Edge effects in an MOS can be significant contributor to threshold or ratio errors.
Firstly, the edge of an MOS is not a brick wall. It extends slightly out or in, and thus 10um width is not exactly 10um, but say 10.5um width. If you just double the width and the edge effects stay constant, then you'd be comparing 10.5um to 20.5um, not exactly a great ratio.
Secondly, other implants nearby can seep in and change your thresholds. I worked in a process where the stop-implant seeped in more then 7um to raise the threshold of the MOS. When MOS were not unit-sized, then the ratio's were horrible.
So for good matching, you absolutely have to have unit sized MOS. Unit sized MOS means all MOS's are groups of the same dimensions, so a 5:1 ratio would involve 5 parallel MOS on one side, and 1 MOS on the other.
In some companies they simply don't allow non-unit sized mirrors out of the designers. I don't believe in such harsh rules because there are situations where you don't care so much about accuracy but do care about area. But if matching is of concern, make sure everything is made from the same MOS.
As a side note, some places even demand from their engineers that the current goes in the same direction. They try to eliminate doping gradients from affecting matching. If you have a common centroid layout then the net gradient effect is averaged out, however, this assumes that every MOS in a given coordinate has the same Vt offset if replaced by another. The problem is that the MOS doesn't occupy a single point, but an area that has a gradient. Furthermore, the MOS is not symemtrical because it has a pinch-off region. This means that depending on what orientation the MOS has, the pinch off region will occupy a different doping intensity, and hence the average doping offset for the active channel is different. If on the other hand the MOS is in the linear region and there is no pinch-off, then you could consider it symetric and the average of the doping is the same regardless of the MOS orientation.