Convergence Issues

[Divya Adiani]

Hi guys,

I hope you're all doing well.

I'm working on a project that aims to assess school predictors of internalising and externalising behaviour in students. To do this, I used MPlus to model different trajectories of internalising and externalising behaviour in males and females separately. Then, I ran binary logistic regression for the internalising model while controlling for covariates (ethnicity, family structure (lone), and SES). I am facing a problem with a specific school variable: school type in the male model in particular. It does not display Wald chi2(14) and Prob> chi2 value for both univariate and multivariate logistic regression, indicating issues with convergence. I have suspected multicollinearity and even perfect separation but could not find anything.

Univariate Stats:

 OR for variable: type

Logistic regression                                     Number of obs =  1,969
                                                        Wald chi2(3)  =      .
                                                        Prob > chi2   =      .
Log pseudolikelihood = -968.97847                       Pseudo R2     = 0.0055

                                   (Std. err. adjusted for 17 clusters in sch)
------------------------------------------------------------------------------
             |               Robust
       class | Odds ratio   std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
        type |
Academy -..  |          1  (base)
Academy S..  |   .8908153   .1432131    -0.72   0.472     .6500446    1.220765
Community..  |   .8912752   .3674544    -0.28   0.780     .3972637    1.999607
Foundatio..  |   .6816894   .1196533    -2.18   0.029     .4832587    .9615977
Voluntary..  |    .586074   .0683146    -4.58   0.000     .4663733    .7364973
             |
       _cons |   .2861446   .0333539   -10.73   0.000      .227702    .3595872
------------------------------------------------------------------------------
Note: _cons estimates baseline odds.

 ( 1)  [class]2.type = 0
 ( 2)  [class]3.type = 0
 ( 3)  [class]4.type = 0
 ( 4)  [class]5.type = 0

           chi2(  4) =   37.76
         Prob > chi2 =    0.0000

For multivariate statistics, I tried running bootstrapping, but it said that 69 out of 100 parameters couldn't be replicated.

 foreach v of varlist type {
  2. 
. bootstrap, reps(100): logistic class i.ethnicity i.lone i.ses i.`v',base vce(cluster sch) nolog
  3. 
. testparm i.`v'
  4. 
. }
(running logistic on estimation sample)

Bootstrap replications (100): xxxxxxxxxx.xxxxx.x.xx.xxx.xxxx....x.xx.xx.xxx.xxxx.xxxxxxx.60.xx.xx.x.70xxxxxx.
> xxxx..xx..x.90xxxxxx.x.100 done
x: Error occurred when bootstrap executed logistic.

Logistic regression                                     Number of obs =  1,568
                                                        Replications  =     31
                                                        Wald chi2(12) = 103.61
                                                        Prob > chi2   = 0.0000
Log pseudolikelihood = -746.29103                       Pseudo R2     = 0.0118

                                    (Replications based on 17 clusters in sch)
------------------------------------------------------------------------------
             |   Observed   Bootstrap                         Normal-based
       class | odds ratio   std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
   ethnicity |
White Bri..  |          1  (base)
White other  |   1.191957   .2404869     0.87   0.384     .8026455    1.770098
Asian/Asi..  |   1.063167   .2474998     0.26   0.792     .6736686    1.677865
Black Bri..  |   .5509081   .1521842    -2.16   0.031     .3205835    .9467106
Chinese/C..  |   .8226271    .266809    -0.60   0.547     .4356385    1.553387
Mixed Eth..  |   1.091711    .239749     0.40   0.689      .709866    1.678953
      Other  |   .8199145    .333577    -0.49   0.626     .3693708    1.820013
             |
        lone |
         No  |          1  (base)
        Yes  |   1.326553   .1189499     3.15   0.002     1.112753    1.581432
             |
         ses |
       High  |          1  (base)
        Low  |   .9944833   .1281618    -0.04   0.966     .7725042    1.280248
             |
        type |
Academy -..  |          1  (base)
Academy S..  |   .8945591   .1751268    -0.57   0.569     .6094952    1.312949
Community..  |   .8684049   .3656499    -0.34   0.738     .3804662    1.982114
Foundatio..  |   .7360338   .1755847    -1.28   0.199     .4611469    1.174779
Voluntary..  |      .6233   .0928465    -3.17   0.002     .4654815    .8346258
             |
       _cons |   .2509813   .0351542    -9.87   0.000     .1907288     .330268
------------------------------------------------------------------------------
Note: _cons estimates baseline odds.
Note: One or more parameters could not be estimated in 69 bootstrap replicates;
      standard-error estimates include only complete replications.

 ( 1)  [class]2.type = 0
 ( 2)  [class]3.type = 0
 ( 3)  [class]4.type = 0
 ( 4)  [class]5.type = 0

           chi2(  4) =   46.49
         Prob > chi2 =    0.0000

I haven't encountered this problem for the female model. Please suggest what I should do. Should I drop the variable altogether?