Quantile Regression: Beyond the Conditional Mean

Ordinary regression describes the average. Quantile regression describes other points in the ranked data, such as the median or the 90th percentile, so you can see whether a relationship changes between low and high outcomes. This guide gives a worked example on real public data, then the steps, assumptions and pitfalls.

Worked Example (Real Data): Household Food Spending in Belgium

Question: Does food spending rise with income at the same rate for low spenders and high spenders? A percentile marks a point in ranked data: the 90th percentile is the level that 90% of households are below.

Data: The Engel data in the R package quantreg: 235 Belgian working-class households, with annual income and annual food expenditure, both in Belgian francs (quantreg manual, "engel"). The data were used in Koenker and Bassett (1982, Econometrica 50, 43-61). Quantile regression itself was introduced in Koenker and Bassett (1978, Econometrica 46, 33-50).

Method: Fit a straight line of food spending on income at the 10th, 50th and 90th percentiles, and fit ordinary least squares (OLS) for comparison. OLS fits the average. The standard errors come from a bootstrap with 1,000 resamples: resample the rows with replacement, refit, and see how much the slope moves. The interval is the slope plus or minus 1.96 standard errors, which is a normal approximation.

library(quantreg)
data(engel)
cat("Rows:", nrow(engel), "\n")
set.seed(417)
ols <- lm(foodexp ~ income, data = engel)
b0 <- coef(ols)
cat(sprintf("OLS: intercept=%.2f slope=%.4f\n", b0[1], b0[2]))
for (tau in c(0.10, 0.50, 0.90)) {
  f <- rq(foodexp ~ income, tau = tau, data = engel)
  b <- coef(f)
  cat(sprintf("tau=%.2f: intercept=%.2f slope=%.4f prediction at income 1000=%.1f\n",
              tau, b[1], b[2], b[1] + b[2] * 1000))
}
cat(sprintf("OLS prediction at income 1000=%.1f\n", b0[1] + b0[2] * 1000))
for (tau in c(0.10, 0.50, 0.90)) {
  f <- rq(foodexp ~ income, tau = tau, data = engel)
  s <- summary(f, se = "boot", R = 1000)
  ci <- s$coefficients[2, 1] + c(-1, 1) * 1.96 * s$coefficients[2, 2]
  cat(sprintf("tau=%.2f slope=%.4f boot SE=%.4f interval %.4f to %.4f\n",
              tau, s$coefficients[2, 1], s$coefficients[2, 2], ci[1], ci[2]))
}

Printed output (R with the quantreg package, run on 2026-11-02):

Rows: 235
OLS: intercept=147.48 slope=0.4852
tau=0.10: intercept=110.14 slope=0.4018 prediction at income 1000=511.9
tau=0.50: intercept=81.48 slope=0.5602 prediction at income 1000=641.7
tau=0.90: intercept=67.35 slope=0.6863 prediction at income 1000=753.7
OLS prediction at income 1000=632.7
tau=0.10 slope=0.4018 boot SE=0.0470 interval 0.3097 to 0.4938
tau=0.50 slope=0.5602 boot SE=0.0338 interval 0.4940 to 0.6264
tau=0.90 slope=0.6863 boot SE=0.0258 interval 0.6358 to 0.7368

Result: Among these households, each extra franc of income goes with about 0.40 francs more food spending at the 10th percentile and about 0.69 francs more at the 90th. The OLS slope of 0.4852 sits between them and describes neither group well. The 10th-percentile and 90th-percentile intervals do not overlap, so the difference in slopes is unlikely to be sampling noise alone. This is a rough check, not a formal test of equal slopes.

Limits: This is one old, small data set, so do not carry the numbers to other households or years. The fit describes association, not cause. The bootstrap interval is a normal approximation; the quantreg package offers other methods. Predictions are only meaningful inside the observed income range (377.1 to 4,957.8 francs).

Second Example (Simulated): Median Versus 90th Percentile With a Known Truth

Question: In made-up weekly demand data, is the fitted value at the 90th percentile higher than the median at the same input? A percentile marks a point in ranked data.

Inputs: Made-up data, seed 417. The code makes 500 rows. It uses x as the input and y as the result. It sets y to a base value plus an x effect and random noise. The noise spread grows as x grows.

