margins after jackknife

[Jean-Michel Galarneau]

Hello Statalisters, I've been attempting to run margins after a jackknifed Firth logistic regression and have been so far unsuccessful.
I've replicated the issue here with the dataset auto and some arbitrarily created variables.

. use https://www.stata-press.com/data/r17/auto.dta,clear
(1978 automobile data)

. gen outcome=0

. replace outcome=1 if price>6000
(23 real changes made)

. tab outcome

    outcome |      Freq.     Percent        Cum.
------------+-----------------------------------
          0 |         51       68.92       68.92
          1 |         23       31.08      100.00
------------+-----------------------------------
      Total |         74      100.00

. *for the sake of argument, pretend make is our cluster
. gen makecluster=1 if _n<=15
(59 missing values generated)

. replace makecluster=3 if _n>15 & _n<=30
(15 real changes made)

. replace makecluster=4 if _n>30 & _n<=45
(15 real changes made)

. replace makecluster=5 if _n>45 & _n<=60
(15 real changes made)

. replace makecluster=6 if _n>60
(14 real changes made)

. tab makecluster,m

makecluster |      Freq.     Percent        Cum.
------------+-----------------------------------
          1 |         15       20.27       20.27
          3 |         15       20.27       40.54
          4 |         15       20.27       60.81
          5 |         15       20.27       81.08
          6 |         14       18.92      100.00
------------+-----------------------------------
      Total |         74      100.00

. logistic outcome i.foreign##c.mpg ,vce(cluster makecluster) coeflegend

Logistic regression                                     Number of obs =     74
                                                        Wald chi2(2)  =      .
                                                        Prob > chi2   =      .
Log pseudolikelihood = -33.976694                       Pseudo R2     = 0.2591

                             (Std. err. adjusted for 5 clusters in makecluster)
-------------------------------------------------------------------------------
      outcome | Coefficient  Legend
--------------+----------------------------------------------------------------
      foreign |
     Foreign  |  -.9883114  _b[1.foreign]
          mpg |  -.4038123  _b[mpg]
              |
foreign#c.mpg |
     Foreign  |   .1612575  _b[1.foreign#c.mpg]
              |
        _cons |   6.383446  _b[_cons]
-------------------------------------------------------------------------------
Note: _cons estimates baseline odds.

. jackknife _b[1.foreign] _b[mpg] _b[1.foreign#c.mpg] _b[_cons],eclass cluster(makecluster) mse:firthlogit outcome i.foreign##c.mpg,
(running firthlogit on estimation sample)

Jackknife replications (5): ..... done

Jackknife results                                           Number of obs = 74
                                                            Replications  =  5

      Command: firthlogit outcome i.foreign##c.mpg
        _jk_1: _b[1.foreign]
        _jk_2: _b[mpg]
        _jk_3: _b[1.foreign#c.mpg]
        _jk_4: _b[_cons]
          n(): e(N)

                             (Replications based on 5 clusters in makecluster)
------------------------------------------------------------------------------
             |              Jknife *
             | Coefficient  std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
       _jk_1 |  -1.424797   3.277778    -0.43   0.686    -10.52537    7.675774
       _jk_2 |  -.3647182   .2051076    -1.78   0.150    -.9341882    .2047518
       _jk_3 |   .1703893   .2085935     0.82   0.460     -.408759    .7495376
       _jk_4 |   5.735816   3.217074     1.78   0.149    -3.196212    14.66784
------------------------------------------------------------------------------

. margins i.foreign,
variable _jk_1 not found
r(111);

Doesn't have to be firthlogit to produce the error but clearly there is something that I am missing in the jackknife part of the command - is there any other way to get at margins after jackknifing the model?

Thanks in advance

[Jean-Michel Galarneau]

I've tested a different way of writing the command and have come up with the solution if anybody else is interested in this.

 jackknife, dots cluster(makecluster) mse:firthlogit outcome i.foreign##c.mpg,

then you can follow up with margins without any issues.

[Federico Tedeschi]

Thank you.
Unfortunately, I found one issue. By adding "eform":

 jackknife, dots cluster(makecluster) mse eform: firthlogit outcome i.foreign##c.mpg,

the model runs, but then, at the end of the replications, I receive the error message:

option eform not allowed

I also tried:

 jackknife, dots cluster(makecluster) mse: firthlogit outcome i.foreign##c.mpg, or

by then getting the analogous error message

 option or not allowed

Although the error code is the same "r(198)", in the latter case there's a difference: the model doesn't run: the error message appears immediately.

[Federico Tedeschi]

Nevertheless, by following the advise here:https://www.statalist.org/forums/for...03#post1478403 , that is, by adding

expression(invlogit(predict(xb)))

to "margins", I was able to get what I wanted, that is: a graph representing probabilities.