Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: gllamm, xtmixed, and level-2 standard errors
From
Stas Kolenikov <[email protected]>
To
[email protected]
Subject
Re: st: gllamm, xtmixed, and level-2 standard errors
Date
Tue, 16 Nov 2010 18:28:54 -0600
What if you run -xtreg, re- or -xtreg, mle-? They will give you the
same model, too.
On Tue, Nov 16, 2010 at 4:27 PM, Trey Causey <[email protected]> wrote:
> Greetings all. I am estimating a two-level, random-effects linear
> model. I know that gllamm is not the most computationally efficient
> option for this, but I am running into some very weird problems. I
> have ~21,000 individuals nested in 16 countries. I have 9
> individual-level predictors (listed as ind1-9) and 2 country-level
> predictors (listed as c1 and c2). When I estimate the model using
> gllamm, here are my results:
>
> . gllamm DV ind1 ind2 ind3 ind4 ind5 ind6 ind7 ind8 ind9 c1 c2,i(id)
> adapt nip(16)
>
> Running adaptive quadrature
> Iteration 0: log likelihood = -22865.024
> Iteration 1: log likelihood = -22841.735
> Iteration 2: log likelihood = -22807.82
> Iteration 3: log likelihood = -22797.118
> Iteration 4: log likelihood = -22794.274
> Iteration 5: log likelihood = -22792.672
> Iteration 6: log likelihood = -22791.582
> Iteration 7: log likelihood = -22791.557
> Iteration 8: log likelihood = -22791.428
> Iteration 9: log likelihood = -22791.426
>
>
> Adaptive quadrature has converged, running Newton-Raphson
> Iteration 0: log likelihood = -22791.426 (not concave)
> Iteration 1: log likelihood = -22791.426 (not concave)
> Iteration 2: log likelihood = -22789.86
> Iteration 3: log likelihood = -22789.371
> Iteration 4: log likelihood = -22788.767
> Iteration 5: log likelihood = -22788.613
> Iteration 6: log likelihood = -22788.604
> Iteration 7: log likelihood = -22788.604
>
> number of level 1 units = 21360
> number of level 2 units = 16
>
> Condition Number = 433.81863
>
> gllamm model
>
> log likelihood = -22788.604
>
> ------------------------------------------------------------------------------
> DV | Coef. Std. Err. z P>|z| [95% Conf. Interval]
> -------------+----------------------------------------------------------------
> ind1 | -.0020515 .000392 -5.23 0.000 -.0028198 -.0012833
> ind2 | -.3839988 .010841 -35.42 0.000 -.4052468 -.3627508
> ind3 | -.079134 .0113476 -6.97 0.000 -.1013749 -.0568931
> ind4 | .0800358 .0109386 7.32 0.000 .0585966 .101475
> ind5 | .0468417 .0048978 9.56 0.000 .0372423
> .0564411
> ind6 | .1685022 .0149735 11.25 0.000 .1391546 .1978497
> ind7 | -.2057474 .0171485 -12.00 0.000 -.2393579 -.1721368
> ind8 | -.093775 .0094251 -9.95 0.000 -.1122479 -.0753021
> ind9 | -.0080367 .0021554 -3.73 0.000 -.0122613 -.0038122
> c1 | .762577 .0802034 9.51 0.000 .6053813 .9197727
> c2 | .1763846 .0664327 2.66 0.008 .0461789 .3065903
> _cons | 1.265279 .1023452 12.36 0.000 1.064686 1.465872
> ------------------------------------------------------------------------------
>
> Variance at level 1
> ------------------------------------------------------------------------------
>
> .49269203 (.00476915)
>
> Variances and covariances of random effects
> ------------------------------------------------------------------------------
>
>
> ***level 2 (id)
>
> var(1): .09866295 (.01101541)
> ------------------------------------------------------------------------------
>
> When I estimate the model using xtmixed or xtreg, the output is
> essentially the same until I get to the country-level predictors; the
> coefficients are slightly different and the standard errors are
> approximately *ten* times smaller:
>
> . xtmixed DV ind1 ind2 ind3 ind4 ind5 ind6 ind7 ind8 ind9 c1 c2 || id:,mle
> Performing EM optimization:
> Performing gradient-based optimization:
> Iteration 0: log likelihood = -22785.965
> Iteration 1: log likelihood = -22785.965
> Computing standard errors:
> Mixed-effects ML regression Number of obs = 21360
> Group variable: id Number of groups = 16
> Obs per group: min = 730
> avg = 1335.0
> max = 2875
>
> Wald chi2(11) = 2296.06
> Log likelihood = -22785.965 Prob > chi2 = 0.0000
> ------------------------------------------------------------------------------
> DV | Coef. Std. Err. z P>|z| [95% Conf. Interval]
> -------------+----------------------------------------------------------------
> ind1 | -.0020472 .0003917 -5.23 0.000 -.002815 -.0012794
> ind2 | -.3840113 .0108422 -35.42 0.000 -.4052615 -.362761
> ind3 | -.0790874 .0113578 -6.96 0.000 -.1013483 -.0568264
> ind4 | .0799408 .0109411 7.31 0.000 .0584966 .101385
> ind5 | .0468955 .0048961 9.58 0.000 .0372994 .0564916
> ind6 | .1686695 .0149734 11.26 0.000 .1393222 .1980167
> ind7 | -.2054921 .0172501 -11.91 0.000 -.2393018 -.1716824
> ind8 | -.0941011 .0093698 -10.04 0.000 -.1124655 -.0757367
> ind9 | -.0079976 .0021584 -3.71 0.000 -.0122279 -.0037672
> c1 | .6718781 .2659761 2.53 0.012 .1505744 1.193182
> c2 | .1812668 .1083347 1.67 0.094 -.0310652 .3935988
> _cons | 1.306302 .2079643 6.28 0.000 .8986998 1.713905
> ------------------------------------------------------------------------------
> ------------------------------------------------------------------------------
> Random-effects Parameters | Estimate Std. Err. [95% Conf. Interval]
> -----------------------------+------------------------------------------------
> id: Identity |
> sd(_cons) | .2033876 .0363049 .1433454 .2885792
> -----------------------------+------------------------------------------------
> sd(Residual) | .7019342 .0033974 .695307 .7086246
> ------------------------------------------------------------------------------
> LR test vs. linear regression: chibar2(01) = 1684.50 Prob >= chibar2 = 0.0000
>
>
> This is obviously a big problem for establishing significance. I have
> read previous threads about this problem with xtlogit but have not
> seen it mentioned for linear models nor I have a seen a solution. It
> is not immediately clear to me why the estimates or standard errors
> should differ at all -- as Rabe-Hesketh and Skrondal say in their
> book, gllamm is not as computationally efficient for linear models but
> the results should be essentially the same. I have replicated this in
> Stata 10 and Stata 11.
>
> Thank you very much.
> Trey
> -----
> Trey Causey
> Department of Sociology
> University of Washington
>
> *
> * 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/
>
--
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
* 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/