rdplot, cmogram and RDD design

[Liu Qiang]

Dear all, I am a freshman and learning Regression Discontinuity Design in recent days. Command rdplot is a newly developed code by very renowned scholars in this field. Now I have several questions concerning this:
1. The confidence intervals. Use the classic data from: https://sites.google.com/site/rdpackages/rdrobust. With the code as below, we can get the picture afterwards:
rdplot demvoteshfor2 demmv,nbins(15 35) graph_options(xlabel(-100(25)100) legend(off) xtitle("demmv") ytitle("demvoteshfor2")) ci(95) shade genvars p(2)



According to the papers of my interest like https://www.sciencedirect.com/scienc...04387815001327 and the cmogram command, I can also draw this picture using collapsed data with tw scatter and the original data with qfitci command:
tw (scatter rdplot_mean_y rdplot_mean_x) ///
(qfitci demvoteshfor2 demmv if demmv<0,fcolor(none)) ///
(qfitci demvoteshfor2 demmv if demmv>=0,fcolor(none)),xline(0) legend(off)



Apparently, the confidence intervals are different. Now I wonder why there exists the difference and which one should be favored? or any one of them is wrong? I guess it's related to the way in which Stata draws the line like the difference between lowess and connected scatter but not pretty sure.
2. I occasionally found this command cmogram developed by Christopher Robert, with whose code I can quickly get a same picture:
cmogram demvoteshfor2 demmv, cut(0) scatter lineat(0) qfitci


It seems that the second and the third method to draw the confidence intervals would be more welcome. But cmogram has a defect. It does not have a weight option like kernal or triangle which is widely applied in an RDD research. Anyone knows how to improve it?
3. It seems thata rdplot cannot be used together with a tw type graph commands. Anyone knows how to reach this goal?
Any statalister interested in this method is welcome. appreciate any comments on related work in advance.

[Steve Samuels]

The qfitci figures are the standard CIs for a quadratic least squares fit. They assume a constant standard deviation on each side of the cutoff.

In contrast, rdplot assumes neither a quadratic model nor constant standard deviations. The data are binned. A different standard deviation is estimated in each bin and the CI based on that SD is centered on the bin mean (Colonico et al., 2017, p.382, 391). I would personally prefer these CIs.

Two-way options can be used: they go inside the graph_options option.


Reference:
Calonico, S., Cattaneo, M. D., Farrell, M. H., and Titiunik, R. (2017). rdrobust: Software for regression-discontinuity designs. Stata Journal, 17(2), 372-404.

[Liu Qiang]

The qfitci figures are the standard CIs for a quadratic least squares fit. They assume a constant standard deviation on each side of the cutoff.

In contrast, rdplot assumes neither a quadratic model nor constant standard deviations. The data are binned. A different standard deviation is estimated in each bin and the CI based on that SD is centered on the bin mean (Colonico et al., 2017, p.382, 391). I would personally prefer these CIs.

Two-way options can be used: they go inside the graph_options option.


Reference:
Calonico, S., Cattaneo, M. D., Farrell, M. H., and Titiunik, R. (2017). rdrobust: Software for regression-discontinuity designs. Stata Journal, 17(2), 372-404.

Really grateful to your clear answer. I have learnt where the difference is.
For the last question, I mean is it possible to absorb rdplot into tw so that we can draw graphs like tw rdplot y x || scatter y x. I do not suppose it's possible for now.

[Steve Samuels]

Maybe graph_options accepts addplot as an option. If so, rdplot will subsume twoway, rather than the other way around.

[Diana Perez Lopez]

You could try using gr line instead of qfitci and use some variables generated using the rdplot option 'genvars'.

[ericmelse]

For those visitors interested, the link to the data mentioned at #1 is broken.
However, all packages and data are available at this Github page.
Note that this page is mentioned at #2 in this Statalist post.