Method: Fit a line at the median and at the 90th percentile. Quantile regression fits a chosen point in ranked data. The code uses a linear program, a set of steps to solve a math task. It cuts pinball loss, a score for errors at a chosen point.

import numpy as np
from scipy.optimize import linprog

SEED = 417
rng = np.random.default_rng(SEED)
n = 500
x = rng.uniform(0, 10, n)
noise = rng.normal(0, 1 + 0.7 * x, n)
y = 20 + 2 * x + noise

X = np.column_stack((np.ones(n), x))

def fit_quantile(X, y, tau):
    n_obs, n_coef = X.shape
    # Variables are beta, positive residual parts, and negative residual parts.
    objective = np.r_[np.zeros(n_coef), np.full(n_obs, tau),
                      np.full(n_obs, 1 - tau)]
    equality = np.column_stack((X, np.eye(n_obs), -np.eye(n_obs)))
    result = linprog(
        objective,
        A_eq=equality,
        b_eq=y,
        bounds=[(None, None)] * n_coef + [(0, None)] * (2 * n_obs),
        method="highs",
    )
    if not result.success:
        raise RuntimeError(result.message)
    return result.x[:n_coef]

print(f"Simulated data, seed {SEED}")
print(f"Rows: {n}")
for tau in (0.50, 0.90):
    beta = fit_quantile(X, y, tau)
    pred = beta[0] + beta[1] * 8
    print(f"Quantile {tau:.2f}: intercept={beta[0]:.3f}, "
          f"slope={beta[1]:.3f}, prediction at x=8={pred:.3f}")

Printed output (Python 3 with numpy and scipy, run on 2026-11-02):

Simulated data, seed 417
Rows: 500
Quantile 0.50: intercept=19.909, slope=1.958, prediction at x=8=35.574
Quantile 0.90: intercept=21.709, slope=2.802, prediction at x=8=44.125

Result: At x=8, the fitted median is 35.574. The fitted 90th percentile is 44.125. The true slope in the simulation is 2 for the average, and the noise grows with x, so the upper fit rises faster than the median fit. That is the pattern the simulation was built to have.

Limit: The data are made up and show only how the method behaves on a known pattern. They do not tell us about real stores. The fit does not show that a change in x causes demand to change.

What is Quantile Regression?

Quantile regression is a statistical technique that extends traditional regression analysis by estimating conditional quantiles of the response variable, rather than just the conditional mean. It can help when you need to understand variability and compare scenarios.

The examples above show how fitted values can differ across quantiles. To see how much your own drivers explain an outcome, and which matter most, try the free linear regression tool.

Ordinary least squares (OLS) regression gives you a single line through your data, representing the average relationship. Quantile regression gives you multiple lines showing how relationships change at different levels of the outcome variable. You might estimate the 10th percentile (lower tail), the 50th percentile (median), and the 90th percentile (upper tail) to see more of the picture.

The Mathematical Foundation

While traditional regression minimizes the sum of squared residuals, quantile regression minimizes an asymmetric weighted sum of absolute residuals. For a given quantile τ (tau), where τ ranges from 0 to 1, the objective function is:

minimize: Σ ρτ(yi - xiβ)

where ρτ(u) = u(τ - I(u < 0))

Here yi is the outcome for row i, xi is its predictors, β is the coefficient vector, and I(u < 0) is 1 when the residual u is negative and 0 otherwise. At τ = 0.5 this is the sum of absolute errors, which gives median regression. At τ = 0.9, a residual above the line costs 0.9 per unit and a residual below costs 0.1, so the fitted line ends up with about 90% of rows below it. Assumption: the quantile of y is a linear function of the predictors.

This asymmetric loss function means that observations above and below the regression line contribute differently to the estimate, allowing the method to target specific quantiles of the conditional distribution.

