ihess {SMPracticals} | R Documentation |
Inverse Hessian matrix, useful for obtaining standard errors
ihess(f, x, ep = 1e-04, ...)
f |
Usually a negative log likelihood |
x |
Usually maximum likelihood estimates for f |
ep |
Step length used to compute numerical second derivatives |
... |
Extra arguments for f, if any |
Matrix of dimension dim(x) times dim(x), containing inverse Hessian matrix of f at x.
This is not needed in R, where hessian matrices are obtained by setting hessian=T in calls to optimisation functions.
Anthony Davison
Based on code written by Stuart Coles of Padova University
# ML fit of t distribution nlogL <- function(x, data) # negative log likelihood { mu <- x[1] sig <- x[2] df <- x[3] -sum(log( dt((data-mu)/sig, df=df)/sig )) } y <- rt(n=100, df=10) # generate t data # this is Splus code.....so remove the #'s for it to work in R # fit <- nlminb(c(1,1,4), nlogL, upper=c(Inf,Inf,Inf), lower=c(-Inf,0,0), # data=y) # fit$parameters # maximum likelihood estimates # J <- ihess(nlogL, fit$parameters, data=y) # sqrt(diag(J)) # standard errors based on observed information # # In this example the standard error can be a bad measure of # uncertainty for the df.