# Bayesian models

## Outline

### Topics

- Bayesian interpretation of probability models: from aleatoric to epistemic.
- Terminology: prior, likelihood, joint.

### Rationale

This week, we will give a new interpretation to the “bag of coin” example (where the uncertainty is “aleatoric”), turning it into a Bayesian model (where the uncertainty is “epistemic”). Mathematically this is the same model but with a different interpretation. This shift of interpretation is the basis of all Bayesian models.

## What is a Bayesian model?

A **Bayesian model** is a probability model equipped with:

- random variable(s) representing the data (we use \(Y\) in this course)
- random variable(s) for the unknown quantities (we use \(X\) in this course).

**Note:** concretely, a Bayesian model is a joint PMF over \(X\) and \(Y\), typically written using the “\(\sim\)” notation introduced last week.

## From coins to rockets

For this week’s rocket example, we will use the exact same probability model as the bag of coin example from last week.

Use the following correspondence to link last week’s example with this week’s:

- \((Y_i = 1)\):
- \(\leftrightarrow\) “\(i\)-th flip is a heads”
- \(\leftrightarrow\) “\(i\)-th launch is a success”

- \((X = k)\):
- \(\leftrightarrow\) “the coin drawn from bag has probability \(p = k/K\) of heads”
- \(\leftrightarrow\) “this type of rocket has probability \(p = k/K\) of success.”

## Terminology

The task of constructing a Bayesian model is often broken down into constructing:

- the
**prior:**the PMF of the unknown \(X\),^{1} - the
**likelihood:**the conditional PMF of observed \(Y\) given \(X\).

## Epistemic vs aleatoric probability

In the bag of coin example, we can replicate the whole “experiment” several times (both prior and likelihood).

In the rocket example, we can only replicate launches (likelihood), not the prior sampling part.

**Epistemic probability**: the part we cannot replicate.**Aleatoric probability**: the part that can be replicated.Bayesian statistics uses both types of probability,

whereas other fields of statistics, e.g. MLE, typically uses only aleatoric.

## Footnotes

if the unknown quantity is continuous, the prior will be expressed using a density. A term that captures both the continuous and discrete case is “distribution” i.e. “prior distribution”.↩︎