well, I indeed ment that the number itself is not being changed.
Say: you have a 1 bit number:
(can only be 0 or 1)
And you convert it to a 2 bit number:
can be 00, 01, 10, 11, so you get 4 values.
Normally when you have a 1 bit value which can only be 0 or 1, the 2 bit value would be reduced to a 1 bit first.
So values 3 and 4 are being set as 1, where the 0 and 1 are being set as 0.
The other way around works too.
1bit 0 would be converted to either 2 bit 1 or 2, while the 1bit 1 would be converted to a 2 bit 3 or 4.
As you can see by this small equasion, its already not precise. In fact, its as presice as the 1 bit value, but we've established that already.
Say you put everything at 2 and 4, and then you start doing calculations with them.
It could be possible that you have some effects that move the 4 bits to a 3, and another effect that blends the 2 and 3 to say a 2,5 where since its only 4 bits, is being rounded to a 3. That means that all the 2's are now set to a 3. Converting it back would make them all a 1.
I know what I wrote here is kind of vague, but I don't know a better way to describe it.
I just say that if you first upsample, and then add effects, just so you are going to downsample it at a later stage, you might risk loosing quality due to the conversion process itself, where a higher quality can be archieved if the conversion doesn't take place at all.