Regression modelStatisticsModel

Quantile Regression (Nonparametric Variants)

Q: How is quantile regression different from ordinary regression?

Ordinary regression models the conditional mean of the outcome, while quantile regression models a chosen conditional quantile — the median or any percentile. This lets you see how predictors affect the lower, middle, and upper parts of the distribution rather than only its average.

Q: What does τ mean and which values should I pick?

τ (tau) is the quantile being modelled, between 0 and 1. τ = 0.5 gives the median; common choices are 0.25, 0.5, and 0.75 to capture the lower quartile, median, and upper quartile. You can fit any set of quantiles relevant to your question.

Q: Why does it not need a normality assumption?

Quantile regression estimates coefficients by minimising an asymmetric absolute-error (check) loss rather than squared error, so it does not assume a particular error distribution. The nonparametric variants go further and avoid a distributional form altogether, which is why they are robust on skewed data and to outliers.

Q: What sample size do I need?

Around 50 observations or more is recommended for stable estimates. Below about 30 the bandwidth selection for the nonparametric variants becomes unreliable, and below roughly 20 the quantile estimates are unstable, so a median-based estimator such as Theil-Sen is preferable.

Also known as: quantile regression, median regression, distribution-free quantile regression, Kantil Regresyon (Nonparametric Varyantlar)

OriginatorKoenker & BassettYear1978Sources2Related methods6

Quantile regression, introduced by Koenker and Bassett in 1978, models a chosen conditional quantile (such as the median or the 25th and 75th percentiles) of a continuous outcome rather than its mean. Its nonparametric variants fit these quantile relationships without assuming a distribution for the errors, making them a robust complement to mean-based regression on skewed data.

Key highlights

Requires no distributional assumption about the errors, so it is robust on skewed data and to outliers.
Estimates the effect of predictors across the whole conditional distribution, not just the mean, revealing how relationships differ in the lower and upper tails.
At the median it minimises absolute deviations, giving a robust central-tendency fit that resists the influence of extreme observations.

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 quantile regression when the outcome is continuous and you care about more than the average — for example the spread, the tails, or how effects differ across low and high values of the response. It is especially suited to skewed distributions and data with outliers, where it serves as a distribution-free alternative to mean regression. A reasonable sample is needed (about 50 observations or more); for the smoothed nonlinear variants the bandwidth choice is critical, and very small samples make quantile estimates unstable.

Strengths & limitations

Strengths

Requires no distributional assumption about the errors, so it is robust on skewed data and to outliers.
Estimates the effect of predictors across the whole conditional distribution, not just the mean, revealing how relationships differ in the lower and upper tails.
At the median it minimises absolute deviations, giving a robust central-tendency fit that resists the influence of extreme observations.

Limitations

Needs a reasonable sample size (around 50 or more) for stable quantile estimates.
For smooth nonlinear quantile curves the bandwidth choice is critical and strongly affects the result.
In small samples (n < 30) the nonparametric bandwidth selection becomes unreliable, and below about 20 observations quantile estimates are unstable.

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

How is quantile regression different from ordinary regression?

Ordinary regression models the conditional mean of the outcome, while quantile regression models a chosen conditional quantile — the median or any percentile. This lets you see how predictors affect the lower, middle, and upper parts of the distribution rather than only its average.

What does τ mean and which values should I pick?

τ (tau) is the quantile being modelled, between 0 and 1. τ = 0.5 gives the median; common choices are 0.25, 0.5, and 0.75 to capture the lower quartile, median, and upper quartile. You can fit any set of quantiles relevant to your question.

Why does it not need a normality assumption?

Quantile regression estimates coefficients by minimising an asymmetric absolute-error (check) loss rather than squared error, so it does not assume a particular error distribution. The nonparametric variants go further and avoid a distributional form altogether, which is why they are robust on skewed data and to outliers.

What sample size do I need?

Around 50 observations or more is recommended for stable estimates. Below about 30 the bandwidth selection for the nonparametric variants becomes unreliable, and below roughly 20 the quantile estimates are unstable, so a median-based estimator such as Theil-Sen is preferable.

Sources

1.
Koenker, R. & Bassett, G. (1978). Regression Quantiles. Econometrica, 46(1), 33-50.
DOI 10.2307/1913643
2.
Koenker, R. (2005). Quantile Regression. Cambridge University Press.
ISBN 978-0521608275

You have read it. What now?

Ask about this method Similar methods

Cite this page

ScholarGate. (2026, June 1). Nonparametric Quantile Regression. ScholarGate. https://scholargate.app/statistics/quantile-regression-np