[BioC] Re: Limma: four suggestions related to weights and background correction

[Gordon Smyth]

At 07:28 PM 9/02/2004, Ramon Diaz-Uriarte wrote:
> > >3. Occasionally, a complet print-tip group might be missing which will
> > > make "normalizeWithinArrays" to fail. We have added the following to that
> > > function, in the print-tip loess case, right before the line "spots <-
> > > spots + nspots" (and substituting the direct call to loessFit):
> > >
> > >******************************************
> > >num.fill <- length(object$M[spots, j])
> > >object$M[spots, j] <- rep(NA, num.fill)
> > >no.objects <-
> > >     try(object$M[spots, j]
> > >         <- loessFit(y, x, w, span = span,
> > >                     iterations = iterations)$residuals)
> > >if(class(no.objects) == "try-error")
> > >     print(paste("WARNING: in print-tip loess normalization",
> > >                 "no samples in array", j, "grid row", gridr,
> > >                 "grid column", gridc))
> > >********************************************
> >
> > That's a good suggestion. I would note though that if this phenomenon
> > occurs for you, you probably shouldn't be using printtiploess normalization
> > in the first place - you should use plain "loess" instead.
>
>
>I didn't make the context clear: we are using this as part of a web interface
>for normalization, and having the normalization functions to return an error
>is undesirable. With the above modification the R run finishes, and the web
>application is happy, but the user is given a warning, letting her choose to
>re-do things otherwise.
>
>However, regarding the type of normalization, I am not sure I see why, if 
>this
>happens occasionally, the user should prefer global loess: if a print-tip
>group in a specific array has been damaged (e.g., a scratch), but the rest of
>the print-tip groups are OK, and the usual assumptions are believed to hold
>for each of the other print-tip groups, then I think print-tip loess is
>justified.

Fair enough. My concern was that print-tip loess should only be used when 
there are a reasonable number of usable points in each print-tip group, say 
100 or more. If you have no usable values in one group, then I worry that 
there might be only a small number of values in another group as well, yet 
enough that there would be no warning. In that case it would be much better 
to use "loess" or "robustspline".

Best regards
Gordon