Definition
Bootstrap consistency means that the conditional distribution produced by the bootstrap approximates the sampling distribution of the statistic as sample size grows, under stated regularity conditions. It is a property of the statistic, data dependence and resampling scheme together. Ordinary iid bootstrap can fail for nonsmooth statistics, strong dependence or boundary problems.
How it is calculated
The bounded-Lipschitz distance between the centered and scaled conditional bootstrap law of a_n(theta-hat-star minus theta-hat) and the sampling law of a_n(theta-hat minus theta) tends to zero in probability, where a_n is a dimensionless normalization rate, usually square root of n.
| Symbol | Meaning | Unit |
|---|---|---|
| d | distance between probability laws | dimensionless |
| BL | bounded-Lipschitz metric | none |
| L | probability law | distribution |
| \mathcal L^* | law conditional on original data | distribution |
| theta | statistic target | statistic units |
| n | sample size | count |
| a | normalization rate, usually square root of n | dimensionless |
Interpretation
A consistent bootstrap is a resampling machine that eventually reproduces the uncertainty shape of the real estimator. The word eventually hides conditions about smoothness, dependence and boundaries. More resamples improve Monte Carlo precision but cannot make an inconsistent resampling scheme consistent.
Assumptions and conventions
-
The statistic and data-generating process satisfy the chosen bootstrap's regularity conditions.
-
The distance metric and normalization are specified.
-
Dependence is handled by iid, block, wild or model resampling as appropriate.
The law distance is dimensionless; theta retains statistic units.
Limitations and common misreadings
-
Bootstrap consistency is not automatic from using many replicates.
-
Consistency of the bootstrap distribution does not ensure a good finite-sample approximation.
References
- Peter Hall. (1992). The Bootstrap and Edgeworth Expansion. Springer Series in Statistics. Chapter 1, Introduction. doi:10.1007/978-1-4612-4384-7. Source

