Just a reminder of an important point raised above:
The non-linearity of the dielectric material is a function of the E field in the dielectric layer, which is the voltage gradient dV/dx, where x is the thickness for parallel-plate construction.
Similarly, non-linearity in resistors depends on the voltage gradient along the length of the resistive material.
The direct result is that for a given material, non-linearity (voltage dependence of capacitance or resistance) will be worse for smaller packages, which will also have lower voltage capability.
At least as long as the above assumptions hold. HV parts tend to be worse than LV parts, because they can be manufactured more consistently. That is, consider a 47uF 6.3V chip with, lord knows, thousands of layers? Whatever it might be. Those layers are extremely thin, 100s nm perhaps, at the limit of manufacturability; and not a single one can have a pinhole failure. There is some practical minimum layer thickness with respect to manufacturing yield. Which isn't to say denser, lower-voltage parts can't be made -- but it does mean you're going to pay through the tooth to use them.
Conversely, higher voltage parts, by the above reasoning, must use thicker dielectric to begin with, and can't fit so many layers into the package. Even large chips don't get much thicker -- compare 2220 and 1210 for example. So they're going for width more than height.
And yet, 470nF 250V X7R parts are available, for example; but the C(0)/C(Vmax) ratio is extraordinary, usually >10, maybe even 20 (i.e. 90-95% reduction at rating). With the lower defect rate of the reduced dielectric area and increased thickness, they can simply run much higher E field before failure occurs.
A corollary: the maximum voltage handling at breakdown, of any random part, might actually be
many times the nominal rating. It's the lowest breakdown, for that series (dielectric thickness and area), for some tolerable manufacturing yield, that determines the ratio between nominal rating and best-case actual breakdown.
I don't know if it's what they actually do, or if there are indeed structural differences between parts -- I would guess a bit of both in practice -- but this might be one explanation for the existence of a range of voltage ratings in otherwise perfectly identical parts (same C(V) and temp curve, same package outline and height, same nominal value). You should therefore prefer lower voltage parts for lower production cost -- at least in such quantities where the cost savings is justified, a reel change cost isn't incurred (or custom reeling can be done), etc. Probably 10-100k's/yr?
Tim