[Rd] bug in plot.acf, ci.type= "ma" (PR#13071)

meermeyer at statistik.uni-wuppertal.de meermeyer at statistik.uni-wuppertal.de
Fri Sep 26 12:15:03 CEST 2008


Full_Name: Martin Meermeyer
Version: 2.51
OS: Win XP SP3
Submission from: (NULL) (132.195.128.210)


I think the confidence-bands in the plot.acf(..., ci.type="ma") must start at
lag 1 due to the Bartlett-approximation. I think so since I´ve calculated these
manually, see the following example:
 
reihe     <- arima.sim(n = 100, list(ma = c(0.5, 0.4)))
reihe.acf <- acf(reihe, ci.type = c("ma"))
reihe.acf <- as.vector(reihe.acf$acf)

# Estimated variances of hat(rho)(2:h) using the Bartlett-approximation
var.acf <- (1 + 2*cumsum(reihe.acf[-1]^2))/100

# Assuming white noise for hat(rho)(1), therefore estimated variance is 1/100 
var.acf <- c(1/100, var.acf)

# Calculate ci-bounds  
ki.acf  <- 1.96*sqrt(var.acf)

# Skip the last one 
ki.acf  <- head(ki.acf, -1)

# Add to the acf, first entry belongs to lag 1.
lines(c(1:20),ki.acf,col=2,lty=2)

#  Numbers identical to actual function 
lines(c(0:19),-ki.acf,col=2,lty=2)



More information about the R-devel mailing list