[R] Parameters for loess to match results from lowess?

[Prof Brian Ripley]

On Tue, 16 May 2006, Stuart Leask wrote:

> I wish to reproduce a locally-weighted fit from (old-fashioned) lowess with
> (new-fangled) loess.

`new-fangled': from 1991!  (That is, well before R.)

> Although ?lowess advises that loess has different defaults, does anyone know
> how to set up loess so it does exactly the same as lowess did?

It is not possible: the algorithms differ considerably in their details.

> Fiddling about suggests setting loess(x, span = 0.8, degree=1) is fairly
> close for my data - but clearly not the same.

You want family="symmetric" as well.

In determining 'local' loess() uses a tricubic weighting, lowess() uses a 
uniform weighting (on reading the code).

> (I can't help feeling that with something basic like this, I've probably
> been unlucky in my archive/help searches so far.)

Did you look at the references on the help pages?  I think that is the 
only place you are likely to find this explained in detail.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595