rdplot - magnitude and statistical significance of discontinuous jump

[Mario Liebensteiner]

Dear all,

I use the command -rdplot- to assess graphically if there is a discontinuous jump in a time-series variable (H, i.e. hydro electricity production) after a particular treatment date (the running variable, "run", indicates 365 days before and after treatment: -365 to +365, treatment is on day 0). I apply a 5th order polynomial fit using the option -p(5)-.
rdplot H run, c(0) msize(tiny) graph_options(xlabel(-365(73)365) title("") xtitle("Day") ytitle("H") graphregion(color(white)) bgcolor(white)) p(5) ci(95) shade binselect(qsmvpr)

This is the plot, which indicates that the discontinuous jump is positive but statistically insignificant:


Using other options for the selection of the bins (-binselect()-), I obtain basically the same picture (i.e. a positive jump), but of course the confidence intervals change.

However, I do not manage to reproduce the underlying regression to obtain the exact magnitude and statistical significance of the jump.

I tried the following command:
rdrobust H run, c(0) p(5)

With the following regression output:

The estimate is, to my surprise, negative (-3.85) and statistically significant (p-value = 0.00). This stands in contrast to what we can see in the plot.

I would be most grateful if anyone could help me with the following questions:

1. How can I obtain the magnitude and statistical significance of the jump indicated by the plot?
2. Why do the two commands (-rdplot- and -rdrobust-) give incompatible (and even contrasting) evidence?

I assume that I have either a severe mistake in my code or an error in my reasoning.

Your help would be most appreciated! Thank you in advance.

Best,
Mario

[Salvatore Lattanzio]

I guess that the difference arises because when using rdrobust you are restricting the sample within the optimal bandwidth, which should not be the case in rdplot.

[Mario Liebensteiner]

Dear Salvatore,
Many thanks for your hint!
Best,
Mario