A commenter on my previous blog entry has drawn my attention to an R function called hetglm() that estimates heteroskedastic probit models. This function is contained in the glmx package. The glmx package is not available on CRAN yet, but thankfully can be downloaded here.

The hetglm() function has a number of computational advantages compared with the crude method outlined in my previous post. The following example replicates the previous analysis showing the speed advantage associated with using the hetglm() function.

rm(list=ls()) # clear ws
library(maxLik) 
library(glmx)
n <- 1000 # no. obs
x1 <- runif(n,-1,1) # predictor 1
x2 <- runif(n,-1,1) # " 2
e1 <- rnorm(n,0,1) # normal error
e2 <- (1 + 0.45*(x1+x2))*e1 # hetero error
y <- ifelse(0.5 + 0.5*x1 -0.5*x2 - e2 >0, 1, 0) # outcome
# estimate normal probit
r1 <- glm(y~x1+x2,family=binomial(link="probit")) 
system.time(ml1 <- maxLik(hll,start=c(0,0,0,0,0))) # maximize
system.time(h1 <- hetglm(y ~ x1 + x2))


# output
> system.time(ml1 <- maxLik(hll,start=c(0,0,0,0,0))) # maximize
   user  system elapsed 
   4.43    0.00    4.59 
> system.time(h1 <- hetglm(y ~ x1 + x2))
   user  system elapsed 
   0.11    0.00    0.11