I have a "continuous" sample, and I want to fit it to particular model using MLE. For this I need to be able to compute its likelihood function, and I am not sure what the correct definition should be. To be specific, here is a very simple version of my situation. (The actual situation has different number of bins with different bin sizes.)

I have collected 20ng of air particles, of which 5ng have particle diameter less than 1$\mu$m, 12ng have particle diameter 1-10$\mu$m, and 3ng have particle diameter bigger than 10$\mu$m. I want to fit a particle size distribution function from this sample.

For this, given an actual size distribution with cdf F, I need to compute the log-likelihood function. Let $F_1=F(1\mu m)$, $F_2=F(10\mu m)-F(1\mu m)$, and $F_3=1-F(10\mu m)$ (the probabilites of landing in each bin.)

Now, instead of this "continuous" sample of 20ng of particles, if I actually had a "discrete" sample of 20N particles, distributed 5N, 12N, 3N in the three bins, the log-likehood function would have been $$5N\log(F_1)+12N\log(F_2)+3N\log(F_3),$$ which scales linearly with N. (In reality, N is close to the Avogardo's number.) So, my first question is:

Q1) What is a correct log-likelihood function in this case? It seems I should divide the above expression by 20N. Is that correct, and is there a reference for this?

Of course, for fitting a particular model (say a lognormal distribution, which has 2 parameters), we will maximize the above formula, and dividing by N doesn't change anything. The main reason is the second question:

Q2) If I am comparing two different models (say a lognormal distribution (2 parameters) vs a sum of 2 lognormal distributions (5 parameters)) using AIC or BIC, what should the formula be? AIC has a term $2kn/(n-k-1)$, while BIC has a term $k\ln(n)$, where n is the sample size. What should n be in my case? (The issue is that, the log-likelihood formula above is linear in N, while the above two terms in AIC and BIC are not linear in n.)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.