margins, dydx(x) at only certain values of x

[Lydia Palumbo]

Dear Statalisters,

I am running a logit

svy, subpop(if marital==1&age>17&age<36&(evunion==2)&jbstat!=7&ivfioall==1): logit event1 ib1.skills##i.wave5  i.agesm i.livingpar i.ukborn i.sex i.educ2 i.biosonsn2 i.london i.religion i.nssecpar , or level(95)

and, then, computing the marginal effects for skills (which is categorical and has 4 levels) this way

margins , level(95)  post atmeans    dydx(ib1.skills) force at(wave5=(1 2 3)) subpop(if marital==1&age>17&age<36&(evunion==2)&jbstat!=7&ivfioall==1)

I am interesting in obtaining the marginal effects only for specific values of i.skills but I am not sure how to do it. If I specify at(skills=(1(1)3), I still get the contrast 4 vs 1.

The reason I am asking it is that I would like to plot the results with marginsplot later and, if I have the contrast 4 vs.1, marginplot is going to plot that one too.

Could you please help me?

Thank you and best,
Lydia

[Jeff Pitblado (StataCorp)]

Marginal effects of a factor variable are the differences of the predictive margins at each level of the factor compared to the base level.
This is also known as a contrast with the reference (base) category.
margins supports contrast operators in the marginslist specification, and you can select which levels to contrast there to.

Here is an example using data from the logit help file.

webuse nhanes2d
* survey data, already setup
svyset
* fit our model
svy: logit highbp height weight age female i.region
* compute the marginal effect of our factor variable
margins, atmeans dydx(region)
* compute the same using contrast syntax
margins r.region, atmeans contrast(nowald effects)
* get the levels of interest
describe region
label list region
* choose contrasts of levels 4 and 3 versus level 1
margins r(4 3 1).region, atmeans contrast(nowald effects)

Here is the log from Stata 17, but should be essentially the same in earlier versions of Stata too:

. webuse nhanes2d

. * survey data, already setup
. svyset

Sampling weights: finalwgt
             VCE: linearized
     Single unit: missing
        Strata 1: strata
 Sampling unit 1: psu
           FPC 1: <zero>

. * fit our model
. svy: logit highbp height weight age female i.region
(running logit on estimation sample)

Survey: Logistic regression

Number of strata = 31                            Number of obs   =      10,351
Number of PSUs   = 62                            Population size = 117,157,513
                                                 Design df       =          31
                                                 F(7, 25)        =      237.00
                                                 Prob > F        =      0.0000

------------------------------------------------------------------------------
             |             Linearized
      highbp | Coefficient  std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
      height |  -.0346967   .0052669    -6.59   0.000    -.0454387   -.0239548
      weight |   .0518302   .0024187    21.43   0.000     .0468972    .0567632
         age |   .0488588   .0019091    25.59   0.000     .0449651    .0527525
      female |  -.4596187   .0571295    -8.05   0.000    -.5761351   -.3431024
             |
      region |
         MW  |  -.1695395   .2367601    -0.72   0.479    -.6524149    .3133359
          S  |  -.1423818    .233099    -0.61   0.546    -.6177903    .3330268
          W  |  -.0362159   .2140672    -0.17   0.867     -.472809    .4003771
             |
       _cons |  -.2934606   .8631403    -0.34   0.736    -2.053847    1.466926
------------------------------------------------------------------------------

. * compute the marginal effect of our factor variable
. margins, atmeans dydx(region)

Conditional marginal effects

Number of strata = 31                            Number of obs   =      10,351
Number of PSUs   = 62                            Population size = 117,157,513
Model VCE: Linearized                            Design df       =          31

Expression: Pr(highbp), predict()
dy/dx wrt:  2.region 3.region 4.region
At: height   = 168.4599 (mean)
    weight   = 71.90064 (mean)
    age      = 42.25264 (mean)
    female   = .5206498 (mean)
    1.region =  .206883 (mean)
    2.region = .2489281 (mean)
    3.region = .2652914 (mean)
    4.region = .2788975 (mean)

------------------------------------------------------------------------------
             |            Delta-method
             |      dy/dx   std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
      region |
         MW  |  -.0379287   .0530809    -0.71   0.480    -.1461879    .0703305
          S  |  -.0319958   .0525213    -0.61   0.547    -.1391136    .0751221
          W  |  -.0082737   .0489926    -0.17   0.867    -.1081948    .0916473
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.

. * compute the same using contrast syntax
. margins r.region, atmeans contrast(nowald effects)

Contrasts of adjusted predictions

Number of strata = 31                            Number of obs   =      10,351
Number of PSUs   = 62                            Population size = 117,157,513
Model VCE: Linearized                            Design df       =          31

Expression: Pr(highbp), predict()
At: height   = 168.4599 (mean)
    weight   = 71.90064 (mean)
    age      = 42.25264 (mean)
    female   = .5206498 (mean)
    1.region =  .206883 (mean)
    2.region = .2489281 (mean)
    3.region = .2652914 (mean)
    4.region = .2788975 (mean)

------------------------------------------------------------------------------
             |            Delta-method
             |   Contrast   std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
      region |
 (MW vs NE)  |  -.0379287   .0530809    -0.71   0.480    -.1461879    .0703305
  (S vs NE)  |  -.0319958   .0525213    -0.61   0.547    -.1391136    .0751221
  (W vs NE)  |  -.0082737   .0489926    -0.17   0.867    -.1081948    .0916473
------------------------------------------------------------------------------

. * get the levels of interest
. describe region

Variable      Storage   Display    Value
    name         type    format    label      Variable label
-------------------------------------------------------------------------------
region          byte    %9.0g      region     Region

. label list region
region:
           1 NE
           2 MW
           3 S
           4 W

. * choose contrasts of levels 4 and 3 versus level 1
. margins r(4 3 1).region, atmeans contrast(nowald effects)

Contrasts of adjusted predictions

Number of strata = 31                            Number of obs   =      10,351
Number of PSUs   = 62                            Population size = 117,157,513
Model VCE: Linearized                            Design df       =          31

Expression: Pr(highbp), predict()
At: height   = 168.4599 (mean)
    weight   = 71.90064 (mean)
    age      = 42.25264 (mean)
    female   = .5206498 (mean)
    1.region =  .206883 (mean)
    2.region = .2489281 (mean)
    3.region = .2652914 (mean)
    4.region = .2788975 (mean)

------------------------------------------------------------------------------
             |            Delta-method
             |   Contrast   std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
      region |
  (S vs NE)  |  -.0319958   .0525213    -0.61   0.547    -.1391136    .0751221
  (W vs NE)  |  -.0082737   .0489926    -0.17   0.867    -.1081948    .0916473
------------------------------------------------------------------------------

[Lydia Palumbo]

That works perfectly.
Thank you very much.