Hello community!
I am having trouble with how to define the cutoff when the running variable is discrete when using the rdrobust package. My running variable is month of birth and the policy I am analyzing was enacted on October. Therefore, any child born after October was affected by this policy. I normalize the month of birth variable X such that it is re-centered around October (October 2001 = 0). The frequency of my running variable is:
There is an overall slight decrease in the number of observations per birth month from -1 (September) to 0 (October) to 1(November). The change does not seem concerning. There is seasonality in the data but it is quite consistent across years. My sample contains almost 2 years before and after the cutoff.
owever, when I run the rddensity command to check for bunching, I get the following result:
.
The p-value suggests that there is in fact disconuity around the cutoff. However, I notice that the first window to the right of the cutoff is 14083, which suggests it is counting the observations from 0 and 1 together. Is this normal?
My concern is that I am not defining the cutoff properly. Maybe I should define cutoff as -0.5, or something where there are basically no observations? Or is this not the right approach?
Any help is appreciated, thanks!
I am having trouble with how to define the cutoff when the running variable is discrete when using the rdrobust package. My running variable is month of birth and the policy I am analyzing was enacted on October. Therefore, any child born after October was affected by this policy. I normalize the month of birth variable X such that it is re-centered around October (October 2001 = 0). The frequency of my running variable is:
Code:
x | Freq. Percent Cum.
------------+-----------------------------------
-25 | 9,037 2.32 2.32
-24 | 8,964 2.30 4.61
-23 | 8,550 2.19 6.81
-22 | 8,452 2.17 8.97
-21 | 9,576 2.45 11.43
-20 | 8,681 2.23 13.65
-19 | 9,936 2.55 16.20
-18 | 9,976 2.56 18.76
-17 | 9,953 2.55 21.31
-16 | 9,524 2.44 23.75
-15 | 9,237 2.37 26.12
-14 | 8,825 2.26 28.38
-13 | 8,755 2.24 30.63
-12 | 7,901 2.03 32.65
-11 | 7,366 1.89 34.54
-10 | 7,518 1.93 36.47
-9 | 7,995 2.05 38.52
-8 | 7,953 2.04 40.56
-7 | 9,045 2.32 42.87
-6 | 8,719 2.24 45.11
-5 | 8,953 2.30 47.41
-4 | 8,813 2.26 49.66
-3 | 8,727 2.24 51.90
-2 | 8,092 2.07 53.98
-1 | 7,868 2.02 55.99
0 | 7,309 1.87 57.87
1 | 6,774 1.74 59.60
2 | 6,961 1.78 61.39
3 | 7,484 1.92 63.31
4 | 7,075 1.81 65.12
5 | 8,276 2.12 67.24
6 | 7,946 2.04 69.28
7 | 8,208 2.10 71.38
8 | 7,907 2.03 73.41
9 | 7,839 2.01 75.42
10 | 7,400 1.90 77.32
11 | 7,003 1.80 79.11
12 | 6,411 1.64 80.76
13 | 5,927 1.52 82.28
14 | 5,679 1.46 83.73
15 | 6,110 1.57 85.30
16 | 5,895 1.51 86.81
17 | 6,782 1.74 88.55
18 | 6,646 1.70 90.25
19 | 6,903 1.77 92.02
20 | 6,145 1.58 93.60
21 | 5,924 1.52 95.12
22 | 5,293 1.36 96.47
23 | 4,948 1.27 97.74
24 | 4,520 1.16 98.90
25 | 4,294 1.10 100.00
------------+-----------------------------------
Total | 390,075 100.00
owever, when I run the rddensity command to check for bunching, I get the following result:
.
Code:
rddensity x, c(0)
Computing data-driven bandwidth selectors.
Point estimates and standard errors have been adjusted for repeated observations.
(Use option nomasspoints to suppress this adjustment.)
RD Manipulation test using local polynomial density estimation.
c = 0.000 | Left of c Right of c Number of obs = 575399
-------------------+---------------------- Model = unrestricted
Number of obs | 344443 230956 BW method = comb
Eff. Number of obs | 200415 162845 Kernel = triangular
Order est. (p) | 2 2 VCE method = jackknife
Order bias (q) | 3 3
BW est. (h) | 23.000 23.000
Running variable: x.
------------------------------------------
Method | T P>|T|
-------------------+----------------------
Robust | -18.6449 0.0000
------------------------------------------
P-values of binomial tests. (H0: prob = .5)
-----------------------------------------------------
Window Length / 2 | <c >=c | P>|T|
-------------------+----------------------+----------
1.000 | 7868 14083 | 0.0000
2.000 | 15960 21044 | 0.0000
3.000 | 24687 28528 | 0.0000
4.000 | 33500 35603 | 0.0000
5.000 | 42453 43879 | 0.0000
6.000 | 51172 51825 | 0.0422
7.000 | 60217 60033 | 0.5977
8.000 | 68170 67940 | 0.5348
9.000 | 76165 75779 | 0.3233
10.000 | 83683 83179 | 0.2182
-----------------------------------------------------
My concern is that I am not defining the cutoff properly. Maybe I should define cutoff as -0.5, or something where there are basically no observations? Or is this not the right approach?
Any help is appreciated, thanks!
