About the Regression Discontinuity Design Prolems

[Alice LIN]

First, how to recognize the type of RDD?
Is that like this: (in my model, the treatment var is Dibao and the running var is lnscore)
gen newdibao=(lnscore>0.5)
tab newdibao dibao
and the result is like this

| dibao
newdibao | 0 1 | Total
-----------+----------------------+----------
0 | 27,335 591 | 27,926
1 | 61,933 996 | 62,929
-----------+----------------------+----------
Total | 89,268 1,587 | 90,855

so, from this table, it is a fuzzy RDD. Is that right?

Second, it is about the graphs after running rdplot command, if the trends of lines are different between before and after the threshold, can I say that this dependent variable cannot be chosen in RDD?

Third, it is about the logic of RDD in the paper. I am confused about the logic while writing a paper using RDD as a model.
Is that "check if it is a sharp or fuzzy RDD -> draw the twoway or plot graph by using twoway or rdplot-> run the regression by using rd command -> do the robustness check (find the discontinuity arount the cutoff, find the optimal bandwidth, and find the sensitivity of the obs...)"

Please tell me, thanks.

[Jeff Wooldridge]

Alice: This is a sharp design. The treatment probability increases from zero to one at the cutoff.

RD methods explicitly allow for different slopes on either side of the cutoff.

In addition to using the bandwidths chosen by rdrobust, you should try larger and smaller bandwidths.

[Alice LIN]

Alice: This is a sharp design. The treatment probability increases from zero to one at the cutoff.

RD methods explicitly allow for different slopes on either side of the cutoff.

In addition to using the bandwidths chosen by rdrobust, you should try larger and smaller bandwidths.

Thanks for your reply.
I thought the slopes on either side of the cutoff should be both downwards or upwards in order to show that the two groups on both sides are comparable. I drew the graphs and they are like below:



they are quite different from the graphs that have shown in some research papers. In my view, graph 1 and 4 do not satisfy the assumption of RDD (1 has different trend, 4 is continuous). Is that right?

In order to get the best bandwidth, the command rdbwselect can be used, I think.

Thanks for your patience.

[Jeff Wooldridge]

I agree that doesn't look promising. Typically the relationship looks monotonic in the running variable. The graphs are a bit hard to see, but in 2, 3, and 4 there appears to be no effect at the jump. In 1, there isn't much data to the left of the jump.

One thing you might do is create the graphs with a narrower band around the cutoff. I doubt that will change much. Sometimes there just isn't an effect.