Choose the Bootstrap Resampling Unit Before Generating Replicates

Identify the independently sampled unit

The ordinary bootstrap resamples observed units with replacement and recalculates a statistic. The unit being resampled is a substantive choice. Rows can represent people, visits, eyes, specimens or clusters; resampling the wrong row type can erase dependence and exaggerate precision.

The SciPy bootstrap documentation describes resampling and a paired option that uses common indices across paired arrays. More complex designs require an appropriate design-aware procedure rather than assuming that one library option covers every correlation structure.

Draw the data hierarchy

List independent sampling units and measurements nested within them. For paired before-and-after data, resample people while retaining both measurements together. For a cluster sample, consider resampling clusters with their constituent observations, under a method appropriate to the sampling design. Time-series dependence may require blocks or a model-based approach rather than independent row resampling.

Define what happens to unequal cluster sizes, strata and model-fitting failures. The bootstrap procedure should reproduce the estimation pipeline, including any transformations or fitting steps that contribute to the reported uncertainty.

Inspect a hypothetical pair sample

Suppose three people have paired values (2,3), (4,6) and (8,9). Their changes are 1, 2 and 1, giving mean change 4/3. A bootstrap index sample of first, first, third should produce paired observations (2,3), (2,3), (8,9), with changes 1, 1 and 1.

Resampling baseline and follow-up arrays independently can create pairs that never existed, changing the dependence structure and therefore the uncertainty calculation. Matching array lengths does not repair this error; the same participant indices must move together when the estimand uses paired data.

Check what the interval can support

Resampling does not create information about populations absent from the observed sample. Very small samples, rare events and unstable estimators can produce degenerate or misleading bootstrap distributions. Inspect the distribution, failure count and sensitivity to the resampling plan before interpreting an interval.

Record the number of replicates, random-stream settings, interval method and independent unit count. The number of bootstrap replicates controls computational precision, not the original study’s sample size. A completed bootstrap analysis should make the dependence-preserving unit visible and show that the procedure resamples the study’s information structure rather than merely drawing rows until the software returns an interval.

Source verification: 6 November 2026. Examples are hypothetical.