Subpop MLM: Xtmixed

[Yan Feng]

Hi all,
I am running a multi-level modeling using complex survey data, which used a stratified, clustered, and unequally sampling design. For my research questions, I have to select a subsample, but due to the weight issues, I think I am not able to create a new tiny dataset which only includes my subsample. For regular regression analyses, I use SVY, subpop (): regress command, but I wonder if it would be possible to use a similar command for MLM with a subpop and continuous outcomes in Stata. If not possible, any thoughts on another approach to deal with it?

Thank you so much!

[Jeff Pitblado (StataCorp)]

meglm allows the svy prefix.

There is an example of a two-level logistic regression model in the manual entry.

Here is a simulated two-level linear regression example:

set seed 12345
* 3 strata
set obs 3
gen sid = _n
* 100 psus per stratum
expand 100
gen uid = _n
gen u = rnormal()
gen w1 = runiformint(2,4)
* 20 individuals per psu
expand 20
gen w2 = runiformint(2,4)
gen x = rnormal()
gen y = 1 - x + u + rnormal()
* select a subpop
gen sub = runiform() > .3

svyset uid, strata(sid) weight(w1) || _n, weight(w2)

svy, subpop(sub) : meglm y x || uid:

Here is a log of the results from running the above in Stata 15:

. set seed 12345

. * 3 strata
. set obs 3
number of observations (_N) was 0, now 3

. gen sid = _n

. * 100 psus per stratum
. expand 100
(297 observations created)

. gen uid = _n

. gen u = rnormal()

. gen w1 = runiformint(2,4)

. * 20 individuals per psu
. expand 20
(5,700 observations created)

. gen w2 = runiformint(2,4)

. gen x = rnormal()

. gen y = 1 - x + u + rnormal()

. * select a subpop
. gen sub = runiform() > .3

. 
. svyset uid, strata(sid) weight(w1) || _n, weight(w2)
Note: Stage 1 is sampled with replacement; further stages will be ignored for
      variance estimation.

      pweight: <none>
          VCE: linearized
  Single unit: missing
     Strata 1: sid
         SU 1: uid
        FPC 1: <zero>
     Weight 1: w1
     Strata 2: <one>
         SU 2: <observations>
        FPC 2: <zero>
     Weight 2: w2

. 
. svy, subpop(sub) : meglm y x || uid:
(running meglm on estimation sample)

Survey: Mixed-effects GLM

Number of strata   =         3                  Number of obs     =      6,000
Number of PSUs     =       300                  Population size   =     55,851
                                                Subpop. no. obs   =      4,165
                                                Subpop. size      =     38,967
                                                Design df         =        297
                                                F(   1,    297)   =    3289.50
                                                Prob > F          =     0.0000

------------------------------------------------------------------------------
             |             Linearized
           y |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
           x |  -.9973098   .0173886   -57.35   0.000     -1.03153   -.9630893
       _cons |   1.129561   .0623665    18.11   0.000     1.006825    1.252297
-------------+----------------------------------------------------------------
uid          |
   var(_cons)|   1.066697   .0882144                      .9064834    1.255227
-------------+----------------------------------------------------------------
     var(e.y)|   .9495035   .0249011                      .9017416     .999795
------------------------------------------------------------------------------

[Yan Feng]

Thank you so much. It was very helpful. However, I wonder if you know how to estimate ICC and AIC. I tried est ic and est gof, but both did not work. If I need to calculate ICC, it seems the formula for meglm is different from the formula for xtmixed models, right?