids, spline-based expansions, and Dirichlet process mixtures that ensure monotonicity and accommodate high-dimensional covariate spaces.
• The approach combines hierarchical modeling with efficient MCMC algorithms to achieve noncrossing, consistent quantile estimates even with complex, irregular, or censored data.

Nonparametric hierarchical Bayesian quantile modeling encompasses a range of methodologies that place flexible, nonparametric priors on quantile functions or related functionals, and structure inference in a hierarchical Bayesian framework. These models allow for robust estimation of quantiles (or entire conditional distributions) in settings featuring nonlinearity, heterogeneity across subpopulations, high dimensionality, and/or distributional irregularity, while additionally supporting shrinkage and borrowing of information in grouped data regimes. This article reviews key frameworks and advances, with particular focus on methodologies that enforce monotonicity, enable simultaneous estimation of multiple quantiles across various indexing sets, and accommodate both finite and infinite-dimensional covariate spaces.

1. Foundational Principles and Model Classes

Nonparametric hierarchical Bayesian quantile models are characterized by the direct placement of nonparametric priors either on quantile functions, the underlying conditional distributions, or the parameters—often within a multilevel structure for handling subpopulations or longitudinal/study-level hierarchies.

A central distinction arises between (a) quantile-based approaches, which impose structure directly on the quantile curve/surface (such as pyramid or spline-based expansions), and (b) distribution-based approaches, which define flexible models for the entire conditional distribution and subsequently derive quantile functionals.

Key model classes include:

• Quantile pyramid models: Construct quantile functions by recursively splitting probability mass via Beta (or Dirichlet) random variables in a multilevel tree, with extensions to covariate-dependent and hierarchical settings (Rodrigues et al., 2016, An et al., 2023, Rodrigues et al., 2017).
• Spline/tensor product expansions: Model quantile or distribution surfaces via spline bases, enforcing monotonicity through constraints on the coefficients, often combined with Dirichlet process or Gaussian process structure for nonparametric flexibility (Das et al., 2016).
• Dirichlet process mixture models: Directly model the (conditional) distribution using flexible DP mixtures, incorporating covariate dependence and enabling quantile extraction post hoc (Bhattacharya et al., 2020, Kobayashi et al., 2016).
• Geometric-measure quantile models: Place priors on quantiles defined with respect to simplex constraints and use geometric measure theory for prior construction and posterior calculation, well-suited to discrete support or censored data settings (Bornn et al., 2016).

Hierarchical Bayesian structure is introduced via shared priors or mixing distributions at the group or population level, allowing quantile functionals of subpopulations to borrow strength from a common distribution, yielding improved stability and interpretability in sparse data regimes.

2. Model Construction and Hierarchical Priors

A variety of construction strategies are used to produce nonparametric and hierarchical structure:

Quantile Pyramid Priors

The quantile pyramid prior recursively divides the quantile function into subintervals, assigning random probabilities using Beta or Dirichlet-distributed split weights. The hierarchical structure is achieved by specifying independent (but possibly exchangeable or partially pooled) quantile pyramids for each group, or, in the dependent quantile pyramid (DQP) framework, by replacing splitting weights with stochastic processes indexed by covariates (Rodrigues et al., 2016, An et al., 2023).

Covariate-dependent extensions include using Gaussian process priors for splitting weights, so that quantile curves vary smoothly over covariate space:

• At each node in the pyramid, the split parameter is a transformation of a GP evaluated at the covariate, ensuring non-crossing by construction (An et al., 2023).

Spline and Tensor Product Models

The spline-based approach expands either the quantile or the distribution function in a tensor-product B-spline basis over quantile level and covariate space (Das et al., 2016). Monotonicity is enforced via constraints (e.g., requiring increments of the basis coefficients to form simplices), with block-Dirichlet priors on increments. Group structure is handled via independent priors on each subpopulation or via hierarchical priors on shared parameters.

Dirichlet Process Mixtures and Dependent Dirichlet Processes

Distributional models using DP or dependent DP mixtures specify the conditional law as a mixture:

$$f(y \mid x) = \int K(y \mid \theta) \, G_x(d\theta)$$

where $G_x$ is a DP or DDP indexed by covariates. Dependence on covariates is introduced via smoothly varying weights or atoms (often modeled as Gaussian process functions of $x$) (Bhattacharya et al., 2020). Hierarchical structure arises by allowing group-specific DPs to be linked via a common base or hyperprior distribution.

3. Posterior Computation and Algorithms

Posterior computation relies on high-dimensional MCMC, often with specialized blocked Gibbs or adaptive Metropolis-Hastings updates tailored to the enriched, constrained parameter spaces.

Quantile Pyramid and Spline-based Models

• MCMC procedures alternate sampling of pyramid sticks or B-spline increments (subject to monotonicity constraints), centering parameters (mean and variance for transformations to the data scale), and group-level hyperparameters (Rodrigues et al., 2016, Das et al., 2016).
• Blocking and adaptive covariance estimation (e.g., from pilot runs) are leveraged for efficiency, with log-difference reparameterizations to improve mixing in high-dimensional settings (Rodrigues et al., 2017).
• For tensor-product models, block-updating over entire simplex blocks allows efficient exploration of the simplex-constrained coefficient space, with acceptance rates controlled by tuning the multiplicative weights in the proposal (Das et al., 2016).

DP Mixture Models

• Blocked Gibbs sampling is enabled via the stick-breaking construction. Latent cluster labels are introduced for both mixture components and subgroups.
• When DDPs are present, atoms or weights are updated via conjugate updates exploiting GP priors, along with stick-breaking weights conditioned on allocation variables (Bhattacharya et al., 2020).
• Truncation of the stick-breaking representation or the number of pyramid levels is handled adaptively, with suggested accelerations including slice sampling, parallelization, or retrospective truncation (Bhattacharya et al., 2020).

