Why is it a bad idea to divide reference voltage only with resistors instead of using an op-amp buffer?

Question

I'm planning on shifting the output of an AD8226 instrumentation amplifier by 'about' 0.5 volts using a 2.5 v precision reference and a voltage divider. According to the datasheet this is a bad idea and a buffer must be used:

I understand that if I don't use a buffer I'll probably end up with a slightly different voltage shift and a slight increase in gain. I'm guessing these can be calibrated out by software without a problem. If yes, then why should I use a buffer? Also, there is a note at the last line about a degradation in CMRR. I'm using the device for reading a 0-10 v (very slowly changing) signal in an industrial environment (single ended). Will the CMRR degradation be significant?

P.S. Voltage divider used has two resistors 20k and 4.7k.

Answer 1

Because "for best performance, the source impedance to the REF terminal should be kept below 2 [ohms]".

If you made a resistive divider to meet that requirement, it would be 4 ohms in each arm, or 8 ohms between power and ground, which would likely consume a lot more power than you'd want to use for this function.

Why does the impedance need to be below 2 ohms?

It's explained in the text you posted:

Answer 2

Have a look at the (simplified) internal schematic of the AD8226:

Obviously having any significant (relative to 50K) impedance in series with the REF pin will negatively affect the common mode rejection. The recommendation of 2 ohms represents a degradation in the matching of 0.004%, or about 88dB. The guaranteed CMRR is 86dB at DC (gain = 1). The gain from the positive input is increased, not the negative input.

Does that affect your application? You'll have to run the numbers and see. There's perhaps little point in spending the money for an instrumentation amplifier if you're going to degrade it as any reasonable values of divider resistors will. By the way, there is no guarantee the resistors R3..R6 will be especially stable or accurate in absolute value. The requirement is for matching to achieve CMRR, not absolute accuracy, so you'll also be degrading stability most likely because of mismatch between tempcos.