0% found this document useful (0 votes)
6 views17 pages

Intro-Bayes Theory

The document introduces Bayesian statistics, contrasting it with Frequentist approaches and emphasizing the importance of updating beliefs based on new data using Bayes' Theorem. It discusses the concept of prior distributions, likelihood, and posterior distributions, highlighting how to select appropriate priors based on existing knowledge and uncertainty. The document also addresses the implications of assumptions in Bayesian modeling and provides references for further reading.

Uploaded by

dominioncitie
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
6 views17 pages

Intro-Bayes Theory

The document introduces Bayesian statistics, contrasting it with Frequentist approaches and emphasizing the importance of updating beliefs based on new data using Bayes' Theorem. It discusses the concept of prior distributions, likelihood, and posterior distributions, highlighting how to select appropriate priors based on existing knowledge and uncertainty. The document also addresses the implications of assumptions in Bayesian modeling and provides references for further reading.

Uploaded by

dominioncitie
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 17

Introduction to

Bayesian Statistics
Camila Caiado
Associate Professor in Statistics
Warm-up

• Think of something you believe to be true.


• You have 1000kr to bet on this truth. If you are right, you receive the full
amount you have bet, otherwise you receive nothing. How much are you
willing to bet?
• If you didn’t bet the full 1000kr, why?
Bayesians vs Frequentists (XKCD)
Bayesians vs Frequentists (XKCD)
Bayesians vs Frequentists

“Using a spade for some jobs and shovel for others


does not require you to sign up to a lifetime of using
only Spadian or Shovelist philosophy, or to believing
that only spades or only shovels represent the One True
Path to garden neatness.” – Ken Rice

• A good statistician should behave like a Bayesian,


always learning, but only use it when it makes sense.
• There is more than just Bayesian and Frequentist
approaches.
Bayes’ Theorem

Conditional probability: the probability of A given B, for 𝑃 𝐵 ≠ 0,


𝑃(𝐴 ∩ 𝐵)
𝑃 𝐴𝐵 = .
𝑃(𝐵)
Similarly, for 𝑃 𝐴 ≠ 0,
𝑃(𝐴 ∩ 𝐵)
𝑃 𝐵𝐴 = 𝐵
𝑃(𝐴)
Bayes’ Theorem: 𝐴
𝑃 𝐵 𝐴 𝑃(𝐴) 𝐴∩𝐵
𝑃 𝐴𝐵 =
𝑃(𝐵)
Medical diagnosis example

Suppose 0.1% of the population have a certain disease.


A diagnostic test is carried out: the outcome of the test
is either positive or negative. It is known that the test is
90% reliable, that is, a person with the disease will
correctly test positive with probability 0.9. A person
without the disease will incorrectly test positive with
probability 0.1.
A person is tested for the disease. What is the
probability that they test positive? What is the probability
that they have the disease if they have tested positive?
Medical diagnosis example
D: person has the disease
𝑃 𝐷 = 0.001
𝑃 𝐷 − = 0.999
T: person tests positive
𝑃 𝑇|𝐷 = 0.9 ⇒ 𝑃 𝑇 𝐷 𝑃 𝐷 = 𝑃 𝑇 ∩ 𝐷 = 0.0009
𝑃 𝑇|𝐷 − = 0.1 ⇒ 𝑃 𝑇 𝐷 − 𝑃 𝐷 − = 𝑃 𝑇 ∩ 𝐷 − = 0.0999
So, the probability of testing positive is
𝑃 𝑇 = 𝑃 𝑇 ∩ 𝐷 + 𝑃 𝑇 ∩ 𝐷 − = 0.1008.
And the probability of having the disease given an positive test:
𝑃 𝑇 𝐷 𝑃(𝐷) 0.009
𝑃 𝐷|𝑇 = = ≈ 0.0893.
𝑃(𝑇) 0.1008
Discussion: What if we test this person again and we observe
another positive result?
Frequentist Treasure Hunt
A pirate has hidden a treasure in an island. You have a map but it isn’t
the most accurate so you and your mates decide to dig holes until you
find it. After digging each hole, you draw a 1 metre circle around the
hole, you have 95% confidence that the circle contains the treasure.
Frequentist Treasure Hunt
Repeat many times, your circles would contain the treasure 95% of
the time.
Say you choose one more location to dig. How do you decide if that’s
a bad choice? What if the pirate was with you and shouted if you are
close or not each time?

Treasure
Bayesian Treasure Hunt
Repeat as many times as you can afford. Sequentially update your
prior based on your previous observations. Use Bayes’ theorem to
update your beliefs about the treasure location.

If the pirate is with you, you might want to include in your model
whether you believe what they are saying or not!
Bayesian Inference

Prior distribution: this distribution describes what you know about the
parameter you want to estimate but excludes information in the data.
Likelihood: includes modelling assumptions and asserts how likely the
observed data are given your parameter.
Posterior distribution: update what we know about the parameter by
combining the prior and the data.
Posterior ∝ Likelihood × Prior
Update your model as more data becomes available and you will eventually
be able to bet on the location of the treasure.
How do you choose a prior?

• The prior should reflect your current knowledge of the problem. If


you don’t know what you are doing, call the experts. That is called
elicitation. If you are the expert and you are unsure, the prior
should reflect your uncertainty.
• The prior contains information external to the current study. If you
have a pilot study, that can be used to inform your prior. If you
don’t have pilot data, can you talk to experts in the field and try to
quantify their knowledge and uncertainty? Design of experiments
exist in the Bayesian world too!
Types of Priors
Subjective: this prior expresses the expert’s personal uncertainty. E.g. using
my personal experience, I think there is a 60% chance of rain today.
Objective and informative: information or data might exist to help formulate
a prior. Historical data and past experiments can help build this type of prior.
E.g. Yesterday rained and I have seen the weather forecast indicating rain, I
think now that there is a 80% chance of rain today.
Today’s posterior is tomorrow’s prior.”
Noninformative: this type of prior expresses ignorance. I have just arrived in
this country and have no idea what the weather is like so the chance of rain
must be 50-50. There is a lot of debate here around improper priors and how
informative an noninformative prior can be.
Garbage in, garbage out

• Like with any statistical tool, a Bayesian model is just as good as


your assumptions and your data.
• If your data is very informative, your likelihood will overshadow
your prior and whether you have succeeded in your elicitation is
irrelevant.
• Being uncertain doesn’t imply ignorance. You (or your expert)
might actually think that extreme values or cases are just as likely
to start with. E.g. for Covid it is unknown what the rate of
asymptomatic infections is so it isn’t invalid to think that the rate
can be between 50% and 80%.
If you like a good argument

• Bayesian epistemology and Bayesian dogmatism


• Bayesianism assumes rationality but humans aren’t always rational.
• Subjective vs Objective Bayes
• Noninformative priors are informative.
• What do you do if your experts think they know more than they actually do?
References and resources

• Gelman et. al, Bayesian Data Analysis


• Carlin and Louis, Bayesian Methods for Data Analysis
• Wakefield, Bayesian and Frequentist Regression Methods
• De Finetti, Theory of Probability: A Critical Introductory Treatment
• De Finetti, Probability, induction and statistics: The art of guessing
• Open University: https://www.open.edu/openlearn/science-maths-
technology/bayesian-statistics/content-section-0?active-tab=description-tab

You might also like