The human speech system can issue a broad range of phonemes, so you can get a lot of different sounds that all have the possibility of becoming roots in their own right. At that point, a language actually using them (so, not oligosynthetic) will give you more "bandwidth", and will naturally be preferred from the beginning of each language family. Other pro-oligosynthetic considerations would arise only much later.
Similarly, we could count in binary positional, but we started using the decimal (in some case, duodecimal or vigesimal) system, and in the beginning it wasn't even positional.
Nowadays (or in the future), one possibility could be some sort of Newspeak imposed from above, for social control or politic/religious reasons. For the language to remain oligosynthetic, though, you would need a very controlled language and media - in short a very controlled society, probably a police or theocratic state.
Another possibility (not for humans) is if you have several signal channels (instead of one as humans do), but each of them has only a few symbols and next to no modulation. It is the case of cuttlefish, that have few "arm" roots and some fifty chromatophoric "suffixes". In such a situation, the longer-ranged or most widely available channels would tend to be used for the roots, if at all possible; the others would supply suffixes. This would allow for a basic meaning to be conveyed even when some channels are unavailable, not unlike what happens with human kinesics.