Running Stata 10.1 all updates.
I am trying to model these data:
DepVar = weight
fixed IndVars:
time
geolocation
treatment
I observe the weight of subjects given 1 of 4 different fractions of a
complex contaminant (control [saline], soluble, insoluble and total). This
variable is called treatment. The contaminant was taken from three
different geographical locations (variable: geolocation). The subjects are
observed over two different time periods before weighing, each subject
belonging to one or the other time periods (variable: time).
The design assigns four subjects to the control treatment condition of each
geolocation and time and six subjects to each other
treatment-geolocation-time condition.
As I see it, I have subjects nested within treatment, treatment nested
within geolocation and geolocation nested within time. There are no
repeated measurements on subjects.
I've tried specifying the model according to suggestions in Marchenko, Y.
2006. Estimating variance components in Stata. The Stata Journal, 6(1):
1-22., Y. Marchenko's answer to David Airey on 22 May 2008 Statlist, the
first edition of Rabe-Hesketh & Skrondal's Multilevel and Longitudinal
Modeling Using Stata, and the longitudinal Stata manual. I keep running
into non-convergence problems with my various formulations of the random
effects.
For example, simplifying the model by only looking at one time period, I
get:
START EXAMPLE:
xi: xtmixed weight i.treatment*i.geoloc if time==24 || _all: i.geoloc ||
_all: i.treatment*i.geoloc, variance
.
.
.
Performing EM optimization:
Performing gradient-based optimization:
numerical derivatives are approximate
flat or discontinuous region encountered
.
.
.
numerical derivatives are approximate
flat or discontinuous region encountered
Hessian has become unstable or asymmetric
FIXED EFFECTS (omitted)
------------------------------------------------------------------------------
Random-effects Parameters | Estimate Std. Err. [95% Conf.
Interval]
-----------------------------+------------------------------------------------
_all: Identity |
var(pm_group) | 4.79e+08 . .
.
-----------------------------+------------------------------------------------
_all: Independent |
var(_Itrea~1) | 8.19e+09 . .
.
var(_Itrea~2) | 8.19e+09 . .
.
var(_Itrea~3) | 8.19e+09 . .
.
var(_Ipm_g~2) | 6.70e+09 . .
.
var(_Ipm_g~3) | 6.70e+09 . .
.
var(_It~a1_2) | 2.46e+10 . .
.
var(_It~a1_3) | 2.46e+10 . .
.
var(_It~a2_2) | 2.46e+10 . .
.
var(_It~a2_3) | 2.46e+10 . .
.
var(_It~a3_2) | 2.46e+10 . .
.
var(_It~a3_3) | 2.46e+10 . .
.
var(_cons) | 2.23e+09 . .
.
-----------------------------+------------------------------------------------
var(Residual) | 2.07e+10 . .
.
------------------------------------------------------------------------------
LR test vs. linear regression: chi2(13) = 0.00 Prob > chi2 =
1.0000
Note: LR test is conservative and provided only for reference.
Warning: convergence not achieved; estimates are based on iterated EM
-------------END EXAMPLE
I get similar failure of convergence with other suggested random effects
specifications. There are no missing data.
It's clear I'm not getting how to specify random effects in nested models
(my first try at nested models), though I've had little difficulty in the
past with -xtmixed- in analyzing panel data over time.
I'd really appreciate some clues about what I'm doing wrong and how to
correct it.
Steve Rothenberg
CINVESTAV
Mexico City
Mexico
*
* 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/
