Bayesian Quantile Regression
Also known as: BQR, Bayesian quantile regression model, asymmetric Laplace Bayesian regression, posterior quantile regression
Bayesian Quantile Regression estimates the full posterior distribution of regression coefficients at any chosen quantile of the outcome. By combining the asymmetric Laplace likelihood with prior distributions over the coefficients, it delivers uncertainty-quantified estimates of conditional quantiles — such as the median, the 10th, or the 90th percentile — without assuming Gaussian errors.
Key highlights
- Estimates any conditional quantile of the outcome, making it ideal for studying distributional tails rather than only the mean.
- Full posterior inference provides credible intervals and predictive distributions without relying on large-sample normality.
- Robust to outliers and non-Gaussian errors: the asymmetric Laplace likelihood down-weights extreme residuals automatically.
- Accommodates prior knowledge and regularisation through the prior distribution, useful in small or high-dimensional samples.
- Can be extended naturally to censored data, longitudinal data, and nonparametric spline terms within a Bayesian framework.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Bayesian Quantile Regression when (1) the outcome distribution is skewed, heavy-tailed, or heteroscedastic so that mean regression misses important features; (2) you care about extreme quantiles such as risk thresholds, top/bottom performers, or value-at-risk; (3) you want formal uncertainty quantification — credible intervals — rather than asymptotic standard errors; or (4) you have prior knowledge to encode through informative priors or need regularisation in high-dimensional settings. Do not use it when the mean is the only target, when the sample is very large and speed matters (MCMC is slower than frequentist QR), or when your outcome is binary or nominal (use Bayesian logistic or multinomial regression instead). Avoid it when the quantile level τ is far into the tail (< 0.05 or > 0.95) and the sample is small — extreme quantile posteriors can be highly sensitive to the prior in that regime.
Strengths & limitations
- Estimates any conditional quantile of the outcome, making it ideal for studying distributional tails rather than only the mean.
- Full posterior inference provides credible intervals and predictive distributions without relying on large-sample normality.
- Robust to outliers and non-Gaussian errors: the asymmetric Laplace likelihood down-weights extreme residuals automatically.
- Accommodates prior knowledge and regularisation through the prior distribution, useful in small or high-dimensional samples.
- Can be extended naturally to censored data, longitudinal data, and nonparametric spline terms within a Bayesian framework.
- Computationally expensive: MCMC sampling is orders of magnitude slower than the linear-programming solution of classical quantile regression.
- The asymmetric Laplace likelihood is a working device, not the true data-generating distribution; posterior credible intervals are approximately valid but not exact.
- Results depend on the choice of prior, particularly for extreme quantiles or small samples — sensitivity analysis is essential.
- Requires careful MCMC diagnostics (R-hat, effective sample size) to confirm convergence, adding analyst burden.
- Simultaneous estimation at multiple quantiles does not automatically guarantee crossing-free quantile curves without additional constraints.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
Is Bayesian Quantile Regression the same as classical quantile regression?
They target the same estimand — conditional quantiles of y given X — but differ in inference. Classical quantile regression solves a linear programme and uses asymptotic standard errors. Bayesian quantile regression uses MCMC to obtain full posterior distributions, enabling exact finite-sample uncertainty quantification and prior regularisation.
Why is the asymmetric Laplace distribution used as the likelihood?
Maximising the asymmetric Laplace log-likelihood at quantile τ is mathematically equivalent to minimising the check-function (pinball) loss that defines the τ-th quantile. This equivalence, established by Yu and Moyeed (2001), allows the quantile estimation problem to be embedded in a proper Bayesian model.
Can I estimate multiple quantiles jointly?
Yes. You can fit separate Bayesian QR models for each τ independently, or use joint quantile regression models (e.g., via simultaneous quantile regression in Stan) that enforce non-crossing constraints and share information across quantile levels.
How many observations do I need?
More than in classical regression, because extreme quantiles are identified by fewer observations. A rough minimum is at least 50–100 observations per predictor for central quantiles (τ near 0.5), and substantially more for extreme quantiles such as τ = 0.05 or 0.95.
Which software implements Bayesian Quantile Regression?
In R: the bayesQR package provides dedicated Gibbs samplers; brms with family = asym_laplace() uses Stan under the hood for flexible model specifications. In Python: PyMC supports custom asymmetric Laplace likelihoods. Stan can implement any formulation directly.
Sources
- 1.Kozumi, H., & Kobayashi, G. (2011). Gibbs sampling methods for Bayesian quantile regression. Journal of Statistical Computation and Simulation, 81(11), 1565–1578.
- 2.Yu, K., & Zhang, J. (2005). A three-parameter asymmetric Laplace distribution and its extension. Communications in Statistics – Theory and Methods, 34(9–10), 1867–1879.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Bayesian Quantile Regression. ScholarGate. https://scholargate.app/statistics/bayesian-quantile-regression