Bayes & statistics
Bootstrap: estimating uncertainty by resampling
How uncertain is an average processing time based on just 30 cases? Resample the same observations and compare the mean, median and the influence of an unusually long case.
The short answer
The nonparametric bootstrap repeatedly samples the observed data with replacement. Variation in the recalculated statistic estimates sampling uncertainty; it does not create new information about unobserved cases.
Explore the interactive exampleFree · in your browser · no sign-up
At a glance
- Every bootstrap sample has the original sample size and can contain observations more than once.
- An interval for a mean is not a range for individual future observations.
- Representativeness, dependence and outliers still matter; more resamples cannot repair unsuitable data.
Thirty completed cases took different amounts of time. Calculating their average is straightforward. The harder question is: How much might that average vary if we observed another comparable sample? Bootstrap addresses this uncertainty. It is not a technique for turning a small dataset into a large collection of genuinely new observations.
Using the observed sample as a model
The nonparametric bootstrap replaces the unknown distribution with the empirical distribution of observed values. Every observation has the same probability of being selected. From the original values, draw another values, this time with replacement. Some original cases are absent, while others occur several times. Efron’s original paper describes this general resampling principle (Efron 1979).
Calculate the same statistic, such as the mean or median, on every resample. After repetitions, the resulting values are . Their spread estimates uncertainty in the original estimator. The star denotes a repetition generated from the sample, not another real measurement.
A tiny example you can calculate
Consider just two processing times: 0 and 2 minutes. This is deliberately artificial. Their mean is 1. Samples of size two have four equally likely ordered pairs: (0,0), (0,2), (2,0) and (2,2). The corresponding bootstrap means are 0, 1, 1 and 2.
This distribution again has mean 1, variance 0.5 and standard error . The tiny example can be enumerated completely. The interactive version instead draws 2,000 actual resamples from synthetic processing times. A fixed random seed makes comparisons reproducible; the chart is calculated rather than supplied as a preset result.
From a histogram to an interval
For the simple 95% percentile interval, sort the bootstrap estimates and take their 2.5th and 97.5th percentiles. The histogram’s centre shows plausible repeated estimates of the statistic, while its width shows estimated uncertainty. Efron compares bootstrap with other ways of estimating standard errors (Efron 1981).
The default synthetic sample of 30 cases has a mean of 10.85 minutes. With 2,000 resamples and the fixed seed, the interval is approximately 9.45 to 12.37 minutes. Different data produces different limits. These numbers explain the tool; they are not a benchmark for any business process.
Interactive: Bootstrap: estimating uncertainty by resampling
Without JavaScript, the two-point example remains fully checkable: bootstrap means 0, 1, 1 and 2, with standard error about 0.707. The calculator uses 2,000 resamples.
Mean, median and one unusually long case
Increase the first processing time to 60 minutes. The mean responds directly because every minute contributes to its sum. The median depends mainly on ordering and often moves less. That does not make the median universally better: the total workload of a queue may depend heavily on long cases.
Choose the statistic before inspecting which answer looks more attractive. Are you describing a typical processing time or planning total capacity? Also document why the long case exists. A valid rare observation is not a measurement error and should not be removed merely because it widens the interval.
What the method assumes
The simple version treats observations as independent and identically distributed. If many cases come from the same team or day, shared influences can violate this assumption. Resampling at group or block level may then be appropriate. A biased sample remains biased after any number of bootstrap repetitions.
The empirical distribution also cannot invent extreme events that have never been observed. Rare failures, maxima and very small samples need additional modelling considerations. More repetitions reduce only the simulation noise of the bootstrap calculation. More suitable original observations can reduce uncertainty about the process itself.
Using the result properly
Report the original sample, statistic, number of resamples and interval method together. A frequentist confidence interval is not a direct posterior probability; Bayesian statistics explains the distinction. To investigate consequences of uncertain processing times, such as a completion deadline, a suitable simulation model is needed next.
Keep uncertainty about the average separate from variation between cases. A narrow mean interval can coexist with widely varying individual durations. This distinction matters when setting service commitments: predicting one customer’s waiting time is a different task from estimating the organisation’s average.
Frequently asked questions
Why sample with replacement?
Without replacement, a sample of the same size would only rearrange the original values. Mean and median would never change. Replacement produces different empirical samples (Efron 1979).
Is the percentile interval always reliable?
No. Small samples, strongly biased estimators and nonsmooth statistics can reduce coverage. The simple percentile interval is an introduction; other bootstrap methods and analytical comparisons may be needed (Efron 1981).
Sources and further reading
- PaperEfron, B. (1979): Bootstrap Methods: Another Look at the Jackknife.
- PaperEfron, B. (1981): Nonparametric estimates of standard error: The jackknife, the bootstrap and other methods.
Content last reviewed on 14 November 2026. Please let us know about any errors: impressum@nyxai.com.