{
  "id": "bootstrap-statistics",
  "title": "Bootstrap (statistics)",
  "lang": "en",
  "source_lang": "en",
  "summary": "The bootstrap is a family of resampling methods for estimating statistical uncertainty by approximating repeated sampling from a population.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Bootstrap method",
    "bootstrapping",
    "bootstrap resampling"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/bootstrap-statistics/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/bootstrap-statistics/",
      "json": "https://aiwiki.page/zh/bootstrap-statistics.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The bootstrap is a family of methods in statistics that use repeated sampling from an estimated population to approximate the uncertainty of a statistical result. Its central idea is to replace an unknown data-generating distribution with an estimate constructed from the observed data, and then reproduce the original analysis on samples drawn from that estimate. The resulting variation provides information about a statistic’s sampling distribution, including its spread, bias, and shape.\n\nIn the ordinary nonparametric bootstrap, an observed sample of size \\(n\\) is treated as a population containing \\(n\\) equally weighted observations. Repeated samples of size \\(n\\) are drawn with replacement, so an observation may appear several times in one resample and not at all in another. More generally, bootstrap samples can be generated from fitted probability models, resampled residuals, blocks of consecutive observations, or weighted versions of the data. These alternatives address different assumptions about how the original observations arose.\n\nBootstrap methods are widely used to estimate standard errors, construct confidence intervals, assess estimator bias, and calibrate statistical tests. They are particularly useful when analytical formulas are difficult to derive. Their validity nevertheless depends on the statistic, sampling design, and resampling scheme: repeatedly analyzing the observed data does not automatically reproduce every important feature of sampling from the underlying population."
    },
    {
      "heading": "Origins and development",
      "text": "Bradley Efron introduced the bootstrap under that name in his 1979 paper “Bootstrap Methods: Another Look at the Jackknife.” The paper connected a computational resampling procedure with a general statistical principle: uncertainty under an unknown distribution could often be approximated by uncertainty under the empirical distribution of the observations. The name alludes to the expression “pulling oneself up by one’s bootstraps,” emphasizing the apparent circularity of using a sample to assess its own reliability.\n\nResampling ideas predated Efron’s formulation. In 1939, Maurice Quenouille’s later subject of work had a precursor in P. C. Mahalanobis’s use of interpenetrating subsamples for studying survey error; Quenouille introduced a systematic bias-reduction procedure in 1949, and John Tukey subsequently developed and popularized it as the jackknife. The jackknife repeatedly deletes observations and examines how the estimate changes. The bootstrap instead generates new samples, usually by sampling with replacement. This distinction gives the bootstrap access to a wider range of distributional features than a simple delete-one calculation.\n\nAn important early attraction was the possibility of estimating uncertainty for complicated statistics without deriving a separate variance formula for each one. The standard error of a sample mean has a simple expression, but the uncertainty of a correlation coefficient, a robust estimator, or a multistage fitting procedure may be substantially harder to calculate. Efron’s approach turned many such problems into repeated executions of an existing algorithm. Increasing computer availability made that substitution of computation for algebra practical.\n\nDuring the 1980s, research moved beyond standard errors to the accuracy of bootstrap confidence intervals and tests. Work by Peter Hall and others developed asymptotic explanations for why appropriately constructed bootstrap procedures can outperform first-order normal approximations. Bias-corrected and accelerated intervals, studentized procedures, and bootstrap methods for dependent observations became major branches of the subject. Efron and Robert Tibshirani’s 1993 book, An Introduction to the Bootstrap, helped establish the method as part of routine statistical practice.\n\nThe bootstrap also contributed to developments outside classical uncertainty estimation. Leo Breiman’s 1996 introduction of bootstrap aggregation connected resampling to predictive ensemble methods. Meanwhile, survey statisticians, econometricians, and researchers working with longitudinal observations developed specialized schemes that preserved features of their sampling designs. The historical development therefore produced a family of related methods rather than one universally applicable algorithm."
