A new article in the Journal of Economic Perspectives by Andrews, Moon, and Shapiro argues that researchers should default to the Bayesian bootstrap, rather than the more familiar nonparametric bootstrap. Conventional nonparametric bootstrapping is equivalent to randomly assigning each observation an integer weight (including 0), whereas Bayesian bootstrapping randomly assigns continuous positive weights instead. This has the convenient property that no observation is ever truly dropped in a given bootstrap replicate, which is handy when you have rare categories that can be altogether absent from some samples.

Bayesian bootstrapping is implemented in R in the bayesboot and fwb packages.

Hat tip to Jason Kerwin.