Welcome to the world of circuit synthesis.
Suppose you want a 20 bit system, with +-5 volts into the ADC. The resistors have 5 PPM temperature coefficients (perhaps these are Vishay metal-film resistors). The resistors, plus the PCB traces and the PCB FR-4 dielectric and the various VDD and GND planes and the metal chassis of the shielding case, provide 100 degree Centigrade per watt thermal resistance. The resistor thermal time constant is 11 milliSeconds; the PCB time constant is several seconds. Can we achieve 20 bits SINAD (signal to noise + distortion)? Can we keep the nonlinearity below 1 bit? Can we keep the self-heating of the resistors below 1 bit, or 1PPM?
For 1PPM, we need 0.2 degree Cent heating. At 100 degree C per watt, and we only budget 0.2 degree, we can only dissipate 2 milliWatts in the resistors.
With 5 volts across the resistors, what value is required?
P = V^2 / R; R = V^2/P = 5*5 / 0.002 = 25 * 500 = 12,500 ohms.
Now................are you able to achieve the Johnson Noise floor needed for 20 bits?
1K ohm in 1 Hz bandwidth is 4 nanoVolts RMS; in 1MHz Bandwidth, expect 4 microVolts.
That 12,500 ohm resistor will generate sqrt(12,500 / 1,000) = sqrt(12.5) ~~3.5X more noise,
or 4uV * 3.5 = 14 uVolts RMS.
Yet the random noise budget for 20 bit system, with 5 volt fullscale, is what?
5uV RMS?
Thus we are boxed in between nonlinearity of thermal heating, and the random noise floor.