Quantile Regression

Definition

Quantile regression (Koenker-Bassett 1978) models the $q$ -th conditional quantile of a response $y$ given covariates $x$ directly, rather than modelling the conditional mean. For quantile $q \in (0,1)$ , the estimator solves $\min_\beta \sum_t \rho_q(y_t - x_t'\beta)$ where $\rho_q(u) = u(q - \mathbb{1}[u<0])$ is the check function. This requires no distributional assumption and allows different parts of the conditional distribution to respond differently to covariates (heterogeneous effects).

Key Ideas

Check function loss: $\rho_q(u) = qu \mathbb{1}[u \geq 0] + (1-q)|u|\mathbb{1}[u < 0]$ . Minimizing its expectation yields the $q$ -th population quantile; the sample analogue solves a linear program.
Coefficient interpretation: $\beta(q)$ measures the marginal effect of $x$ on the $q$ -th conditional quantile — unlike ordinary least squares (OLS) which averages effects, quantile regression allows effects to be strong in tails and weak at the center, or vice versa.
Asymptotic theory: Under mild regularity, $\sqrt{T}(\hat\beta(q) - \beta^*(q)) \to \mathcal{N}(0, q(1-q)f_u^{-2}(0)(X'X)^{-1}X'\Sigma X (X'X)^{-1})$ where $f_u(0)$ is the density of residuals at zero. Standard errors require density estimation at zero.
Monotonicity problem: With multiple quantiles estimated separately, the fitted quantile curves $x'\hat\beta(q)$ can cross in finite samples (non-monotone in $q$ at given $x$ ), violating the definition of a cumulative distribution function (CDF). This is a fundamental inconsistency of the frequentist approach at finite $T$ .
Limitation: Provides estimates at a finite grid of quantile indices; does not recover the full conditional distribution $p(y|x)$ from which any functional can be computed.

How It Works

Classical Frequentist Approach

The check function minimization is equivalent to a linear program and can be solved exactly for any sample size. Bootstrap standard errors are standard. The Angrist-Chernozhukov-Fernandez-Val (2004) framework analyzes quantile regression under misspecification, showing the estimator recovers best linear projections for each quantile.

Bayesian Alternative — Smoothly Mixing Regressions

Geweke-Keane (2007) propose the Smoothly Mixing Regressions (SMR) model as a Bayesian counterpart that:

Recovers the full $p(y|x)$ : From Markov Chain Monte Carlo (MCMC) output, posterior quantiles are computed by embedding $P(y \leq c | x, \text{data})$ in root-finding.
Guarantees monotonicity: Since posterior quantiles are derived from a proper CDF at each $x$ , they are monotone in $q$ by construction.
Provides uncertainty quantification: Posterior distributions over the entire quantile surface, not just point estimates.
Out-of-sample superiority: For S&P 500 daily returns, SMR achieves a predictive log-likelihood 57 points higher than a t-generalized autoregressive conditional heteroskedasticity (t-GARCH)(1,1) model.

Applied Quantile Regression — Earnings Example

Buchinsky (1994) studied U.S. wage structure 1963–1987 using quantile regression, documenting that returns to education and experience differ substantially across the earnings distribution — effects are stronger at higher quantiles. Geweke-Keane (2007) revisit a similar question (1993 Panel Study of Income Dynamics (PSID) cross-section) and find with SMR: earnings quantiles almost always increase in education; age profiles vary sharply by quantile; dispersion is roughly constant for men ages 25–55 with 10–15 years of education, then rises sharply.

Why It Matters

Quantile regression is the standard tool for documenting heterogeneous treatment effects and distributional changes over time or across groups (rising returns to education, wage inequality changes).
The monotonicity problem motivates Bayesian full-distribution approaches such as SMR, which recover quantiles as a by-product of a well-defined $p(y|x)$ model.
Access to the full conditional distribution enables additional functionals beyond quantiles: tail probabilities, Gini coefficients, option prices, poverty headcount measures.

Open Questions

Nonparametric Bayesian approaches (Dirichlet process) can model $p(y|x)$ without specifying the number of mixture components, at the cost of harder computation.
Extending SMR and quantile regression to multivariate outcomes raises the curse of dimensionality in covariate specification.
Panel quantile regression (allowing individual fixed effects) remains an active area without a settled asymptotic theory.