ICC coefficient with cross-classified modeling

[Maarten Huurne]

Hi,

I'm wondering whether it is possible to get an intraclass coefficient after a cross-classified regression. I used the following syntax:

mixed var1 var2 if var3>=5 || _all: R.ResponseID || profileid:

Using the following commands returns an error:

. iccvar
only _cons can have random effects
r(9);

. estat icc
estat icc not allowed after crossed-effects models
r(321);

Thanks in advance.

[Johannes Muller]

Maarten, have you found an answer since?
My hunch is that there is no way with mixed or meglm, but would be delighted to hear there is a way!

[Joseph Coveney]

would be delighted to hear there is a way!

There is.

.ÿ
.ÿversionÿ15.1

.ÿ
.ÿclearÿ*

.ÿ
.ÿsetÿseedÿ`=strreverse("1490777")'

.ÿ
.ÿ//ÿRaters
.ÿquietlyÿsetÿobsÿ10

.ÿgenerateÿbyteÿridÿ=ÿ_n

.ÿgenerateÿdoubleÿrid_uÿ=ÿrnormal(0,ÿsqrt(0.25))

.ÿ
.ÿtempfileÿraters

.ÿquietlyÿsaveÿ`raters'

.ÿ
.ÿ//ÿTargets
.ÿdropÿ_all

.ÿquietlyÿsetÿobsÿ100

.ÿgenerateÿbyteÿtidÿ=ÿ_n

.ÿgenerateÿdoubleÿtid_uÿ=ÿrnormal(0,ÿ1)

.ÿ
.ÿcrossÿusingÿ`raters'

.ÿ
.ÿ//ÿScores
.ÿgenerateÿdoubleÿwhatever1ÿ=ÿruniform()

.ÿgenerateÿdoubleÿwhatever2ÿ=ÿruniform()

.ÿ
.ÿgenerateÿdoubleÿscoÿ=ÿrid_uÿ+ÿtid_uÿ+ÿrnormal(0,ÿsqrt(0.25))

.ÿ
.ÿ*
.ÿ*ÿBeginÿhere
.ÿ*
.ÿmixedÿscoÿc.w*ÿ||ÿ_all:ÿR.tidÿ||ÿrid:ÿ,ÿnogroupÿnostderrÿnolrtestÿnolog

Mixed-effectsÿMLÿregressionÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿNumberÿofÿobsÿÿÿÿÿ=ÿÿÿÿÿÿ1,000

ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿWaldÿchi2(2)ÿÿÿÿÿÿ=ÿÿÿÿÿÿÿ0.42
Logÿlikelihoodÿ=ÿ-912.96748ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿProbÿ>ÿchi2ÿÿÿÿÿÿÿ=ÿÿÿÿÿ0.8107

------------------------------------------------------------------------------
ÿÿÿÿÿÿÿÿÿscoÿ|ÿÿÿÿÿÿCoef.ÿÿÿStd.ÿErr.ÿÿÿÿÿÿzÿÿÿÿP>|z|ÿÿÿÿÿ[95%ÿConf.ÿInterval]
-------------+----------------------------------------------------------------
ÿÿÿwhatever1ÿ|ÿÿÿ.0364551ÿÿÿ.0580891ÿÿÿÿÿ0.63ÿÿÿ0.530ÿÿÿÿ-.0773974ÿÿÿÿ.1503076
ÿÿÿwhatever2ÿ|ÿÿÿ.0097939ÿÿÿ.0566459ÿÿÿÿÿ0.17ÿÿÿ0.863ÿÿÿÿÿÿ-.10123ÿÿÿÿ.1208178
ÿÿÿÿÿÿÿ_consÿ|ÿÿÿ.3024975ÿÿÿÿ.178453ÿÿÿÿÿ1.70ÿÿÿ0.090ÿÿÿÿ-.0472639ÿÿÿÿ.6522588
------------------------------------------------------------------------------

