Deleting people from rolls would be unwise, since it is easier to detect when the people affected go to vote and detect they are not allowed.
In a busy precint, removing one of each ten means that, while the first person affected is still trying to solve it/argue with election officials, the next one who has been deleted will arrive and join in... in few time you will have an actual crowd of angry people at most of the precincts, getting nationwide attention. If you are doing something criminal, usually you do not want to attract nationwide attention1.
And yes, in Michigan you could still vote with a provisional ballot.
A different tactic would be adding to the rolls people who are not allowed to vote, or even unexistent people (so people with fake IDs would vote twice or more). Detecting these is more of a problem, since it does not involve innocent people. Still, it is a significant risk due to:
You need a lot of people to effectively go and knowingly vote illegally, risking someone reporting the information to the authorities.
Voter rolls are to be verified and are. Anyone could check why so many people share the same address, or why some guys have an address in a corn field. Of course, the more voters you add, the more risk.
And those cases are detected. Although I would say that this case was more one of registration workers trying to "meet objectives" with made up registers than of attemptin at falsifying the election2, it was indeed detected and punished.
Third, and more subtle, could be changing the census data in order to get a red or blue state more electors due to a higher population. This would be somewhat related to option #2, differing in that:
If you are in the right position you need less people to take part (less risk of people talking out), but
You need to fake a LOT more of data, increasing the risk of the wrong data being detected.
For all of these, it has to be considered that:
a) Your manipulations are likely to show inconsistencies with other data (e.g., way more voters enrolled that the census numbers would predict).
b) You will not know beforehand if the risk is worth it. If someone knew a year before that Michigan race was going to be that close and would be the one that decided the POTUS, maybe he could consider the risk of being caught modifying a few hundreds or thousands of votes justified by the possibility of changing the government. But if he does not know if those hundred or thousands are going to be decissive or just "noise", there is way less incentive to attempt fraud.
1At least I don't want to. YMMV.
2Unless someone planned to show at the poll station with a "Mickey Mouse" ID.