report var(residual) / melogit vs. xtmixed ?

[urban kim]

l am working about cross-classified multilevel model

dependent variable : immigration y/n (binary, immigration : 1, none : 0), level1 : personal variable, level2 : move-out reigon(varlist : depid_), level2: move-in region(varlist : arrid_)



meqrlogit immigration(binary) || all : R.depid_(move-out region) || arrid_(move-in region):, var


i use "var" option. but, didn't report the result of variance(residual). what's the problem?



can i use xtmixed model instead of melogit model in binary dependent variable?

[Clyde Schechter]

In a logistic model, the residual variance is always pi²/3.

[Weiwen Ng]

Also, xtmixed is the old name for mixed. Both fit hierarchical linear models. If you fit one to binary data, that's like fitting a fancier linear probability model. Some people will get nervous if you do that, because we do have logistic regression. LPMs actually work fairly well, though.

[Menglei Gu]

Professor Schechter, I am working on a study using "meologit" in Stata. But when I type "estat ICC" after "meologit dv || Country:", it prompts "requested action not valid after most recent estimation command". I tried to compute ICC manually, but I don't know how to compute the within-group variation, since the code "meologit dv || Country:" only returns the between-group variation. I also tried the "meologit dv || Country:, var", but it turns out "option var not allowed" . So what should I do? Is the residual variance "pi2/3" the within-group variance in my situation?

[Clyde Schechter]

Well, yes, the within-group variance in an ologit model is always pi²/3. But what Stata is trying to tell you is that the ICC is not definable after the model you have run. I can't think of any reason this would be true if your code is truly -meologit dv || Country:"-, so I'm wondering if you are presenting here a simplified or edited version of your actual -meologit- command. In particular, the ICC is not defined in models with crossed-random effects or non-independent covariance structure.

So please post back showing the exact -meologit- code you ran, the output that Stata gave you from that command, and the -estat icc- command and Stata's response to that. Show everything, exactly, without any editing. Place it between code delimiters for legibility.

Added: -meologit- does not have a -var- option. It reports the variance components as variances, not standard deviations in any case (at least in recent versions of Stata).

[Menglei Gu]

Thanks for your reply, and here are the codes and output (I don't understand the "code delimiters" you said , instead I bold the code). Looking forward to your reply, thanks.

. meologit dv || Country:

Fitting fixed-effects model:

Iteration 0: log likelihood = -16279.597
Iteration 1: log likelihood = -16279.597

Refining starting values:

Grid node 0: log likelihood = -14921.29

Fitting full model:

Iteration 0: log likelihood = -14921.29
Iteration 1: log likelihood = -14826.973
Iteration 2: log likelihood = -14826.52
Iteration 3: log likelihood = -14826.516
Iteration 4: log likelihood = -14826.516

Mixed-effects ologit regression Number of obs = 11,521
Group variable: Country Number of groups = 74

Obs per group:
min = 12
avg = 155.7
max = 1,645

Integration method: mvaghermite Integration pts. = 7

chi2() = .
Log likelihood = -14826.516 Prob > chi2 = .
------------------------------------------------------------------------------
dv | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
/cut1 | .0753714 .1448334 -.2084968 .3592396
/cut2 | 1.38835 .1454075 1.103356 1.673343
/cut3 | 1.870417 .1459492 1.584361 2.156472
/cut4 | 2.472127 .1470642 2.183886 2.760367
/cut5 | 3.296253 .1499327 3.00239 3.590115
/cut6 | 3.964193 .1542821 3.661805 4.26658
-------------+----------------------------------------------------------------
Country |
var(_cons)| 1.47951 .265326 1.041043 2.10265
------------------------------------------------------------------------------
LR test vs. ologit model: chibar2(01) = 2906.16 Prob >= chibar2 = 0.0000

. estat icc
requested action not valid after most recent estimation command
r(321);

[Clyde Schechter]

I'm stumped. The regression results look plausible, and the model has no crossed-effects or exotic covariance structure. The estimation converged. But for some reason, you cannot get -estat icc- to run. I might ask you to post an example data set that reproduces your problem, but I suspect that would not resolve the issue. I have checked on my own installation that -estat icc- runs following a model similar to yours. One question: what version of Stata are you running? I don't know, but it may be that some earlier versions do not allow -estat icc- after -meologit-. I suggest you type -help estat postestimation- and see if that page contains a link to estat icc. If it doesn't, that would suggest your version of Stata doesn't support this calculation.

If there is a link to estat icc there, then you shouldn't be encountering this difficulty. Then I would try running -update all- to make sure that your Stata is the latest possible update of your version (and that all the components are consistent with each other.) If that doesn't solve the problem, I would refer it to technical support.

[Leonardo Guizzetti]

-estat icc- support following -meologit- was added in the 16oct2019 update for Stata 16:

. help whatsnew16

-------- update 16oct2019 --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
...
     3. estat icc is now supported after meoprobit, meologit, and meglm with the ordinal family and probit or logit links.
...

[Menglei Gu]

Thanks Professor Schechter and Leonardo. I used Stata 15, that explained why the "estat icc" did not work after "meologit". After updating Stata, finally it works.
Besides, I used the pi2/3 (3.29) to compute the ICC, the two results are almost the same. Thanks a lot, Professor Schechter.