7
$\begingroup$

Sometimes I have seen likelihood written as $L(\mu,\sigma |y)$ and sometimes as $L(y|\mu,\sigma)$.

I have been told that in the first case it means that there is a pre-assumed model depicting the probability density of $\mu$ and $\sigma$ parameterized by $y$. In the second case it means the pre-assumed model depicts the probability density of $y$, parameterized by $\mu$ and $\sigma$

I have seen the $|$ symbol used before in Bayes theory and read it as "given that"

I am from a programming background (C#) and tend to think of parameters as inputs to a function.

How should I be thinking about this word in statistics?

$\endgroup$
3
  • 1
    $\begingroup$ I am not aware of the latter usage of likelihood. But let others comment on this. $\endgroup$ Commented Oct 10, 2022 at 19:21
  • 2
    $\begingroup$ $L(y \mid \mu,\sigma)$ looks unconventional as a likelihood, and you would more usually see a parametrised density as $f(y \mid \mu,\sigma)$ or $p(y \mid \mu,\sigma)$ $\endgroup$
    – Henry
    Commented Oct 10, 2022 at 20:01
  • 1
    $\begingroup$ Related question $L(\theta;x)=f(x;\theta)$ vs. $L(\theta;x)\propto f(x|\theta)$ $\endgroup$ Commented Oct 11, 2022 at 17:17

2 Answers 2

7
$\begingroup$

“Parametrized by” means that the function $f$ of $x$ has additional parameters $\Theta$, so we can write it as $f(x;\Theta)$. In such a case, we call the function by $x$ given a fixed value of $\Theta$.

Likelihood has two meanings, the traditional on, and Bayesian. Traditionally, likelihood is written as

$$ L(\theta|x) \propto P(x|\theta) $$

The vertical bar $|$ is used on the right-hand side to denote conditional probability and is a slight abuse of notation on the left-hand side. People write it as $L(\theta|x)$ to show that we keep the data $x$ fixed, but we evaluate the function for different parameters $\theta$. In a Bayesian setting, you usually would not see the left-hand side notation, just the right-hand side will be called the likelihood (here you might have seen $L$ used instead of $P$). See Wikipedia entry on likelihood seems ambiguous for more discussion.

$\endgroup$
5
  • $\begingroup$ I suspect many Bayesians are happy with $L(\theta \mid x)$ though some would say $L(\theta \mid x) \propto P(x \mid \theta)$. A key point of such a likelihood is that it is function of $\theta$ but does not need to sum or integrate to $1$ over $\theta$. $\endgroup$
    – Henry
    Commented Oct 10, 2022 at 19:22
  • $\begingroup$ Hm. That is right, Tim. I never encountered likelihood to be written in that expression. $\endgroup$ Commented Oct 10, 2022 at 19:23
  • $\begingroup$ @Henry the multiplicative constant. That sums it up. $\endgroup$ Commented Oct 10, 2022 at 19:23
  • $\begingroup$ @User1865345 If I flip a biased coin with parameter $p$ of heads and see $H,T,H,T,H,H$ then the likelihood for a particular value of $p$ is the same as if I record $4$ heads and $2$ tails in total. It does not matter whether I write this as $p^4(1-p)^2$ or $15 p^4(1-p)^2$ $\endgroup$
    – Henry
    Commented Oct 10, 2022 at 19:35
  • $\begingroup$ Exactly @Henry. I agreed with your initial as well this comment. Nothing of any disagreement here. I just mentioned the multiplicative constant which doesn't matter. $\endgroup$ Commented Oct 10, 2022 at 19:37
1
$\begingroup$

What is parameter? What is a parametric model?

Definition $1.$ Let $(\Omega,\mathfrak F) $ be a probability space. The set of probability measures $\{ \mathbb P_\theta:~{\boldsymbol \theta\in\Theta}\}$ indexed by a parameter $\boldsymbol\theta$ is a parametric family if and only if $\Theta\in\mathbb R^n,~n\in\mathbb Z^{>0}$, and each probability measure is known when $\boldsymbol\theta$ is known. Here $\Theta$ is parameter space.

Remark $1.$ A parametric model assumes the population comes from a parametric family.

Example $1.$ A parametric family of $n\in\mathbb Z^{>0}$ dimensional normal distributions indexed by $( \boldsymbol\mu, \mathbf \Sigma) $ is

$$\{\mathcal N_n(\boldsymbol \mu,\mathbf \Sigma): \boldsymbol\mu\in\mathbb R^n,\mathbf \Sigma\in\mathcal M_n\}.\tag 1$$

So, basically the probability measure is labeled or indexed by parameter(s) and the primary objective of inference is to draw information about the parameter $\theta$ in order to know the probability measure (if it is $\sigma$-finite, then the set is identified by the densities; so the problem is to infer about the parameter to draw information about the associated pdf).

$\bullet$ $\mathcal L(\boldsymbol\theta|\mathbf y) $ denotes, given that $\mathbf Y= \mathbf y$ is realised or observed, the likelihood of parameter $\boldsymbol\theta.$


References and further read:

$\rm [I]$ Mathematical Statistics, Jun Shao, Springer Science$+$Business Media, $2003, $ section $2.1.2, $ p. $94.$

$\rm [II]$ Testing Statistical Hypotheses, E. L. Lehmann, Joseph P. Romano, Springer Science$+$Business Media, $2005, $ section $1.1, $ p. $3.$

$\rm [III]$ Statistical Inference, George Casella, Roger L. Berger, Wadsworth, $2002,$ section $6.3, $ p. $290.$

$\endgroup$

Your Answer

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

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