------------------------------------------------------------------------------
ÿÿRandom-effectsÿParametersÿÿ|ÿÿÿEstimateÿÿÿStd.ÿErr.ÿÿÿÿÿ[95%ÿConf.ÿInterval]
-----------------------------+------------------------------------------------
_all:ÿIdentityÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ|
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿvar(R.tid)ÿ|ÿÿÿ.8506707ÿÿÿÿÿÿÿÿÿÿ.ÿÿÿÿÿÿÿÿÿÿÿÿÿ.ÿÿÿÿÿÿÿÿÿÿÿ.
-----------------------------+------------------------------------------------
rid:ÿIdentityÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ|
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿvar(_cons)ÿ|ÿÿÿ.2147508ÿÿÿÿÿÿÿÿÿÿ.ÿÿÿÿÿÿÿÿÿÿÿÿÿ.ÿÿÿÿÿÿÿÿÿÿÿ.
-----------------------------+------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿvar(Residual)ÿ|ÿÿÿ.2437365ÿÿÿÿÿÿÿÿÿÿ.ÿÿÿÿÿÿÿÿÿÿÿÿÿ.ÿÿÿÿÿÿÿÿÿÿÿ.
------------------------------------------------------------------------------

.ÿ
.ÿ//ÿVarianceÿcomponents
.ÿtempnameÿVrÿVtÿVe

.ÿscalarÿdefineÿ`Vt'ÿ=ÿexp(_b[lns1_1_1:_cons])^2ÿ//ÿTargets

.ÿscalarÿdefineÿ`Vr'ÿ=ÿexp(_b[lns2_1_1:_cons])^2ÿ//ÿRaters

.ÿscalarÿdefineÿ`Ve'ÿ=ÿexp(_b[lnsig_e:_cons])^2ÿ//ÿResidual

.ÿ
.ÿdisplayÿinÿsmclÿasÿtextÿ"ICCÿ=ÿ"ÿasÿresultÿ%04.2fÿ`Vt'ÿ/ÿ(`Vt'ÿ+ÿ`Vr'ÿ+ÿ`Ve')
ICCÿ=ÿ0.65

.ÿ
.ÿexit

endÿofÿdo-file


.

I could have sworn that there is a way to directly refer to the variance components as a regression coefficient, that is, _b[<something>] in one fell swoop, but I cannot find it at the moment. You can get there also via estat sd and so on, but that's no more convenient than just squaring the exponential as I've done above.

[Johannes Muller]

Dear Joseph,

Thank you very much!
This is great!
I am now however still struggling with the translation of this to the case of dichotomous dependent variables.
`Ve' won't be there, and generally I don't seem to be able to call the scalars in the same way after running melogit.
Do you have any suggestions how to pursue there?

Best

. ***ADJUSTMENT
. replace sco = 0 if sco < 0
(402 real changes made)

. replace sco = 1 if sco > 0
(598 real changes made)

. 
. * Begin here
. *
. 
. 
. . melogit sco c.w* || _all: R.tid || rid: , nogroup   nolrtest nolog
note: crossed random-effects model specified; option intmethod(laplace) implied

Mixed-effects logistic regression               Number of obs     =      1,000

Integration method:     laplace

                                                Wald chi2(2)      =       2.59
Log likelihood = -422.21191                     Prob > chi2       =     0.2741
------------------------------------------------------------------------------
         sco |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
   whatever1 |  -.1950722   .4040661    -0.48   0.629    -.9870271    .5968827
   whatever2 |   .6084129   .3975913     1.53   0.126    -.1708517    1.387677
       _cons |   .8416372   .7204371     1.17   0.243    -.5703935    2.253668
-------------+----------------------------------------------------------------
_all>tid     |
   var(_cons)|   10.14345   2.260062                      6.554347    15.69791
-------------+----------------------------------------------------------------
rid          |
   var(_cons)|   3.062974   1.464869                      1.199653    7.820434
------------------------------------------------------------------------------

[Joseph Coveney]

I am now however still struggling with the translation of this to the case of dichotomous dependent variables. Do you have any suggestions how to pursue there?

