1

I was looking at AIC, which is given by AIC = 2K - ln(L).

However, to my understanding, and observation, L = Log-Likelihood can be negative.

So in the case where L is negative, is AIC not applicable ?

Otherwise, ln(L) would give complex number, and I don't see complex number would be a useful value for this case.

Joon
  • 11
  • 1
  • You seem to be confused by the definitions. If L is the **likelihood** (a product of probabilities, so to speak) it will be always positive. Then ln(L) is the **loglikelihood**, which can (and actially will, since 0$\leq$prob$\leq$1) be negative. And then it doesn't really matter what are the values of AIC themselves, [only the differences matter](https://stats.stackexchange.com/questions/232465/how-to-compare-models-on-the-basis-of-aic/232494#232494). – corey979 Apr 01 '21 at 10:40

1 Answers1

1

Log-likelihood can be negative, indeed it will be if the likelihood is essentially a product of probabilities less than $1$. Thus $-\ln(L)$ will be positive, and so too will be $2k-\ln(L)$.

But the likelihood will be more than $0$, so the log-likelihood will be a real number.

So you do not need to worry.

Likelihoods are usually stated as being proportional up to a multiplicative constant, so log-likelihoods are really stated up to an additive constant. You do not need to worry about this either, so long as you use consistent calculations for comparisons of likelihoods .

Henry
  • 30,848
  • 1
  • 63
  • 107