Any change in leverage based on distance from center is proportionally offset by the distance the motor has to travel to achieve the same amount of total pitch/roll deflection. Eg., roughly speaking, the motor/prop on an arm 2x longer than another will have 2x more torque but will also need to travel 2x further to achieve the same overall tilt angle.
OTOH, since there are 3 motors/arms per side on the roll axis, this makes it heavier than the pitch axis, and may require a bit more power. So increasing the output a bit might help.
There are many other factors which could affect one axis over another, or even just one motor over another. Unless your setup is very unique somehow, I'd tend to start with what has worked for most people (the default hexa mix in QGC). Not saying it's perfect for every situation, but it was mathematically derived, if that makes you feel any better.
Ultimately the "divergence" is the inability of the PID controller to keep up with real-world physics. So either changing the physics or adjusting the controller inputs should fix the problem. It would be cheaper to test this on a rubber-band setup like in some of Menno's tuning videos.
-Max