Conditional autoregression distribution #4504
Conversation
|
This looks like a great start! |
|
Regarding tests, yeah, pointwise is probably good enough here if we don't have a reference implementation. |
|
The test that I wrote for this distribution doesn't fit neatly into the framework already set up for testing against Scipy because while the CAR logp is defined to be that of a multivariate normal with a specific covariance structure, the implementation here isn't exactly equal - it differs from the MVN logp by an additive constant. |
|
I think all of the suggested changes have been addressed. Is the preferred workflow to individually mark each conversation as 'resolved' once the changes have been made, or should that be done by the reviewer? |
|
Best to let @ricardoV94 resolve them. |
Added reference for CAR distribution Running precommit checks Adding CAR test and updating CAR logp Extending CAR test to try multiple shapes Docstring change for CAR logp Adding bounds to logp for CAR distribution
…to car-distribution
| ) | ||
|
|
||
| def random(self, point=None, size=None): | ||
| raise NotImplementedError("Sampling from a CAR distribution is not supported.") |
There was a problem hiding this comment.
Should we turn this into a feature request?
There was a problem hiding this comment.
Sure, I've got a few leads on efficient methods for doing this. We can track that discussion at #4518.
|
Thanks for the reviews, @ricardoV94 and @twiecki! |
* Add conditional autoregression distribution (CAR)
* Add conditional autoregression distribution (CAR)
After many months of dragging my feet, I finally spent some time cleaning up the CAR implementation that @junpenglao showed in an example notebook (Issue #3689). I have modified that code to allow for multiple observations of a CAR-distributed vector conditional on a single fixed value for the adjacency matrix W.
This is definitely a work in progress and I am humbly asking for some assistance in devising a testing strategy for this distribution. Right now, all I can think of is to do pointwise tests on the
logpfor fixed values of the inputs.