    },
    {
      "heading": "The empirical-distribution principle",
      "text": "Suppose \\(X_1,\\ldots,X_n\\) are independent observations from an unknown probability distribution \\(F\\). A population quantity of interest can often be written as \\(\\theta=T(F)\\), where \\(T\\) is a statistical functional. Its sample analogue is \\(\\hat\\theta=T(F_n)\\), where \\(F_n\\) is the empirical distribution assigning probability \\(1/n\\) to each observed value, counting repeated values with their multiplicities. Its cumulative distribution function is\n\n\\[F_n(x)=\\frac{1}{n}\\sum_{i=1}^{n}\\mathbf{1}\\{X_i\\leq x\\},\\]\n\nwhere \\(\\mathbf{1}\\) denotes an indicator function.\n\nThe bootstrap applies the same substitution at the level of sampling distributions. The unknown distribution of \\(\\hat\\theta-\\theta\\) under sampling from \\(F\\) is approximated by the conditional distribution of \\(\\hat\\theta^*-\\hat\\theta\\), where \\(X_1^*,\\ldots,X_n^*\\) are sampled independently from \\(F_n\\) and \\(\\hat\\theta^*=T(F_n^*)\\). A star marks a quantity computed in the bootstrap world. Conditional on the observed data, \\(F_n\\) is treated as fixed, and randomness comes from generating the resample.\n\nThis reasoning is a distributional form of the plug-in principle. The same principle estimates a population mean by the sample mean or a population variance by a sample-based analogue. The bootstrap extends the substitution to an entire repeated-sampling experiment. Its usefulness depends not merely on whether \\(F_n\\) resembles \\(F\\), but on whether the operation that maps a distribution to the statistic’s sampling behavior preserves that resemblance.\n\nFor the mean, the calculation is especially transparent. Write \\(\\bar X=n^{-1}\\sum_iX_i\\). Under the empirical distribution,\n\n\\[\\operatorname{E}_*(\\bar X^*)=\\bar X, \\qquad \\operatorname{Var}_*(\\bar X^*) =\\frac{1}{n^2}\\sum_{i=1}^{n}(X_i-\\bar X)^2.\\]\n\nHere \\(\\operatorname{E}_*\\) and \\(\\operatorname{Var}_*\\) denote conditional expectation and variance over bootstrap samples. If \\(s^2\\) is the usual sample variance with denominator \\(n-1\\), the bootstrap variance is \\((n-1)s^2/n^2\\), compared with the conventional estimate \\(s^2/n\\). The small finite-sample difference disappears asymptotically.\n\nThe empirical distribution also makes the method’s limits visible. It places no probability on values absent from the sample. Nevertheless, it can approximate uncertainty for many central statistics because those statistics depend mainly on well-represented features of the distribution. Estimation of extreme tails, unseen categories, or support boundaries is more demanding. The ordinary bootstrap does not reconstruct information about such features simply by repeating observed values."
