I'm trying to get the shape and scale parameters for this data using the optim function in R.

incomeData = data.frame(L = c(850,rep(1000,24),rep(2001,112),rep(3001,267),rep(4001,598),rep(5001,1146)),
       U = c(999,rep(2000,24),rep(3000,112),rep(4000,267),rep(5000,598),rep(10000,1146)),
       Interval = c(1,rep(2,24),rep(3,112),rep(4,267),rep(5,598),rep(6,1146)))

The maximum likelihood function is defined as this:


F is the cumulative gamma function evaluated in the upper and lower bound of the income interval with shape = shape and scale = scale.

And for the initial values of the parameters I'm using the methods of moments:

mu: mean of the middle points of the invervals.


Where s^2 is the variance of the middle points of the intervals.

The initial parameters were calculated using the method of moments

incomeData$middle = (incomeData$U+incomeData$L)/2 # middle point of the interval

middlePointMean = mean(incomeData$middle) # mean of the middle points
middlePointVar = var(incomeData$middle) # variance of the middle points

initialPar1 = middlePointVar/(middlePointMean^2) # initial shape parameter (this was suggested)
initialPar2 = initialPar1/middlePointMean # initial scale parameter

This is the code I used to run the optimization

 # The likelihood function for this problem is defined by the product of the difference between the 
 # cumulative gamma evaluated in the upper bound of the interval - the cumulative gamma evaluated in 
 # the lower bound of the interval. 
 logLikelihood = function(par){

 ub = incomeData$U 
 lb = incomeData$L 

  # I'm applying sum instead of prod since the log of a product would be the sum
 logLike = sum(pgamma(ub,shape = par[1],scale = par[2]) - 
 pgamma(lb,shape = par[1],scale = par[2]))



optim(par = c(initialPar1,initialPar2),fn = logLikelihood,
method = "L-BFGS-B",lower = 0.00001,upper=.99999)

I get these results

[1] 1.014180e-01 1.737418e-05

[1] 0

function gradient 
  1       1 

[1] 0


When I test the results with those parameters the values are too low and I can't plot the distribution nor the likelihood function and it doesn't make sense to me. This is supposed to give the proability of falling in a particular income interval. What I'm doing wrong?

  • $\begingroup$ Your code uses the sum of the likelihoods (not the loglikelihoods) and calls it logLike. $\endgroup$ May 22, 2021 at 23:15

1 Answer 1


It's a bit strange that your data don't include any intervals at the ends, say $<850$ or $>10000$. I modified your approach and got some sensible results, I think. Instead of the individual data, I used the grouped data with the frequencies of each interval. Further, I used the log of the shape and rate of the gamma distribution for fitting for numerical purposes. Lastly, your code doesn't calculate the log-likelihood, just the likelihood so I changed that as well.

Here's the code:

# Grouped data
dat <- data.frame(
  U = c(999, 2000, 3000, 4000, 5000, 10000)
  , L = c(850, 1000, 2001, 3001, 4001, 5001)
  , f = c(1, 24, 112, 267, 598, 1146)

# Log-likelihood
logLikelihood = function(par, data){
  df <- function(f, low, up, par1, par2) {
    f*log(pgamma(up, par1, par2) - pgamma(low, par1, par2))
  shape = exp(par[1])
  rate = exp(par[2])  
  -sum(df(dat$f, dat$L, dat$U, shape, rate))

# Fit
res <- optim(par = c(3, -6), fn = logLikelihood, data = dat, control = list(reltol = 1e-15))

# Results
[1] 11.264619015  0.002143716

So the fitted gamma distribution has shape $11.265$ and rate $0.00214$. The density looks like this:


The mean of the gamma distribution is $11.265/0.00214=5254.7$ which is not too far from the mean of the grouped data ($5837.3$).

Here is the code to produce the graph:

par(mar = c(5.5, 5.5, 0.1, 0.5), mgp=c(4.5,1,0))
curve(dgamma(x, exp(res$par)[1], exp(res$par)[2]), from = 0, to = 15000, n = 1000L, lwd = 2, col = "steelblue", ylab = "Density", xlab = "Income", xaxt = "n", las = 1)
axis(1, at = seq(0, 15000, 2500))
  • $\begingroup$ As you said, I also think that the grouped data works better. I was checking the code and if I remove the exponential from the shape and rate and the control parameter in the optim function, the results don't change that much. This is not a big deal is it, or there might be some implications? Also, sorry for my inexperience with R graphs, but how did you plot the density? $\endgroup$
    – Seb
    May 23, 2021 at 0:57
  • $\begingroup$ @Seb I added the R code I used to produce the density plot. The reason I used the log-parameters for optimization is that it removes any boundaries and hopefully makes the fitting easier and more stable. As you said: The fit also works on the scale of the original parameters. $\endgroup$ May 23, 2021 at 6:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.