Swammerdami
Squadron Leader
In another thread, there seemed to be disagreement about how to calculate simple probabilities when there are three or more outcomes to consider.
It seemed best to start a math thread focusing on a very simple example depicting the confusion.
The three outcomes are S1, S2, and Not-S. But suppose we only care whether EITHER S1 or S2 is the outcome; we have no desire to distinguish between the two S cases.
I claim, with a simple demonstration as shown below, that to develop the best estimates we often need to handle the separate outcomes separately. But some in the thread seemed to think that "Bayesian analysis"* would magically glean what can be gleaned.
* - The Bayes Theorem is a very trivial fact, trivially proven. While very useful it is so trivial an analyst will often invoke it without naming it. (When your hear a layman invoking "Bayesian analysis," as though with pride and awe, the grasp is too often over-confident.
I've prepared a table showing the probabilities needed for a computation. We are given prior probabilities for each of the three outcomes. We are given two clues, A and B, assumed to be independent and their probabilities conditioned on each outcome. I show the correct conclusions ("Swammi") and the inferior results ("Carrier"). I don't know or care whether Richard Carrier's grasp of simple math is really this bad, but in the other thread his method was described this way.
With all the net probabilities 50-50, when S1 and S2 are NOT separated, A and B provide us with NO useful information. "Carrier" doesn't even see the two final columns in the table. The posterior probabilities are the same as the priors. Note that each "50%" can be considered EXACT (50.0000%) -- there is no "range" of estimates that might mitigate the flaw.
When you look at S1 and S2 separately, you can see that S1 increases the chance of A and decreases the chance of B. S2 vice versa. If A were true and B false (or vice versa) the chance of Not_S would fall, but with both (or neither) true, Not_S is most likely.
There are many VERY smart posters at IIDB, but those with a good mathematical bent are in short supply! I hope every Infidel who feels qualified shows up and comments.
It seemed best to start a math thread focusing on a very simple example depicting the confusion.
The three outcomes are S1, S2, and Not-S. But suppose we only care whether EITHER S1 or S2 is the outcome; we have no desire to distinguish between the two S cases.
I claim, with a simple demonstration as shown below, that to develop the best estimates we often need to handle the separate outcomes separately. But some in the thread seemed to think that "Bayesian analysis"* would magically glean what can be gleaned.
* - The Bayes Theorem is a very trivial fact, trivially proven. While very useful it is so trivial an analyst will often invoke it without naming it. (When your hear a layman invoking "Bayesian analysis," as though with pride and awe, the grasp is too often over-confident.
I've prepared a table showing the probabilities needed for a computation. We are given prior probabilities for each of the three outcomes. We are given two clues, A and B, assumed to be independent and their probabilities conditioned on each outcome. I show the correct conclusions ("Swammi") and the inferior results ("Carrier"). I don't know or care whether Richard Carrier's grasp of simple math is really this bad, but in the other thread his method was described this way.
With all the net probabilities 50-50, when S1 and S2 are NOT separated, A and B provide us with NO useful information. "Carrier" doesn't even see the two final columns in the table. The posterior probabilities are the same as the priors. Note that each "50%" can be considered EXACT (50.0000%) -- there is no "range" of estimates that might mitigate the flaw.
When you look at S1 and S2 separately, you can see that S1 increases the chance of A and decreases the chance of B. S2 vice versa. If A were true and B false (or vice versa) the chance of Not_S would fall, but with both (or neither) true, Not_S is most likely.
There are many VERY smart posters at IIDB, but those with a good mathematical bent are in short supply! I hope every Infidel who feels qualified shows up and comments.
Outcomes | Not-S | S | S1 | S2 |
---|---|---|---|---|
Prior probabilities | 50% | 50% | 25% | 25% |
P(A|.) | 50% | 50% | 90% | 10% |
P(B|.) | 50% | 50% | 10% | 90% |
A&B true --> Posterior probabilities (Carrier) | 50% | 50% | - | - |
A&B true --> Posterior probabilities (Swammi) | 73.5% | (26.5%) | 13.2% | 13.2% |