Statalist


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

RE: st: estimation using gllamm, oprobit model fails to converge


From   Cameron McIntosh <cnm100@hotmail.com>
To   STATA LIST <statalist@hsphsun2.harvard.edu>
Subject   RE: st: estimation using gllamm, oprobit model fails to converge
Date   Tue, 11 Aug 2009 09:37:13 -0400

Hi Frank, 
 
Yes, I can understand your wanting to use oprobit, so you can report your results in terms of the change in the probability of a more severe level of force, given a one-unit increase in a predictor. 
 
As suggested in the output, I would try to increase the # of quadrature points and/or impose your own starting values (perhaps the point estimates from the final iteration). I am not a Stata programmer but I imagine that would be pretty easy to do.
 
Cam 

----------------------------------------
> From: fjgallo@mac.com
> To: statalist@hsphsun2.harvard.edu
> Subject: Re: st: estimation using gllamm, oprobit model fails to converge
> Date: Mon, 10 Aug 2009 23:30:24 -0400
>
> Hi Cam,
>
> Thank you for your response. I am glad to see you participating on
> this list, besides SEMNET and MULTILEVEL. Below, I pasted output from
> the -xtmixed- run (preferable with mixed linear models), which I
> mentioned suggests groups matter. The variable police force is scaled
> along a severity metric (minimum force 1 to maximum force 9), ratings
> obtained from a panel of officers. Varieties of force weighted with
> average ratings. I agree with the no value-added in practice, but have
> concerns about theoretical violations (maybe with reviewers). Your
> thoughts? Thank you.
>
> Best,
> Frank
>
>
> xtmixed pforce || pd:, mle variance
>
> Performing EM optimization:
>
> Performing gradient-based optimization:
>
> Iteration 0: log likelihood = -3790.7576
> Iteration 1: log likelihood = -3790.7576
>
> Computing standard errors:
>
> Mixed-effects ML regression Number of obs
> = 3300
> Group variable: pd Number of groups
> = 16
>
> Obs per group: min
> = 22
> avg
> = 206.2
> max
> = 696
>
>
> Wald chi2(0)
> = .
> Log likelihood = -3790.7576 Prob> chi2
> = .
>
> ------------------------------------------------------------------------------
> pforce | Coef. Std. Err. z P>|z| [95% Conf.
> Interval]
> -------------
> +----------------------------------------------------------------
> _cons | 3.365989 .0380829 88.39 0.000
> 3.291348 3.44063
> ------------------------------------------------------------------------------
>
> ------------------------------------------------------------------------------
> Random-effects Parameters | Estimate Std. Err. [95% Conf.
> Interval]
> -----------------------------
> +------------------------------------------------
> pd: Identity |
> var(_cons) | .0177083 .0073768 .
> 0078269 .040065
> -----------------------------
> +------------------------------------------------
> var(Residual) | .5776353 .0142484 .
> 5503734 .6062477
> ------------------------------------------------------------------------------
> LR test vs. linear regression: chibar2(01) = 102.86 Prob>= chibar2
> = 0.0000
>
>
>
>
>
>
>
>
> On Aug 10, 2009, at 10:18 PM, Cameron McIntosh wrote:
>
> Hi Frank,
> You use an ordered probit model, yet your DV appears to have 23
> categories (22 thresholds).
> Could this be the problem? Perhaps simply modeling it as continuous
> would be more appropriate. With 23 categories, I don't think modeling
> the latent response variate y* offers much value-added over modeling
> the observed y.
> Cam
> ----------------------------------------
>> From: fjgallo@mac.com
>> To: statalist@hsphsun2.harvard.edu
>> Subject: st: estimation using gllamm, oprobit model fails to converge
>> Date: Mon, 10 Aug 2009 18:57:28 -0400
>>
>> Hi All,
>>
>> I used -gllamm- to run a Random Intercepts-Only Model. Below is the
>> output. The DV is ordinal, and believed to have a continuous latent
>> continuum. I am teaching myself multilevel modeling, Stata, and the
>> _gllamm- command. I am using Stata Version 11. Would the below failure
>> to converge suggest that there is little variability between j groups