Why Quantile Regression Matters for Data-Driven Decisions

When you're making business decisions, knowing the average isn't always enough. Consider these scenarios:

• Risk Management: Analysts often want worst-case scenarios, not just average losses
• Supply Chain Planning: Planners may want a high-demand estimate to reduce stockouts
• Performance Analysis: Understanding both high and low performers reveals different intervention strategies
• Quality Control: Manufacturing teams may watch both tails of the distribution

When to Use Quantile Regression: A Step-by-Step Decision Framework

Determining whether quantile regression is the right approach for your analysis requires systematic evaluation. Follow this step-by-step methodology to assess your needs:

Step 1: Evaluate Your Decision-Making Requirements

Start by asking what decisions you need to make. If you're only interested in average behavior and your decisions are symmetric around the mean, traditional regression may suffice. However, quantile regression is worth considering when:

• You need to manage tail risks or extreme scenarios
• Different stakeholders care about different parts of the distribution
• Your decisions are asymmetric (e.g., the cost of understocking differs from overstocking)
• You're setting targets or thresholds for different performance levels

Step 2: Assess Your Data Characteristics

Examine your data for these indicators that favor quantile regression:

Heteroscedasticity: If the variance of your errors changes across the range of your predictors, quantile regression does not assume constant variance, so you can fit separate quantiles. Transforming the data or using weighted least squares are other options.

Outliers or Heavy Tails: The method uses absolute error, not squared error. Large errors in y can still affect the fit. Points far from the main x range can also matter. Check points that may sway the fit before drawing a result. If your data contains legitimate extreme values, they stay in the fit.

Non-Normal Distributions: When your response variable is skewed or has a complex distribution, quantile regression does not require normal errors.

Step 3: Consider the Complexity of Relationships

Quantile regression is useful when relationships between variables are not uniform. For instance, a predictor might go with larger changes among high outcomes than among low ones, as income does with food spending in the Engel example. If you suspect this, quantile regression can show it.

Key Assumptions and Requirements

One of quantile regression's major advantages is its relatively relaxed assumption structure compared to OLS regression. Knowing them helps you apply the method appropriately and check your results.

Main Assumptions

1. Linearity in Quantiles: The relationship between predictors and the specified quantile of the response must be linear. This doesn't mean the overall relationship is linear: different quantiles can have different linear relationships. You can assess this by examining residual plots for each quantile of interest.

2. Independent Observations (or an inference method that allows for dependence): Like most regression techniques, quantile regression requires observations to be independent. Violations occur with time series data, clustered data, or repeated measures, which need specialized extensions or different standard errors.

Diagnostic, not an assumption: quantile crossing. Estimated quantile lines for different quantiles should not cross inside the range of your data. For instance, the 25th percentile line should not cross the 75th percentile line. Some crossing can come from sampling variability. Repeated crossing is a sign the model may be misspecified.

Assumptions NOT Required

Understanding what quantile regression doesn't assume is equally important:

• Normality: No assumption about the distribution of errors is needed
• Homoscedasticity: The spread of errors can change across the predictor range, and that change is itself informative
• Symmetry: The distribution can be skewed or asymmetric
• Absence of Outliers: Extreme values are not ruled out, but check the points that may sway the fit

Sample Size Considerations

Quantile regression needs enough rows near the quantile you fit. The further into a tail you go, the fewer rows inform the fit, so estimates there are less stable.

Step-by-Step Methodology for Implementing Quantile Regression

These steps give a repeatable order of work.

Step 1: Define Your Quantiles of Interest

Start by identifying which quantiles matter for your decisions. Do not just default to standard choices. Pick quantiles that match the decision:

• Risk Analysis: Focus on lower quantiles (0.05, 0.10) for downside risk
• Opportunity Analysis: Examine upper quantiles (0.90, 0.95) for upside potential
• Comprehensive Understanding: Use quartiles (0.25, 0.50, 0.75) for a balanced view
• Specific Thresholds: Target quantiles that align with business metrics or regulatory requirements

Step 2: Prepare and Explore Your Data

Before fitting models, conduct thorough exploratory analysis:

Examine the unconditional distribution of your response variable using histograms, density plots, and quantile-quantile plots. This reveals skewness, multimodality, or other features that quantile regression can accommodate.

Create scatterplots of the response against each predictor, looking for evidence of heteroscedasticity or varying relationships across the response range. A fan shape, where points spread out as x grows, suggests quantile regression may show something OLS hides.

