Bayesian updating - How to update?

285 Views Asked by At

Consider that we have two sets of dices. One set contains fair dices while the other set contains dices in which the probability of $6$ is $1/2$. The fair dices correspond to $99\%$ of the whole sample, which means the loaded dices occur with a frequency of $1\%$.

If I pick a dice at random and throw 3 sixes in a row, what is the probability that I have a loaded dice?

It is clear that Bayes rule can be applied:

$$ P( \mathrm{Loaded} \, | \, \mathrm{3 six} ) = \frac{ P( \mathrm{3 six} \, | \, \mathrm{Loaded} ) P ( \mathrm{Loaded} ) }{ P( \mathrm{3 six} )} $$

my question however is about the prior probability. After I've thrown one six, I can update my beliefs whether the die is loaded or not. Do I use $P(\mathrm{Loaded}) = 1/100$ or $P(\mathrm{Loaded}) = P( \mathrm{Loaded} \, |\, \mathrm{2 six})$, which to be computed needs $P(\mathrm{Loaded} \, | \, \mathrm{1 six})$?

These give different results. What is their interpretation? Which has the correct interpretation as probability of having a loaded die?

This calculation is an example in the book Richard Durbin, Sean R. Eddy, Anders Krogh, Graeme Mitchison - Biological Sequence Analysis - Probabilistic Models of Proteins and Nucleic Acids - Cambridge University Press (1999) in section 1.3. They compute $P(\mathrm{Loaded} \, | \, \mathrm{3 six})$ using directly $P(\mathrm{Loaded}) = 0.01$. They do not compute the probability of obtaining one six, then proceed to use that as a prior for the probability of obtaining two sixes, and so on. Is their procedure wrong? What is exactly what they are computing?

1

There are 1 best solutions below

2
On BEST ANSWER

First, let's update it three times. Let $L$ be the event that we chose the loaded die, and let $S$ be the event that we roll a $6$.

For the first update:

$$\Pr(L|S)=\frac{\Pr(L\wedge S)}{\Pr(S)}=\frac{\frac12\cdot\frac1{100}}{\frac1{100}\cdot\frac12+\frac{99}{100}\cdot\frac16}=\frac1{34}$$

In the second update, we use $\frac1{34}$ as the prior.

$$\Pr(L|S)=\frac{\Pr(L\wedge S)}{\Pr(S)}=\frac{\frac12\cdot\frac1{34}}{\frac1{34}\cdot\frac12+\frac{33}{34}\cdot\frac16}=\frac1{12}$$

We do the same thing a third time, using $\frac1{12}$ as the prior:

$$\Pr(L|S)=\frac{\Pr(L\wedge S)}{\Pr(S)}=\frac{\frac12\cdot\frac1{12}} {\frac1{12}\cdot\frac12+\frac{11}{12}\cdot\frac16}=\frac3{14}$$

The posterior probability that the die is loaded is $\frac3{14}.$

Alternatively, we just update once. Let $S_3$ be the event that we roll three consecutive sixes.

$$\Pr(L|S_3)=\frac{\Pr(L\wedge S_3)}{\Pr(S)}=\frac{\frac18\cdot\frac1{100}}{\frac1{100}\cdot\frac18+\frac{99}{100}\cdot\frac1{216}}=\frac3{14},$$

the same answer as before.