There isn't a standard protocol for initial statistical testing of large microarray data prior to R analysis. We usually do quantile or z normalization procedures but I am just wondering when do we really need to use quantile normalization or z normalization in a given dataset?
Remember that the whole reason for normalizing in the first place is to bring the final whole data distribution into alignment for application of a particular statistical tests. So for example, if using ANOVA, the assumption is the data is normally distributed (although ANOVA is reasonably robust to deviations from that), and a quantile normalization and log transformation usually achieves that.
But, array data is highly variable, so often one size does not fit all. You may have some big differences in background, in labeling efficiency, or you may have batch effects to deal with, or a host of other issues that may crop up. So, you need to explore your data sometimes to find a normalization that actually works as intended with your particular data. MA plots and PCA can be useful to see what your data tells you about how well a particular normalization is working, or not.
See if this may help
http://libra.msra.cn/Publication/6055145/systematic-evaluation-of-scaling-methods-for-gene-expression-data
Hi Ryan,
Good that you ask this question. Most of the time it is really not clear why and when one should (further) normalize the data.
If I would be to provide a simple answer I would say perform an MA plot (http://en.wikipedia.org/wiki/MA_plot) "test" and see how your data behaves. Exploratory analysis can empower you a lot in your decision making.
For further (yet beginner's level) guidance I would point to the work of Rafael Irizarry. Here you can find a very good introductory video to data normalization (for microarray data) : https://www.youtube.com/watch?v=OZEQF4YrivY .
Some of these principles can be applied to mass spectrometry based proteomic data.
Ilian
Remember that the whole reason for normalizing in the first place is to bring the final whole data distribution into alignment for application of a particular statistical tests. So for example, if using ANOVA, the assumption is the data is normally distributed (although ANOVA is reasonably robust to deviations from that), and a quantile normalization and log transformation usually achieves that.
But, array data is highly variable, so often one size does not fit all. You may have some big differences in background, in labeling efficiency, or you may have batch effects to deal with, or a host of other issues that may crop up. So, you need to explore your data sometimes to find a normalization that actually works as intended with your particular data. MA plots and PCA can be useful to see what your data tells you about how well a particular normalization is working, or not.
quantro: a data-driven approach to guide the choice of an appropriate normalization method, this paper is related and interesting. Article quantro: A data-driven approach to guide the choice of an ap...
- Badges
- Science topic