Step 3: Fit Multiple Quantile Models

Fit quantile regression models for each quantile of interest using the same set of predictors. Most statistical software packages offer quantile regression functionality:

# Python example using statsmodels (illustrative: replace "data" and the column names with your own)
import statsmodels.formula.api as smf

# Fit models for different quantiles
q25_model = smf.quantreg('sales ~ advertising + price + seasonality', data).fit(q=0.25)
q50_model = smf.quantreg('sales ~ advertising + price + seasonality', data).fit(q=0.50)
q75_model = smf.quantreg('sales ~ advertising + price + seasonality', data).fit(q=0.75)

Fitting models at several quantiles shows how relationships vary across the distribution.

Step 4: Check Model Assumptions

Check the assumptions that matter:

Linearity Check: Plot residuals against fitted values for each quantile. Random scatter is consistent with linearity; a pattern suggests transformations or nonlinear terms may help.

Quantile Crossing: Plot the estimated quantile functions together. They should maintain their ordering without systematic crossing. Minor crossing due to sampling variability is acceptable, but widespread crossing suggests model problems.

Coefficient Stability: Examine how coefficients change across quantiles. Smooth, gradual changes are common; erratic jumps may indicate overfitting or too little data at extreme quantiles.

Step 5: Compare Across Quantiles

Compare results across quantiles. Create coefficient plots showing how each predictor's effect varies across the distribution. This shows which factors go with larger changes at the low or high end of the outcome.

Interpreting Quantile Regression Results for Better Decisions

Read each result as a statement about one quantile.

Understanding Coefficients

A quantile regression coefficient represents the change in a specific quantile of the response variable for a one-unit change in the predictor, holding other variables constant. In the Engel example, the median slope of 0.5602 means that, among households at the same income, the fitted median food spending is 0.5602 francs higher per extra franc of income.

Read each coefficient as a statement about one quantile only. It may differ from the effect at other quantiles.

Comparing Effects Across Quantiles

If fitted slopes differ across quantiles, as they do in the examples above (0.4018 at the 10th percentile and 0.6863 at the 90th in the Engel data; 1.958 at the median and 2.802 at the 90th in the simulation), the input is associated with different changes in different parts of the outcome. Such differences can inform planning, but they do not show cause.

Interpreting Confidence Intervals

Standard errors and confidence intervals for quantile regression coefficients are typically obtained through bootstrap methods or asymptotic theory. Intervals at extreme quantiles are often wider because fewer rows inform the fit, although in the Engel data the 10th-percentile interval is the widest and the 90th-percentile interval the narrowest. When you decide from an extreme quantile, account for that uncertainty.

Statistical Significance vs. Practical Significance

As with any analysis, distinguish between statistical and practical significance. A coefficient might be statistically significant but too small to matter for decisions, or large enough to be important despite marginal statistical significance (especially with small samples).

Visualizing Results for Stakeholders

Effective communication of quantile regression results often requires visualization:

• Coefficient Plots: Show how each predictor's effect changes across quantiles with confidence bands
• Quantile Regression Curves: Plot multiple quantile lines overlaid on the data
• Effect Comparison Charts: Bar charts comparing coefficient magnitudes across quantiles
• Prediction Intervals: Show the range of predicted values at different quantiles for specific scenarios

Common Pitfalls and How to Avoid Them

These are common problems to watch for.

Pitfall 1: Overfitting at Extreme Quantiles

Extreme quantiles are estimated from fewer effective observations, making them prone to overfitting. A model with many predictors might fit the 95th percentile closely in-sample but do worse on new data.

Solution: Use regularization techniques like penalized quantile regression, employ cross-validation to assess out-of-sample performance, and consider simplifying models for extreme quantiles by including only the most important predictors.

Pitfall 2: Ignoring Quantile Crossing

When estimated quantile functions cross, predictions become nonsensical (e.g., the predicted 75th percentile is lower than the 25th percentile). This often results from model misspecification or insufficient data.

Solution: Use non-crossing quantile regression methods that impose monotonicity constraints, check for crossing during model validation, and consider simpler models or transformations if crossing persists.

Pitfall 3: Misinterpreting Changes Across Quantiles

Different coefficients across quantiles do not automatically mean the effects differ. Sampling variability can produce apparent differences that aren't statistically significant.

