Margins, Non-parametric Regression

[Lindsey Buck]

Hi all,

My code looks like this:
"npregress kernel test01 test00
npgraph
margins
marginsplot"

Basically, I want to have a graph of given an x in test00, what is the predicted test01. My professor told me I should do this using margins. However, when I do margins and marginsplot, I only get one point and I think I need a point for many values. Specifically, my professor told me to break test00 in to ten different bins for some scores (1-2, 2-3, 3-4.....) and then predict based on these. Can someone help? I have tried playing around with different things on the margins command but I am getting nowhere.
I want a graph that shows if you have a test00 in between 1-2, your predicted test01 is......and shows this for all the bins I create.
Thank you!
Best,
Lindsey

[FernandoRios]

Hi Lindsey,
To get the plots that you want you need a code like the one below

clear all
webuse dui, clear
npregress kernel citations fines,
margins, at(fines=(8(.5)12)) vce(bootstrap)
marginsplot, name(g1) title(npreg)

Alternatively, you can do something like this too:

gen qfines=ceil(fines/.5)*5
reg citations i.qfines
margins,over(qfines)
marginsplot, name(g2) title(Dummies)

graph combine g1 g2

HTH

[Lindsey Buck]

Thank you so much!!

[Lindsey Buck]

I am now attempting to bootstrap it so I can get a an actual confidence interval:


"/* beginning of Bootstrap loop */
global numberBSreps=100
forvalues i = 1/$numberBSreps {


preserve
bsample
drop if treat==0
drop if test00<-1.3
drop if test00>1.6
npregress kernel test00 test01
margins, at (test00=(-1.3(.3)1.6))
marginsplot, name(Treatment) title(Treatment Group)
restore

/* now we're done, go do another BS step */"

Do you know what I am doing wrong? I want a graph that shows the confidence interval as well from bootstrapping!

[Jeff Wooldridge]

Lindsey: It looks to me like you don't want to do a smoothed kernel regression but rather a bin regression. Is that correct? This may seem clunky, but the way I would do that, and get standard errors for the change in the mean of test01 for different bin levels, is the following. I assume there are K total bins.

reg test01 bin1 bin2 bin3 ... binK, nocons vce(robust)
lincom b2 - b1
lincom b3 - b2
...
lincom bK - bK_1

These will give the "marginal" effects along with standard errors and confident intervals.