I'm hoping to get advice on the difference between two predictive margins outputs at their observed value (asobserved), after running a mixed-effect tobit regression.

Here is a sample / modified regression output, where the dependent variable (score) is censored from above and below.

HTML Code:
. metobit score ib2.sex##ib1.socioeconomic_status ||site:, ul(33) ll(0) nolog

Mixed-effects tobit regression                  Number of obs     =      2,517
                                                   Uncensored     =      1,908
Limits: lower = 0                                  Left-censored  =        605
        upper = 33                                 Right-censored =          4

Group variable:            site                 Number of groups  =         18
                                                Obs per group:
                                                              min =          1
                                                              avg =      139.8
                                                              max =        544

Integration method: mvaghermite                 Integration pts.  =          7

                                                Wald chi2(5)      =     157.08
Log likelihood = -6742.3941                     Prob > chi2       =     0.0000
------------------------------------------------------------------------------------------
                   score |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------------------+----------------------------------------------------------------
                   1.sex |  -1.086671   .8398094    -1.29   0.196    -2.732667    .5593255
                         |
    socioeconomic_status |
                      2  |   4.746687   .6094876     7.79   0.000     3.552114    5.941261
                      3  |   2.188796   .5561682     3.94   0.000     1.098726    3.278866
                         |
sex#socioeconomic_status |
                    1 2  |   1.175162   .9674324     1.21   0.224    -.7209706    3.071295
                    1 3  |   .1404838   .9092797     0.15   0.877    -1.641672    1.922639
                         |
                   _cons |   3.722792   1.754563     2.12   0.034     .2839111    7.161673
-------------------------+----------------------------------------------------------------
site                     |
               var(_cons)|   48.08904    17.2385                      23.81856    97.09051
-------------------------+----------------------------------------------------------------
             var(e.score)|   39.85155   1.345925                        37.299    42.57878
------------------------------------------------------------------------------------------
LR test vs. tobit model: chibar2(01) = 1258.16        Prob >= chibar2 = 0.0000
I'm interested in predicting the scores, fixing certain values of interest and leaving the other values at their observed value (i.e., Stata's default, asobserved function). In the margins output below, I deliberately exclude the "predict(ystar(0,33))" extension. The predicted scores match up with the regression output posted above. For example, the coefficient for sex (-1.087) can be back-calculated by subtracting the predicted scores between socioeconomic_status#sex (3.72 - 2.63), given the coefficient for sex in the regression is the difference at the base-level of "1" for socio-economic status:

HTML Code:
. margins i.socioeconomic_status##i.sex, asobserved post

Predictive margins                              Number of obs     =      2,517
Model VCE    : OIM

Expression   : Marginal linear prediction, predict()

------------------------------------------------------------------------------------------
                         |            Delta-method
                         |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------------------+----------------------------------------------------------------
    socioeconomic_status |
                      1  |   3.292355   1.726483     1.91   0.057    -.0914902      6.6762
                      2  |   8.504531    1.69834     5.01   0.000     5.175847    11.83322
                      3  |   5.536797   1.688287     3.28   0.001     2.227816    8.845778
                         |
                     sex |
                      1  |   5.706036   1.694424     3.37   0.001     2.385027    9.027045
                      2  |    6.37486   1.687879     3.78   0.000     3.066679    9.683041
                         |
socioeconomic_status#sex |
                    1 1  |   2.636121   1.804892     1.46   0.144    -.9014013    6.173644
                    1 2  |   3.722792   1.754563     2.12   0.034     .2839111    7.161673
                    2 1  |   8.557971   1.721823     4.97   0.000      5.18326    11.93268
                    2 2  |   8.469479   1.711487     4.95   0.000     5.115027    11.82393
                    3 1  |   4.965401    1.70449     2.91   0.004     1.624661    8.306141
                    3 2  |   5.911588   1.693488     3.49   0.000     2.592413    9.230762
------------------------------------------------------------------------------------------
The second predicted margins output includes the "predict(ystar(0,33))" extension into the command:

HTML Code:
. margins i.socioeconomic_status##i.sex, predict(ystar(0,33)) asobserved post

Predictive margins                              Number of obs     =      2,517
Model VCE    : OIM

Expression   : E(score*|0<score<33), predict(ystar(0,33))

------------------------------------------------------------------------------------------
                         |            Delta-method
                         |     Margin   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------------------+----------------------------------------------------------------
    socioeconomic_status |
                      1  |   5.619221   1.097938     5.12   0.000     3.467302     7.77114
                      2  |   9.421321   1.381066     6.82   0.000     6.714481    12.12816
                      3  |   7.142504   1.216324     5.87   0.000     4.758552    9.526455
                         |
                     sex |
                      1  |   7.325288   1.223074     5.99   0.000     4.928107     9.72247
                      2  |   7.791398   1.259778     6.18   0.000     5.322279    10.26052
                         |
socioeconomic_status#sex |
                    1 1  |   5.204548   1.100674     4.73   0.000     3.047266     7.36183
                    1 2  |   5.891215   1.145994     5.14   0.000     3.645108    8.137321
                    2 1  |   9.464794   1.402626     6.75   0.000     6.715697    12.21389
                    2 2  |   9.392806   1.390159     6.76   0.000     6.668143    12.11747
                    3 1  |   6.732649   1.193513     5.64   0.000     4.393406    9.071892
                    3 2  |   7.411337   1.242576     5.96   0.000     4.975931    9.846742
------------------------------------------------------------------------------------------
The values in some cases are quite different.

I expected that when I added in the predict(ystar, 0,33) extension, my margins results would have been those observed in the first margins output, which matches up with the censored regression output posted. What am I not understanding?