Dear Matt,

[quote]
Is a dataset of 6 allele sequences enough to say anything?  I realize that the answer may depend on the data itself (i.e., how many variable sites, how much variation, etc.)
[/quote]

Generally speaking, there is little information in a 6-sequence alignment to detect selection at an individual site (but there may be enough to test for selection on the proportion of sites in the gene - just not which individual site is under selection). If the sequences are really divergent, you may also need to worry about alignment quality in variable regions (this depends on the gene of course) and saturation (i.e. where the rates are basically infinite).

[quote]
Which method (FEL, SLAC, REL) is best for analyzing small data sets in general?  The impression I got from reading the "Not So Different..." MBE paper was that REL is best, but not as conservative as the other methods.  What is the minimum number of sequence you would use?
[/quote]

Run REL - also take a look at the J Mol Evol paper (http://www.springerlink.com/content/x471k2187q966135/) for how to run REL tests when false positive rates are a concern.

[quote]For the above methods, is it possible to calculate the power to detect pos/neg selection on a site-by-site basis?  
[/quote]

Again, the JME paper talks about that to an extent.

Cheers,
Sergei