general question about homogeneity of variances between microarray groups
2
0
Entering edit mode
Guido Leoni ▴ 200
@guido-leoni-3328
Last seen 10.1 years ago
European Union

Dear list

I'm performing some microarrays analysis for a simple case(15 microarrays), control(3 microarrays) experiment design. Don't ask me the reason for which i have a so unbalanced dataset ;-)

In order to detect differentially expressed genes I wish to perform a LIMMA analysis...but checking the omogeneity of variances with bartlett test I observ a difference statistically significative between cases and controls.

According to your experience: Is a good idea before doing a parametric analysis checking the variances
utilizing Bartlett test? In my case a non parametric test(like SAM) might be better than LIMMA?

thak you for any tips
Best
Guido

limma • 1.2k views
ADD COMMENT
0
Entering edit mode
@james-w-macdonald-5106
Last seen 8 hours ago
United States
Hi Guido, On 7/31/2012 5:13 AM, Guido Leoni wrote: > Dear list > I'm performing some microarrays analysis for a simple case(15 microarrays) > , control(3 microarrays) experiment design. > Don't ask me the reason for which i have a so unbalanced dataset ;-) > In order to detect differentially expressed genes I wish to perform a LIMMA > analysis...but checking the omogeneity of variances with bartlett test I > observ a difference statistically significative between cases and controls. If you are using limma, you will be shrinking your variance estimates towards an overall value, so I doubt that the unbalanced design will be a big concern. > According to your experience: > Is a good idea before doing a parametric analysis checking the variances > utilizing Bartlett test? That might be an issue if you are doing a handful of tests, but I don't see the applicability when you are doing thousands. > In my case a non parametric test(like SAM) might be better than LIMMA? SAM isn't non-parametric. The biggest difference between limma and SAM is that you are using permutation to construct a null distribution in the case of SAM, but you use a conventional t-distribution for limma. However, in both cases you are estimating parameters and comparing to a null distribution. Best, Jim > thak you for any tips > Best > Guido > > [[alternative HTML version deleted]] > > _______________________________________________ > Bioconductor mailing list > Bioconductor at r-project.org > https://stat.ethz.ch/mailman/listinfo/bioconductor > Search the archives: http://news.gmane.org/gmane.science.biology.informatics.conductor -- James W. MacDonald, M.S. Biostatistician University of Washington Environmental and Occupational Health Sciences 4225 Roosevelt Way NE, # 100 Seattle WA 98105-6099
ADD COMMENT
0
Entering edit mode
@gordon-smyth
Last seen 18 minutes ago
WEHI, Melbourne, Australia

No, using Bartlett's test to check whether a t-test is ok is a very bad idea.  Bartlett's test is well known to be highly sensitive to non-normality, so it is very likely to give significant results as a result of small deviations from normality rather than genuine differences in variances.  By contrast, the two-sided t-test that limma does is quite robust against both non-normality and inequality of variances.

George Box had a few choice words more than half a century ago for what you propose, in a famous paper in Biometrika in 1953.  He said it was like setting out  to sea in a rowing boat to check if the weather was calm enough for an ocean liner to leave port.

As James MacDonald has already said, SAM is not non-parametric and it assumes equal variances just like limma.

Even non-parametric tests like the Wilcoxon 2-sample test still assume equal variances.

This is not to say that you shouldn't be checking your data.  But exploratory methods like plotMDS() are much more relevant than Bartlett's test, and solutions like arrayWeights() in limma are much better than switching to another type of test if you really do have a meaningful difference in variabilities between groups.

Best wishes
Gordon

ADD COMMENT

Login before adding your answer.

Traffic: 700 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6