    },
    {
      "heading": "Resampling algorithms, standard errors, and bias",
      "text": "For the ordinary bootstrap, the basic computational procedure begins with the original estimate \\(\\hat\\theta\\). For each replication \\(b=1,\\ldots,B\\), indices \\(I_1^{(b)},\\ldots,I_n^{(b)}\\) are sampled independently and uniformly from \\(\\{1,\\ldots,n\\}\\). The statistic is then calculated from \\(X_{I_1^{(b)}},\\ldots,X_{I_n^{(b)}}\\), producing \\(\\hat\\theta^{*(b)}\\). These repeated bootstrap samples approximate the conditional bootstrap distribution.\n\nFor a scalar estimate, the bootstrap standard error is\n\n\\[\\widehat{\\operatorname{se}}_{\\mathrm{boot}} = \\left[ \\frac{1}{B-1} \\sum_{b=1}^{B} \\left(\\hat\\theta^{*(b)}-\\bar\\theta^*\\right)^2 \\right]^{1/2}, \\qquad \\bar\\theta^*=\\frac{1}{B}\\sum_{b=1}^{B}\\hat\\theta^{*(b)}.\\]\n\nFor a vector-valued estimate, the corresponding calculation yields a covariance matrix. This permits joint uncertainty assessment for several coefficients or derived quantities, rather than treating each component in isolation.\n\nThe same replicates estimate bias:\n\n\\[\\widehat{\\operatorname{bias}}_{\\mathrm{boot}} =\\bar\\theta^*-\\hat\\theta.\\]\n\nSubtracting this estimate gives the bias-corrected estimator \\(2\\hat\\theta-\\bar\\theta^*\\). Such correction is not automatically an improvement. Estimating bias introduces additional variability, and a reduction in bias can be outweighed by an increase in variance. The relevant criterion may be mean squared error, which combines both. For some smooth estimators, analytical examination of the bootstrap expectation reveals useful finite-sample corrections; for irregular estimators, a numerical bias estimate can itself be unstable.\n\nA simple example illustrates the resampling mechanism. From observations \\(2,4,7,9\\), one possible resample is \\(4,4,2,9\\). Both samples contain four observations, but the resample repeats one value and omits another. Recomputing the mean, median, or another statistic across many such samples measures its sensitivity to changes in empirical frequencies. It does not generate four newly observed population members; it simulates the consequences of sampling from an estimated population.\n\nThe number of distinct observations in a bootstrap sample is usually smaller than \\(n\\). A particular observation is omitted with probability \\((1-1/n)^n\\), so the expected number of distinct sampled indices is\n\n\\[n\\left[1-\\left(1-\\frac{1}{n}\\right)^n\\right],\\]\n\napproximately \\(0.632n\\) for large \\(n\\). This does not mean that bootstrap samples should ordinarily be enlarged to contain \\(n\\) distinct cases. Repetition is an essential feature of sampling with replacement. Enforcing distinctness would change the sampling mechanism and therefore the uncertainty being approximated."
    },
    {
      "heading": "Confidence intervals and their calibration",
      "text": "A bootstrap standard error can be inserted into a normal-approximation interval,\n\n\\[\\hat\\theta\\pm z_{1-\\alpha/2} \\widehat{\\operatorname{se}}_{\\mathrm{boot}},\\]\n\nwhere \\(z_p\\) is a quantile of the standard normal distribution. This approach uses the bootstrap only to estimate scale. It can work well for approximately symmetric sampling distributions, but it does not directly account for skewness, parameter boundaries, or the dependence of standard error on the parameter.\n\nLet \\(q_p^*\\) denote the \\(p\\)-quantile of the bootstrap estimates. The percentile interval is\n\n\\[[q_{\\alpha/2}^*,q_{1-\\alpha/2}^*].\\]\n\nIts endpoints transform naturally under increasing transformations: applying a logarithm to positive bootstrap estimates gives the logarithm of the original percentile endpoints, apart from numerical quantile conventions. The basic, or reverse-percentile, interval instead approximates the distribution of estimation error:\n\n\\[[2\\hat\\theta-q_{1-\\alpha/2}^*, \\;2\\hat\\theta-q_{\\alpha/2}^*].\\]\n\nThese intervals use the same replicates but embody different approximations. Neither automatically has accurate coverage merely because many bootstrap samples were generated.\n\nThe studentized bootstrap, often called bootstrap-\\(t\\), uses\n\n\\[T^*=\\frac{\\hat\\theta^*-\\hat\\theta}{\\hat s^*},\\]\n\nwhere \\(\\hat s^*\\) estimates the standard error within a resample. If \\(t_p^*\\) is its bootstrap quantile and \\(\\hat s\\) is the original standard-error estimate, the interval is\n\n\\[[\\hat\\theta-t_{1-\\alpha/2}^*\\hat s,\\; \\hat\\theta-t_{\\alpha/2}^*\\hat s].\\]\n\nStudentization can make the distribution less dependent on unknown features of the population. It may, however, require a second layer of resampling when no direct formula for \\(\\hat s^*\\) is available, substantially increasing computation.