Suggestions for the visualization of a triple interaction effect

[Simon Schillebeeckx]

I am looking to create some figure(s) that visualize a triple interaction effect I am observing in the data.
I reprint the basic results below

xtreg fwd log_dom t_s_d c.log_dom#c.t_s_d p_pria_timew p_pria_usew c.log_dom#c.p_pria_usew c.p_pria_usew#c.t_s_degree_cent c.log_dom#c.p_pria_usew#c.t_s_d log_other t_k_s team_t_1st col_difpairs team_m_patcount_5y p_classes dif_cpc p_cpc_1 p_pria p_claims p_inv f_inv_prod_5y f_search f_acap f_dar f_emp i.p_appy i.p_gry if compustat == 1 , fe

Fixed-effects (within) regression               Number of obs     =     40,282
Group variable: fid                             Number of groups  =        109

R-sq:                                           Obs per group:
     within  = 0.0831                                         min =          1
     between = 0.0728                                         avg =      369.6
     overall = 0.0679                                         max =      8,495

                                                F(36,40137)       =     101.00
corr(u_i, Xb)  = -0.2782                        Prob > F          =     0.0000

-------------------------------------------------------------------------------------------------------------
                                        fwd |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
--------------------------------------------+----------------------------------------------------------------
                                  log_dom_t |   .5544467   .0850644     6.52   0.000     .3877186    .7211748
                            t_s_degree_cent |    .052213     .02431     2.15   0.032     .0045649    .0998612
                                            |
              c.log_dom_t#c.t_s_degree_cent |  -.0276355   .0060148    -4.59   0.000    -.0394246   -.0158464
                                            |
                               p_pria_timew |  -1.091934   .0910784   -11.99   0.000     -1.27045   -.9134187
                                p_pria_usew |   .3448452   .1871209     1.84   0.065    -.0219161    .7116065
                                            |
                  c.log_dom_t#c.p_pria_usew |   .1115617    .063443     1.76   0.079     -.012788    .2359115
                                            |
            c.p_pria_usew#c.t_s_degree_cent |   .1329657   .0234215     5.68   0.000     .0870591    .1788722
                                            |
c.log_dom_t#c.p_pria_usew#c.t_s_degree_cent |  -.0285318   .0052981    -5.39   0.000    -.0389161   -.0181474

Normally, for a simple interaction effect I would use something like

marginscontplot log_dom_t t_s_d, at1(0(.25)6) at2(0 5.7 14.3 22.9)

So I am wondering
1) Is it possible to run marginscontplot for specific values of my variable p_pria_usew (e.g. by using by, for, or over)? This should create multiple graphs that exhibit the effects
2) Is there a way to create a 3D plot (with fwd on the Y axis, log_dom_t on the X-axis, and t_s_degree_cent on the Z-axis for three different values of p_pria_usew (e.g. mean +/- 1 standard deviation
3) If either 1 or 2 is possible, would the solution be different if I would run a negative binomial regression instead of a simple OLS?

Thanks heaps in advance!

Simon

[Maarten Buis]

Here is an example from a workshop I gave at last German Stata Users's meeting: http://maartenbuis.nl/workshops/grap...l#slide57.smcl

All materials from the workshop can be accessed here: http://maartenbuis.nl/workshops/

[Simon Schillebeeckx]

HI Maarten,

Thanks a lot (bedankt!). This is quite a helpful resource. I managed to replicate your first approach with my data to create a plot (see below).
This is the code I used (for three continuous variables)

xtnbreg fwd10 c.log_dom##c.t_s_d##c.p_pria_usew p_pria_timew f_emp f_acap f_dar f_search f_inv_prod_5y  p_inv log_other t_k_s team_t_1st col_difpairs team_m_patcount_5y p_classes dif_cpc p_cpc_1 p_pria p_claims priordum ts i.p_appy i.p_gry if compustat == 1 , fe
est sto A5
        margins, dydx(log_dom_t) at(t_s_d = (0 5.66 14.4 23) p_pria_usew = (-1.5 -.9 0 .9 1.8 3.6))
        _marg_save, saving(temp, replace)
        use temp, clear
        twoway contour _margin _at2 _at3, ccut(-1(.5)6) ztitle(effect of team domain experience) name(margins3d_try1, replace)
        graph export "C:\User\...\margins3d_try1.png", as(png) replace

Could you help me make sense of the interpretation please? Specifically, it is unclear to me how this plot relates to the response variable fwd.




I was actually hoping to create an image that was more like the one you can see on this link
https://www.researchgate.net/figure/...fig2_321711274

Basically, a combination of three graphs in which a 3-dimensional surface connects log_dom_t (X axis) and t_s_d (Z axis) to the response fwd (Y axis).
One graph would show this for high value of moderator p_pria_usew, and the others for an average and a low value of that moderator.
Is this possible with Stata or should I turn to R or some other program?

[Dimitriy V. Masterov]

Adrian Mander wrote a command called surface that could produce the 3D plot you seek.

[Simon Schillebeeckx]

Thanks Dimitriy, I will look at that today.

If you could help me locate the exact file that would be helpful. I did not find it by using

findit surface

but found this link. Is that what you had in mind?
http://fmwww.bc.edu/RePEc/bocode/s/surface.ado

[Dimitriy V. Masterov]

Yes, that is one. Try

ssc install surface