Unfortunately it seems so that there are no existing publications/statistics from av vendors. Reffered to [A] and the mentioned entropy analysis, 90% of the used samples are obfuscated by a polymorphic technique.
[A] Toward Generic Unpacking Techniques for Malware Analysis with
Quantification of Code Revelation - 2009
Perhaps a own analysis could help with a bigger set of malware samples. The following "evalualtion" of me was done with the public available kaggle malware set. Even if it is not a clear classifcation of the used obfuscation technique of the families, hopefully this approach could help you out or point you to the right direction.
I did the classifaction with the help of different statistics like entropy value, chi-square distribution and a pi approximation.
NORM COMPR ENCR
-----------------------------------------
581 691 260 Ramnit
2475 3 0 Lollipop
0 10 2932 Kelihos_ver3
6 51 418 Vundo
3 22 18 Simda
233 260 258 Tracur
387 4 7 Kelihos_ver1
524 667 37 Obfuscator
779 219 15 Gatak
NORM = non-obfuscated characteristics (i.e. not compressed/encrypted)
COMPR = compressed or packed characteristics
ENCR = encrypted characteristics