Bootstrapping is the practice of estimating properties of an estimator (such as its variance) by measuring those properties when sampling from an approximating distribution. One standard choice for an approximating distribution is the empirical distribution of the observed data.
It may also be used for constructing hypothesis tests. It is often used as an alternative to inference based on parametric assumptions when those assumptions are in doubt, or where parametric inference is impossible or requires very complicated formulas for the calculation of standard errors.
The advantage of bootstrapping over analytical method is its great simplicity - it is straightforward to apply the bootstrap to derive estimates of standard errors and confidence intervals for complex estimators of complex parameters of the distribution, such as percentile points, proportions, odds ratio, and correlation coefficients.
The disadvantage of bootstrapping is that while (under some conditions) it is asymptotically consistent, it does not provide general finite sample guarantees, and has a tendency to be overly optimistic.
Example: Fisher's iris data
To introduce the basic ideas and the value of the method, a rather artificial example will be used. Fisher's famous iris data are available via the classic data sets page. In this example, only the species virginica and versicolor are considered. The analysis was performed in R.
The species can be modelled by logistic regression as a function of sepal length (that is, the other variables available are ignored). Fitting the logistic regression model by maximum likelihood gives the following parameter estimates and their standard errors:
It is known that maximum likelihood estimators are asymptotically normally distributed. We can check this assumption using a bootstrap procedure as follows:
- Sample n observations with replacement from the original data, where n is the number of observations.
- Fit the logistic regression model by maximum likelihood.
- Repeat steps 1 and 2 many, B, times.
- Use the sampling distribution of the estimates thus computed to be an approximation to the 'true', population sampling distribution.
The plot below contains kernel density plots of the two parameters in the model, as estimated from 10 000 bootstrap samples.
The distributions of the parameter estimates are clearly not normal. That is, the asymptotic assumptions about the maximum likelihood estimates cannot be relied on, and quantities such as confidence intervals and hypothesis tests that rely on those assumptions will be suspect.
One way to estimate confidence intervals from bootstrap samples is to take the and quantiles of the estimated values. These are called bootstrap percentile intervals. In this case, for the intercept and for sepal length, the bootstrap 95% percentile intervals are (-20.02, -7.08) and (1.26, 3.20) respectively. These can be contrasted with the asymptotic intervals derived from the maximum likelihood estimates plus or minus 1.96 standard errors: (-18.26, -6.87) and (1.10, 2.93). The intervals from the asymptotic theory are apparently too narrow (as well as being symmetric).
This simple bootstrap method is not the only way of making improved inferences over the asymptotic approach. Other bootstrap schemes are available, as are approaches based on likelihood or Bayesian considerations. (In fact, the simple bootstrap scheme used here can be quite easily criticized.)
There are more complicated bootstraps for sampling without replacement, two-sample problems, regression, time series, hierarchical sampling, mediation analyses, and other statistical problems.
How many bootstrap samples is enough?
In the example with Fisher's iris data, 10 000 bootstrap samples were used. However, no explanation was given for this number. It seems that the number of bootstrap samples recommended in the literature has increased as available computing power has increased. Whereas a few years ago, 10 000 samples would have seemed excessive, the above example ran in just a few minutes.
As a general guideline, 1000 samples is often enough for a first look. However, if the results really matter, as many samples as is reasonable given available computing power and time should be used.
Types of bootstrap scheme
In univariate problems, it is usually acceptable to resample the individual observations with replacement. However, in small samples, a parametric bootstrap approach might be preferred, and for some problems a smooth bootstrap will likely be preferred.
For regression problems, various other alternatives are available.
Under this scheme, a small amount of (usually normally distributed) zero-centered random noise is added on to each resampled observation. This is equivalent to sampling from a kernel density estimate of the data.
In this case, a parametric model is fit to the data, often by maximum likelihood, and samples of random numbers are drawn from this parametric model. Then, the quantity, or estimate, of interest is calculated from these samples.
In regression problems, case resampling refers to the simple scheme of resampling individual cases - often rows of a data set. For regression problems, so long as the data set is fairly large, this simple scheme is often acceptable (and this is the method used in the iris example above). However, the method is open to criticism.
In regression problems, the explanatory variables are often fixed, or at least observed with more control than the response variable. Also, the range of the explanatory variables defines the information available from them. Therefore, to resample cases means that each bootstrap sample will lose some information. As such, alternative bootstrap procedures should be considered.
