W-Estimator Robust Regression (Welsch / Tukey Bisquare)
Also known as: Tukey bisquare M-estimator, Welsch M-estimator, redescending M-estimator, W-Tahmin Edici (Welsch / Tukey Bisquare)
The W-estimator is a family of robust M-estimator variants for linear regression that use the Tukey bisquare and Welsch weight functions, introduced in the line of work going back to Beaton and Tukey (1974). Because its weights fall rapidly toward zero as a residual grows, it resists outliers more strongly than the Huber M-estimator.
Key highlights
- Stronger resistance to outliers than the Huber M-estimator because its redescending weights fall to exactly zero for large residuals.
- Tolerates a high fraction of symmetric outliers — up to roughly half the sample.
- Keeps the familiar linear-model interpretation: coefficients read just like OLS coefficients but are fitted to the clean bulk of the data.
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 the W-estimator when you want a linear regression of a continuous outcome that stays reliable in the presence of outliers in the response direction. It assumes the residuals are roughly symmetric and tolerates a contaminated fraction of up to about 50% of the data; a sample of at least 30 observations is recommended. It is less appropriate when the data contain high-leverage points in the predictor space, where a high-breakdown method is needed, and it can fail to converge in very small samples (n < 20).
Strengths & limitations
- Stronger resistance to outliers than the Huber M-estimator because its redescending weights fall to exactly zero for large residuals.
- Tolerates a high fraction of symmetric outliers — up to roughly half the sample.
- Keeps the familiar linear-model interpretation: coefficients read just like OLS coefficients but are fitted to the clean bulk of the data.
- Can be inadequate when high-leverage points are present, where a high-breakdown estimator is required instead.
- May suffer convergence problems in very small samples (n < 20).
- Assumes a symmetric residual distribution; strongly skewed errors undermine the weighting.
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 does the W-estimator differ from the Huber M-estimator?
Both down-weight large residuals, but the Huber weight only flattens the loss for big residuals while the W-estimator's bisquare and Welsch weights redescend all the way to zero. That makes the W-estimator reject gross outliers outright, giving stronger resistance at the cost of more delicate convergence.
What does the tuning constant c control?
The constant c sets where a residual stops contributing. A common choice for the Tukey bisquare is c = 4.685, which trades a small loss of efficiency under clean Gaussian data for robustness; smaller c rejects outliers more aggressively.
When should I prefer an MM- or S-estimator instead?
When the contamination is in the predictor space (high-leverage points) rather than only in the response, the W-estimator can be inadequate. A high-breakdown method such as the MM-estimator, often started from an S-estimator, is the recommended alternative.
What sample size do I need?
At least about 30 observations is recommended. Below roughly 20 the iteratively reweighted fit can fail to converge, and a simpler robust method such as Theil-Sen is preferable.
Sources
- 1.Beaton, A. E. & Tukey, J. W. (1974). The Fitting of Power Series, Meaning Polynomials, Illustrated on Band-Spectroscopic Data. Technometrics, 16(2), 147-185.
- 2.Maronna, R. A., Martin, R. D., Yohai, V. J. & Salibián-Barrera, M. (2019). Robust Statistics: Theory and Methods (with R) (2nd ed.). Wiley.ISBN 978-1119214687
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). W-Estimator. ScholarGate. https://scholargate.app/statistics/w-estimator