Specifying Lowess Smoothing Parameters
main topic
You can specify two parameters to define the lowess smoother:
· Degree of smoothing:
A lowess smoother generally works best when the fraction (f) of points
is large enough to give a smooth fit without distorting the underlying
relationship between the variables. Cleveland [2]
suggests that you make f as large as possible, but maintain unrelatedness
in a separate lowess plot of the y-value residuals versus the x-values.
· Number
of steps: To limit the influence of outliers on the smoothed y-values,
you can set the number of iterations of smoothing. Each step reduces the
weights given to outliers in the next iteration of weighted linear regression,
based on the size of residuals in the previous lowess step. For more details,
refer to step 4 of the lowess method.
When you set the number of steps to 0, step 4 of the lowess method is
eliminated entirely. Cleveland [2]
suggests that two robust steps adequately smooth outlier effects for most
data.