Another approach to bootstrapping in regression problems is to resample residuals. The method proceeds as follows.
- Fit the model and retain the fitted values and the residuals , .
- For each pair, , in which is the (possibly multivariate) explanatory variable, add a randomly resampled residual, , to the response variable
- Refit the model and retain the quantities of interest (often the estimated parameters).
- Repeat steps 2 and 3 many, B, times.
This scheme has the advantage that it retains the information in the explanatory variables. However, a question arises as to which residuals to resample. Raw residuals are one option, another is studentized residuals (in linear regression). Whilst there are arguments in favour of using studentized residuals, in practice it often makes little difference and it is easy to run both schemes and compare the results against each other.
This is the same as resampling residuals but with the additional step that each randomly resampled residual is randomly multiplied by 1 or -1. This method assumes that the 'true' residual distribution is symmetric and can offer advantages over simple residual sampling for smaller sample sizes.
Application to testing for mediation
Bootstrapping is becoming the most popular method of testing mediation   because it does not require the normality assumption to be met, and because it can be effectively utilized with smaller sample sizes (N < 20). However, mediation continues to be (perhaps inappropriately) most frequently determined using (1) the logic of Baron and Kenny  or (2) the Sobel test.
Example: smoothed bootstrap
Newcomb's speed of light data are used in the book Bayesian Data Analysis by Gelman et al. and can be found via the classic data sets page. Some analysis of these data appears on the robust statistics page.
The data set contains two obvious outliers so that, as an estimate of location, the median is to be preferred over the mean. Bootstrapping is a method often employed for estimating confidence intervals for medians. However, the median is a discrete statistic, and this fact shows up in the bootstrap distribution.
In order to smooth over the discreteness of the median, we can add a small amount of random noise to each bootstrap sample. We choose for sample size .
Histograms of the bootstrap distribution and the smooth bootstrap distribution appear below. The bootstrap distribution is very jagged because there are only a small number of values that the median can take. The smoothed bootstrap distribution overcomes this jaggedness.
Although the bootstrap distribution of the median looks ugly and intuitively wrong, confidence intervals from it are not bad. The simple 95% percentile interval is (26, 28.5) for the simple bootstrap and (25.98, 28.46) for the smoothed bootstrap.
Relationship to other resampling methods
For more details see bootstrap resampling.
References and further reading
- Bradley Efron (1979). "Bootstrap Methods: Another Look at the Jackknife". The Annals of Statistics. 7 (1): 1&ndash, 26.
- Bradley Efron (1994). An Introduction to the Bootstrap. Chapman & Hall/CRC. Unknown parameter
|coauthors=ignored (help) 
- Efron, B. (1981). Nonparametric estimates of standard error: The jackknife, the bootstrap and other methods. Biometrika, 68, 589-599.
- Efron, B. (1982). The jackknife, the bootstrap, and other resampling plans. Society of Industrial and Applied Mathematics CBMS-NSF Monographs, 38.
- Diaconis, P. & Efron, B. (1983). Computer-intensive methods in statistics. Scientific American, May, 116-130.
- Efron, B., & Tibshirani, R. J. (1993). An introduction to the bootstrap. New York: Chapman & Hall, software.
- Mooney, C Z & Duval, R D (1993). Bootstrapping. A Nonparametric Approach to Statistical Inference. Sage University Paper series on Quantitative Applications in the Social Sciences, 07-095. Newbury Park, CA: Sage
- Edgington, E. S.(1995). Randomization tests. New York: M. Dekker.
- Davison, A. C. and Hinkley, D. V. (1997): Bootstrap Methods and their Applications, software.
- Simon, J. L. (1997): Resampling: The New Statistics.
- Moore, D. S., G. McCabe, W. Duckworth, and S. Sclove (2003): Bootstrap Methods and Permutation Tests
- Hesterberg, T. C., D. S. Moore, S. Monaghan, A. Clipson, and R. Epstein (2005): Bootstrap Methods and Permutation Tests, software.
- Bootstrap Sampling Tutorial: Introduction to Bootstrap sampling, a tutorial using MS Excel.
- Bootstrap tutorial from ICASSP 99: Tutorial from a signal processing perspective