\n\nThe bias-corrected and accelerated, or BCa, interval adjusts percentile levels using a bias-correction constant \\(z_0\\) and an acceleration constant \\(a\\). Typically,\n\n\\[z_0=\\Phi^{-1}\\!\\left( \\frac{\\#\\{\\hat\\theta^{*(b)}<\\hat\\theta\\}}{B} \\right),\\]\n\nwhere \\(\\Phi\\) is the standard normal cumulative distribution function. For delete-one jackknife estimates \\(\\hat\\theta_{(i)}\\), with average \\(\\bar\\theta_{(\\cdot)}\\), a common acceleration estimate is\n\n\\[a= \\frac{\\sum_i(\\bar\\theta_{(\\cdot)}-\\hat\\theta_{(i)})^3} {6\\left[\\sum_i(\\bar\\theta_{(\\cdot)}-\\hat\\theta_{(i)})^2\\right]^{3/2}}.\\]\n\nA nominal percentile level \\(p\\) is replaced by\n\n\\[p_{\\mathrm{adj}} = \\Phi\\left[ z_0+\\frac{z_0+z_p}{1-a(z_0+z_p)} \\right].\\]\n\nThe adjusted lower and upper quantiles form the interval.\n\nBCa and studentized procedures can achieve higher-order coverage accuracy under suitable smoothness and moment assumptions. Their advantages are not universal: discrete statistics, unstable jackknife estimates, ties, or nearly degenerate standard errors can create difficulties. An interval’s nominal confidence level describes repeated-sampling coverage of the procedure; it is not a posterior probability that the particular realized interval contains the fixed parameter."
    },
    {
      "heading": "Parametric bootstraps and regression models",
      "text": "The parametric bootstrap replaces the empirical distribution with a fitted member of a specified probability model. If observations are modeled by \\(F_\\eta\\), the parameter is estimated as \\(\\hat\\eta\\), often through maximum likelihood estimation. Each bootstrap dataset is generated from \\(F_{\\hat\\eta}\\), after which the entire fitting procedure is repeated. For Poisson observations, for example, replicate datasets can be generated using the fitted rate; for a normal model, they can be generated using fitted location and scale parameters.\n\nThis approach can reproduce features that an empirical distribution cannot, such as continuous variation beyond observed values or the sampling mechanism of a latent-variable model. Its additional structure can improve efficiency when the model is appropriate. The same structure creates vulnerability to misspecification. A normal parametric bootstrap cannot reveal the consequences of heavy tails that the fitted model excludes. Parametric and nonparametric bootstrap procedures therefore make different assumptions rather than representing universally better and worse versions of one method.\n\nIn linear regression, the resampling choice depends partly on whether predictor values are regarded as random or fixed. The pairs bootstrap resamples complete observations \\((X_i,Y_i)\\), preserving the association between a response and its predictors. Under independent sampling of observational units, it approximates the joint randomness of both. Resampling predictor and response columns independently would destroy their relationship and generate a different statistical experiment.\n\nThe residual bootstrap instead fits a model such as\n\n\\[Y_i=x_i^{\\mathsf T}\\beta+\\varepsilon_i,\\]\n\nand constructs responses\n\n\\[Y_i^*=x_i^{\\mathsf T}\\hat\\beta+\\varepsilon_i^*.\\]\n\nHere residuals are typically centered before resampling, and leverage adjustments may be used. The predictors remain fixed. Ordinary residual resampling relies on errors having an appropriate common distribution, usually including constant variance. If error variability changes across predictor values, pooling residuals can erase the heteroskedasticity that determines coefficient uncertainty.\n\nA wild bootstrap addresses this problem by multiplying fitted residuals by random variables with mean zero and variance one:\n\n\\[Y_i^*=x_i^{\\mathsf T}\\hat\\beta+\\tilde e_iV_i.\\]\n\nThe adjusted residual \\(\\tilde e_i\\) retains observation-specific scale, while the multiplier \\(V_i\\) introduces resampling variation. Different multiplier distributions and leverage corrections produce different variants. Related score and multiplier methods extend this principle beyond least squares.\n\nFor generalized linear models, a parametric bootstrap can simulate responses from the fitted conditional distribution at the observed predictor values. In multistage analyses, nuisance-parameter estimation and model fitting are normally repeated within each replicate when their uncertainty belongs to the target analysis. Holding those steps fixed instead answers a conditional question and can omit important sources of variation."
