tilt.boot: Non-parametric Tilted Bootstrap

Description

This function will run an initial bootstrap with equal resampling probabilities (if required) and will use the output of the initial run to find resampling probabilities which put the value of the statistic at required values. It then runs an importance resampling bootstrap using the calculated probabilities as the resampling distribution.

Usage

tilt.boot(data, statistic, R, sim = "ordinary", stype = "i", 
          strata = rep(1, n), L = NULL, theta = NULL, 
          alpha = c(0.025, 0.975), tilt = TRUE, width = 0.5, 
          index = 1, ...)

Value

An object of class "boot" with the following components

t0: The observed value of the statistic on the original data.
t: The values of the bootstrap replicates of the statistic. There will be sum(R) of these, the first R[1] corresponding to the uniform bootstrap and the remainder to the tilted bootstrap(s).
R: The input vector of the number of bootstrap replicates.
data: The original data as supplied.
statistic: The statistic function as supplied.
sim: The simulation type used in the bootstrap(s), it can either be "ordinary" or "balanced".
stype: The type of statistic supplied, it is the same as the input value stype.
call: A copy of the original call to tilt.boot.
strata: The strata as supplied.
weights: The matrix of weights used. If R[1] is greater than 0 then the first row will be the uniform weights and each subsequent row the tilted weights. If R[1] equals 0 then the uniform weights are omitted and only the tilted weights are output.
theta: The values of theta used for the tilted distributions. These are either the input values or the values derived from the uniform bootstrap and alpha.

Arguments

data: The data as a vector, matrix or data frame. If it is a matrix or data frame then each row is considered as one (multivariate) observation.
statistic: A function which when applied to data returns a vector containing the statistic(s) of interest. It must take at least two arguments. The first argument will always be data and the second should be a vector of indices, weights or frequencies describing the bootstrap sample. Any other arguments must be supplied to tilt.boot and will be passed unchanged to statistic each time it is called.
R: The number of bootstrap replicates required. This will generally be a vector, the first value stating how many uniform bootstrap simulations are to be performed at the initial stage. The remaining values of R are the number of simulations to be performed resampling from each reweighted distribution. The first value of R must always be present, a value of 0 implying that no uniform resampling is to be carried out. Thus length(R) should always equal 1+length(theta).
sim: This is a character string indicating the type of bootstrap simulation required. There are only two possible values that this can take: "ordinary" and "balanced". If other simulation types are required for the initial un-weighted bootstrap then it will be necessary to run boot, calculate the weights appropriately, and run boot again using the calculated weights.
stype: A character string indicating the type of second argument expected by statistic. The possible values that stype can take are "i" (indices), "w" (weights) and "f" (frequencies).
strata: An integer vector or factor representing the strata for multi-sample problems.
L: The empirical influence values for the statistic of interest. They are used only for exponential tilting when tilt is TRUE. If tilt is TRUE and they are not supplied then tilt.boot uses empinf to calculate them.
theta: The required parameter value(s) for the tilted distribution(s). There should be one value of theta for each of the non-uniform distributions. If R[1] is 0 theta is a required argument. Otherwise theta values can be estimated from the initial uniform bootstrap and the values in alpha.
alpha: The alpha level to which tilting is required. This parameter is ignored if R[1] is 0 or if theta is supplied, otherwise it is used to find the values of theta as quantiles of the initial uniform bootstrap. In this case R[1] should be large enough that min(c(alpha, 1-alpha))*R[1] > 5, if this is not the case then a warning is generated to the effect that the theta are extreme values and so the tilted output may be unreliable.
tilt: A logical variable which if TRUE (the default) indicates that exponential tilting should be used, otherwise local frequency smoothing (smooth.f) is used. If tilt is FALSE then R[1] must be positive. In fact in this case the value of R[1] should be fairly large (in the region of 500 or more).
width: This argument is used only if tilt is FALSE, in which case it is passed unchanged to smooth.f as the standardized bandwidth for the smoothing operation. The value should generally be in the range (0.2, 1). See smooth.f for for more details.
index: The index of the statistic of interest in the output from statistic. By default the first element of the output of statistic is used.
...: Any additional arguments required by statistic. These are passed unchanged to statistic each time it is called.

References

Booth, J.G., Hall, P. and Wood, A.T.A. (1993) Balanced importance resampling for the bootstrap. Annals of Statistics, 21, 286--298.

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

Hinkley, D.V. and Shi, S. (1989) Importance sampling and the nested bootstrap. Biometrika, 76, 435--446.

Examples

Run this code

# Note that these examples can take a while to run.

# Example 9.9 of Davison and Hinkley (1997).
grav1 <- gravity[as.numeric(gravity[,2]) >= 7, ]
grav.fun <- function(dat, w, orig) {
     strata <- tapply(dat[, 2], as.numeric(dat[, 2]))
     d <- dat[, 1]
     ns <- tabulate(strata)
     w <- w/tapply(w, strata, sum)[strata]
     mns <- as.vector(tapply(d * w, strata, sum)) # drop names
     mn2 <- tapply(d * d * w, strata, sum)
     s2hat <- sum((mn2 - mns^2)/ns)
     c(mns[2]-mns[1],s2hat,(mns[2]-mns[1]-orig)/sqrt(s2hat))
}
grav.z0 <- grav.fun(grav1, rep(1, 26), 0)
tilt.boot(grav1, grav.fun, R = c(249, 375, 375), stype = "w", 
          strata = grav1[,2], tilt = TRUE, index = 3, orig = grav.z0[1]) 


#  Example 9.10 of Davison and Hinkley (1997) requires a balanced 
#  importance resampling bootstrap to be run.  In this example we 
#  show how this might be run.  
acme.fun <- function(data, i, bhat) {
     d <- data[i,]
     n <- nrow(d)
     d.lm <- glm(d$acme~d$market)
     beta.b <- coef(d.lm)[2]
     d.diag <- boot::glm.diag(d.lm)
     SSx <- (n-1)*var(d$market)
     tmp <- (d$market-mean(d$market))*d.diag$res*d.diag$sd
     sr <- sqrt(sum(tmp^2))/SSx
     c(beta.b, sr, (beta.b-bhat)/sr)
}
acme.b <- acme.fun(acme, 1:nrow(acme), 0)
acme.boot1 <- tilt.boot(acme, acme.fun, R = c(499, 250, 250), 
                        stype = "i", sim = "balanced", alpha = c(0.05, 0.95), 
                        tilt = TRUE, index = 3, bhat = acme.b[1])