George gmc107 at york dot ac dot uk wrote:
> I am experiencing exactly the same problems as those in Wiji
> Arulampalam's posting (see email subject title). I am trying to
> estimate a RE probit ( with a lagged dependent variable) and I am
> using Wooldridge's 2005 ( Journal of Applied Econometrics) solution
> to the initial conditions problem that still permits to estimate the
> model using the xtprobit command. However, while I can actually get
> proper estimations with convergent values ( eg full rank variance
> covariance matrix) and standard errors for two of my models (
> male1997-2002, female1991-1996) this does not seem to be the case
> with my (male1996-2002 and female 1997-2002) models. The maximum
> likelihood estimation for the male sample 1991-1996 stops at rho=0.1
> when fitting the the full model and then the results are such that
> no standard errors are reported for lnsig2u , sigma_u and rho (
> although s.es are reported for all the rest) Note that, the balanced
> panels do not seem to share any markedly different characteristics (
> eg descriptive stats of explanatory variables are reasonably
> similar, correlation coefficients between variables are also
> genarally low and similar for both time periods with regard to both
> male and female sub samples).
>
> For the female 1997-2002 sample though, the maximisation process
> does proceed normally to rho=0.2 but then the estimates are again
> such that no s.es are given for rho lnsig2u, sigma_u ( though in
> this case I get something like : Likelihood-ratio test of rho=0:
> chibar2(01) = 113.35 Prob >= chibar2 = 0.000 for the LR test whereas
> in the previous one when it does not proceed to rho=0.2 I get a Prob
> >= chibar2 = 1.000.
Weihua Guan replied to Wiji' Arulampalam's original message here:
http://www.hsph.harvard.edu/cgi-bin/lwgate/STATALIST/archives/statalist.0209/date/article-269.html
Weihua's response may be of interest. He notes that using the -from-
option got convergent results, although -quadchk- indicated the
quadrature technique used to evaluate the likelihood was not stable.
George does not state whether he is using Stata 8 or 9 -- in Stata 9
xtprobit re (as well as the random effects models for xtlogit,
xtcloglog, xtpoisson, xtintreg and xttobit) were updated to use
adaptive quadrature, which usually provides more robust results.
I would suggest trying the -from- option and increasing the number of
quadrature points as two possible things that might help here.
If George wants to email me privately with his data and the commands
he's using, I can take a closer look at it to see if anything else
comes to mind.
--Jean Marie
[email protected]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
