[R] 'asymmetric span' for 2D loess?

[Prof Brian Ripley]

On Mon, 2 Jun 2008, Tokuyasu, Taku wrote:

> Hello,
>
> I am interested in performing a 2D loess smooth on microarray data, i.e.
> log2 ratios on a 2D grid, using different spans in the horizontal and
> vertical directions (the immediate reason being that replicate spots are
> laid out in the horizontal direction).  Is it possible to do this in R?
> Functions like loess(stats) seem to apply the same span for all
> predictors, which carries over to functions like ma2D(marray).

See the next comment.  'span' applies to 2D distances, and I think you 
need to rescale your inputs so Euclidean distance is appropriate.

> As an elementary second question, are there circumstances where one
> expects to see a substantial difference in the fits between say loess(y
> ~ x1 + x2) and loess(y ~ x1 * x2) with an interaction term (and if so,
> what are they)?

>From the help page:

  formula: a formula specifying the numeric response and one to four
           numeric predictors (best specified via an interaction, but
           can also be specified additively).  Will be coerced to a
           formula if necessary.

So the two versions both specify smoothing in 2D.

You can use things like lo(x1) + lo(x2) in some gam fits.

It often helps to read the primary sources, in this case chapters in the 
White Book (Chambers & Hastie, nominally 1992, actually published in 
1991).

>
> Many thanks,
>
> _Taku
>
> Taku A. Tokuyasu, PhD
> UCSF Helen Diller Family Comprehensive Cancer Center
> San Francisco, CA 94143-0808

-- 
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