Re: st: Question about how streg using the constant-only model for starting values

[Steve Samuels <sjsamuels@gmail.com>]

I
It disregards the three constant model.  The starting iteration is based on the values from the full
model with:
         _t:        _t:        _t:      ln_p:
         age    smoking      _cons      _cons

Steve

On Dec 12, 2013, at 8:18 PM, Jonathan Eggleston wrote:

Thank you for your response.  I'm still a bit confused about how Stata generates the number for iteration 0 for the full model.  Does it use the estimates from the constant-only model at all to generate the numbers for iteration 0, or does it completely disregard these numbers for some reason?

On 12/12/2013 05:45 PM, Steve Samuels wrote:
> See Formulas and Methods in the Manual entry for -streg- (p. 387, Stata
> 13 Survival Manual). the frailty parameter multiplies S(t) and h(t) in
> the log-likelihood, and these terms are functions of the regression
> coefficients and the two other constants, and so are computed first. The
> frailty parameter is fitted last: the value (0.92) from the constants only
> model  is not useful, might not even  be feasible, as a starting
> value for the full model, where the eventual estimate is 0.33.
> 
> Steve
> 
> 
> On Dec 12, 2013, at 3:55 PM, Jonathan Eggleston wrote:
> 
> I have a question about how Stata uses the constant-only model in streg.  My understanding is that Stata uses the constant-only model to generate starting values for the main model.  I've used the "trace" option to see how this works, but I'm confused about how this works when I use frailty.  If I do
> 
> webuse bc, clear
> stset t, fail(dead)
> 
> and then I run
> 
> streg age smoking, d(weibull)  trace
> 
> I get this as part of my output
> 
> -------------+----------------------------------------------------------------
> Iteration 4:
> Coefficient vector:
>           _t:      ln_p:
>        _cons      _cons
> r1  -.6860931  -.3380965
>                                                   log likelihood = -138.83047
> 
> -------------+{hline 64
> 
> 
> Fitting full model:
> 
> --------------------------------------------------------------------------------------------------------------------------
> Iteration 0:
> Parameter vector:
>           _t:        _t:        _t:      ln_p:
>          age    smoking      _cons      _cons
> r1          0          0  -.6860931  -.3380965
> 
>                                                   log likelihood = -138.83047
> 
> 
> So without frailty, it makes sense how Stata uses the constant-only model.  However, when I add frailty with the command
> 
> streg age smoking, d(weibull) frailty(gamma) trace
> 
> I get this for the output
> 
> Iteration 4:
> Parameter vector:
>           _t:      ln_p:    ln_the:
>        _cons      _cons      _cons
> r1  -2.368818   .2681029   .9247814
> 
>                                                   log likelihood = -136.01201
> --------------------------------------------------------------------------------------------------------------------------
> 
> Fitting full model:
> 
> --------------------------------------------------------------------------------------------------------------------------
> Iteration 0:
> Parameter vector:
>           _t:        _t:        _t:      ln_p:    ln_the:
>          age    smoking      _cons      _cons      _cons
> r1   .1644213   .9056537  -11.20242   .3633523          0
> 
>                                                   log likelihood = -85.933969
> 
> 
> Without a clear connection between these two iterations.  Does anyone know how Stata uses the constant-only model with frailty? Thank you.
> 
> 
> 


-- 
Jonathan Eggleston
PhD Candidate
Department of Economics
University of Virginia

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


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