Kernel Regressions: NPREGRESS

[Lindsey Buck]

I have two non-parametric plots. I am told to:

"Try to produce something similar to Figure 3, Panel A of the published article. That is, take the difference between these two non-parametric plots, and plot them out. I think you will need to calculate the difference based on a grid of X-values of the 2000 test score (the x variable) and calculate the conditional expectation of the Y-value. Then graph these in a connected line). Note, you probably won’t be able to exactly replicate their results because you are using a nonparametric kernel regression whereas they use a fan locally weighted regression. Try to get as close as you can. To get a confidence interval on this graph, bootstrap using the following procedure: one: draw a bootstrap sample from the pooled (treatment and control) distribution two: estimate the regression for the treatment group three: estimate the regression for the control group four: Estimate the difference for each of the grid points of x-values, and save this. Five: Repeat until you estimate the standard error at each point on the graph: then graph the confidence intervals."

I don't have any idea how ot being at this. Can someone help? I am trying to imitate the following graph:

[FernandoRios]

Hi Lindsay,
I think you can do this relatively easily, but you to provide more information about what your model looks like. In specific, I'm confused about what you call "Fan Regression".
A reference would help understand what you want to do.
From the information you provide, it seems that your model is somthing like this:

y=b0(z)+b1(x)*sex+b1(x)*treatment+b2(x)*gender+b3(x)*treatment*gender+error

where x is your variable in the horizontal axis.
If that is the case, you only need the Bandwidth to estimate the local regression:
something like this:

* Instead of zero, you can use any of the points along X
gen dx=x-0
regress y  (c.treatment c.gender c.treatment#c.gender)##c.dx [pw=normalden(x,0,bw)]
* where bw is your bandwidth

You can also create a wrapper program to do the Bootstrapping.

program mybsreg, eclass
regress y  (c.treatment c.gender c.treatment#c.gender)##c.dx [pw=normalden(x,0,bw)]
end
bootstrap: mybsreg

HTH
Fernando

[Lindsey Buck]

Thank you!!