MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

The paper, "The Multinomial-Poisson Transformation" by S. Baker (see presents "likelihood kernels" for multinomial variables, and then transforms these kernels to Poisson.

What is a likelihood kernel and how does it relate to the multinomial or Poisson PDF? I gather from the paper that the likelihood kernel would have the same maximum as the distribution itself, but would it have the same variance?

share|cite|improve this question
up vote 1 down vote accepted

I'm about 90% sure that in this case it means the equivalence class to which the likelihood function belongs, where two functions are equivalent precisely if either is a positive scalar multiple of the other. Notice that $$ \Pr(Y_1=y_1\ \&\ \cdots\ \&\ Y_J=y_j) = \frac{(y_1+\cdots+y_J)!}{y_1!\cdots y_J!}\cdot \prod_{j=1}^J \left(\frac{e^{\beta_j}}{e^{\beta_1}+\cdots+ e^{\beta_J}}\right)^{y_j} $$ As a function of $\beta_1,\ldots,\beta_J$, this is the likelihood function $$ L(\beta_1,\ldots,\beta_J) = \text{constant}\cdot \prod_{j=1}^J \left(\frac{e^{\beta_j}}{e^{\beta_1}+\cdots+ e^{\beta_J}}\right)^{y_j} $$ where "constant" in this case means not depending on $\beta_1,\ldots,\beta_J$. For many purposes, the values of such "constants" don't matter. The celebrated "likelihood principle" is a more-or-less philosophical position saying all statistical inferences should be based only on what I am surmising this author means by the "likelihood kernel". Certainly the function he's written is the simplest representative of the right equivalence class, which is what any sensible person would use.

Two contexts in which the statement that the value of that constant doesn't matter can be made into a precise mathematical statement and demonstrated to be true are (1) maximum-likelihood estimation and (2) finding a posterior probability distribution (in this case a distribution of $\beta_1,\ldots,\beta_J$) by multiplying the prior probability distribution by the likelihood function and then normalizing (Bayes' theorem).

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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