Uniform Inference Across Quantile Levels

Why Inference Must Account for a Range of Quantiles

A quantile-regression coefficient is indexed by the quantile level. For example, the association between a predictor and the conditional 25th percentile of an outcome need not equal its association with the conditional median. When the scientific question concerns a range of quantiles, inference at each quantile separately is not enough: a collection of pointwise intervals does not generally have the advertised coverage for the whole curve.

Uniform inference treats the estimated coefficient as a process over a prespecified interval of quantile levels. It provides a way to construct a band that covers the selected coefficient function across that interval with a stated asymptotic probability. The target here is the quantile-regression coefficient function itself, not a causal effect.

Definition: Let $Y_i$ be an outcome and $X_i$ a $p$-vector of regressors, including any intercept. On a quantile interval $\mathcal T=[\tau_L,\tau_U]$, where $0<\tau_L<\tau_U<1$, suppose the conditional quantile is linear: $$ Q_{Y\mid X}(\tau\mid X_i)=X_i'\beta(\tau), \qquad \tau\in\mathcal T. $$ The vector-valued function $\beta(\tau)$ is the quantile-regression coefficient process. Its estimator $\hat\beta(\tau)$ is computed at each $\tau$ using the quantile-regression check-loss criterion.

A simultaneous band concerns a specified component, or specified collection of components, of this process. For a single coefficient $j$, write $\beta_j(\tau)$ and its estimator $\hat\beta_j(\tau)$. The interval $\mathcal T$, the coefficient or coefficients of interest, and the desired confidence level should be chosen before interpreting the band.

Conditions for Uniform Approximation

The standard process approximation applies under regularity conditions stronger than those needed for inference at one fixed quantile. A useful regular case has independent and identically distributed observations; a correctly specified linear conditional quantile over $\mathcal T$; and a conditional density that is continuous and positive near the relevant conditional quantiles. The interval stays away from 0 and 1. Regressors have suitable finite moments, the matrix below is nonsingular uniformly over $\mathcal T$, and empirical-process conditions ensure the score process behaves regularly.

Define the conditional density at the quantile and the matrix $J(\tau)=E[f_{Y\mid X}(X'\beta(\tau)\mid X)XX']$. The conditional density must be well behaved uniformly over the quantile range, and the eigenvalues of $J(\tau)$ must remain bounded away from zero. These conditions exclude, for this standard approximation, irregular cases such as an atom at a target quantile. They also ensure the coefficient is locally identified and the sandwich standard errors can be consistently estimated.

Under these conditions, the estimator admits a uniform asymptotic linear representation. In the formula, $I_i(\tau)=1\{Y_i\leq X_i'\beta(\tau)\}$, and $J(\tau)^{-1}$ maps the quantile score into coefficient units.

\sqrt n\{\hat\beta(\tau)-\beta(\tau)\} = J(\tau)^{-1}\frac{1}{\sqrt n}\sum_{i=1}^n X_i\{\tau-I_i(\tau)\}+o_p(1), \qquad \text{uniformly for }\tau\in\mathcal T.

“Uniformly” means that the approximation error is negligible over the entire interval, rather than only at one fixed $\tau$. The leading term converges to a mean-zero Gaussian process. Its covariance links different quantile levels: estimates at two levels are generally dependent because they are calculated from the same observations.

\operatorname{Cov}\{G(\tau),G(s)\} = J(\tau)^{-1} E[XX'\{\min(\tau,s)-\tau s\}] J(s)^{-1}, \qquad \tau,s\in\mathcal T.

From the Process to a Simultaneous Band

For a selected coefficient $j$, let $e_j$ be the vector with 1 in position $j$ and 0 elsewhere. The asymptotic variance of its estimate at $\tau$ is $\sigma_j^2(\tau)=e_j'\operatorname{Cov}\{G(\tau),G(\tau)\}e_j$. An estimated standard error for $\hat\beta_j(\tau)$ is denoted $\widehat{SE}_j(\tau)$; it estimates the standard deviation of the coefficient estimator, not of its $\sqrt n$-scaled version.

A simultaneous band uses a critical value for the largest absolute studentized error over the interval. In practice, a multiplier procedure estimates the distribution of that supremum. It uses estimated influence contributions, multiplies each observation's contribution by an independent mean-zero, variance-one random weight, and recomputes the maximum over quantile levels. Repeating this many times gives a simulated reference distribution. The chosen critical value is its desired quantile.

T_n=\sup_{\tau\in\mathcal T} \left| \frac{\sqrt n\{\hat\beta_j(\tau)-\beta_j(\tau)\}} {\widehat{\sigma}_j(\tau)} \right|, \qquad \widehat{\sigma}_j(\tau)=\sqrt n\,\widehat{SE}_j(\tau).

If $c_{1-\alpha}$ is the estimated $1-\alpha$ quantile of the supremum reference distribution, the band is $\hat\beta_j(\tau)\pm c_{1-\alpha}\widehat{SE}_j(\tau)$. Its asymptotic interpretation is that the probability the entire band contains the coefficient function on $\mathcal T$ approaches $1-\alpha$, under the stated model and regularity conditions. With a numerical grid, the maximum is computed on that grid; reliable approximation to a continuum band requires a sufficiently fine grid and suitable process regularity.

