Generally speaking, noise falls into at least two categories. One, of course, is a totally random phenomenon, resulting from the sum of all unwanted signals within the system. some of these are likely to be periodic so accumulating and averaging brings with it the risk that one will bias the result with cumulative error.

Another is inherent and not-necessarily unbiased or "balanced" effects on the conversion system. I'm not sure what the Philips folks meant by "unadjusted" error, but there could be noise in the reference, noise on the supply as well as internally generated irregularities within the ADC as well as within the MCU that can affect the outcome.

Not only that, but this figure is likely based on a nearly perfect reference and a nearly perfect supply, combined with a running and possibly very slow program that does as much as it can, meaning as little as it can get by with, to minimize the effects on the ADC. While I'm sure that there are people who know much more than I about the inner workings of this particular MCU, and particularly about the specifics of this ADC, there's reason to question whether one can make use of more than five or six bits of its output. The figures on page 67 (section 13.2) of the datasheet suggest that the best one can rely on without extensive oversampling and averaging is one part in 32, or 5 bits.

If you know that your reference voltage is absolutely correct and stabile, and that the analog supply is clean and stabile, then you can take the value from the ADC, knowing how it behaves, which you likely will have to acquire through extensive and careful experimentation.

Truncating the likely error bits will not eliminate some "flickering," as Kai has pointed out, and averaging has the hazzard I mentioned above, that it accumulates systematic error. Averaging a truncated value carries with it the additional risk that a small error, because it causes a large transition, as in the move from 01111111b to 10000000b looks like a 4x larger variation when truncated to 6 bits. Totally random flicker will disappear in averaging. Unfortunately, it is often difficult to predict what else will be happening in the system when one is reading an ADC, hence, periodic effects can have significant impact.

One way to skirt these matters is to vary the sample rate. That will make the periodic effects appear random, while having no effect on the impact of statistically random events. It will not negate the effect of those two bits of "unadjusted error", nor will it negate the effect of those two bits of offset error.

I guess that's an "exercise for the student."

RE