I was just doing some thought on the transformation for likelihoods to log likelihoods for binary response models and I realized what I am sure many people have realized before that the transformation is non-standard.

For instance, given a response pattern $x \in 0,1$ and a parameter set $\theta$ with $P$ defined as the probability $x=1$, the likelihood of a binary response is given as


It is fairly straightforward to break the above separately as $l(\theta;x=1)=P$ and $l(\theta;x=0)=(1-P)$. The next step people take for various reasons is to take the log likelihood


Which we can easily see is equivalent to taking the log of $l(\theta;x)$ observing that x can only be $1$ or $0$. However, this is a non-standard transformation ($\log(ab + cd) \ne a\log(b) + c\log(d)$ generally) and only achieved through reasoning (which I do not have a problem with). I am just wondering if there is some law, rule, or lemma in which the above transformation falls into?


The likelihood function is actually:

$l(\theta;x) = P^x(1-P)^{1-x}$

So if $x=0$ it reduces to:

$l(\theta;x=0) = P^0(1-P)^{1-0} = (1-P)$

if $x=1$ it reduces to:

$l(\theta;x=1) = P^1(1-P)^{1-1} = P$

Taking the logarithm of the likelyhood function results in:

$\ln(l(\theta;x)) = \ln(P^x(1-P)^{1-x}) = x \ln(P) + (1-x) \ln(1-P)$

| cite | improve this answer | |
  • $\begingroup$ Yep, that answers it. $\endgroup$ – Francis Smart Jul 30 '14 at 13:06

Your Answer

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

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