Schedule for categorical outcome estimation

4 replies [Last post]
Steve's picture
Offline
Joined: 07/30/2009

Now that the closed beta is released we can get back to the categorical outcomes estimation. What is the status and expected completion date on the backend? Are there front end issues that need to be dealt with?

Steve's picture
Offline
Joined: 07/30/2009
The ordinal first plan sounds

The ordinal first plan sounds good.

There are, indeed, front end issues to discuss.

I'll start a thread over in the front-end forum.

And let's put categorical front-end issues on the agenda for next week's meeting (Friday, 8/14), if that's agreeable with everyone.

tbrick's picture
Offline
Joined: 07/31/2009
There're still open questions

There're still open questions about both the user experience and the front-end/back-end interface for ordinal models. I have most of a first back-end implementation for pure-ordinal estimation already prepared, so once the specifications are ironed out, it's just a matter of implementing those, generating some tests, and debugging.

We need to define how thresholds are represented in the front end and how they're passed to the back-end, as well as answering questions about ordinal levels that don't appear in the data.

I'm with Mike in recommending we start with a pure ordinal function. The joint ordinal/continuous estimation can follow afterwards.

Paras's picture
Offline
Joined: 07/31/2009
I am not sure if the planned

I am not sure if the planned version does this, but having the following two options in the front end, even if these are not implemented for a while may be useful.
(a) "variable-type" tag for each DV: {continuous, binary, ordinal, count,...}
(b) "error-distrtibution" for the DV (or "link-function"): {logistic, probit for binary & ordinal; poisson for count...}.

neale's picture
Offline
Joined: 07/31/2009
As soon as we give Tim a bit

As soon as we give Tim a bit of breathing space, it should not take too long. There's a bit of a trade-off depending on whether we go for the joint continuous and ordinal FIML straight out of the box, or do just ordinal first, then add - with some loss of efficiency - the joint part. Given the proximity of open beta, I'd say we should go for straight ordinal to begin with, and have people beat on it while coding the joint piece. All depends on Tim's other responsibilities over the next couple of months - afaik he has everything necessary outlined by me, but will likely want a bit of fine tuning here and there. An option for specifying what the highest category is (not necessarily what is observed in the dataset) are very important. They're critical for ML estimation of factor scores, for example.