[BioC] [limma] Strange results in contrasts with dye-swap

[Gordon K Smyth]

On Fri, 5 Aug 2011, mjonczyk at biol.uw.edu.pl wrote:

> Dear Gordon,
>
> thank you for clarifying this issue.
>
>
>>> Results:
>>>
>>> TEST WITHOUT DYE EFFECT
>>>> summary(test.ug)
>>>   c2 - k2 c4 - k4 c6 - k6 c8 - k8 c10 - k10 c12 - k12 c14 - k14
>>> -1     115      96     377     141       263       175       265
>>> 0    43082   43048   42326   42973     42694     42842     42752
>>> 1      196     249     690     279       436       376       376
>>>
>>>
>>> TEST WITH DYE EFFECT
>>>> summary(test.ug.d)
>>>      p2    p4    p6    p8   p10   p12   p14
>>> -1   246   217   686   317   530   316   526
>>> 0  42797 42755 41594 42530 42103 42446 42239
>>> 1    350   421  1113   546   760   631   628
>>>
>>> TEST WITH DYE EFFECT AND CONTRAST FOR DYE EFFECT
>>>> summary(test.ug.d2)
>>>   DyeEffect    p2    p4    p6    p8   p10   p12   p14
>>> -1      6660   452   402  1103   593   858   539   883
>>> 0      30803 42353 42261 40659 41853 41359 41888 41564
>>> 1       5930   588   730  1631   947  1176   966   946
>>>
>
>> The lesson here is that you should not include DyeEffect in decideTests()
>> with your other contrasts.  When using method="global", you should only
>> include contrasts that are closely comparable to one another, and about
>> which you will be making conclusions as a group.
>>
>>> *OTHER QUESTIONS*
>>> 1. Is second model (WITH DYE EFFECT) correct?
>>
>> Fine.
>>
> Ok, so I use this model and "global" method.
>>
>>
>>> 4. Should I include biol-replication effect in analysis (as block)?
>>
>> If the biol replicates seem to vary randomly, and are only slightly
>> different, then I would suggest that you enter them as a random effect
>> using block instead.  If there are large differences between the biol
>> reps, in particular if one rep is different to the others, then including
>> them in the design matrix as you have done is safer and probably better.
>
> The second scenario is true in this experiment, from PCA I know that differences
> between replications isn't negligible.
> So replication effect is already included, just because in target frame (and
> consequently design matrix) each combination of samples on array is repeated
> four times?

Yes, the analysis is adjusting for any batch effect between your 
replicates.

Best wishes
Gordon

> Best regards,
> Maciej
>>> ________________________________________________________
>>> Maciej Jończyk, MSc
>>> Department of Plant Molecular Ecophysiology
>>> Institute of Plant Experimental Biology
>>> Faculty of Biology, University of Warsaw
>>> 02-096 Warszawa, Miecznikowa 1

______________________________________________________________________
The information in this email is confidential and intend...{{dropped:4}}