PIC spec sheets. On PIC18F452/252, they're Fig23-28 & 23-29, pg 302/302.
Actually, looking at them again, the inaccuracy goes up 3.5x LSBs when Vdd AND VrefH drop to 2.5V.
If you drop VrefH from 5V to an external 2.5V PVR, the inaccuracy only goes up 2.4x in terms of LSBs. The LSB is 2x finer, so the net inaccuracy is now 1.2x, or 20% poorer. Also, the spec only bothers to list an offset err of +/-1.5V at Vdd=Vref=5.0V, but logically that offset error is going to be a fixed DC voltage (+/-7.33 mV) and will not decrease will a decrease in Vref. I could be wrong, but I doubt it. Anybody want to do a precision experiment? Or maybe I should just ask Microchip to be sure.
Of course, if this is versus using an LM7805 with a tolerance of +/-4%, that's +/- 41 full scale codes right there so it dwarfs the internal PIC inaccuracies IF you're measuing an absolute voltage. However, some sensors- many barometric ones included- are ratiometric, they do NOT give absolute voltage but rather a proportion of Vin. This totally cancels out the inaccuracy of the reg. That's the situation for any thermistor being pulled up to 5V through a resistor, for example. Even if the reg gave 4.8V and VrefH is 4.8V, the thermistor will have no additional error from the reg.
In fact, if you've got a ratiometric input, an external voltage reference will reintroduce the regulator error into the calculation since the sensor still gives output as a ratio of the regulator voltage.