This function provides a simple front-end to the boot function in the
boot package that is tailored to bootstrapping based on regression models. Whereas boot is very general and therefore
has many arguments, the Boot function has very few arguments.
Boot(object, f=coef, labels=names(f(object)), R=999,
method=c("case", "residual"), ncores=1, …)# S3 method for default
Boot(object, f=coef, labels=names(f(object)),
R=999, method=c("case", "residual"), ncores=1,
start = FALSE, …)
# S3 method for lm
Boot(object, f=coef, labels=names(f(object)),
R=999, method=c("case", "residual"), ncores=1, …)
# S3 method for glm
Boot(object, f=coef, labels=names(f(object)),
R=999, method=c("case", "residual"), ncores=1, …)
# S3 method for nls
Boot(object, f=coef, labels=names(f(object)),
R=999, method=c("case", "residual"), ncores=1, …)
A regression object of class "lm", "glm" or
"nls". The function may work with other regression objects that support
the update method and have a subset argument. See discussion of
the default method in the details below.
A function whose one argument is the name of a regression object that will
be applied to the updated regression object to compute the statistics of
interest. The default is coef, to return regression coefficient
estimates. For example, f = function(obj) coef(obj)[1]/coef(obj)[2] will
bootstrap the ratio of the first and second coefficient estimates.
Provides labels for the statistics computed by f. Default
labels are obtained from a call to f, or generic labels if f
does not return names.
Number of bootstrap samples. The number of bootstrap samples actually computed may be smaller than this value if either the fitting method is iterative and fails to converge for some boothstrap samples, or if the rank of a fitted model is different in a bootstrap replication than in the original data.
The bootstrap method, either “case” for resampling
cases or “residuals” for a residual bootstrap. See the details
below. The residual bootstrap is available only for lm and nls
objects and will return an error for glm objects.
Arguments passed to the boot function, see
boot.
Should the estimates returned by f be passed as starting
values for each bootstrap iteration? Alternatively, start can be a
numeric vector of starting values. The default is to use
the estimates from the last bootstrap iteration as starting values
for the next iteration.
A numeric argument that specifies the number of cores for parallel processing for unix systems. If less than or equal to 1, no parallel processing wiill be used. Note in a Windows platform will produce a warning and set this argument to 1.
See boot for the returned value of the structure returned by this function.
Boot uses a
regression object and the choice of method, and creates a function that is
passed as the statistic argument to the boot function in the boot package. The argument R is also passed to boot. If ncores is greater than 1, then the parallel and ncpus arguments to boot are set appropriately to use multiple codes, if available, on your computer. All other arguments to boot are kept at their default values unless you pass values for them.
The methods available for lm and nls objects are “case” and
“residual”. The case bootstrap resamples from the joint distribution
of the terms in the model and the response. The residual bootstrap fixes the
fitted values from the original data, and creates bootstraps by adding a
bootstrap sample of the residuals to the fitted values to get a bootstrap
response. It is an implementation of Algorithm 6.3, page 271, of
Davison and Hinkley (1997). For nls objects ordinary residuals are used
in the resampling rather than the standardized residuals used in the lm
method. The residual bootstrap for
generalized linear models has several competing approaches, but none are
without problems. If you want to do a residual bootstrap for a glm, you
will need to write your own call to boot.
For the default object to work with other types of regression models, the model must have methods for the the following generic functions: residuals(object, type="pearson") must return Pearson residuals; fitted(object) must return fitted values; hatvalues(object) should return the leverages, or perhaps the value 1 which will effectively ignore setting the hatvalues. In addition, the data argument should contain no missing values among the columns actually used in fitting the model, as the resampling may incorrectly attempt to include cases with missing values. For lm, glm and nls, missing values cause the return of an error message.
An attempt to fit using a bootstrap sample may fail. In a lm or
glm fit, the bootstrap sample could have a different rank from the original
fit. In an nls fit, convergence may not be obtained for some bootstraps.
In either case, NA are returned for the value of the function f.
The summary methods handle the NAs appropriately.
Fox and Weisberg (2017) cited below discusses this function and provides more examples.
Davison, A, and Hinkley, D. (1997) Bootstrap Methods and their Applications. Oxford: Oxford University Press.
Fox, J. and Weisberg, S. (2019) Companion to Applied Regression, Third Edition. Thousand Oaks: Sage.
Fox, J. and Weisberg, S. (2019) Bootstrapping Regression Models in R, https://socialsciences.mcmaster.ca/jfox/Books/Companion/appendices/Appendix-Bootstrapping.pdf.
Weisberg, S. (2014) Applied Linear Regression, Fourth Edition, Wiley Wiley, Chapters 4 and 11.
Functions that work with
boot objects from the boot package are
boot.array,
boot.ci,
plot.boot and
empinf. Additional
functions in the car package are
summary.boot,
confint.boot, and
hist.boot.
# NOT RUN {
m1 <- lm(Fertility ~ ., swiss)
betahat.boot <- Boot(m1, R=199) # 199 bootstrap samples--too small to be useful
summary(betahat.boot) # default summary
confint(betahat.boot)
hist(betahat.boot)
# Bootstrap for the estimated residual standard deviation:
sigmahat.boot <- Boot(m1, R=199, f=sigmaHat, labels="sigmaHat")
summary(sigmahat.boot)
confint(sigmahat.boot)
# }
Run the code above in your browser using DataLab