    },
    {
      "heading": "Dependence, clustering, and sampling design",
      "text": "Ordinary case resampling assumes that the resampled units are interchangeable in the way required by the original sampling mechanism. This is often inappropriate for repeated measurements, geographic observations, household surveys, or time series. Dependence contributes covariance terms to uncertainty. For a sample mean,\n\n\\[\\operatorname{Var}(\\bar X) = \\frac{1}{n^2} \\sum_{i=1}^{n}\\sum_{j=1}^{n} \\operatorname{Cov}(X_i,X_j).\\]\n\nResampling individual observations independently generally removes the off-diagonal dependence represented by these terms. Increasing the number of replicates cannot repair that change in the statistical experiment.\n\nFor clustered observations, a cluster bootstrap resamples whole clusters and retains their internal records. In a study of students within schools, independently sampled schools may be the appropriate resampling units rather than students. If treatment is assigned at school level, the school-level structure is particularly important. The effective amount of independent information is tied to the number of clusters; a large number of students within a handful of schools does not create a large collection of independent clusters.\n\nTime-series procedures often use a block bootstrap. The moving-block bootstrap samples blocks of consecutive observations and concatenates them to construct a new series. Overlapping blocks provide more candidate segments than a partition into nonoverlapping blocks. Circular versions wrap the end of the series to its beginning, while the stationary bootstrap uses randomly varying block lengths. These schemes preserve dependence within blocks, although dependence across artificial joins is modified.\n\nBlock length creates a statistical trade-off. Very short blocks omit relevant serial dependence, whereas very long blocks leave relatively few independently resampled pieces. In many consistency results, block length \\(\\ell\\) increases with sample size while satisfying \\(\\ell/n\\to0\\). The appropriate rate and data-driven selection criterion depend on the statistic and dependence structure. A method suitable for a short-memory stationary series need not work for a strongly trending process, a structural break, or long-range dependence.\n\nComplex surveys require resampling schemes that respect stratification, unequal inclusion probabilities, and multistage selection. Survey bootstrap methods may resample primary sampling units within strata and use adjusted replicate weights. Ordinary resampling of final records need not reproduce the design variance, especially when finite-population corrections or sampling without replacement matter.\n\nPreserving structure can also mean retaining matched pairs, entire family units, or complete multivariate records. In spatial data, blocks or model-based simulations may represent local dependence, but spatial geometry and boundary effects complicate their construction. The common principle is that the resampling unit and resampling distribution must represent the source of randomness relevant to the inferential target, rather than being chosen solely for computational convenience."
