Output table for marginal and impact effects

[Anne Wanyonyi]

Hello,

I have run a probit model followed by the

mfx

command to get the marginal and impact effects.

I then used this commands

eststo

esttab C:\Users\Wanyonyi\Desktop\DHS\output.csv

However, the output table has estimates from the main probit model and not the the marginal and impact effect estimates.

Kindly, what can I do to get the output table that has the marginal and impact effect estimates?

[Clyde Schechter]

-mfx- has been superseded by the -margins- command. Use -margins- with the -saving()- option to save that output. Then -use- that saved file, clean it up to your taste, and then you can make a .csv file out of it with -export delimited-.

Alternatively, you can use -margins- with the -post- option, and then I think -esttab- will handle it after that. I'm not sure of this as I don't use -esttab- myself.

[Anne Wanyonyi]

Hi Clyde,

What's the command I can use for the -margins -with the -post -option?

What do you normally use instead of esttab?

[Clyde Schechter]

The command itself is called -margins-. There are extensive examples of how to use it in the [R] manual. In addition, there is a really clearly written handout by Richard Williams at https://www3.nd.edu/~rwilliam/stats/Margins01.pdf.

To give you more specific advice about applying it to your situation, you would have to show us your regression command and results.

As for what I use instead of -esttab-, the answer is: nothing in particular. When I am in the midst of a project and want to share regression outputs with colleagues on the project I just send them my log files (as .smcl if they run Stata, as .pdf if they don't.) When it comes time to make a publication-ready table, I usually have narrowed down the results that will be presented to a small number of manageable tables. Often the regression results just as Stata shows them, or with light editing, do the trick (especially now that we have -set cformat- and lots of options that control the display of the regression table in the regression commands themselves. If something more complicated is needed, I tend to just write some custom-code for the circumstances.

I have nothing particularly against -esttab-. It's just that it has a zillion options, and my need for it is sufficiently infrequent that I can never remember what they are and how they work anyway. For me, it's less work to do light hand-edits to the Stata output directly or write a little custom code than it is to figure out how to configure an -esttab- command.

[Anne Wanyonyi]

My command is:

probit hv121 hv206 hv105 agesq hv104 hv220 headagesq hv219  hdensity mwall wood sourceofwater sourceofwatersq hv009  hv025, cluster(hhid)

The output is:

Iteration 0:   log pseudolikelihood = -2568.8374  
Iteration 1:   log pseudolikelihood = -2306.3589  
Iteration 2:   log pseudolikelihood = -2285.9021  
Iteration 3:   log pseudolikelihood = -2285.8463  
Iteration 4:   log pseudolikelihood = -2285.8463  

Probit regression                               Number of obs     =     10,573
                                                Wald chi2(14)     =     316.31
                                                Prob > chi2       =     0.0000
Log pseudolikelihood = -2285.8463               Pseudo R2         =     0.1102

                                  (Std. Err. adjusted for 4,860 clusters in hhid)
---------------------------------------------------------------------------------
                |               Robust
          hv121 |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
----------------+----------------------------------------------------------------
          hv206 |  -.2737755   .0924836    -2.96   0.003    -.4550401   -.0925109
          hv105 |   .4563149   .0601365     7.59   0.000     .3384496    .5741803
          agesq |  -.0201105   .0030597    -6.57   0.000    -.0261074   -.0141136
          hv104 |   .0128462   .0399903     0.32   0.748    -.0655333    .0912258
          hv220 |   .0039563   .0127487     0.31   0.756    -.0210307    .0289434
      headagesq |  -.0000529   .0001216    -0.43   0.664    -.0002913    .0001855
          hv219 |  -.2067014   .0654437    -3.16   0.002    -.3349686   -.0784341
       hdensity |   2.231987   .2648749     8.43   0.000     1.712842    2.751132
          mwall |   .1730469   .0608503     2.84   0.004     .0537825    .2923112
           wood |  -.0565939   .0986922    -0.57   0.566     -.250027    .1368392
  sourceofwater |  -.3416941   .0540198    -6.33   0.000     -.447571   -.2358172
sourceofwatersq |   .0289682   .0094729     3.06   0.002     .0104017    .0475346
          hv009 |  -.0440757   .0135297    -3.26   0.001    -.0705935    -.017558
          hv025 |  -.0377929   .0630255    -0.60   0.549    -.1613206    .0857348
          _cons |  -.9383623   .3985501    -2.35   0.019    -1.719506   -.1572185
---------------------------------------------------------------------------------

Then for marginal and impact effects:

mfx

Marginal effects after probit
      y  = Pr(hv121) (predict)
         =  .95424258
------------------------------------------------------------------------------
variable |      dy/dx    Std. Err.     z    P>|z|  [    95% C.I.   ]      X
---------+--------------------------------------------------------------------
   hv206*|  -.0307527      .01212   -2.54   0.011  -.054504 -.007001   .162679
   hv105 |   .0438372      .00594    7.38   0.000   .032193  .055481    9.7678
   agesq |   -.001932       .0003   -6.42   0.000  -.002522 -.001342   102.161
   hv104*|    .001234      .00384    0.32   0.748  -.006293  .008761   .494562
   hv220 |   .0003801      .00122    0.31   0.756  -.002019  .002779   43.1677
headag~q |  -5.08e-06      .00001   -0.44   0.663  -.000028  .000018   1974.99
   hv219*|  -.0220596      .00773   -2.85   0.004  -.037205 -.006914   .201362
hdensity |   .2144223       .0233    9.20   0.000   .168756  .260089   .317888
   mwall*|   .0155706      .00513    3.04   0.002    .00552  .025621   .270027
    wood*|  -.0052675       .0089   -0.59   0.554  -.022704  .012169    .83212
source~r |  -.0328258      .00531   -6.19   0.000  -.043228 -.022424   .515182
source~q |   .0027829      .00091    3.05   0.002   .000996   .00457   .932687
   hv009 |  -.0042343      .00132   -3.21   0.001  -.006816 -.001652   7.09827
   hv025*|  -.0036817      .00623   -0.59   0.555  -.015892  .008528   .282512
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

