You would need to lower both R1 and R2 to maintain the same ratio, but really this isn't the right part for precision DC measurements using high impedance dividers.

The datasheet says this part has a typical input bias current of 80nA @ 25C. You have an input impedance of 10M||2M = 1.66M, 1.66M * 80nA = 0.133v. Assuming your Arduino reference is 5v, your scaling factor would report this as 0.8v, close to what you are seeing. The worst case input bias current (i.e. what you should be designing for) is 200nA @ 25C.

Also note that the output can only get within about 50mV of the rails, irrespective of the divider so there will always be an offset to deal with.

Some problems like this can be resolved with software calibration, but input bias current is tricky since it can vary with temperature and common mode voltage.