Likelihood Method
Introduction
The likelihood method estimates parameters using an assumed distribution on the data and a randomly sampled dataset from this distribution. To estimate the parameters, one can use either standard methods of calculus or a computer. This page will cover the aspects of calculus behind the likelihood method.
The goal is to find the most likely value of the parameter(s) given a
set of random variables,
The logic underlying the likelihood method goes like this. Set up the likelihood function. The maximum likelihood estimator is the argument of the likelihood function that maximizes the likelihood function. Often, this is written as
to denote that the best guess is the maximal argument to the
likelihood function given the data
The likelihood function is defined relative to the density
function
Intuition (Bernoulli)
The intuition behind the product of the density functions goes like
this. Imagine you have
Now, imagine that you don't know that the value for
Next, since we know that Bernoulli distribution is an
appropriate model of coin flips, write this probability using
the density function of a Bernoulli distribution. Since the
Bernoulli distribution's density function maps
The last step in understanding the setup of the likelihood
function is to recognize that until we observe data such as
The discussion above captures the intuition behind the setup
of the likelihood function. From here the main differences are
notational and a conceptual understanding of how we can treat
this product as a function of the unknown parameter
To get from
to the general definition of the likelihood function, we generalize
the unknown parameter
Once we have
as a function of the unknown parameter
In an attempt to bring the general likelihood function back down to
earth, consider the following plot depicting the scenario introduced
above: the observations
Intuition (Normal)
Consider three data from a Normal distribution with parameter values
that I'm keeping secret:
With only three data points, displayed on the plot below as empty
circles, you are unlikely to guess exactly the values of
Try it; slide
Example
The last way we'll' demonstrate the maximum likelihood method is by
walking through an example. Suppose you have a sample of
for
To find the maximum likelihood estimate of
The goal is to find the value
Both humans and computers have difficulty working with products and exponents of functions. Therefore, it is common take the natural log of the likelihood function. This is so common, the log of the likelihood hood function has its own name, the log-likelihood function. The log-likelihood function is written as
where we've used properties of the natural log
Below is a plot of the log-likelihood
Recall from calculus that we can find local maxima/minima by
differentiating a function, setting the derivative equal to
zero, and solving for the variable of interest. In this
scenario, the variable of interest is the unknown parameter,
Often it's helpful to simplify the log-likelihood function to
aid differentiation. In this case, the most helpful
simplification is to realize that the first term within the
sum is constant with respect to
The symbol
To find the maximum of
Next, set the derivative equal to zero and solve for
Manipulate the expression until you find a solution for the
parameter of interest. At this point, we put a hat over the
parameter to recognize that it is our best guess of the unknown
parameter based on the random variables
The maximum likelihood estimator