Calculating Marginal Effects

[Hylke Dijkstra]

Dear all,

I have had some trouble calculating one of the marginal effects of my regression. I am running the following regression:

xtreg manemp lnpop lnpop2 lngdppc lngdppc2 _2000s _2010s 1._2000s#1.la 1._2010s#1.la, fe

My main variables of interest are _2000s and _2010s, which are two decadal dummies and the interaction term between the indicator variable la, which is an indicator variable for a specific region, Latin America, and the two decadal dummies. I do this in order to estimate how much of the period dummies vary by region (in this case Latin America). My results are as follows:

esttab, drop(_cons lnpop lnpop2 lngdppc lngdppc2)

----------------------------
                      (1)  
                   manemp  
----------------------------
_2000s           -0.00286  
                  (-1.51)  

_2010s            0.00128  
                   (0.42)  

1._2000s#1~a     -0.00849**
                  (-2.68)  

1._2010s#1~a      -0.0132***
                  (-3.97)  
----------------------------
N                    1479  
----------------------------
t statistics in parentheses
* p<0.05, ** p<0.01, *** p<0.001

I am quite happy with these results, but now I am trying to calculate the marginal effect of the specific region, Latin America, compared to the full sample. This is in fact pretty simple, as I can just add the coefficients for_2000s (-0.00286) and for _2000s X la (-0.00849) which should be equal to -0,01176. I want to compute the marginal effects for both decadal dummies, specifically when my indicator variable, in this case, la, is equal to 1. I want to compute the marginal effects using:

margins dydx(varlist)

I have tried to get the results I wanted using the different options in the stata manual (https://www.stata.com/manuals13/rmargins.pdf) with trial and error, but I have not been able to get the results I want just yet. Does anyone have some advice perhaps?

A second question, perhaps unrelated, I am dealing with panel data on 51 countries over 29 years. I want to run the specified regressions and obtain the marginal effects for 7 different regions (Latin America being one) and for 51 different countries. So this means I want to run 58 regressions. Is there a somewhat efficient way to do this?

Thank you in advance.

Best,

Hylke Dijkstra

[Hylke Dijkstra]

I have been able to obtain some of the results that I want by using:

xtreg manemp lnpop lnpop2 lngdppc lngdppc2 _2000s _2010s 1._2000s#1.la 1._2010s#1.la, fe
mfx, dydx

which gives me:

Marginal effects after xtreg
      y  = Linear prediction (predict)
         =  .11394801
------------------------------------------------------------------------------
variable |      dy/dx    Std. err.     z    P>|z|  [    95% C.I.   ]      X
---------+--------------------------------------------------------------------
   lnpop |  -.2077072      .04591   -4.52   0.000  -.297689 -.117725   17.0983
  lnpop2 |   .0063001      .00138    4.55   0.000   .003588  .009012   294.518
 lngdppc |   .3723604      .01862   20.00   0.000   .335861   .40886   8.60209
lngdppc2 |   -.022366      .00108  -20.72   0.000  -.024481  -.02025   75.1394
  _2000s*|  -.0113541      .00311   -3.66   0.000  -.017441 -.005267   .344828
  _2010s*|  -.0119193      .00381   -3.12   0.002  -.019396 -.004443   .310345
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

I am quite happy with this result, except for the computation of the standard error. So, I would like to obtain the same results using:

margins

I have attempted to do run the same regression using a different interaction for a different region, in this case, Advanced Asia. The results are as follows:

xtreg manemp lnpop lnpop2 lngdppc lngdppc2 _2000s _2010s 1._2000s#1.aa 1._2010s#1.aa, fe
mfx, dydx

esttab, drop(_cons lnpop lnpop2 lngdppc lngdppc2)

----------------------------
                      (1)   
                   manemp   
----------------------------
_2000s           -0.00114   
                  (-0.63)   

_2010s            0.00226   
                   (0.78)   

1._2000s#1~a      -0.0295***
                  (-7.70)   

1._2010s#1~a      -0.0374***
                  (-8.58)   
----------------------------
N                    1479   
----------------------------
t statistics in parentheses
* p<0.05, ** p<0.01, *** p<0.001

When I run the same command as before:

mfx, dydx

It gives me the following results:

Marginal effects after xtreg
      y  = Linear prediction (predict)
         =  .11527156
------------------------------------------------------------------------------
variable |      dy/dx    Std. err.     z    P>|z|  [    95% C.I.   ]      X
---------+--------------------------------------------------------------------
   lnpop |  -.1592691      .04503   -3.54   0.000  -.247528  -.07101   17.0983
  lnpop2 |   .0047516      .00136    3.50   0.000   .002088  .007415   294.518
 lngdppc |   .3113827       .0198   15.72   0.000   .272567  .350199   8.60209
lngdppc2 |  -.0185762      .00116  -16.04   0.000  -.020846 -.016306   75.1394
  _2000s*|  -.0011438       .0018   -0.63   0.526  -.004676  .002389   .344828
  _2010s*|    .002264       .0029    0.78   0.434  -.003412   .00794   .310345
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

This is confusing to me because it seems to report the coefficients for just the decadal dummies..

Does anyone have some thoughts on this (preferably using the margins command)?

[Leonardo Guizzetti]

Please note our policy on cross-posting in the FAQ.

This was cross posted to Reddit at https://www.reddit.com/r/stata/comme...ginal_effects/