Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: GLLAMM earnings model convergence issue


From   Brent Gibbons <brent.gibbons@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   st: GLLAMM earnings model convergence issue
Date   Tue, 16 Aug 2011 12:13:22 -0400

Statalist users, I am working on a model that has a dependent variable of earnings in a period (for those receiving disability income so the maxes aren't too high and there are a fair number of zeros). This is experimental data with a treatment and control group that was measured over time (8 periods).

I am using the gllamm estimation for its the flexibility in being able to control for multiple levels of random coefficients along with being able to control for time fixed effects. The levels go from each observation (level 1) to the ID (level 2) to the location (level 3). The best choice for the family is the poisson with log link, however we cannot use the scale option for the variance, which is desirable. So I am trying to use the gamma family with log link as shown below:

eq het: location_2-location_2200
gllamm earnings baseline_earnings treatment_dummy period_(1...8) treatment_x_period, i(id location) family(gamma) link(log) s(het)

Note: the poisson version of the above code does converge, without the scale option of course. Right now, I am just trying to get the gamma model to run; with or without the scale option, I am not having any luck. I have been receiving the following error:

Without adaptive quadrature:  initial values not feasible, (error occurred in ML computation), (use trace option and check correctness of initial model)
With adaptive quadrature: Log-likelihood cannot be computed

Besides adaptive quadrature and varying the number of quadrature points, I have tried the following:

-running the poisson model and using the results as the initial values for the gamma model
-trying options init and nocons 
-the only option that had some success was setting the coefficient of variation for level 1, lns1, to 0 along with the poisson results - this allowed it to get past the initial values, but still failed to converge.

All suggestions are greatly appreciated. I am new to this model and am still learning its capabilities. Thanks, Brent
*
*   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/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index