Speed Group
Microarray Page
Index to our site
Avoid assuming normality
The collection of log ratios from a single microarray experiment is typically quite unlike a random sample from a single normal population. This is particularly so when a lot (say > 10%) of genes are differentially expressed. To see this, do a normal probability plotof your log ratios. Chances are it will look like Figure 1a below, which comes from a replicated experiment in which we believe at most 7 or 8 spots have meaningfully different intensities. Typically, as here, the plot exhibits longer tails than would a normal sample, i.e., (relatively ) too many large and small log ratios. This remains true for square root and higher root ratios, see Figure 1b,c,d.
Conclusion. It is dangerous to use normal statistical theory to guide your selection of differentially expressed genes. The normal thinking which says that about 68% should be within 1 standard deviation (SD), 95% within 2 SDs and 99% within 3 SDs of the mean does not apply, even when no differential expression is present.
Figure 1. Quantile-Quantile plot of transformed intensity ratios from a single array (Apo AI ko 9).
last updated March 07, 2000
zarray@stat.berkeley.edu
