MPU9250 temperature compensation

Hi,

I have a question regarding the temperature compensation of the MPU9250 accelerometers.
I took several readings from 0 to 20°C ambient temperature for all axis. The x and y axis showed a near linear temperature dependency. But the z-axis showed a nonlinear dependency which is NOT reproducible! This means, the data for the z axis had different readings at the same temperature on different days.
And these differences are huge, like 15 LSB in 16g mode.

The x and y axis are showing less the 0.05% error over temperature from 0 to 20°C ambient.
I need the z-axis be as good.

Is there anything I am missing? Is there another dependency? like moisture? Is the device damaged?

Kind regards