Selecting the Best when Selection is Hard

In dynamic promotion contests, where performance measurement is noisy and ordinal, selection can be improved by biasing later stages in favor of early leaders. Even in the worst-case scenario, where noise swamps ability differences in determining relative performance, optimal bias is i) strictly positive; ii) locally insensitive to changes in the heterogeneity-to-noise ratio. A close relationship with expected optimal bias under cardinal information helps explain this surprising result. Properties i) and ii) imply that the simple rule of setting bias as if in the worst-case scenario achieves most of the potential gains in selective efficiency from biasing dynamic rank-order contests.