-margins- after -melogit-

[Clyde Schechter]

I'm experiencing a problem with -margins- after -melogit-. When I ask for -margins A#B-, Stata responds with the output appropriate for -margins A B-, that is, separate margins for A and B variables, but nothing for the combinations thereof.

. webuse nlswork, clear
(National Longitudinal Survey.  Young Women 14-26 years of age in 1968)

. melogit union i.nev_mar##i.collgrad || idcode:

Fitting fixed-effects model:

Iteration 0:   log likelihood = -10441.017  
Iteration 1:   log likelihood = -10422.908  
Iteration 2:   log likelihood = -10422.892  
Iteration 3:   log likelihood = -10422.892  

Refining starting values:

Grid node 0:   log likelihood = -8696.0695

Fitting full model:

Iteration 0:   log likelihood = -8696.0695  
Iteration 1:   log likelihood = -8039.0834  
Iteration 2:   log likelihood =   -7882.09  
Iteration 3:   log likelihood = -7858.7862  
Iteration 4:   log likelihood = -7858.2813  
Iteration 5:   log likelihood = -7858.2803  
Iteration 6:   log likelihood = -7858.2815  
Iteration 7:   log likelihood = -7858.2822  

Mixed-effects logistic regression               Number of obs     =     19,227
Group variable:          idcode                 Number of groups  =      4,150

                                                Obs per group:
                                                              min =          1
                                                              avg =        4.6
                                                              max =         12

Integration method: mvaghermite                 Integration pts.  =          7

                                                Wald chi2(3)      =      50.71
Log likelihood = -7858.2822                     Prob > chi2       =     0.0000
----------------------------------------------------------------------------------
           union |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
       1.nev_mar |   .1612591   .1059014     1.52   0.128    -.0463039     .368822
      1.collgrad |   .9411026   .1342232     7.01   0.000       .67803    1.204175
                 |
nev_mar#collgrad |
            1 1  |  -.3990216   .2271745    -1.76   0.079    -.8442754    .0462322
                 |
           _cons |  -2.804856   .0815507   -34.39   0.000    -2.964692   -2.645019
-----------------+----------------------------------------------------------------
idcode           |
       var(_cons)|   8.429619   .5000171                      7.504426    9.468875
----------------------------------------------------------------------------------
LR test vs. logistic model: chibar2(01) = 5129.22     Prob >= chibar2 = 0.0000

. margins nev_mar#collgrad

Predictive margins                              Number of obs     =     19,227
Model VCE    : OIM

Expression   : Marginal predicted mean, predict()

------------------------------------------------------------------------------
             |            Delta-method
             |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     nev_mar |
          0  |   .2251414   .0046998    47.90   0.000     .2159299    .2343529
          1  |   .2316429   .0076755    30.18   0.000     .2165991    .2466866
             |
    collgrad |
          0  |   .2120307   .0049563    42.78   0.000     .2023166    .2217448
          1  |   .2855803   .0098943    28.86   0.000     .2661878    .3049727
------------------------------------------------------------------------------

.

This doesn't happen with -mixed-. I haven't tried any of the other -me- commands yet. It didn't used to happen with version 14.

My setup is:

. about

Stata/MP 15.0 for Windows (64-bit x86-64)
Revision 29 Jun 2017
Copyright 1985-2017 StataCorp LLC

Total physical memory:     8269900 KB
Available physical memory: 5062552 KB

Single-user 2-core Stata perpetual license:
       Serial number:  REDACTED
         Licensed to:  Clyde Schechter
                       Albert Einstein College of Medicine

. update query
(contacting http://www.stata.com)

Update status
    Last check for updates:  07 Jul 2017
    New update available:    none         (as of 07 Jul 2017)
    Current update level:    29 Jun 2017  (what's new)

Possible actions

    Do nothing; all files are up to date.

I can't see that I'm doing anything wrong. Can others reproduce this problem on their installations?

