The log-likelihood of the gamma distribution with scale parameter 1 can be written as:
(α−1)s−nlogΓ(α)
where alpha is the shape parameter and s=∑logXi
is the sufficient statistic.
Randomly draw a sample of n = 30 with a shape parameter of alpha = 4.5. Using newton_search
and make_derivative
, find the maximum likelihood estimate of alpha. Use the moment estimator of alpha, i.e., mean of x as the initial guess. The log-likelihood function in R is:
x <- rgamma(n=30, shape=4.5)
gllik <- function() {
s <- sum(log(x))
n <- length(x)
function(a) {
(a - 1) * s - n * lgamma(a)
}
}
I have created the make_derivative function as follows:
make_derivative <- function(f, h) {
(f(x + h) - f(x - h)) / (2*h)
}
I also have created a newton_search
function that incorporates the make_derivative
function; However, I need to use newton_search
on the second derivative of the log-likelihood function and I'm not sure how to fix the following code in order for it to do that:
newton_search2 <- function(f, h, guess, conv=0.001) {
set.seed(2)
y0 <- guess
N = 1000
i <- 1; y1 <- y0
p <- numeric(N)
while (i <= N) {
make_derivative <- function(f, h) {
(f(y0 + h) - f(y0 - h)) / (2*h)
}
y1 <- (y0 - (f(y0)/make_derivative(f, h)))
p[i] <- y1
i <- i + 1
if (abs(y1 - y0) < conv) break
y0 <- y1
}
return (p[(i-1)])
}
Hint: You must apply newton_search
to the first and second derivatives (derived numerically using make_derivative
) of the log-likelihood. Your answer should be near 4.5.
when I run newton_search2(gllik(), 0.0001, mean(x), conv = 0.001)
, I get double what the answer should be.