Since EAX is very flexible with regard to the length of the calculated MAC (what they call "tag" is the MAC value right?), and the EAX paper as well as other documentations carefully avoid suggesting any value for the length of the MAC, is there anything to take into consideration before setting for a value?
Is the length just "proportional" with collision resistance or are there also other more subtle consequences depending on the MAC length?
While the EAX mode permits truncating the tag to any length $\tau$ between 0 and $n$ bits, where $n$ is the block size of the underlying block cipher, this should only be taken as a statement that truncating the tag in EAX mode does not introduce any security issues beyond the obvious (an attacker only needs $2^\tau$ attempts to froge a $\tau$-bit tag by brute force). Still, in the absence of a particular space or bandwidth constraint that would require a shorter tag, I see no reason not to use a full-length tag ($\tau = n$).
(I suspect the main reason for introducing a variable tag length into EAX was because CCM had one; see the table in figure 4 of the EAX paper you cited. EAX was designed as a replacement for CCM, and so one of the main design goals was to be able to do anything CCM did, only better. If EAX had been specified with a fixed tag length, someone might have cited that "missing feature" as a reason to use CCM over EAX.)
