In principle, this manipulation could invalidate the RD design if savvy restaurateurs heap directly above each Yelp rounding threshold; the assignment of restaurants to either side of the threshold would then no longer be quasi-random.
However, we show that a restaurateurs incentive to leave a fake review does not change discontinuously at the Yelp rounding threshold, and a rich set of robustness checks on the density of restaurants and on reviewer characteristics confirm that restaurants are not manipulating ratings in a manner that causes them to land right above the Yelp rounding threshold.