Estimate sampling variability without inventing new information

Try

A sample contains four independent values: 1,3,6,10. Imagine drawing four times with replacement from these observed units. A resample might be 1,1,6,10, with mean 4.5. This is not a new field sample from the population; it is a draw from the empirical distribution supported on the recorded observations.

Predict

Predict why sampling without replacement with the same sample size would produce no variation in a symmetric statistic such as the mean. Every resample would contain all four original values. Replacement creates varied multiplicities, approximating how a statistic might vary under repeated sampling from the estimated distribution.

Experiment

Compare several possible resamples, without treating this tiny list as a reliable confidence calculation.

Resample	Mean
1,1,6,10	4.5
3,3,6,6	4.5
1,3,10,10	6
6,6,6,10	7

The original mean is 20/4=5. A large collection of correctly generated resamples describes a bootstrap distribution of the mean. Its spread can estimate sampling uncertainty under suitable assumptions. These illustrative rows were chosen to explain the mechanism, not generated as an independent simulation or used to estimate quantiles.

Understand

The ordinary nonparametric bootstrap samples n units with replacement from an original sample of n units and recomputes the statistic. Repeating this many times approximates a distribution under the empirical population. It can estimate standard errors and support interval methods when the original sampling design and the statistic make the approximation appropriate.

A simple percentile interval takes specified quantiles of the bootstrap statistic distribution, such as the 2.5th and 97.5th percentiles for an approximate central 95% interval. Its coverage is not guaranteed merely because those percentiles are available. Bias, skew, small samples, parameter boundaries and nonsmooth or extreme statistics can undermine the approximation. More advanced bootstrap intervals address some problems under their own conditions.

Resample the correct units. For paired observations, keep each pair intact or resample its difference. Independent resampling of before and after columns destroys their association. Clustered or time-dependent data may require cluster or block methods; the ordinary independent-unit bootstrap does not automatically handle them. Design awareness remains necessary.

The bootstrap cannot recover unobserved population structure reliably from a poor sample. If rare large values never entered the sample, resampling observed values will not invent them. It cannot repair volunteer selection, systematic measurement error or outcome-related dropout. Increasing the number of resamples reduces Monte Carlo error conditional on the data, not the original sample’s bias or lack of information.

Ordinary bootstrap estimation is not automatically a null hypothesis test. Its distribution is generally centered near the observed statistic rather than the proposed null value. A bootstrap test needs an appropriate null-generating method, centering or other construction justified for that problem. Counting ordinary resamples above zero does not by itself produce a valid test of mean zero.

Record the resampling scheme, statistic, number of resamples, random seed when relevant, interval method and limitations. Reproducibility makes the computation inspectable; it does not establish statistical validity by itself.

Apply

Specify a bootstrap plan for a paired time study, naming what is sampled and what statistic is recomputed. Then explain how more resamples could stabilize a reported percentile while leaving biased recruitment untouched. In the written practice, interpret supplied percentiles as an approximate procedure, not proof that the fixed parameter has been assigned a posterior probability.

Continue

This unit supplies a substantial inference foundation, not a complete AP examination course or advanced statistics sequence. Independent projects, complex survey methods, regression inference and deeper resampling theory remain separate work. Use the method-selection lesson when a new dataset does not fit a familiar template.