Computational complexity scales as $O(nT)$ for methods fitting $T$ quantiles to $n$ data points. Careful design is required in high-dimensional covariate settings, where convex hull or knot selection for splines and pivotal points for pyramids are optimized for stability and computational tractability (Rodrigues et al., 2017).

4. Properties: Monotonicity, Noncrossing, and Consistency

A primary concern is the joint behavior of multiple quantile curves with respect to monotonicity (in $\tau$) and noncrossing (across covariates or groups).

• Noncrossing enforcement: Pyramid-based and B-spline tensor-product models guarantee noncrossing for all $\tau$ by construction, via recursive partitioning (pyramids) or ordering constraints on spline coefficients (Rodrigues et al., 2016, An et al., 2023, Das et al., 2016). For pyramid regression in high dimensions, convex-hull arguments ensure noncrossing across the relevant predictor space (Rodrigues et al., 2017).
• Posterior support and consistency: Quantile pyramid models (at sufficient depth) and DP mixtures possess full Kullback–Leibler support under mild regularity, enabling posterior consistency for the quantile functionals (Rodrigues et al., 2016, Bhattacharya et al., 2020). Rate results and Hellinger strong consistency are available for models with increasing complexity as sample size grows (Rodrigues et al., 2016).

A summary of key monotonicity mechanisms is as follows:

Model Class	Enforces Monotonicity By	Noncrossing Guarantee
Quantile pyramids, DQP	Pyramid recursion	Yes, for all $\tau$ by design
Spline (tensor-product) approaches	Ordered simplex constraints	Yes, for all $\tau$, all $x$
DP/DDP mixtures	Post hoc in quantile output	No, unless extra joint modeling

5. Hierarchical Extensions and Grouped Data Modeling

Nonparametric hierarchical Bayesian quantile models enable flexible shrinkage and information pooling for quantile estimation in multilevel or grouped settings:

• Hierarchical quantile modeling: In the geometric-measure approach, subpopulations are modeled via group-specific quantiles linked through a mixing distribution, with the mixing weights themselves endowed with a nonparametric Dirichlet prior, supporting data-driven shrinkage and uncertainty quantification (Bornn et al., 2016).
• Group-adaptive pyramids/splines: For grouped or longitudinal data, one can place independent or partially pooled quantile pyramids, possibly sharing hyperparameters or covariate processes, to support borrowing across subgroups without imposing strict parametric structure (Rodrigues et al., 2016, An et al., 2023).
• Hierarchical structure for censored/noisy data: Explicit samplers are provided for models where observations may be right-censored or reported in bins/quantile intervals, with group-specific Dirichlet priors easily adapted to truncated sampling schemes (Bornn et al., 2016, Das et al., 2016).

This hierarchy permits robust inference in settings with variable sample sizes across groups and substantial skew or censoring, with reduction of extreme quantile variability for underrepresented groups through Bayesian shrinkage.

6. Applications, Practical Considerations, and Empirical Performance

Methods reviewed have been applied to problems including hurricane wind speeds, sports performance metrics, AIDS patient outcomes, and extreme value analysis. Key findings:

• Empirical superiority in tail estimation: Pyramid and tensor-product quantile models display improved root mean squared error and credible interval coverage relative to frequentist and semiparametric Bayesian competitors, particularly at extreme quantiles (Rodrigues et al., 2016, Rodrigues et al., 2017).
• Stability under censoring and sparse data: Hierarchical models stabilize estimates for underrepresented or censored subpopulations, adjusting uncertainty to reflect both observed and latent data structures (Bornn et al., 2016, Das et al., 2016).
• Scalability considerations: Block-adaptive MCMC, log-difference parameterizations, and Metropolis-within-Gibbs procedures are critical for efficient posterior computation, but complex hierarchical models remain computationally intensive for large $n$ or high-dimensional covariates (Rodrigues et al., 2017).

Hyperparameter selection (such as Beta or Dirichlet concentration, pyramid depth, spline knot placement) is important for balancing flexibility and overfitting; recommended guidelines are documented in individual works (Rodrigues et al., 2016, Bornn et al., 2016, Rodrigues et al., 2017).

7. Extensions and Ongoing Developments

Active areas of methodological and theoretical extension include:

• Multivariate quantile functionals and depth-based indices: Beyond geometric quantiles, alternative notions (directional, depth-based) can be incorporated as the primary target of the nonparametric prior (Bhattacharya et al., 2020).
• Higher-dimensional covariate and response spaces: Product GPs, multivariate kernels, and multi-output quantile pyramids (for vector-valued $Y$) are developed to accommodate complex dependencies (Bhattacharya et al., 2020, An et al., 2023).
• Computational acceleration: Variational Bayes, GPU-based MCMC, and efficient block-sampling strategies are areas of current exploration (Bhattacharya et al., 2020).
• Posterior contraction rates and uniform consistency: Theoretical work continues to extend contraction and uniformity results to more general settings, including high-dimensional predictors and multivariate responses (Rodrigues et al., 2016, Bhattacharya et al., 2020).

Nonparametric hierarchical Bayesian quantile modeling thus provides a flexible, theoretically grounded, and empirically validated toolkit for conditional quantile inference under complex data-generating regimes, maintaining simultaneous estimation, monotonicity, noncrossing, and adaptive shrinkage across grouped or high-dimensional structures (Rodrigues et al., 2016, An et al., 2023, Bhattacharya et al., 2020, Bornn et al., 2016, Rodrigues et al., 2017, Das et al., 2016, Kobayashi et al., 2016).