Theorem: Under the regularity conditions for the uniform linear representation and consistent estimation of its covariance process, the studentized coefficient process converges to a Gaussian process on $\mathcal T$. If the critical value consistently estimates the $1-\alpha$ quantile of its supremum absolute value, then the resulting band has asymptotic simultaneous coverage $1-\alpha$ over $\mathcal T$.

Worked Bands and Their Interpretation

Worked Example: A Simultaneous Band for One Coefficient

Suppose a hypothetical wage analysis estimates the coefficient on years of education over $\mathcal T=[0.25,0.75]$. The coefficient is measured in thousands of dollars per additional year of education. At the median, suppose $\hat\beta_j(0.50)=0.150$ and $\widehat{SE}_j(0.50)=0.045$. Assume a multiplier procedure for the full interval gives $c_{0.95}=2.35$.

The median-level slice of the simultaneous 95% band is $$ 0.150\pm 2.35(0.045) =0.150\pm0.10575. $$ Rounding the lower endpoint to four decimal places using ties-to-even gives $0.0442$, and the upper endpoint is $0.2558$. Thus the slice is $[0.0442,\,0.2558]$ thousand dollars per year of education.

This endpoint is wider than a pointwise 95% interval using 1.96: $0.150\pm1.96(0.045)= [0.0618,\,0.2382]$. The pointwise interval targets coverage at $\tau=0.50$ alone. The band uses a larger critical value to account for searching across the entire quantile interval, and its simultaneous coverage statement applies to the whole coefficient curve.

Worked Example: Covering Two Coefficients Together

In a hypothetical analysis, two coefficients are selected in advance: the coefficient on education, measured in thousands of dollars per year, and the coefficient on experience, measured in thousands of dollars per year. At a particular quantile, their estimates are $1.20$ and $-0.40$, with standard errors $0.30$ and $0.12$. Suppose a multiplier procedure based on the maximum over both coefficients and the quantile interval gives a 95% critical value of 2.70.

For education, the margin is $2.70(0.30)=0.810$, so the interval at this quantile is $[1.20-0.810,\,1.20+0.810]=[0.390,\,2.010]$ thousand dollars per year. For experience, the margin is $2.70(0.12)=0.324$, so the interval is $[-0.40-0.324,\,-0.40+0.324]=[-0.724,\,-0.076]$ thousand dollars per year.

These intervals are slices of a joint band only if the critical value was computed for the maximum over both selected coefficients and all quantile levels in the target range. Using separate coefficient-specific critical values would not, in general, give 95% simultaneous coverage for both curves together.

Worked Example: A Band at an Interior Quantile

Suppose a hypothetical regression estimates a coefficient on a binary indicator, measured in outcome units, at $\tau=0.40$. The estimate is $0.320$ and its standard error is $0.070$. A multiplier procedure over the prespecified interval produces a simultaneous 95% critical value of 2.45. The margin is $2.45(0.070)=0.1715$.

The band slice at this quantile is $[0.320-0.1715,\,0.320+0.1715]=[0.1485,\,0.4915]$ outcome units. The interpretation is that this is the value of the simultaneous band at the 40th conditional quantile. Under the assumptions and procedure used, the band is designed to cover the selected coefficient function over the whole target interval with approximately 95% probability in repeated samples. It does not state that there is a 95% probability that this fixed interval contains the parameter.

Why Uniform Inference Matters

A researcher may inspect many quantile-specific estimates and then emphasize whichever level appears most striking. Pointwise intervals do not account for this across-level search. A simultaneous band addresses the broader question: which coefficient functions remain compatible with the data over the prespecified range, while controlling the probability of missing any point on that range at the nominal asymptotic rate.

The band is only as meaningful as its target and assumptions. Report the quantile interval, the selected coefficient or coefficients, the standard-error method, and how the supremum critical value was obtained. If inference is performed on a finite grid, state the grid; a grid band is not automatically a continuum band. The stated coverage is asymptotic, so finite-sample performance can differ, particularly with small samples, sparse tail observations, or an unstable density estimate.

Common Mistake

Wrong reasoning: “Every pointwise 95% interval covers its own coefficient, so the intervals cover the entire coefficient curve with 95% probability.”

Correction: Each pointwise statement concerns one fixed quantile. Across several quantiles, the chance that every interval covers can be lower than 95%; dependence across quantiles does not remove this multiplicity issue. A uniform band instead calibrates a critical value using the maximum studentized deviation over the chosen quantile range (and over all selected coefficients, if joint coverage is desired).

Key takeaway: Uniform inference is inference about a coefficient process, not a stack of unrelated pointwise intervals. Define the quantile range and coefficients of interest, use a supremum-based critical value, and interpret coverage as applying simultaneously over that target.

Check Your Understanding

Try these questions without looking back.

In one sentence, distinguish a pointwise confidence interval at a fixed quantile from a simultaneous band over a quantile interval.
Why does the covariance of coefficient estimates at two quantile levels generally involve both $\tau$ and $s$?
A coefficient estimate is 0.80, its standard error is 0.10, and a simultaneous critical value is 2.50. Calculate the band slice at that quantile and state its units if the coefficient is measured in percentage points.
What additional maximum must be included when the goal is simultaneous coverage for two coefficient curves over the same quantile range?
Why does a band computed on a finite grid require care before it is described as covering a continuum of quantile levels?