R-Listers:

I am doing a quasi-maximum likelihood estimation and I get a "subscript 
out of bound" error message,  Typically I would think this means that a 
subscript used in the function is literally out of bounds however I don't 
think this is the case.  All I change in the code is a constant, that is 
hard-wired in (not data dependent and not parameter dependent), 
furthermore, the constant is not used in any subscripting.

Sorry I cannot provide a toy example, but the likelihood function looks 
like this:

lnL <- function(theta, gsvr, gsvR) { 
            # theta[1]= mu, theta[2]=gamma, theta[3]=kappa1, 
theta[4]=kappa2
            nn <- 252
            d <- 0:nn
            wd <- exp((theta[3] * d + theta[4] * d^2))/(sum(exp(theta[3] * 
d + theta[4] * d^2)))
            sigsq <- numeric(length(gsvR$ret))
            x <- numeric(length(gsvR$ret)

        # Below this line can be specified differently
             x[1] <- 1
            lenR <- length(gsvR$ret)
            for (i in 1:(lenR-1)) {
                    x[i+1] <- sum(gsvR$rw[1:i]) + 1
                    rsq <- (gsvr$ret[(x[i]):(nn+x[i])])^2
                    sigsq[i] <- 22 * sum(wd * rsq)
            }
 
           if((nn+x[lenR]) < length(gsvr$ret)) rsq <- 
(gsvr$ret[(x[lenR]):(nn+x[lenR])])^2 else rsq <- NA
           sigsq[length(gsvR$ret)] <- 22 * sum(wd * rsq)
           sigsq <- na.omit(sigsq)
            sigsq <- sigsq[-1]
           # Above this line can be specified differently

            Ret <- gsvR$ret[1:length(sigsq)]
            mymu <- (theta[1]+theta[2]*(sigsq))
            n <- length(wd)#/22 
            ll <- numeric(length(sigsq))
            for (j in 1:length(sigsq)) {
            ll[j] <- -(n/2) * log(2*pi) - (n/2) * log(sigsq[j]) - 0.5 * 
sum(((Ret - mymu[j])^2)/(sigsq[j]))
            } 
            l <- mean(ll)
            -l
       }
#########Alternative specitication################
            for (i in 1:(length(gsvR$ret))) {
            x[i] <- sum(gsvR$rw[1:i]) + 1
            rsq <- (gsvr$ret[(x[i]):(nn+x[i])])^2
            sigsq[i] <- 22 * sum(wd * rsq)
 
            }
 
            sigsq <- na.omit(sigsq)
###############################################



The constant that is changed is "n" when the n <- length(wd)/22 the 
optimization converges using both nlm and optim(with the default method), 
however with n <- length(wd) the functions returns the subscript out of 
bounds error message.  Maximizing the likelihood with the alternative 
specification replacing the  code marked in the original function allows 
the optimization to converge.  It should be noted that the original 
likelihood function and the alternative specification return the same 
likelihood value when evaluated at the initial values.

I realize that in the original specification the variable sigsq could be 
all NAs and na.omit(sigsq) would produce a zero length vector, upon which 
sigsq[-1] would be out of bounds, however the original specification 
converges when n<-length(wd)/22, in which case the aforementioned case 
would still be true.  Plus both specifications evaluate the initial values 
when submitted line by line (rather than in the function), using both 
n<-length(wd)/22 and n<-length(wd).

Could the error "subscript out of bounds" mean something different?  Is 
there likely to be a bug? 

I am using:

> version
         _ 
platform i386-pc-mingw32
arch     i386 
os       mingw32 
system   i386, mingw32 
status 
major    1 
minor    8.0 
year     2003 
month    10 
day      08 
language R 


Thanks for reading, and/or replying to this.

Jason Higbee
Research Associate
Federal Reserve Bank of St. Louis
T: 314.444.7316
F:314.444.8731
        [[alternative HTML version deleted]]

______________________________________________
[EMAIL PROTECTED] mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html

Reply via email to