    },
    {
      "heading": "Asymptotic justification and higher-order accuracy",
      "text": "Bootstrap validity is commonly formulated as convergence of a conditional distribution. Suppose an estimator has a normalization \\(a_n\\), often \\(\\sqrt n\\), such that \\(a_n(\\hat\\theta-\\theta)\\) has a limiting distribution. A representative consistency requirement is\n\n\\[\\sup_t \\left| \\Pr_*\\!\\left\\{a_n(\\hat\\theta^*-\\hat\\theta)\\leq t\\right\\} - \\Pr_F\\!\\left\\{a_n(\\hat\\theta-\\theta)\\leq t\\right\\} \\right| \\xrightarrow{P}0.\\]\n\nThe notation \\(\\xrightarrow{P}\\) denotes convergence in probability over repeated original samples. This formulation assumes conditions under which the uniform comparison is appropriate; more general theory uses other metrics on probability distributions.\n\nFor the sample mean, bootstrap consistency follows under familiar finite-variance conditions. The ordinary central limit theorem describes the population sampling distribution, while its conditional bootstrap counterpart describes resampling from the empirical distribution. The empirical variance converges to the population variance, causing the two normalized distributions to approach the same normal limit. This example explains why the discrete nature of the empirical distribution does not prevent successful approximation of a continuous limiting law.\n\nMany more complicated estimators are asymptotically linear:\n\n\\[\\hat\\theta-\\theta = \\frac{1}{n}\\sum_{i=1}^{n}\\psi(X_i) + o_P(n^{-1/2}),\\]\n\nwhere \\(\\psi\\) has mean zero and finite variance. The leading term behaves like a sample average, and the remainder becomes negligible at the \\(\\sqrt n\\) scale. Influence-function representations of this kind explain bootstrap validity for broad classes of estimators. The functional delta method generalizes ordinary differentiation to functionals of probability distributions and provides another central route to such results.\n\nFirst-order consistency means that approximation error vanishes asymptotically. It does not establish a particular finite-sample error rate or guarantee superiority over every alternative. Under stronger smoothness, moment, and distributional assumptions, asymptotic expansions can describe corrections involving skewness and higher cumulants. A bootstrap procedure may estimate these corrections automatically, whereas a simple normal approximation ignores them.\n\nThis is the basis of higher-order accuracy. Properly studentized or calibrated bootstrap procedures can, in suitable regular problems, reduce the order of coverage or distributional error relative to a first-order approximation. Exact rates depend on the interval, whether coverage is one-sided or two-sided, and the assumptions used. Percentile intervals do not inherit the best available rate merely from being bootstrap intervals.\n\nThe distinction between pointwise and uniform validity is also important. A method may work for every fixed distribution within a regular class yet perform poorly along sequences approaching a boundary or degeneracy. Near-zero variances, weakly identified parameters, or nearly tied model choices can make asymptotic approximations nonuniform. Bootstrap theory therefore concerns both the estimator’s local mathematical behavior and the range of populations over which an approximation is claimed."
    },
    {
      "heading": "Failure cases and alternatives",
      "text": "A classic failure concerns estimation of a support endpoint. Suppose \\(X_1,\\ldots,X_n\\) are uniform on \\([0,\\theta]\\), and \\(\\hat\\theta=M_n=\\max_iX_i\\). Every ordinary bootstrap observation is at most \\(M_n\\), and the resampled maximum equals \\(M_n\\) whenever the largest original observation is selected. Assuming distinct observations,\n\n\\[\\Pr_*(M_n^*=M_n) = 1-\\left(1-\\frac1n\\right)^n \\longrightarrow 1-e^{-1}.\\]\n\nThus the bootstrap distribution of \\(n(M_n-M_n^*)\\) retains a point mass of approximately \\(0.632\\) at zero.\n\nThe actual endpoint error has different limiting behavior. Specifically, \\(n(\\theta-M_n)/\\theta\\) converges to an exponential distribution, which has no point mass at zero. The bootstrap discrepancy persists as the sample grows. This example demonstrates that convergence of the empirical distribution to the population distribution is not sufficient: the endpoint functional is too sensitive to information beyond the observed maximum.\n\nOther failures arise from discontinuous or insufficiently smooth statistical operations. The ordinary bootstrap may be inconsistent for change-point estimators, some constrained estimators at boundaries, and procedures whose limiting distributions change abruptly with the underlying parameter. Model selection introduces similar difficulties when small data changes switch which variables are included. Refitting a selection algorithm within every bootstrap replicate reproduces its observed instability, but does not by itself establish valid confidence intervals after selection.