Re: st: Competing Risk for repeated event nominal dependent variables

[Mike Lacy <Michael.Lacy@colostate.edu>]

>Date: Sun, 28 Feb 2010 11:55:44 -0500
>From: "David A. Cort" <dcort@soc.umass.edu>
>Subject: st: Competing Risk for repeated event nominal dependent variables
>
>Dear Listserv,
>
>I am attempting to fit discrete-time event history model where the
>outcome is nominal and can be repeated over time. The social process is
>residential mobility. Instead of wanting to know the risk of moving into
>a community or neighborhood (Allison-type model), I'm interested in the
>risk of moving into a specific type of neighborhood. The dependent
>variable therefore has multiple categories (4 to be specific) for
>neighborhood type and time is discretized (into months). Any help
>concerning how STATA 10 can handle this type of setup would be very helpful.

I have looked for the same kind of thing, and have not run across any 
settled "practical" advice, so others' views here would be most 
welcome (i.e., needed). That being said, here's what I have come 
across up to this point.

1)  Converting to a person-period file, with k outcomes for each 
case, and then using multinomial logit with "stay" as the base category.

2)  The problem with 1) that it will almost certainly fail the IIA 
test. One approach is outlined in
Hill D H; Axinn W G; Thornton  A. 1993. "Competing hazards with 
shared unmeasured risk factors."
Sociological Methodology 23:245-77.
What they suggest, as best I understand it, amounts to a nested logit 
model, with (in your case), all the "move states" in one nest, and 
"stay" in the other. The first stage is move vs. stay, and the second 
stage is "which move," given "move."  When I tried something like 
this, I ran into issues with wanting/needing to use some of the same 
variables as predictors at both states (e.g., income influences the 
decision to move at all, and given the choice to move, it influences 
the kind of move), which -nlogit- seemed not to tolerate.

3) I have seen approaches using multinomial logit with a random 
intercept. I don't know how good a solution this is. In principle, 
the model should be estimable with -gllamm- or -mixlogit-, though 
perhaps not easily with a file as large as one is likely to get after 
conversion into person-period format. Perhaps someone can comment 
from on the feasibility as well as desirability of this approach.

4) I recently saw some suggestions (can't find them at the moment) to 
use multinomial logit, with a "discrete factor" rather than random 
intercept.  (I can't find a published source.  I found an online Rand 
working paper by Z. Nazarov at http://ssrn.com/abstract=1533001 
.  Although the application is different, the sense of "discrete 
factor" here is as in Mroz, T. A. 1999. "Discrete factor 
approximations in simultaneous equation models:Estimating the impact 
of a dummy endogenous variable on a continuous outcome." Journal of 
Econometrics, 92, 233-274.

This seems an interesting approach, and I would be particularly 
interested if anyone has ideas or pointers to "discrete factor" 
analyses using Stata.

Regards,
Mike Lacy, Dept. of Sociology, Colorado State University, Fort 
Collins, CO 80523 


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/