You need to look at how the coefficient vector is labeled.

.ÿ
.ÿ*
.ÿ*ÿBeginÿhere
.ÿ*
.ÿassertÿ!missing(sco)

.ÿquietlyÿreplaceÿscoÿ=ÿscoÿ>ÿ0

.ÿ
.ÿmelogitÿscoÿc.w*ÿ||ÿ_all:ÿR.tidÿ||ÿrid:ÿ,ÿdifficultÿnogroupÿnolrtestÿnolog
note:ÿcrossedÿrandom-effectsÿmodelÿspecified;ÿoptionÿintmethod(laplace)ÿimplied

Mixed-effectsÿlogisticÿregressionÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿNumberÿofÿobsÿÿÿÿÿ=ÿÿÿÿÿÿ1,000

Integrationÿmethod:ÿÿÿÿÿlaplace

ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿWaldÿchi2(2)ÿÿÿÿÿÿ=ÿÿÿÿÿÿÿ2.47
Logÿlikelihoodÿ=ÿ-422.21192ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿProbÿ>ÿchi2ÿÿÿÿÿÿÿ=ÿÿÿÿÿ0.2907
------------------------------------------------------------------------------
ÿÿÿÿÿÿÿÿÿscoÿ|ÿÿÿÿÿÿCoef.ÿÿÿStd.ÿErr.ÿÿÿÿÿÿzÿÿÿÿP>|z|ÿÿÿÿÿ[95%ÿConf.ÿInterval]
-------------+----------------------------------------------------------------
ÿÿÿwhatever1ÿ|ÿÿ-.1940105ÿÿÿ.3992221ÿÿÿÿ-0.49ÿÿÿ0.627ÿÿÿÿ-.9764713ÿÿÿÿ.5884504
ÿÿÿwhatever2ÿ|ÿÿÿ.6082802ÿÿÿ.4122425ÿÿÿÿÿ1.48ÿÿÿ0.140ÿÿÿÿ-.1997003ÿÿÿÿ1.416261
ÿÿÿÿÿÿÿ_consÿ|ÿÿÿ.8411393ÿÿÿÿ.740902ÿÿÿÿÿ1.14ÿÿÿ0.256ÿÿÿÿÿ-.611002ÿÿÿÿ2.293281
-------------+----------------------------------------------------------------
_all>tidÿÿÿÿÿ|
ÿÿÿvar(_cons)|ÿÿÿ10.14312ÿÿÿ2.283375ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ6.524584ÿÿÿÿÿ15.7685
-------------+----------------------------------------------------------------
ridÿÿÿÿÿÿÿÿÿÿ|
ÿÿÿvar(_cons)|ÿÿÿ3.062783ÿÿÿ1.502697ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ1.17082ÿÿÿÿ8.012025
------------------------------------------------------------------------------

.ÿ
.ÿ//ÿVarianceÿcomponents
.ÿtempnameÿvrÿvtÿve

.ÿscalarÿdefineÿ`vt'ÿ=ÿ_b[/:var(_cons[_all>tid])]

.ÿscalarÿdefineÿ`vr'ÿ=ÿ_b[/:var(_cons[rid])]

.ÿscalarÿdefineÿ`ve'ÿ=ÿ_pi^2ÿ/ÿ3

.ÿ
.ÿdisplayÿinÿsmclÿasÿtextÿ"ICCÿ=ÿ"ÿasÿresultÿ%04.2fÿ`vt'ÿ/ÿ(`vt'ÿ+ÿ`vr'ÿ+ÿ`ve')
ICCÿ=ÿ0.61

.ÿ
.ÿexit

endÿofÿdo-file


.

[Johannes Muller]

That's fantastic, thank you very much!!