32lowess(1)

# 32lowess(1) - Smoothing part 3 I Penalized splines is not...

This preview shows pages 1–5. Sign up to view the full content.

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Smoothing - part 3 I Penalized splines is not the only way to estimate f ( x ) when y = f ( x ) + I Two others are kernel smoothing and the Lowess (Loess) smoother I I’ll only talk about Lowess I Penalized splines and Lowess have same goal. I Lowess is more ad-hoc. Only practical difference I’ve found is that Lowess includes an outlier detection/downweighting step I Lowess is LOcally WEighted polynomial regreSSion I Original paper is Cleveland, W.S. (1979) Robust locally weighted regression and smoothing scatterplots. JASA 74:829-836 c 2011 Dept. Statistics (Iowa State University) Stat 511 section 31 1 / 15 Lowess concepts I any function, f ( x ) , is closely approximated by β + β 1 x over a small enough range of x I so pick a point, x , may or may not be an x in the data I fit a linear regression to the subset of the data with x values near x I predict f ( x ) from ˆ β + ˆ β 1 x I repeat for as many x ’s as needed to estimate f ( x ) I A couple of refinements: I Use weighted regression with weights dependent on distance from x I “Robustify” the fit by downweighting the influence of regression outliers (far vertically from ˆ f ( x )) c 2011 Dept. Statistics (Iowa State University) Stat 511 section 31 2 / 15 I A simple algorithm that doesn’t work well: I Consider all points within specified distance d of x I Fit OLS regression to these points I Imagine fitting the curve to 5 points around x I Then shifting to x + . 01 that now includes 6 points I that 6th point is within d of x + . 01 but not within d of x I Result is a very “jumpy” fitted curve because that 6’th point has a very large influence on the fitted value I Lowess incorporates weighting so points ≈ d from x have some influence, but not much. I Influence increases as x gets close to a data point I Also hard to specify an appropriate d : depends on range of x values c 2011 Dept. Statistics (Iowa State University) Stat 511 section 31 3 / 15 Lowess details I What is “close to...
View Full Document

{[ snackBarMessage ]}

### Page1 / 15

32lowess(1) - Smoothing part 3 I Penalized splines is not...

This preview shows document pages 1 - 5. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online