Here is a simple hierarchical model to attack this problem:
Instead of assuming a logistic change in failure probability, this model assumes that it is piecewise constant with an unknown change point:
To summarize all the marginal likelihood for different models we tried on the challenger data:
Description |
Blang model |
Marginal likelihood |
---|---|---|
Basic logistic regression |
Challenger.bl |
-18.8 |
Spike-and-slab logistic regression |
ChallengerSpiked.bl |
-17.7 |
Spike-and-slab logistic regression (modified prior) |
ChallengerSpiked2.bl |
-17.2 |
Change point model |
ChallengerChangePoint.bl |
-13.9 |
So we can see that the change point model outperforms the other ones according to the marginal likelihood criterion.
Here is an implementation of a simple model to approach the German Tank problem.
Read the example given in this wikipedia article, In the code below, I am using the same observation as in the wikipedia example (summarized with the sufficient statistics that 4 serial numbers were sampled, and that the maximum observed in these 4 is the serial number '60'):
We obtain a credible interval from 61 to 91.
The implementation of the likelihood, SerialSampling
, is shown below: