Replicate margins results in a coefplot graph

[Federico Requena]

Dear all,

I am a new user of Stata, my back up is R, but now I need make some coefplot (SSC) graph. I wana replicate margins value after a logit regression. I did read that i can use post in a margins command but for others consideration I wanna use transform in a coefplot (SSC):

sysuse nlsw88, clear

logit union grade i.south
margins i.south


Predictive margins                                       Number of obs = 1,876
Model VCE: OIM

Expression: Pr(union), predict()

------------------------------------------------------------------------------
             |            Delta-method
             |     Margin   std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
       south |
  Not south  |    .297394   .0138465    21.48   0.000     .2702553    .3245327
      South  |   .1732535   .0134751    12.86   0.000     .1468428    .1996643
------------------------------------------------------------------------------

I did read that I should use :

margins i.south, post

But I would like use some like:

coefplot (., transform(* = (exp(@)/( 1+exp(@))))), drop(_cons) mlabel baselevels

coefplot (., transform(* = invlogit(@))), drop(_cons) mlabel citype(logit) baselevels

Values not coincide with margins:

Thanks

[Andrew Musau]

In general, there is no way to recover predictive margins using only the coefficient vector without the data (unless the model contains no other covariates, or you are willing to evaluate at arbitrary fixed values such as means). This is precisely why the margins command exists: it uses the estimation sample to generate predictions for each observation and then averages them appropriately (and computes standard errors via the delta method). For details, see the PDF manual entry of margins.

[Federico Requena]

Thanks Andrew Musau, My idea is not just a whim, here I can explain why i post it:

My idea is use (, if(@ll>0 | @ul<0)) to get only probabilities significant.

But if I use margins I can not select coefficients significants, because in margins table all coefficients are significants (except age==46), but in logit table only age==36 and 40 are significants.

This is the reason to my idea to transforms coefficient to get some coefficients.

------------------------------------------------------------------------------
       union | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
         age |
         34  |          0  (base)
         35  |  -.6802438   .3717162    -1.83   0.067    -1.408794    .0483066
         36  |  -.8699093   .3781728    -2.30   0.021    -1.611114   -.1287042
         37  |  -.2789024   .3680752    -0.76   0.449    -1.000317    .4425117
         38  |  -.4297407   .3715141    -1.16   0.247    -1.157895    .2984135
         39  |  -.2578905   .3670603    -0.70   0.482    -.9773154    .4615344
         40  |  -.7810485   .3801542    -2.05   0.040    -1.526137     -.03596
         41  |  -.5280255   .3713805    -1.42   0.155    -1.255918    .1998669
         42  |  -.4597705    .386396    -1.19   0.234    -1.217093    .2975518
         43  |  -.3713595   .3808608    -0.98   0.330    -1.117833    .3751139
         44  |  -.5572821   .3867026    -1.44   0.150    -1.315205    .2006411
         45  |   -.501552   .4453043    -1.13   0.260    -1.374332    .3712283
         46  |   .6190392   1.452546     0.43   0.670    -2.227899    3.465977
             |
       _cons |  -.6190392   .3314968    -1.87   0.062    -1.268761    .0306826
------------------------------------------------------------------------------

sysuse nlsw88, clear

logit union i.age

margins i.age, post


Adjusted predictions                                     Number of obs = 1,878
Model VCE: OIM

Expression: Pr(union), predict()

------------------------------------------------------------------------------
             |            Delta-method
             |     Margin   std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
         age |
         34  |        .35   .0754155     4.64   0.000     .2021883    .4978117
         35  |   .2142857   .0283152     7.57   0.000      .158789    .2697824
         36  |   .1840796   .0273356     6.73   0.000     .1305028    .2376564
         37  |   .2894737   .0329017     8.80   0.000     .2249876    .3539598
         38  |   .2594595   .0322273     8.05   0.000     .1962952    .3226237
         39  |   .2938144   .0327036     8.98   0.000     .2297166    .3579122
         40  |   .1978022   .0295271     6.70   0.000     .1399302    .2556742
         41  |   .2410256   .0306287     7.87   0.000     .1809946    .3010567
         42  |   .2537313   .0375909     6.75   0.000     .1800546    .3274081
         43  |   .2708333   .0370325     7.31   0.000     .1982509    .3434157
         44  |   .2357143   .0358721     6.57   0.000     .1654063    .3060223
         45  |   .2459016   .0551353     4.46   0.000     .1378384    .3539649
         46  |         .5   .3535534     1.41   0.157    -.1929519    1.192952
------------------------------------------------------------------------------


coefplot (, if(@ll>0 | @ul<0)), drop(_cons) mlabel citype(logit) baselevels

[Andrew Musau]

The margins table is not directly testing whether each coefficient is significant. It's testing whether the predicted probability is significantly different from 0. If your goal is to highlight significant logit coefficients, you need to work with the logit output, not margins. If you want to highlight significant predicted probabilities, margins is fine - but then it is not testing coefficient significance, it's testing predicted probabilities. So I cannot follow exactly what you are after. From a coding perspective, if I want to focus on interpreting marginal effects of statistically significant coefficients (at the 5% level), I'd filter them out as follows:

sysuse nlsw88, clear
logit union i.age

local list
levelsof age if e(sample), local(ages)
foreach a of local ages{
    if abs(`=_b[`a'.age]/_se[`a'.age]')>=1.96 & `=_b[`a'.age]/_se[`a'.age]'<. {
        local list `list' `a'.age
    }
}
margins `list'

Res.:

. logit union i.age

Iteration 0:  Log likelihood = -1046.6242  
Iteration 1:  Log likelihood = -1038.8365  
Iteration 2:  Log likelihood =  -1038.771  
Iteration 3:  Log likelihood =  -1038.771  

Logistic regression                                     Number of obs =  1,878
                                                        LR chi2(12)   =  15.71
                                                        Prob > chi2   = 0.2051
Log likelihood = -1038.771                              Pseudo R2     = 0.0075

------------------------------------------------------------------------------
       union | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
         age |
         35  |  -.6802438   .3717162    -1.83   0.067    -1.408794    .0483066
         36  |  -.8699093   .3781728    -2.30   0.021    -1.611114   -.1287042
         37  |  -.2789024   .3680752    -0.76   0.449    -1.000317    .4425117
         38  |  -.4297407   .3715141    -1.16   0.247    -1.157895    .2984135
         39  |  -.2578905   .3670603    -0.70   0.482    -.9773154    .4615344
         40  |  -.7810485   .3801542    -2.05   0.040    -1.526137     -.03596
         41  |  -.5280255   .3713805    -1.42   0.155    -1.255918    .1998669
         42  |  -.4597705    .386396    -1.19   0.234    -1.217093    .2975518
         43  |  -.3713595   .3808608    -0.98   0.330    -1.117833    .3751139
         44  |  -.5572821   .3867026    -1.44   0.150    -1.315205    .2006411
         45  |   -.501552   .4453043    -1.13   0.260    -1.374332    .3712283
         46  |   .6190392   1.452546     0.43   0.670    -2.227899    3.465977
             |
       _cons |  -.6190392   .3314968    -1.87   0.062    -1.268761    .0306826
------------------------------------------------------------------------------

.
.
.
. local list

.
. levelsof age if e(sample), local(ages)
34 35 36 37 38 39 40 41 42 43 44 45 46

.
. foreach a of local ages{
  2.
.     if abs(`=_b[`a'.age]/_se[`a'.age]')>=1.96 & `=_b[`a'.age]/_se[`a'.age]'<. {
  3.
.         local list `list' `a'.age
  4.
.     }
  5.
. }

.
. margins `list'

Adjusted predictions                                     Number of obs = 1,878
Model VCE: OIM

Expression: Pr(union), predict()

------------------------------------------------------------------------------
             |            Delta-method
             |     Margin   std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
         age |
         36  |   .1840796   .0273356     6.73   0.000     .1305028    .2376564
         40  |   .1978022   .0295271     6.70   0.000     .1399302    .2556742
------------------------------------------------------------------------------

[Federico Requena]

Good! Thanks Andrew Musau !