Marginsplot: y scale (plot) different of scale of original variable

[Rocio Aguilar]

Dear Stata users,
I have a problem with the range of y scale's marginsplot. I have a (0 100) dependent variable, but the y-scale of the marginsplot has a different range:

. regress  INCOME    c.MONTER1##c.MONTER2 , r


. margins, at( MONTER1= (200 1202222) MONTER2=(0 0.5)) vsquish

 . marginsplot, xsize(13) ysize(9) scale(0.8)  noci noseparator recast(line) plotopts(lwidth(medthick) lcolor(black) cmissing(n)  connect(direct))  plot1opts( lpattern(dash_dot_dot)) plotregion(fcolor(white)) graphregion(fcolor(white)) xtitle(" ")   legend(order(1 "Low MONTER2" 2 "High MONTER2") size(medsmall) ) clegend(region(fcolor(white) lcolor(white))) title("") xlabel(200 "Low MONTER1"   1150000 "High MONTER2") yline(0) yline(100) 

 . marginsplot, xsize(13) ysize(9) scale(0.8)  noci noseparator recast(line) plotopts(lwidth(medthick) lcolor(black) cmissing(n)  connect(direct))  plot1opts( lpattern(dash_dot_dot)) plotregion(fcolor(white)) graphregion(fcolor(white)) xtitle(" ")   legend(order(1 "Low MONTER2" 2 "High MONTER2") size(medsmall) ) clegend(region(fcolor(white) lcolor(white))) title("") xlabel(200 "Low MONTER1"   1150000 "High MONTER2") yline(0) yline(100) yscale(r(0 100)) ylabel(0(10)100) 

Sorry, but I don't understant why the range of de y scale of the marginsplot is not (0 100) as the range of the dependent variable.
I would appreciate yours advices.
Thanks in advance,
Rocio

[Maarten Buis]

What margins shows is the predicted outcome, and apperently your model predicts outcomes way outside the range of your dependent variable. Your model is a linear regression (regress), and that model does not impose a constraint on the values of the outcome; it just plots a straight line through your data. My guess is that most of your data is concentrated around the low end of monter1, so your regression tries to fit a straight line that fits best in this lower part, and it fits pretty awfully at the higher end of monter1. So the problem is not margins or marginsplot. The problem is that your model is wrong. Since it just the lower bound that seems to be the problem you could use poisson with robust standard errors https://blog.stata.com/2011/08/22/us...tell-a-friend/ . Alternative, you can use a model that respects both the upper and lower bound by dividing your dependent variable by 100 and use fracreg.

[George Ford]

Maarten has good ideas. When I need to predict interval data, I often use:

glm y x , family(binomial) link(logit)

[Maarten Buis]

For the solution proposed by George Ford you typically need to add the vice(robust) option. In that case it is the same as fracreg.

[Rocio Aguilar]

Thank you very much for your answers. However, in the case of using fracreg, I think it is necessary logit or probit after fracreg. I get the following results:

. quiet  fracreg  INCOME10 c.MONTER1##c.MONTER2, vce(robust)
. quiet  margins, at( MONTER1= (200 1202222) MONTER2=(0 0.5)) vsquish
. marginsplot, xsize(13) ysize(9) scale(0.8)  noci noseparator recast(line) plotopts(lwidth(medthick) lcolor(black) cmissing(n)  connect(direct))  plot1opts( lpattern(dash_dot_dot)) plotregion(fcolor(white)) graphregion(fcolor(white)) xtitle(" ")   legend(order(1 "Low MONTER2" 2 "High MONTER2") size(medsmall) ) clegend(region(fcolor(white) lcolor(white))) title("") xlabel(200 "Low MONTER1"   1150000 "High MONTER2") yline(0) yline(100) yscale(r(0 100)) ylabel(0(10)100)

 . marginsplot, xsize(13) ysize(9) scale(0.8)  noci noseparator recast(line) plotopts(lwidth(medthick) lcolor(black) cmissing(n)  connect(direct))  plot1opts( lpattern(dash_dot_dot)) plotregion(fcolor(white)) graphregion(fcolor(white)) xtitle(" ")   legend(order(1 "Low MONTER2" 2 "High MONTER2") size(medsmall) ) clegend(region(fcolor(white) lcolor(white))) title("") xlabel(200 "Low MONTER1"   1150000 "High MONTER2") yline(0) yline(100)

Thank you very much for your advice, but I don't think it will help me in my case.

Please help if something is wrong.


In the case of glm, the results are also strange.

. quiet glm INCOME10 c.MONTER110##c.MONTER210, family(binomial) link(logit) vce(robust)
. quiet margins, at( MONTER1= (2 12000) MONTER2=(0 0.003)) vsquish
. marginsplot, xsize(13) ysize(9) scale(0.8)  noci noseparator recast(line) plotopts(lwidth(medthick) lcolor(black) cmissing(n)  connect(direct))  plot1opts( lpattern(dash_dot_dot)) plotregion(fcolor(white)) graphregion(fcolor(white)) xtitle(" ")   legend(order(1 "Low MONTER2" 2 "High MONTER2") size(medsmall) ) clegend(region(fcolor(white) lcolor(white))) title("") xlabel(200 "Low MONTER1"   1150000 "High MONTER2") yline(0) yline(100) yscale(r(0 100)) ylabel(0(10)100)

. marginsplot, xsize(13) ysize(9) scale(0.8)  noci noseparator recast(line) plotopts(lwidth(medthick) lcolor(black) cmissing(n)  connect(direct))  plot1opts( lpattern(dash_dot_dot)) plotregion(fcolor(white)) graphregion(fcolor(white)) xtitle(" ")   legend(order(1 "Low MONTER2" 2 "High MONTER2") size(medsmall) ) clegend(region(fcolor(white) lcolor(white))) title("") xlabel(200 "Low MONTER1"   1150000 "High MONTER2")

Thanks you very much for you advice.
Best

[Maarten Buis]

Why do you think it won't help?

[Rocio Aguilar]

Dear Maarten, I think that it is not the same graph.
Thanks for you interest.
Rocio

[Maarten Buis]

Ofcourse, when done correctly you will get a better graph.