[BioC] Limma: bad spots flagged out?

Gordon K Smyth smyth at wehi.EDU.AU
Tue Feb 21 13:44:05 CET 2006


Although you don't say exactly, I assume that your question is why the two results MA.1 and MA.2
are slightly different.

Zero weights in the loess normalization are almost but not quite the same as NA.  Observations
which are NA really are ignored entirely.  Observations which have zero weight are given no weight
in the local regressions, but are used by the loess function to determine the span neighborhood of
each local point.  Hence zero weights in loess are nearly but not quite the same as missing
observations.  This behavior is a characteristic of the loess() function which is used by
normalizeWithinArrays().

The effect of zero weights is usually close enough to treating the value as missing, unless the
propotion of zero weights is exceptionally high.  The difference is seldom important in practice,
and doesn't seem important in your case.

Best wishes
Gordon

> Date: Mon, 20 Feb 2006 12:58:15 +0100
> From: Ana Conesa <aconesa at ivia.es>
> Subject: [BioC] Limma: bad spots flagged out?
> To: bioconductor at stat.math.ethz.ch
> Message-ID: <7.0.0.16.0.20060220124132.02020ff0 at ivia.es>
> Content-Type: text/plain; charset="iso-8859-1"
>
>
>    Dear list,
>    I   have   a  doubt  about  the  real  use  of  spots  weights  during
>    normalization in limma. According with the documentation spots weights
>    equal  to  0  are  ignored  during  normalization (and other posterior
>    analyses)  , and although these spots are not removed they do not have
>    any  influence on the rest. I had observed some strange behavior on my
>    normalized  data  and made a tried to make a check on this. What I did
>    was  to  replace  weights==0 spots by NA and redo analysis. I have
>    found  the  results  do  no   spots  are  adequately ignored by limma or I made a conceptual
> mistake
>    in this check. This is exactly the code I used for checking:
>    > RG.b <- backgroundCorrect(RG)
>    > MA.1 <- normalizeWithinArrays(RG.b)
>    > RG.b$G[RG.b$weights==0] <- NA
>    > RG.b$R[RG.b$weights==0] <- NA
>    > MA.2 <- normalizeWithinArrays(RG.b)
>    > MA.1$M[2,]
>             BFN33        S50.05        S65.02        S65.03        S71.02
>    S75.01      S77.03     control     control     control     control
>    -0.96794887     0.06715693    -0.08766477    -0.50161127   -1.25216169
>    -0.80724650    -0.61351625    -0.80751427    -0.49960303   -0.66912129
>    0.27447918
>    > MA.2$M[2,]
>             BFN33        S50.05        S65.02        S65.03        S71.02
>    S75.01      S77.03     control     control     control     control
>    -0.94108221     0.04803980    -0.08030398    -0.49226094            NA
>    -0.79588160             NA    -0.77182854             NA   -0.61426784
>    0.18076174
>    Thank you
>    Ana
>
>       O @@@@@     Ana Conesa, PhD.
>      @@@ O @@ O @   Centro de Gen?mica
>       @  O  @@@@  O  @   Instituto Valenciano de Investigaciones Agrarias
>    (IVIA)
>      @@@ O @@@@        @@@@ O @     46113 Moncada (Valencia) SPAIN
>         ||       Tel. +34 963424000 ext.70161; Fax. +34 963424001
>         ||       email: aconesa at ivia.es



More information about the Bioconductor mailing list