GSEA: Why is signal2noise suggested only for 3+ samples?

Question

[bipin]

The GSEA manual states that the signal2noise ranking metric should be used only in case of 3+ samples per condition.

What is the reason for this restriction?

I do understand that the denominator has standard deviations for the input values and it may be more meaningful in case of 3+ samples but how is it unsuitable in case of 2 samples?

[h.mon]

Short and informative title, please.

Edit: Much better now, thanks.

Answer 1

[h.mon]

I guess it is partially arbitrary, but the confidence intervals improve a lot with 3 samples, compared to 2. Suppose a SD of 1, then you have the following confidence intervals:

N=2     95% CI of the SD        0.45 to 31.91
N=3     95% CI of the SD        0.52 to 6.28
N=4     95% CI of the SD        0.57 to 3.73
N=5     95% CI of the SD        0.60 to 2.87