[Martyn Sherriff]

Hello Clyde,
I cannot even get that far:

. webuse nlswork, clear
(National Longitudinal Survey.  Young Women 14-26 years of age in 1968)

. melogit union i.nev_mar##i.collgrad || idcode:

Fitting fixed-effects model:

Iteration 0:   log likelihood = -10441.017  
Iteration 1:   log likelihood = -10422.908  
Iteration 2:   log likelihood = -10422.892  
Iteration 3:   log likelihood = -10422.892  

Refining starting values:

Grid node 0:   log likelihood = -8696.0695

Fitting full model:

Iteration 0:   log likelihood = -8696.0695  
Iteration 1:   log likelihood = -8039.0834  
Iteration 2:   log likelihood =   -7882.09  
Iteration 3:   log likelihood = -7858.7862  
Iteration 4:   log likelihood = -7858.2813  
Iteration 5:   log likelihood = -7858.2803  
Iteration 6:   log likelihood = -7858.2815  
Iteration 7:   log likelihood = -7858.2822  

Mixed-effects logistic regression               Number of obs     =     19,227
Group variable:          idcode                 Number of groups  =      4,150

                                                Obs per group:
                                                              min =          1
                                                              avg =        4.6
                                                              max =         12

Integration method: mvaghermite                 Integration pts.  =          7

                                                Wald chi2(3)      =      50.71
Log likelihood = -7858.2822                     Prob > chi2       =     0.0000
----------------------------------------------------------------------------------
           union |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
         nev_mar |
              0  |          0  (base)
              1  |   .1612591   .1059014     1.52   0.128    -.0463039     .368822
                 |
        collgrad |
              0  |          0  (base)
              1  |   .9411026   .1342232     7.01   0.000       .67803    1.204175
                 |
nev_mar#collgrad |
            0 0  |          0  (base)
            0 1  |          0  (base)
            1 0  |          0  (base)
            1 1  |  -.3990216   .2271745    -1.76   0.079    -.8442754    .0462322
                 |
           _cons |  -2.804856   .0815507   -34.39   0.000    -2.964692   -2.645019
-----------------+----------------------------------------------------------------
idcode           |
       var(_cons)|   8.429619   .5000171                      7.504426    9.468875
----------------------------------------------------------------------------------
LR test vs. logistic model: chibar2(01) = 5129.22     Prob >= chibar2 = 0.0000

. margins nev_mar#collgrad
could not calculate numerical derivatives -- discontinuous region with missing values encountered
r(459);

Stata/IC 15.0 for Windows (64-bit x86-64)
Revision 29 Jun 2017
Copyright 1985-2017 StataCorp LLC

Total physical memory:     8263816 KB
Available physical memory: 3788244 KB

Stata/IC 15.0 for Windows (64-bit x86-64)
Revision 29 Jun 2017
Copyright 1985-2017 StataCorp LLC

Total physical memory:     8263816 KB
Available physical memory: 3788244 KB

Single-user Stata perpetual license:
       

. update query
(contacting http://www.stata.com)

Update status
    Last check for updates:  08 Jul 2017
    New update available:    none         (as of 08 Jul 2017)
    Current update level:    29 Jun 2017  (what's new)

Possible actions

    Do nothing; all files are up to date.

Martyn

[Roman Mostazir]

Hi Clyde,

Here is the output for margins from Stata.SE (14.2). The main model output is same as yours but clearly the 'margins' interaction contradict. I think Stata Corp people need to have a look at it:

melogit union i.nev_mar##i.collgrad || idcode:


Mixed-effects logistic regression               Number of obs     =     19,227
Group variable:          idcode                 Number of groups  =      4,150

                                                Obs per group:
                                                              min =          1
                                                              avg =        4.6
                                                              max =         12

Integration method: mvaghermite                 Integration pts.  =          7

                                                Wald chi2(3)      =      50.71
Log likelihood = -7858.2822                     Prob > chi2       =     0.0000
----------------------------------------------------------------------------------
           union |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
                 |
       1.nev_mar |   .1612591   .1059014     1.52   0.128    -.0463039     .368822
      1.collgrad |   .9411026   .1342232     7.01   0.000       .67803    1.204175
                 |
nev_mar#collgrad |
            1 1  |  -.3990216   .2271745    -1.76   0.079    -.8442754    .0462322
                 |
           _cons |  -2.804856   .0815507   -34.39   0.000    -2.964692   -2.645019
-----------------+----------------------------------------------------------------
idcode           |
       var(_cons)|   8.429619   .5000171                      7.504426    9.468875
----------------------------------------------------------------------------------
LR test vs. logistic model: chibar2(01) = 5129.22     Prob >= chibar2 = 0.0000

