Definition
Bootstrap variance is the sample variance of a statistic computed across bootstrap replicates. If theta-star-b denotes B replicate estimates, a common estimate uses the B-minus-1 denominator. It estimates sampling variability under the chosen bootstrap scheme, not the variance of the raw observations unless the statistic is the raw observation itself.
How it is calculated
Bootstrap variance estimate = sum over B bootstrap statistic estimates of squared deviations from their bootstrap mean, divided by B minus 1.
| Symbol | Meaning | Unit |
|---|---|---|
| Var | variance operator | squared statistic units |
| theta | target statistic | statistic units |
| B | number of bootstrap replicates | count |
| \bar\theta^* | bootstrap replicate mean | statistic units |
-
Compute the statistic for each bootstrap resample.
-
Average the replicate statistics.
-
Sum their squared deviations from that average and divide by the number of replicates minus one.
-
Take the square root when a bootstrap standard error is required.
Worked example
Synthetic example.
-
inputs:
- bootstrap estimates: 1, 2, 3, 4
-
result:
-
bootstrap mean: 2.5
-
sample variance: 1.6666666666666667
-
Interpretation
Each bootstrap replicate produces one plausible value of the statistic. The spread of those values is the uncertainty estimate. If the resampling mechanism is wrong, precise-looking spread can still describe the wrong sampling process.
Assumptions and conventions
-
The bootstrap resampling scheme is appropriate for the data dependence and statistic.
-
B >= 2.
-
B is large enough for the variance estimate to be numerically stable.
-
The B-minus-1 convention is used consistently.
Squared units of the statistic; B is dimensionless.
Variants
- Bootstrap variance is not automatically population variance.
Limitations and common misreadings
-
Bootstrap variance is not automatically population variance.
-
The denominator B versus B-minus-1 is a convention that must be stated.
References
- B. Efron. (1979). Bootstrap Methods: Another Look at the Jackknife. The Annals of Statistics. Sections 1–2. doi:10.1214/aos/1176344552. Source