\n\nLack of ordinary differentiability is not a universal reason for failure. The sample median can be bootstrapped consistently when the population has a unique median and a positive, sufficiently regular density there. Problems arise when such regularity is absent—for example, when a quantile is nonunique or the distribution is locally flat in a way that changes its convergence rate. The relevant question concerns the population-level functional, not simply whether the statistic’s finite-sample formula contains an absolute value or an ordering operation.\n\nHeavy tails create another set of limitations. With infinite population variance, the sample mean need not have a standard \\(\\sqrt n\\) normal limit, and the ordinary bootstrap can fail to reproduce its correct limiting law. Rare categories and extreme tail probabilities pose related information problems: an empirical distribution cannot assign mass to unobserved categories without additional assumptions. Outliers can strongly affect both the original estimate and its bootstrap distribution.\n\nAlternatives include model-based tail methods, smoothing, subsampling, and the \\(m\\)-out-of-\\(n\\) bootstrap, which draws \\(m<n\\) observations with replacement. In certain nonregular problems, validity can be recovered when \\(m\\to\\infty\\) but \\(m/n\\to0\\), with appropriate rescaling. Subsampling typically draws smaller subsets without replacement. These methods require choices about subset size, normalization, and target distribution; they do not remove the need to understand why the ordinary bootstrap fails."
    },
    {
      "heading": "Hypothesis tests, prediction, and related methods",
      "text": "For hypothesis testing, a bootstrap must usually approximate a distribution satisfying the null hypothesis. Resampling directly from the empirical distribution generally reproduces a population centered near the observed estimate, not necessarily the hypothesized value. For a test of \\(H_0:\\mu=\\mu_0\\), one possible construction uses shifted observations\n\n\\[Y_i=X_i-\\bar X+\\mu_0,\\]\n\nthen resamples from the \\(Y_i\\) values and recomputes the chosen test statistic. A parametric bootstrap can instead simulate from a model fitted under the null restriction.\n\nFor a statistic whose larger values indicate stronger disagreement with the null, the bootstrap tail probability is estimated from the proportion of null-generated statistics at least as large as the observed one. With independent simulations and suitable calibration, a common finite-simulation expression is\n\n\\[\\hat p= \\frac{1+\\sum_{b=1}^{B}\\mathbf{1}\\{T_b^*\\geq T_{\\mathrm{obs}}\\}} {B+1}.\\]\n\nThis avoids reporting a zero simulated p-value, but does not make an approximate bootstrap test exact. Its validity still depends on the null simulation scheme, treatment of nuisance parameters, and definition of extremeness.\n\nA permutation test answers a different resampling question. It rearranges labels or assignments under an exchangeability or randomization assumption, rather than sampling observations with replacement from an estimated distribution. Some permutation tests have exact finite-sample justification under their null assumptions. Bootstrap tests generally rely on approximating an unknown sampling law. The distinction matters even when both methods are implemented by repeatedly recomputing the same statistic.\n\nIn prediction, resampling observations and refitting a model assesses variation in the fitted rule. A prediction interval for a future outcome additionally requires the future outcome’s own randomness. The distribution of bootstrap fitted means alone omits that component. Likewise, evaluating each fitted model on the same data used to fit it generally produces optimistic performance estimates. Bootstrap optimism correction and out-of-bag evaluation address this problem in different ways.\n\nAn observation omitted from a bootstrap training sample is called out of bag for that replicate. Its omission probability approaches \\(e^{-1}\\), making out-of-bag predictions available across many replicates. The .632 bootstrap estimate combines apparent training error and out-of-bag error with weights approximately \\(0.368\\) and \\(0.632\\); the .632+ method modifies the weighting to address some overfitting settings. These approaches differ from cross-validation in their training-sample construction and correction logic.\n\nBootstrap aggregation, or bagging, averages predictions from bootstrap-fitted models to reduce instability. Its objective is predictive performance rather than confidence-interval coverage. The Bayesian bootstrap, introduced by Donald Rubin in 1981, instead assigns random Dirichlet weights to observed data values, producing a distribution over weighted empirical distributions. Although closely related computationally, its interpretation belongs to Bayesian inference rather than the ordinary frequentist repeated-sampling argument."
