Forum: open-discussion

RE: censreg. two requests, 1 suggestion [ Reply ]
By: Arne Henningsen on 2014-04-16 12:14

Hi Paul

Regarding point 1: What do the warnings say about vcov? It might be the case that the final Hessian is (nearly) singular so that the variance covariance matrix cannot be calculated. I would try to use different optimization methods (BFGS, BFGSR, BHHH, NM) and would try to set argument "finalHessian" to "BHHH".

Regarding point 2: I agree that a predict method would be nice. IIRC, there are different approaches to calculate predicted values of censored regression models with random individual effects. Do you know a reference that describes how to do this?

Regarding point 3: I agree that offering the user to use another way to specify the panel structure could be a good idea.

You may want to register points 2 and 3 via the feature request "tracker" at R-Forge. You are welcome to implement these features yourself (with my help if you need it). Otherwise, I will try to do it myself as soon as I have time for this.

In the future, please post unrelated questions in different threads. Thanks a lot!

Best wishes,
Arne

censreg. two requests, 1 suggestion [ Reply ]
By: Paul Johnson on 2014-04-16 03:45

[forum:40980]

Hi, Arne

I'm fitting a model with 40000 rows and about 4 observations per case. Your CensReg model produces believable estimates with 40 parameters within 4 hours. That part is nice. I see some elements needing your advice.

1. What do you recommend when the model converges after 8 iterations, but the summary has "inf" for standard errors? Comments at bottom warn about vcov. This happend after seemingly minor specification changes.

If you can point at troubleshooting info, I would be grateful.

2 I need a predict method. How about it? I see people have proposed them to you. A regression without predict is not complete. I'm guessing plm/pdata makes that difficult.

Are you thinking on what to do about confidence intervals on predicted values? I don't know theory on incorporating censoring or mixed effscts there. I understand blups in general. Not in this case, tho.

It looks to me like relying on plm is a mistake. What benefit do you get? You require user to spawn a new dataset. It puts a big complication in the user interface. It does not tolerate regression models with constructed variables like ns(x, 6) because they dont show up in pdata. So I have to explicitly msnufacture those columns. Then doing predicted valuse for particular cases is very tough.

Is there some other R package people are using for this kind of model? I see Stata has xttobit, I'm a little interested to see how that works...