A Bootstrap Sample Can Repeat One Observation and Leave Another Out

Decorative acrylic columns arranged on grid paper as a visual metaphor for charts, not actual measured data.
Conceptual illustration created with AI; not a research result or laboratory photograph.

Resampling with replacement can look strange the first time you see it: one original observation appears twice while another is absent. In a basic nonparametric bootstrap, that repetition is part of the procedure.

Try a tiny labelled example

Imagine three observed values on cards: A has 2, B has 5, and C has 9. Draw a card, record it, replace it, and repeat until you have three draws. One possible resample is A, A, C; another is B, C, B.

Calculate the mean for each illustrative resample. They need not match, even though both come from the same three original observations. This toy exercise explains replacement; two resamples are not enough for a real uncertainty analysis.

Remember what resampling can and cannot add

NIST describes bootstrap resampling as a way to explore uncertainty in a statistic. It does not collect new observations or automatically repair a biased original sample.

Real applications require a resampling scheme appropriate to the data, especially when observations are dependent. In your notes, separate the original sample size from the number of resamples. Repeating a computation many times is different from having measured many more independent cases.