Diagnosing And Repairing Quantile Crossing

Why Crossing Matters

A conditional quantile curve describes an ordered part of an outcome distribution at each covariate value. If \(\tau_1<\tau_2\), the conditional \(\tau_1\)-quantile cannot exceed the conditional \(\tau_2\)-quantile at the same covariates. Yet separately estimated quantile regression curves can violate this ordering. Such quantile crossing is not just a visual inconvenience: it makes the fitted curves incompatible with a single conditional distribution.

Crossing can arise because each quantile is estimated separately, with its own sampling variation and fitted coefficient vector. It can also reflect a restrictive model form: separate linear approximations may fit different parts of a conditional distribution but fail to preserve their order throughout the covariate range. The diagnostic question is therefore not simply whether two curves cross somewhere on a plot. It is whether, and where, the estimated quantiles violate the required ordering over the covariate domain relevant to the analysis.

For linear quantile regression, the estimated curve at level \(\tau\) is commonly written \(\widehat Q(\tau\mid x)=x'\widehat\beta_\tau\), where \(x\) is the chosen regressor vector, including an intercept if the model has one. This is a model for a conditional quantile, not a claim that changes in \(x\) cause changes in the outcome. The basic crossing condition for two fitted levels is immediate:

The left-hand condition specifies the required order of the quantile levels; the right-hand condition says the estimated lower quantile is numerically above the estimated higher quantile at that covariate value. Equality is not crossing, though it can indicate that fitted curves touch.

Diagnosing Crossing Over a Covariate Domain

A useful diagnosis starts by naming the range of covariates where the fitted model will be interpreted or used. Checking only the sample mean of a regressor can miss violations elsewhere; checking far outside the observed support can instead emphasize extrapolations with little empirical basis. For each ordered pair of fitted quantile levels, calculate the difference \(D(x)=\widehat Q(\tau_2\mid x)-\widehat Q(\tau_1\mid x)\). A negative value is a crossing, zero is a touch, and a positive value is correctly ordered.

In practice, evaluate these differences over a grid of covariate values covering the target domain, or over a suitable grid of design points when there are several regressors. Report more than a yes-or-no flag when possible: the region with negative differences, the most negative difference, and the proportion of grid points with a violation help describe the severity and location of crossing. Grid summaries are diagnostics, not universal guarantees. A fine grid can reveal problems between observed covariate values, but it does not prove that a nonlinear fitted curve is ordered at every point between grid values.

For a linear model with a single regressor, the difference between two fitted quantiles is also linear, so solving for its zero can locate a crossing boundary directly. For multiple regressors, a grid or domain-specific optimization can map the region. If the covariate domain is a compact polytope and the difference is affine, checking its values at the domain's vertices is sufficient to verify its sign throughout that polytope: every point in the polytope is a convex combination of vertices, and an affine function at that point is the same combination of its vertex values. That shortcut does not apply to general nonlinear differences.

Repair 1: Monotone Rearrangement

Monotone rearrangement repairs a set of fitted values by sorting them in increasing order across quantile levels at each fixed covariate value. For a discrete ordered grid of levels \(\tau_1<\cdots<\tau_J\), calculate \(\widehat Q(\tau_j\mid x)\) for all \(j\), sort those \(J\) values, and assign the smallest to \(\tau_1\), the next-smallest to \(\tau_2\), and so on. The result is nondecreasing on the grid at that \(x\). Repeat separately at each covariate value where predictions are needed.

For a continuous range of quantile levels, the increasing rearrangement is the nondecreasing function obtained by sorting the values of the original curve over that range, with respect to a specified weighting over quantile levels. It is not generally the same as replacing each crossed pair by its average. Rearrangement changes the assignment of predicted values to quantile levels, while preserving the distribution of those values across the rearranged range. Its pointwise implementation does not generally correspond to one common vector of linear regression coefficients across all covariate values.

Repair 2: Constrained Quantile Estimation

Constrained estimation addresses ordering during estimation rather than sorting predictions afterward. For levels \(\tau_1<\cdots<\tau_J\), one standard approach jointly estimates coefficient vectors by minimizing the sum of the quantile check losses, subject to noncrossing inequalities. The check loss is \(\rho_\tau(u)=u\{\tau-\mathbf{1}(u<0)\}\), where \(u\) is a residual. With observed outcomes \(Y_i\) and regressors \(x_i\), a sample-design constraint is \(x_i'\beta_{\tau_j}\leq x_i'\beta_{\tau_{j+1}}\) for each observation \(i\) and each adjacent pair of levels.

The constraints here guarantee ordering at the observed design points. They do not necessarily guarantee ordering at every point in a broader prediction domain. To guarantee ordering over a specified domain, impose the inequalities there, using a finite set of constraints when appropriate. For linear fitted quantiles on a compact polytope, it is enough to impose each adjacent-level inequality at every vertex of the polytope. The inequalities are linear in the coefficients, and the check-loss objective is convex and piecewise linear, so this formulation can be represented as a constrained optimization problem.

Choosing and Interpreting a Repair

The two approaches answer related but distinct questions. Rearrangement is a direct way to obtain ordered predictions from fitted curves already in hand. It is convenient when prediction at many covariate values is the main goal, but it may not preserve the original regression form or the original sample fitting criterion. Constrained estimation builds the ordering requirement into a joint fit and can retain a specified parametric form, but it changes the optimization problem and may move several fitted curves to satisfy the restrictions.

Whichever method is used, state the domain where ordering is intended, which quantile levels are included, and where the guarantee applies. Inference should also reflect the estimator actually used: uncertainty calculations for separately fitted curves do not automatically describe the sampling behavior of rearranged or constrained estimates. The aim is not to conceal a misspecified model, but to produce coherent quantile predictions and make the repair transparent.

Common Mistakes

• Checking only at the average covariate value. Correct ordering at one point does not imply ordering across the analysis domain. Evaluate the full relevant range or use a valid domain-wide argument.
• Treating touching curves as crossing. Equality satisfies the nondecreasing requirement. A strict reversal, not a tie, is a crossing.
• Assuming a grid proves global order. A grid diagnoses the points it contains. For a general nonlinear difference, it cannot rule out a violation between grid values.
• Calling rearrangement a constrained quantile regression refit. Sorting predictions does not minimize the original check-loss objective under constraints. State clearly which repair produced the reported estimates.
• Assuming sample-point constraints guarantee order everywhere. They guarantee order only where imposed, unless additional structure justifies a broader conclusion.