Unfortunately, our grant has run out, so we can only afford to sample 10 people:
We’ll use this data to estimate the probability of unemployment in two ways
Maximimum-likelihood (frequentist) estimation:
Posterior (Bayesian) estimation:
Posterior (Bayesian) estimation:
(E)
Posterior (Bayesian) estimation:
(E, E)
Posterior (Bayesian) estimation:
(E, E, E)
Posterior (Bayesian) estimation:
(E, E, E, U)
Posterior (Bayesian) estimation:
(E, E, E, U, U)
Posterior (Bayesian) estimation:
(E, E, E, U, U, E, E, E, U, E)
Posterior (Bayesian) estimation:
(E, E, E, U, U, E, E, E, U, E)
Maximum likelihood:
Posterior:
500 samples; uniform prior (click to animate)
500 samples; “informative” prior (click to animate)
The “posterior” is represented as a conditional probability distribution (the probability of varying values of
Bayes’ rule is a simple formula that allows us to ‘flip’ a conditional probability
And for our unemployment model this becomes
The posterior probability is our answer. It tells us everything we know about the probability of unemployment rate (
The prior probability is everything we claim to know about the probability of unemployment (
The evidence is just the average probability of seeing our sample across all possible values of
Fortunately we can almost always ignore it.
The likelihood is where our model lives.
In reality we know
In our model we know
The probability of getting
The likelihood is where our model lives.
In this case, a binomial distribution is a good choice. Given a particular probabily of unemployment
In practice, we rarely need to calculate the “evidence” (the denominator) in Bayes’ formula:
The posterior probability is proportional to (
Figures by Peter McMahan (source code)
Derrick Mercer, CC BY-SA 2.0, via Wikimedia Commons