    },
    {
      "heading": "Computation, implementation, and scientific applications",
      "text": "A bootstrap calculation has two distinct approximation errors. Statistical approximation error comes from replacing the true data-generating process with a fitted or empirical one. Monte Carlo error comes from using only \\(B\\) simulated replicates to represent the resulting bootstrap distribution. Increasing \\(B\\) reduces the second error, not the first. A misspecified model, an inappropriate resampling unit, or an inconsistent bootstrap procedure remains problematic even with millions of replicates.\n\nThe computational precision required depends on the target. A central spread measure can often be estimated with fewer replicates than an extreme tail quantile. For approximately normal bootstrap estimates and independent replicates, the relative Monte Carlo standard error of the estimated standard deviation is approximately\n\n\\[\\frac{1}{\\sqrt{2(B-1)}}.\\]\n\nAt \\(B=2{,}000\\), this is about \\(1.6\\%\\). For a simulated tail probability \\(p\\), the standard error of the corresponding proportion is approximately \\(\\sqrt{p(1-p)/B}\\). Quantile precision also depends on the density near the endpoint, so a fixed replicate count does not imply equal accuracy across all interval problems.\n\nBootstrap computation is naturally suited to parallel computing because replicates can usually be fitted independently. Reproducibility requires controlled random-number generation, especially when work is distributed among processors. Batched computation and running moment calculations can reduce memory use when only a standard error is required. Quantile intervals need either retained replicate estimates or an appropriate quantile-estimation scheme; storing every resampled dataset is generally unnecessary.\n\nThe full analysis pipeline determines what a replicate must reproduce. If preprocessing, tuning, variable selection, or missing-data imputation is estimated from the sample, holding it fixed can exclude its contribution to uncertainty. Repeating all steps may better represent the intended experiment, although discontinuities and selection effects can still invalidate ordinary interval methods. Optimization failures also matter: silently dropping unsuccessful fits can change the bootstrap distribution by preferentially removing difficult samples.\n\nApplications range from uncertainty in biological dose-response curves to confidence intervals for economic coefficients and error estimates for complex engineering models. In phylogenetics, bootstrap support is often calculated by resampling alignment columns and reconstructing trees. The frequency with which a clade appears measures its stability under that site-resampling procedure; it is not generally a posterior probability that the clade is true. Dependence among sites and model assumptions affect its interpretation.\n\nBootstrap procedures quantify uncertainty represented by the chosen resampling experiment. They do not automatically account for selection bias, unmeasured confounding, measurement-system errors, or population changes absent from that experiment. Reporting therefore distinguishes the estimate being studied, the resampling unit, the model or design assumptions, the number of replicates, and the interval or test construction. These details determine which sources of variation the numerical results actually describe."
    }
  ],
  "links": [
    "statistics",
    "sampling-distribution",
    "standard-error",
    "confidence-interval",
    "estimator-bias",
    "probability-distribution",
    "empirical-distribution",
    "indicator-function",
    "sample-mean",
    "variance",
    "bootstrap-sampling",
    "covariance-matrix",
    "mean-squared-error",
    "normal-distribution",
    "maximum-likelihood-estimation",
    "linear-regression",
    "generalized-linear-model",
    "time-series",
    "convergence-in-probability",
    "central-limit-theorem",
    "asymptotic-expansion",
    "median",
    "statistical-hypothesis-testing",
    "null-hypothesis",
    "p-value",
    "permutation-test",
    "prediction-interval",
    "cross-validation",
    "bagging",
    "bayesian-inference",
    "parallel-computing",
    "missing-data-imputation",
    "selection-bias",
    "confounding"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}