I want an output that has the estimates from the mfx command as it is easier and more relevant to interpret marginal and impact effects than standardized probit

[Clyde Schechter]

OK. There are two problems: you didn't use factor variable notation in your -probit-, so you can't use -margins-. And it looks to me like your model is mis-specified.

The misspecifications arise in connection with the terms agesq and headagesq. If I'm understanding the variable names correctly, these appear to be quadratic terms, but they have no linear terms included along with them. That is almost always a model misspecification: there are rare circumstances where a quadratic term with no linear is appropriate--but I'd be very surprised if that's the case here. So you need to build in the linear terms. And then the quadratics also need to be constructed using factor variable notation, so that -margins- will know that those terms are in fact the squares of the linear terms and will handle them appropriately.

So I would do this as follows:

probit hv121 i.hv206 hv105 c.age##c.age i.hv104 hv220 ///
    c.headage##c.headage i.hv219  hdensity i.mwall i.wood ///
    c.sourceofwater##c.sourceofwater hv009  i.hv025, cluster(hhid)

Then I would get the marginal effects as follows:

margins, dydx(*) saving(marginal_effects, replace)

The final step is to -use marginal_effects-, rename the variables to something nicer than what -margins- gave them, drop any that you don't need, and then -export delimited- to create a .csv.

Or, as I said earlier, it may be that if you do

margins, dydx(*) post

-esttab- will get you what you want. It's certainly worth a try.

[Anne Wanyonyi]

Thank you so much Clyde for the correction, I hadn't noticed my model was mis-specified.
After introducing the factor variable notation, I noticed that it drops the first category when the variable is a categorical dummy. Is there a way I can alter it so as to have a different category as the base category.


If you don't mind, why is factor variable notation important in probit when one wants to use margins?

[Rich Goldstein]

see "help fvvarlist##bases" for how to choose the base level using factor variable notation

margins understands factor variable notation and therefor will correctly deal with all "pieces" of a variable (several indicator variables or polynomials are 2 examples); thus, if you use factor variable notation then margins will do what you expect - if you don't margins will be dealing with each piece as a separate variable with is (almost) never what you want

[Enrique Pinzon (StataCorp)]

Anne,

Using factor variable notation, when you have a nonlinear model like probit, is important to get the marginal effects you want. Suppose the variable x1 is categorical with two values and when introducing it in your list of covariates instead of typing

i.x1

you type

x1

The average marginal effect ( dydx(x1)), in the first case, is the average over the values of the covariates for

\begin{equation}
\Phi\left(\beta_0 + 1 + \beta_2x_2 \ldots \right) - \Phi\left(\beta_0 + 0 + \beta_2x_2 \ldots \right)
\end{equation}

In the second case it is the average over the values of the covariates for:

\begin{equation}
\phi \left(\beta_0 + \beta_1x_1 + \beta_2x_2 + \ldots \right)\beta_1
\end{equation}

These two results are numerically and qualitatively different. You can find a little of extra information about this topic in
http://blog.stata.com/2016/07/12/eff...-interpreting/

[Anne Wanyonyi]

Thank you for all for the helpful comments.

I need to extend my model and include other variables.
Initially the probit model was focusing on impact of electrification on school attendance at the household level. But now I want to include variables which capture child labour. I am using occupation of the children's parents but collapsing them to the village level from the household level.

Occupation is a categorical variable. But after collapsing,I end up with the five different categories. I am wondering how I can use the factor variable notation in this case as the occupation variable from which the 5 different categories are generated 'disappears' after collapsing.

These are the categories after collapsing the household father's occupation to the village level.

 des occfather1 occfather2 occfather3 occfather4 occfather5

              storage   display    value
variable name   type    format     label      variable label
-----------------------------------------------------------------------------------------------------------------------------
occfather1      float   %8.0g                 (mean) occfather1
occfather2      float   %8.0g                 (mean) occfather2
occfather3      float   %8.0g                 (mean) occfather3
occfather4      float   %8.0g                 (mean) occfather4
occfather5      float   %8.0g                 (mean) occfather5

The variable which 'disappeared' after use of collapse that I could have used with the factor variable notation is:

tab mv717

             occupation (grouped) |      Freq.     Percent        Cum.
----------------------------------+-----------------------------------
                      not working |      1,732       22.91       22.91
professional/technical/managerial |        528        6.99       29.90
      agriculture - self employed |      2,545       33.67       63.57
           household and domestic |      1,009       13.35       76.91
                   skilled manual |        556        7.36       84.27
                 unskilled manual |      1,189       15.73      100.00
----------------------------------+-----------------------------------
                            Total |      7,559      100.00

I don't know whether I can ignore factor variable notation for only this variable and whether it will impact my results

[Attaullah Shah]

For exporting output from the margins command, you can also try asdoc (from SSC).

ssc install asdoc
webuse margex
regress y i.sex i.group
asdoc margins sex, replace