could you check out the SDCCH average duration ? do you notice an increase here as well ?
if it is the case, then you could simply perform a drivetest in a cell with this problem, and see where is the delay : during authentication ? ciphering ? tmi reallocation ?
a sdcch phase which is too long will lead to more ocupancy –> higher congestion.
you confirm that the sdcch erlang has increased above the sdcch capacity of the cell, right ?