The typical advise in this type of situation is to estimate a
simpler model. Whith these random effects models, that usually
start with only the random constant and gradually add random
coefficients, till the model breaks, and the model before that
is the most complicated model that is possible with your data.
Also, not specifying the -cov(unstr)- option results in a model
that is usualy much easier to estimate.
Hope this helps,
Maarten
--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany
http://www.maartenbuis.nl
--------------------------
--- On Wed, 6/1/10, Ada Lo <[email protected]> wrote:
> From: Ada Lo <[email protected]>
> Subject: st: standard errors failed to be calculated in xtmixed
> To: [email protected]
> Date: Wednesday, 6 January, 2010, 1:48
> Hi all,
>
> I'm trying to see if the rate of change of var X covary
> with the rate
> of change of var Y across time using xtmixed model.
>
> Runnnig this command in STATA11, none of the standard
> errors for the
> random parts were generated. I read in an earlier thread
> suggesting a
> translation of the variables with a mean or minimum value.
> I have done
> that to both BMI and memory but the results were the same.
> Why? In
> fact, can I still trust the estimates in the model anyway?
>
> Origianl command and output
> xtmixed varY c.varX##i.agegp1 c.Time##i.agegp1 || StudyID:
> Time varX,
> cov(unstr) var mle
>
> Performing EM optimization:
>
> Performing gradient-based optimization:
>
> Iteration 0: log likelihood = -4950.0532
> Iteration 1: log likelihood = -4946.0478
> Iteration 2: log likelihood = -4945.7244
> Iteration 3: log likelihood =
> -4945.0769 (not concave)
> Iteration 4: log likelihood = -4945.0712
> Iteration 5: log likelihood = -4945.0652
> Iteration 6: log likelihood = -4945.0614
> Iteration 7: log likelihood = -4945.0608
> Iteration 8: log likelihood = -4945.0606
>
> Computing standard errors:
> standard-error calculation failed
>
> Mixed-effects ML regression
> Number
> of obs = 1301
> Group variable: StudyID
>
> Number of groups =
> 482
>
> Obs per group: min =
> 1
> avg = 2.7
> max = 3
>
>
> Wald chi2(5) =
> 235.11
> Log likelihood = -4945.0606
> Prob
> > chi2 = 0.0000
>
>
> varY
> Coef. Std. Err.
> z P>z
> [95% Conf. Interval]
>
> varX .0710077 .0934028
> 0.76 0.447
> -.1120585 .2540739
> 1.agegp1 -23.19086 5.407541
> -4.29 0.000
> -33.78945 -12.59228
>
> agegp1#c.varX
> 1 .4936027
> .19496
> 2.53 0.011
> .1114881 .8757173
>
> Time .8146081 .100225
> 8.13 0.000
> .6181708 1.011045
>
> agegp1#
> c.Time
> 1 -.9569501 .1776534
> -5.39 0.000
> -1.305144 -.6087557
>
> _cons 77.36584 2.625491
> 29.47 0.000
> 72.21997 82.51171
>
>
>
> Random-effects Parameters
> Estimate Std. Err.
> [95% Conf. Interval]
>
> StudyID: Unstructured
> var(Time) .7128062
> .
> .
> .
> var(varX) .0130577
> .
> .
> .
> var(_cons) 178.6541
> .
> .
> .
> cov(Time,varZ) -.020441
> .
> .
> .
> cov(Time,_cons) -2.190107
> .
> .
> .
> cov(varZ,_cons) -1.40149
> .
> .
> .
>
> var(Residual) 55.52195
> .
> .
> .
>
> LR test vs. linear regression:
> chi2(6)
> = 407.56 Prob > chi2 =
> 0.0000
>
> Note: LR test is conservative and provided only for
> reference.
>
> ***********************************************************
>
> Now, using varZ and varY, in the exact same command as
> above, I got
> the output below. What does Hessian is not negative
> semidefinite
> really mean? What can I do about my data to make the model
> converge?
>
> Iteration 10: log likelihood = -4787.6438
> (backed up)
> numerical derivatives are approximate
> nearby values are missing
> numerical derivatives are approximate
> nearby values are missing
> Hessian is not negative semidefinite
>
> Mixed-effects ML regression
> Number
> of obs = 1290
> Group variable: StudyID
>
> Number of groups =
> 481
>
> Obs per group: min =
> 1
> avg = 2.7
> max = 3
>
>
> Wald chi2(5) =
> 100.38
> Log likelihood = -4797.623
> Prob
> > chi2 = 0.0000
>
>
> varY
> Coef. Std. Err.
> z P>z
> [95% Conf. Interval]
>
> varZ .1605253 6.491569
> 0.02 0.980
> -12.56272 12.88377
> 1.agegp1 -12.57239 9.058903
> -1.39 0.165
> -30.32752 5.182731
>
> agegp1#c.varZ
> 1
> 5.441863 11.21308
> 0.49 0.627
> -16.53536 27.41909
>
> Time .0161628 .0922692
> 0.18 0.861
> -.1646814 .1970071
>
> agegp1#
> c.Time
> 1 -.4144446 .1644543
> -2.52 0.012
> -.7367692 -.0921201
>
> _cons 51.92391 5.174088
> 10.04 0.000
> 41.78288 62.06493
>
>
>
> Random-effects Parameters
> Estimate Std. Err.
> [95% Conf. Interval]
>
> StudyID: Unstructured
> var(Time) .5704303
> .
> .
> .
> var(varZ) 113.3907
> .
> .
> .
> var(_cons) 99.90904
> .
> .
> .
> cov(Time,varZ) -1.132285
> .
> .
> .
> cov(Time,_cons) -.6162719
> .
> .
> .
> cov(varZ,_cons) -53.27241
> .
> .
> .
>
> var(Residual) 47.8776
> .
> .
> .
>
> LR test vs. linear regression:
> chi2(6)
> = 394.06 Prob > chi2 =
> 0.0000
>
> Note: LR test is conservative and provided only for
> reference.
> Warning: convergence not achieved; estimates are based on
> iterated EM
> ********************************************************************************************
>
> Your help would be much appreciated.
>
> Regards,
>
> Ada
> *
> * 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/
>
*
* 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/
