Pairs cluster bootstrap
The pairs cluster bootstrap resamples whole observed groups with replacement and recomputes a statistic while retaining each selected group's members.
For G observed groups, sample G group identities with replacement, retaining every member each time a group is selected. With unequal group sizes, the resulting row count can vary. Recalculate the chosen statistic and its denominator in each resample; a pooled device proportion is not generally an equally weighted mean of rack proportions. [1] [3]
Whole-group resampling preserves observed within-group dependence. Inference still requires a defensible sampling design and independent sampling units at the chosen level. Rack boundaries may be crossed by shared power, firmware or time effects; a rack label does not establish independence. Few clusters can make inference unreliable despite many resamples. [1]
The pairs variant differs from wild cluster bootstrap, which uses cluster-level perturbations rather than redrawing whole groups. Neither the name nor the choice of groups guarantees confidence-interval coverage. [1] Grouping is also distinct from stratification, which samples within designated strata. [2]
Relations
- related
- Bootstrap, Confidence interval, Design effect
Citations
- Cameron and Miller (2015): A Practitioner's Guide to Cluster-Robust Inference, sections II.F and VI
- rsample: Group Bootstraps and grouped sampling with replacement
- rsample: Common Resampling Patterns, grouped resampling and unequal group sizes
Sources & caveats
- Cameron and Miller discuss statistical methodology, mainly regression, not measured storage reliability. Applying the grouping cautions to drives and racks is a methodological interpretation.