Solution: Conduct formal tests for equality of coefficients across quantiles, examine confidence intervals to see if they overlap substantially, and focus interpretation on clear, consistent patterns rather than isolated differences.

Pitfall 4: Inappropriate Extrapolation

Quantile regression estimates are most reliable within the range of observed data. Extrapolating beyond this range, especially for extreme quantiles, can produce unreliable predictions.

Solution: Clearly identify the range of your data and avoid making predictions outside it, use domain knowledge to assess the plausibility of extrapolations, and consider alternative methods like extreme value theory for tail predictions.

Pitfall 5: Skipping Median Regression

Analysts sometimes jump straight to extreme quantiles without considering that median regression (τ = 0.5) is a simple place to start and can be compared with OLS.

Solution: Start with median regression as an alternative to OLS, compare median and mean regression results to assess sensitivity to outliers, and then expand to other quantiles based on specific analytical needs.

Best Practices for Quantile Regression Analysis

These habits make results easier to read and to check.

1. Start with Exploratory Visualization

Before fitting any models, visualize your data thoroughly. Create scatterplots with multiple quantile regression lines overlaid to see if relationships vary across the distribution. This can guide how you specify the model.

2. Fit a Grid of Quantiles

Rather than analyzing just one or two quantiles, fit models for a grid spanning the distribution (e.g., τ = 0.1, 0.2, ..., 0.9). This provides a comprehensive view and helps detect patterns in how effects change across quantiles. You can then focus detailed interpretation on the most relevant quantiles for your decisions.

3. Compare with OLS Regression

It is a good habit to fit a standard OLS model alongside your quantile regressions. A gap between median regression and OLS can point to skew or outliers, and comparing upper and lower quantiles with OLS shows how much a mean-only fit leaves out. In the Engel data the OLS slope (0.4852) sits between the 10th-percentile and 90th-percentile slopes.

4. Test for Coefficient Differences

When you observe different coefficients across quantiles, test whether these differences are statistically significant. Many statistical packages provide tests for equality of coefficients across quantiles, helping distinguish meaningful heterogeneity from sampling noise.

5. Use Bootstrap for Inference

Bootstrap methods provide robust standard errors and confidence intervals for quantile regression, especially with smaller samples or complex models. Use enough bootstrap replications that the estimates stop changing.

6. Check on Holdout Data

Assess model performance on data not used for estimation. For quantile regression, a useful check is calibration: about 25% of new observations should fall below the 25th percentile prediction, for instance.

7. Communicate Uncertainty

When presenting results to stakeholders, include uncertainty estimates. Report intervals alongside predictions, especially for extreme quantiles where fewer rows inform the fit.

8. Document Your Quantile Choices

Be explicit about why you chose specific quantiles to analyze and report. Linking quantile choices to business objectives or decision thresholds makes your analysis more transparent and actionable.

Related Techniques and When to Use Them

Other methods answer related questions.

Linear Regression

Standard linear regression is a reasonable choice when you care about average outcomes and OLS assumptions such as constant error spread hold. It is simpler to explain and tends to give more precise estimates when its assumptions hold. See our linear regression guide.

Use linear regression when: You need to predict means, your data is well-behaved without heteroscedasticity, and stakeholders primarily care about average relationships.

Robust Regression

Robust regression methods (like M-estimation or MM-estimation) downweight outliers. They sit between OLS and quantile regression.

Use robust regression when: You want mean estimates that aren't distorted by outliers, but you don't need to understand behavior at different quantiles.

Expectile Regression

Expectile regression is similar to quantile regression but uses squared rather than absolute residuals, which can make it more precise but more sensitive to outliers. It estimates expectiles (generalizations of the mean) rather than quantiles.

Use expectile regression when: You want the efficiency benefits of squared loss while still exploring heterogeneity beyond the mean, and your data does not have severe outliers.

Generalized Additive Models for Location, Scale, and Shape (GAMLSS)

GAMLSS models the entire distribution by specifying models for multiple parameters (location, scale, shape). It is a parametric alternative to quantile regression.

Use GAMLSS when: You want to model the full distribution, you can specify an appropriate parametric family, and you need smooth predictions across all quantiles.