>> on the DV? "or" Did I do something wrong in the model specification? I
>> found that this model - xtmixed pforce || pd:, mle variance -
>> converged and yielded a significant between-group difference that
>> suggested groups mattered. I would greatly appreciate any guidance and
>> resources. I have been using Rabe-Hesketh & Skrondal's (2008) book for
>> Stata. Thank you.
>>
>> Best,
>> Frank
>>
>>
>> . gllamm pforce, i(pd) nip(12) link(oprobit) adapt trace
>>
>> General model information
>> ------------------------------------------------------------------------------
>>
>> dependent variable: pforce
>> ordinal responses: oprobit
>> equations for fixed effects
>> _cut11: _cons
>> _cut12: _cons
>> _cut13: _cons
>> _cut14: _cons
>> _cut15: _cons
>> _cut16: _cons
>> _cut17: _cons
>> _cut18: _cons
>> _cut19: _cons
>> _cut110: _cons
>> _cut111: _cons
>> _cut112: _cons
>> _cut113: _cons
>> _cut114: _cons
>> _cut115: _cons
>> _cut116: _cons
>> _cut117: _cons
>> _cut118: _cons
>> _cut119: _cons
>> _cut120: _cons
>> _cut121: _cons
>> _cut122: _cons
>>
>>
>> Random effects information for 2 level model
>> ------------------------------------------------------------------------------
>>
>>
>>
>> ***level 2 (pd) equation(s):
>>
>> standard deviation of random effect
>> pd1: _cons
>>
>> number of level 1 units = 3300
>> number of level 2 units = 16
>>
>> Initial values for fixed effects
>>
>>
>> Iteration 0: log likelihood = -2735.2811
>>
>> Ordered probit estimates Number of obs
>> = 3300
>> LR chi2(0)
>> = 0.00
>> Prob> chi2
>> = .
>> Log likelihood = -2735.2811 Pseudo R2
>> = 0.0000
>>
>> ------------------------------------------------------------------------------
>> pforce | Coef. Std. Err. z P>|z| [95% Conf.
>> Interval]
>> -------------
>> +----------------------------------------------------------------
>> -------------
>> +----------------------------------------------------------------
>> _cut1 | -1.583387 .035338 (Ancillary parameters)
>> _cut2 | 1.196465 .0285592
>> _cut3 | 1.19802 .0285798
>> _cut4 | 1.202704 .0286423
>> _cut5 | 1.269557 .0295796
>> _cut6 | 1.271259 .0296046
>> _cut7 | 1.276389 .0296804
>> _cut8 | 1.464599 .0328633
>> _cut9 | 1.466823 .032906
>> _cut10 | 1.524945 .0340702
>> _cut11 | 1.529823 .0341722
>> _cut12 | 1.674974 .0375413
>> _cut13 | 1.678071 .0376208
>> _cut14 | 1.684313 .037782
>> _cut15 | 1.806059 .0412223
>> _cut16 | 1.809953 .0413422
>> _cut17 | 1.947163 .0460173
>> _cut18 | 2.262989 .0610329
>> _cut19 | 2.349713 .0665442
>> _cut20 | 2.361894 .0673782
>> _cut21 | 3.236012 .2017793
>> _cut22 | 3.428888 .2713744
>> ------------------------------------------------------------------------------
>> ------------------------------------------------------------------------------
>>
>>
>> start running on 10 Aug 2009 at 17:55:00
>>
>> Running adaptive quadrature
>> ------------------------------------------------------------------------------
>> Iteration 0 of adaptive quadrature:
>> Initial parameters:
>>
>> _cut11: _cut12: _cut13: _cut14: _cut15:
>> _cut16: _cut17: _cut18: _cut19: _cut110: _cut111:
>> _cons _cons _cons _cons _cons
>> _cons _cons _cons _cons _cons _cons
>> y1 -1.583387 1.196465 1.19802 1.202704 1.269557 1.271259
>> 1.276389 1.464599 1.466823 1.524945 1.529823
>>
>> _cut112: _cut113: _cut114: _cut115: _cut116:
>> _cut117: _cut118: _cut119: _cut120: _cut121: _cut122:
>> _cons _cons _cons _cons _cons
>> _cons _cons _cons _cons _cons _cons
>> y1 1.674974 1.678071 1.684313 1.806059 1.809953 1.947163
>> 2.262989 2.349713 2.361894 3.236012 3.428888
>>
>> pd1:
>> _cons
>> y1 .5
>>
>> Updated log likelihood:
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 0
>> 0 0 0 0 0 Convergence
>> not achieved: try with more quadrature points
>> finish running on 10 Aug 2009 at 17:55:31
>>
>>
>
> *
> * 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/
_________________________________________________________________
Stay in the loop and chat with friends, right from your inbox!
http://go.microsoft.com/?linkid=9671354
*
*   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   |   What's new   |   Site index