One thing to watch out for is the voltage co-efficient of resistance, which can be surprisingly high in surface-mount high-resistance cermet or metal-oxide “thick film” units. “Thin film” (similar to through-hole metal film) is better, but hard to find in high-megohm values in small packages. Even in 10 volt circuits, I had to specify 1206 premium parts (instead of 0805) where I needed 50 megohm feedback resistors. It is hard to obtain the value of this spec for high-megohm units, specified as ppm/V. If you need 0.1% linearity at 100 V signal level, you need better than 10 ppm/V. This is in addition to the self-heating plus temperature co-efficient, which is also important, although self-heating improves at higher resistance due to lower power dissipation.
At the internal level, for a given resistive material, the deviation from the Ohmic linear approximation is a function of the linear voltage gradient (V/mm) which gets worse for physically short smd packages.