/*Margins*/

 margins nev_mar#collgrad

Adjusted predictions                            Number of obs     =     19,227
Model VCE    : OIM

Expression   : Marginal predicted mean, predict()

----------------------------------------------------------------------------------
                 |            Delta-method
                 |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
nev_mar#collgrad |
            0 0  |   .2094762   .0052294    40.06   0.000     .1992268    .2197255
            0 1  |   .2896667   .0107121    27.04   0.000     .2686714    .3106619
            1 0  |   .2226864   .0086199    25.83   0.000     .2057918     .239581
            1 1  |   .2685346   .0167344    16.05   0.000     .2357357    .3013335
----------------------------------------------------------------------------------

[Mukesh Kumar]

Hi Clyde,
Just wondering if this issue is sorted. I am facing the similar problem.
melogit stpn i.pm_centile26 sex vacynpcv13 smoker if week==26 & actualperruntime>=18|| cluster: , vce(robust) nolog

Mixed-effects logistic regression Number of obs = 437
Group variable: cluster Number of groups = 16

Obs per group:
min = 1
avg = 27.3
max = 53

Integration method: mvaghermite Integration pts. = 7

Wald chi2(12) = 413.21
Log pseudolikelihood = -165.61797 Prob > chi2 = 0.0000
(Std. Err. adjusted for 16 clusters in cluster)
------------------------------------------------------------------------------
| Robust
stpn | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
|
pm_centile26 |
2 | .1963519 .4183176 0.47 0.639 -.6235356 1.016239
3 | 2.076125 .7616225 2.73 0.006 .5833727 3.568878
4 | .7694861 .5451704 1.41 0.158 -.2990282 1.838
5 | 1.385389 .389097 3.56 0.000 .6227727 2.148005
6 | 1.01557 .4163496 2.44 0.015 .1995394 1.8316
7 | .7431474 .4297225 1.73 0.084 -.0990933 1.585388
8 | 1.198104 .6095183 1.97 0.049 .0034698 2.392738
9 | 1.394681 .4566441 3.05 0.002 .4996748 2.289687
10 | 1.739653 .871677 2.00 0.046 .0311979 3.448109
|
sex | -.5470586 .36124 -1.51 0.130 -1.255076 .1609588
vacynpcv13 | .2427846 1.260505 0.19 0.847 -2.22776 2.713329
smoker | .0914405 .3799396 0.24 0.810 -.6532273 .8361084
_cons | 1.490247 1.252858 1.19 0.234 -.9653099 3.945804
-------------+----------------------------------------------------------------
cluster |
var(_cons)| 4.21e-34 1.17e-33 1.86e-36 9.55e-32
------------------------------------------------------------------------------

.

. margins pm_centile26
could not calculate numerical derivatives -- discontinuous region with missing values encountere

[Joseph Coveney]

Clyde, that was fixed in the July 20th update:

12. margins has the following fixes:
. . .
b. margins, when an interaction was specified in marginslist after dfactor, gsem, mecloglog, meglm, meintreg, melogit, menbreg, menl, meologit, meoprobit, mepoisson, meprobit,
mestreg, metobit, mgarch, nl, nlsur, sem, spivregress, spregress, spxtregress, sspace, or ucm, produced margins for each factor variable separately instead of for the level
combinations specified by the interactions. This has been fixed.

What I don't understand is why your -update query- still leaves you with the June 29th update and not with anything later. By the way, there's a September 5th update now.

[Clyde Schechter]

Joseph, my original post predates the July 20th update. I did get that update when it came out, and for me the problem is now solved. (And I'm now running the September 5 update.)

Mukesh, this is not related to the problem originally posted and would have better been done as a new thread. Be that as it may, you may have an easy way out. Looking at your -melogit- output, the cluster-level variance component is, for all practical purposes zero. So you can forget about the cluster level in your model and just do a plain -logit-. You may have better luck with -margins- following that.

[Joseph Coveney]

Thought the thread was contemporary and didn't notice the date, sorry.