Scaling Methods

Residual scale estimation during robustness iterations.

Overview

When iterations > 0, LOWESS computes robustness weights by comparing each residual to the current residual scale estimate. The scaling_method parameter controls how that scale is measured.

The robustness weight for point $i$ is:

\[w_i = B\!\left(\frac{|r_i|}{6 \cdot \hat{\sigma}}\right)\]

where $B$ is the bisquare function and $\hat{\sigma}$ is the scale estimate. A larger $\hat{\sigma}$ makes the algorithm more tolerant of large residuals; a smaller one makes it more aggressive.

MethodFormulaRobustnessSpeed
"mad"Median of |residuals − median(residuals)|Very robustModerate
"mar"Median of |residuals|RobustFast
"mean"Mean of |residuals|Less robustFastest

Scaling Methods Comparison


MAD — Median Absolute Deviation (Default)

\[\hat{\sigma} = \text{median}(|r_i - \text{median}(r_i)|)\]

First centers residuals at their median, then takes the median of the absolute deviations. Double use of the median makes it highly resistant to extreme outliers. This is the standard choice for robust regression.

Use when: Data may contain outliers (default for most applications).

using FastLOWESS
using Random, Statistics

rng = MersenneTwister(42)
x = collect(range(0, 2π, length=100))
y = sin.(x) .+ randn(rng, 100) .* 0.3

model = Lowess(; iterations=3, scaling_method="mad")
result = fit(model, x, y)
println("First smoothed value (MAD scaling): ", result.y[1])
First smoothed value (MAD scaling): 0.4426033958974517

MAR — Median Absolute Residual

\[\hat{\sigma} = \text{median}(|r_i|)\]

Uses the uncentered median — unlike MAD it does not subtract the residual median first. Still robust (median-based) but slightly less resistant than MAD when residuals are systematically shifted. Faster than MAD in practice because it requires only one partial sort.

Use when: Speed matters and data have minimal systematic bias in residuals.

using FastLOWESS
using Random, Statistics

rng = MersenneTwister(42)
x = collect(range(0, 2π, length=100))
y = sin.(x) .+ randn(rng, 100) .* 0.3

model = Lowess(; iterations=3, scaling_method="mar")
result = fit(model, x, y)
println("First smoothed value (MAR scaling): ", result.y[1])
First smoothed value (MAR scaling): 0.44291545433676993

Mean — Mean Absolute Residual

\[\hat{\sigma} = \frac{1}{n}\sum_i |r_i|\]

Arithmetic mean of absolute residuals. Non-robust: a single extreme outlier inflates $\hat{\sigma}$, causing the algorithm to under-downweight it. Fastest to compute (no sort required). Useful when data are believed to be clean and speed is a priority.

Use when: Clean data with no outliers; maximum computation speed required.

using FastLOWESS
using Random, Statistics

rng = MersenneTwister(42)
x = collect(range(0, 2π, length=100))
y = sin.(x) .+ randn(rng, 100) .* 0.3

model = Lowess(; iterations=3, scaling_method="mean")
result = fit(model, x, y)
println("First smoothed value (mean scaling): ", result.y[1])
First smoothed value (mean scaling): 0.4461390971801288

Choosing a Scaling Method

SituationRecommended Method
General purpose, possible outliers"mad" (default)
Speed matters; residuals have minimal systematic bias"mar"
Clean data, no outliers"mean"

See Robustness for a broader discussion of outlier handling.