I found out that boot
function of the boot package is not working with complex numbers. I am trying to bootstrap a data by taking the eigenvalue of the bivariate matrix. The problem with the eigenvalue is that, it often returns complex numbers, and by that it (boot
) gives error. Is there a way to avoid complex numbers?
Here is my codes,
Data <- read.table('http://ubuntuone.com/6n1igcHXq4EnOm4x2zeqFb', header = FALSE)
Mat <- cbind(Data[["V1"]],Data[["V2"]])
Data.ts <- as.ts(Mat)
Below are some functions needed,
library(mvtnorm)
var1.sim <- function(T, n.start=100, phi1=matrix(c(0.7,0.2,0.2,0.7),nr=2),
err.mu=c(0,0), err.sigma2=matrix(c(1,0.5,0.5,1), nr=2),
errors=NULL) {
e <- rmvnorm(n.start + T, err.mu, err.sigma2) # (n.start+T) x 2 matrix
y <- matrix(0, nrow=n.start+T, ncol=2)
if (!is.null(errors) && is.matrix(errors) && ncol(errors) == 2) {
rows <- nrow(errors)
if (rows < n.start + T) {
# replace last nrow(errors) errors
e[seq.int(n.start+T-rows+1,n.start+T),] <- errors
} else {
e <- errors[seq.int(n.start+T+1, rows)]
}
}
for (t in seq.int(2, n.start + T)) {
y[t,] <- phi1 %*% y[t-1,] + e[t,]
}
return(ts(y[seq.int(n.start+1,n.start+T),]))
}
########
coef.var1 <- function(var.fit) {
k <- coef(var.fit)
rbind(k[[1]][,"Estimate"], k[[2]][,"Estimate"])
}
And here is the main method,
library(vars)
library(boot)
y.var <- VAR(Data.ts, p=1, type="none")
y.resid <- resid(y.var)
rm(y.boot)
y.boot <- boot(y.resid, R=100, statistic=function(x,i) {
resid.boot <- x[i,]
y.boot1 <- var1.sim(T=nrow(x), errors=resid.boot)
min(eigen(coef.var1(VAR(y.boot1, p=1, type="none")))$values)
}, stype="i")
y.boot$t
y.ci <- boot.ci(y.boot, type="norm", conf=0.95)$normal[2:3]
list(t=y.boot$t,ci=y.ci)
The problem occurs in y.boot
object, particularly this line
min(eigen(coef.var1(VAR(y.boot1, p=1, type="none")))$values)
When the obtain minimum eigenvalue is complex, then boot
will return this error
Error in min(eigen(coef.var1(VAR(y.boot1, p = 1, type = "none")))$values) :
invalid 'type' (complex) of argument
Otherwise, there is no problem. Now, it would be safe if this 100 bootstraps is performed once, but I am going to loop this actually about 100 times too. So, there is a big chance that complex values will occur in these loops. Hence, we will obtain the above error again.
Is there a way to avoid these complex values?