Quantile Forests and Machine Learning Approaches

Random forests and other machine learning methods can be extended to predict quantiles, combining quantile regression's goals with machine learning's flexibility for nonlinear relationships and interactions.

Use quantile forests when: Relationships are highly nonlinear, you have many predictors with complex interactions, and interpretability is less critical than prediction accuracy.

Combining Methods

Techniques can be combined. You might use quantile regression to identify that heterogeneity exists, then use machine learning to capture complex functional forms, or employ quantile regression for inference and interpretation alongside more complex predictive models.

Implementing a Data-Driven Decision Process with Quantile Regression

One way to place quantile regression inside a decision process:

Step 1: Define Decision Objectives and Constraints

Clarify what decision you're making and what factors constrain it. Inventory decisions face storage costs and stockout costs; hiring decisions balance talent acquisition against budget; pricing decisions weigh revenue against demand.

Step 2: Identify Decision-Relevant Quantiles

Map your decision context to specific quantiles. Risk-averse decisions focus on lower quantiles; opportunity-seeking decisions emphasize upper quantiles; balanced approaches examine multiple quantiles representing different scenarios.

Step 3: Build and Validate Models

Follow the methodology above and check the assumptions that matter.

Step 4: Translate Statistical Outputs to Decision Inputs

Convert coefficient estimates and predictions into quantities that matter for decisions: expected costs, probability of meeting targets, resource requirements, or risk metrics.

Step 5: Conduct Sensitivity Analysis

Examine how decisions change under different quantile choices, coefficient uncertainty, or model specifications. Robust decisions remain sound across plausible scenarios.

Step 6: Monitor and Update

After implementing decisions, track outcomes and compare them to predictions at various quantiles. This shows how well the models hold up and when relationships change.

Frequently Asked Questions

What is the main difference between quantile regression and ordinary least squares regression?

While ordinary least squares (OLS) regression estimates the conditional mean of the response variable, quantile regression estimates conditional quantiles. This means OLS gives you one average relationship, while quantile regression reveals how relationships vary across different parts of the distribution, which helps when you want to understand variability in outcomes.

When should I use quantile regression instead of linear regression?

Use quantile regression when you need to understand relationships beyond the average, when you want to keep extreme values in the fit, when errors are non-normally distributed or heteroscedastic, or when different quantiles show different relationships with predictors. It is often used for risk assessment, quality control, and decisions under uncertainty.

How do I interpret quantile regression coefficients?

A quantile regression coefficient represents the change in a specific quantile of the response variable for a one-unit change in the predictor. For example, at the 75th percentile (τ=0.75), a coefficient of 2.5 means that a one-unit increase in the predictor is associated with a 2.5-unit increase in the 75th percentile of the response variable.

What are the key assumptions of quantile regression?

Quantile regression assumes linearity in quantiles (the relationship between predictors and the chosen quantile is linear) and independent observations. Estimated quantile lines should also not cross inside the data range, which is a check rather than an assumption. Unlike OLS, it does not assume normal errors or constant error spread.

Can quantile regression handle outliers better than standard regression?

The method uses absolute error, not squared error. Large errors in y can still affect the fit. Points far from the main x range can also matter. Check points that may sway the fit before drawing a result.

Conclusion

Quantile regression estimates chosen points of the ranked outcome instead of only the average. In the Engel data, the food-spending slope was 0.4018 at the 10th percentile and 0.6863 at the 90th, while OLS reported a single 0.4852.

Key points to remember:

• Match your quantile choices to your decision rather than using defaults
• Compare several quantiles, and report intervals
• Be cautious at extreme quantiles, where fewer rows inform the fit
• Differences across quantiles describe association, not cause

Sources

• Koenker, R. and Bassett, G. (1978). Regression Quantiles. Econometrica 46(1), 33-50. doi:10.2307/1913643
• Koenker, R. and Bassett, G. (1982). Robust Tests for Heteroscedasticity Based on Regression Quantiles. Econometrica 50(1), 43-61. doi:10.2307/1912528
• quantreg R package manual (Engel data set). cran.r-project.org

Want to try it on your own file? Upload a CSV and a tested statistical method runs on it. The same data gives the same numbers on a re-run.