we start with basic model assesment:
the first model "base" includes only type of mask.
Then "base_E" I add ethnic group.
Then "base_E_S" I add gender.
Then "base_E_S_A" I add age.
Then "base_E_S_A_B" I add BMI. 

For performance we look at BIC which behaves better than AIC for huge data sets.

Akaike's information criterion and Bayesian information criterion

-----------------------------------------------------------------------------
       Model |          N   ll(null)  ll(model)      df        AIC        BIC
-------------+---------------------------------------------------------------
        base |      9,592          .  -5716.536       9   11451.07   11515.59
      base_E |      9,592          .  -5697.238      12   11418.48    11504.5
    base_E_S |      9,592          .  -5683.342      13   11392.68   11485.88
  base_E_S_A |      9,592          .  -5619.072      17   11272.14   11394.01
base_E_S_A_B |      9,592          .  -5612.042      20   11264.08   11407.46
-----------------------------------------------------------------------------

The smaller the BIC the better is the model. The winner is "base_E_S_A" as we 
do not need BMI.

Here are the details for the best model:


Mixed-effects logistic regression               Number of obs     =      9,592
Group variable:              id                 Number of groups  =      5,544

                                                Obs per group:
                                                              min =          1
                                                              avg =        1.7
                                                              max =          5

Integration method: mvaghermite                 Integration pts.  =          7

                                                Wald chi2(15)     =     261.82
Log likelihood = -5619.0725                     Prob > chi2       =     0.0000
----------------------------------------------------------------------------------
            PASS | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
           MODEL |
        3M 1863  |          1  (base)
       3M 1863+  |   .9325028   .0914436    -0.71   0.476     .7694478    1.130111
       3M 1873V  |   1.528529   .2086359     3.11   0.002      1.16974    1.997368
        3M 8833  |    1.53392   .1563768     4.20   0.000     1.256106    1.873178
Cardinal RFP3FV  |   .5125445   .0596051    -5.75   0.000      .408078    .6437541
          FSM18  |   .8064716    .140645    -1.23   0.217     .5729852    1.135102
         Other*  |   1.012831   .0877289     0.15   0.883     .8546885    1.200234
        3M 1873  |   .8933268   .1301806    -0.77   0.439     .6713801    1.188645
                 |
               E |
          white  |          1  (base)
          black  |   .6613418    .086852    -3.15   0.002     .5112582    .8554835
          asian  |   .5965401   .0520652    -5.92   0.000     .5027452    .7078339
          mixed  |   .7923941   .0893187    -2.06   0.039     .6353218    .9882997
                 |
             SEX |
         Female  |          1  (base)
           Male  |   1.571049   .1452736     4.89   0.000     1.310629    1.883214
                 |
             AGE |
18-24 years old  |          1  (base)
25-34 years old  |   3.410393   .3997582    10.47   0.000      2.71037    4.291215
35-44 years old  |    2.55482   .3124884     7.67   0.000     2.010236    3.246934
45-54 years old  |   2.749823    .338829     8.21   0.000     2.159837    3.500971
 55+  years old  |   3.457362   .5043512     8.50   0.000     2.597609    4.601675
                 |
           _cons |   1.433216   .1664456     3.10   0.002     1.141453    1.799556
-----------------+----------------------------------------------------------------
id               |
       var(_cons)|   2.485466   .2160378                      2.096143    2.947098
----------------------------------------------------------------------------------
Note: Estimates are transformed only in the first equation.
Note: _cons estimates baseline odds (conditional on zero random effects).
LR test vs. logistic model: chibar2(01) = 524.33      Prob >= chibar2 = 0.0000


you can clearly see which models perform better and worse than the reference, 
significantly. The results are also clear for age and sex and ethnic group. 

Of course, the question is if there is any interaction between model and ethnic group, 
or sex or age. So, that would mean that the effect of a mask would be different 
in different ethnic groups. This is what I have focussed now on. The result is that
there is no eveidence of interaction.


Akaike's information criterion and Bayesian information criterion

-----------------------------------------------------------------------------
       Model |          N   ll(null)  ll(model)      df        AIC        BIC
-------------+---------------------------------------------------------------
  base_E_S_A |      9,592          .  -5619.072      17   11272.14   11394.01
     model_E |      9,592          .  -5595.729      38   11267.46   11539.87
     model_S |      9,592          .  -5617.152      24    11282.3   11454.35
     model_A |      9,592          .  -5584.465      45   11258.93   11581.52
-----------------------------------------------------------------------------
Note: BIC uses N = number of observations. See [R] BIC note.

The last three models are the interaction models with ethnic group, sex,
and age, respectively. Top line is our best model from above. You see the BIC 
is worse for all interaction models. 

So, there are clear and significant effects of type of mask, ethnic group, 
sex and age on fit, but these effects do not interact and work independently.
Nice and clear result.