abc.ci: Nonparametric ABC Confidence Intervals

Description

Calculate equi-tailed two-sided nonparametric approximate bootstrap confidence intervals for a parameter, given a set of data and an estimator of the parameter, using numerical differentiation.

Usage

abc.ci(data, statistic, index=1, strata=rep(1, n), conf=0.95, 
       eps=0.001/n, …)

Arguments

data

A data set expressed as a vector, matrix or data frame.

statistic

A function which returns the statistic of interest. The function must take at least 2 arguments; the first argument should be the data and the second a vector of weights. The weights passed to statistic will be normalized to sum to 1 within each stratum. Any other arguments should be passed to abc.ci as part of the …{} argument.

index

If statistic returns a vector of length greater than 1, then this indicates the position of the variable of interest within that vector.

strata

A factor or numerical vector indicating to which sample each observation belongs in multiple sample problems. The default is the one-sample case.

conf

A scalar or vector containing the confidence level(s) of the required interval(s).

eps

The value of epsilon to be used for the numerical differentiation.

...

Any other arguments for statistic. These will be passed unchanged to statistic each time it is called within abc.ci.

Value

A length(conf) by 3 matrix where each row contains the confidence level followed by the lower and upper end-points of the ABC interval at that level.

Details

This function is based on the function abcnon written by R. Tibshirani. A listing of the original function is available in DiCiccio and Efron (1996). The function uses numerical differentiation for the first and second derivatives of the statistic and then uses these values to approximate the bootstrap BCa intervals. The total number of evaluations of the statistic is 2*n+2+2*length(conf) where n is the number of data points (plus calculation of the original value of the statistic). The function works for the multiple sample case without the need to rewrite the statistic in an artificial form since the stratified normalization is done internally by the function.

References

Davison, A.C. and Hinkley, D.V. (1997) Bootstrap Methods and Their Application, Chapter 5. Cambridge University Press.

DiCiccio, T. J. and Efron B. (1992) More accurate confidence intervals in exponential families. Biometrika, 79, 231--245.

DiCiccio, T. J. and Efron B. (1996) Bootstrap confidence intervals (with Discussion). Statistical Science, 11, 189--228.

Examples

Run this code

# NOT RUN {
# 90% and 95% confidence intervals for the correlation 
# coefficient between the columns of the bigcity data

abc.ci(bigcity, corr, conf=c(0.90,0.95))

# A 95% confidence interval for the difference between the means of
# the last two samples in gravity
mean.diff <- function(y, w)
{    gp1 <- 1:table(as.numeric(y$series))[1]
     sum(y[gp1, 1] * w[gp1]) - sum(y[-gp1, 1] * w[-gp1])
}
grav1 <- gravity[as.numeric(gravity[, 2]) >= 7, ]
abc.ci(grav1, mean.diff, strata = grav1$series)
# }

Run the code above in your browser using DataCamp Workspace