The cost of winsorizing is highest when treatment effects are concentrated in the tails. Medical claims are the canonical example: interventions often move the right tail, so winsorizing imputes away the very variation that identifies the effect!
Paul Goldsmith-PinkhamI don't think there's much of a good justification to winsorize data -- almost all "robust statistic" approaches would point to either censoring or something else. Winsorizing is basically imputing data with very limited justification in order to maintain statistical power