Dear STATA list,
I am trying to determine if my (i) regression results are valid using xttobit and (ii) how to interpret my results due to
a major shift in likelihood functions.
I have been using a random effects tobit model to a measure left-censored variable [a propensity to commit
delinquent acts among a panel of respondents]. My initial model provides negative log-likehood with an individual-level random effect that seems reasonable:
>Random-effects tobit regression Number of obs = 2401
>Group variable: aid Number of groups = 804
>Random effects u_i ~ Gaussian Obs per group: min = 2
> avg = 3.0
> max = 3
> Wald chi2(3) = 91.53
>Log likelihood = -4211.3782 Prob > chi2 = 0.0000
.
.
.
> /sigma_u | 2.079586 .2733142 7.61 0.000 1.5439 2.615272
> /sigma_e | 5.835778 .1588994 36.73 0.000 5.524341 6.147215
>-------------+----------------------------------------------------------------
> rho | .1126776 .0284796 .0662654 .1786233
>------------------------------------------------------------------------------
However, after adding several variables, the log-likelihood becomes highly positive and the
individual-level random effect that is of the order of E-39
>Random-effects tobit regression Number of obs = 2294
>Group variable: aid Number of groups = 768
>
>Random effects u_i ~ Gaussian Obs per group: min = 2
> avg = 3.0
> max = 3
>
> Wald chi2(9) = 300.30
>Log likelihood = 52807.689 Prob > chi2 = 0.0000
.
.
.
> /sigma_u | 4.17e-39 1.06e-40 39.19 0.000 3.96e-39 4.38e-39
> /sigma_e | 5.793114 .140685 41.18 0.000 5.517377 6.068852
>-------------+----------------------------------------------------------------
> rho | 5.19e-79 3.65e-80 4.52e-79 5.95e-79
>------------------------------------------------------------------------------
Both models converge to a concave solution and I am estimating about 35 quadrature points for the
Gauss-Hermite function. Other than the random effects, the estimated coefficients for Beta's are
consistent with theory/prior research. Its just the change in likelihood and a highly significant random effect
near zero from the first to the second models above that leave me scratching my head.
Thanks for your time and help!
Mike Roettger
*
* 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/
