Versión de HTML Básico
Lorena Moreno
102
Analiti a, Revista de análisis estadístico, Vol. 13 (1), 2017
Table 3:
CJX Manipulation test
Cutoff c = 28.2
Left of c Right of c
Number of obs
2.139
4.182
Eff. Number of obs
652
1.012
Order loc. poly. (p)
2
2
Order BC (q)
3
3
Bandwidths (hl,hr)
estimated estimated
Bandwidth values
6,465
8,895
Bandwidth scales
0,5
0,5
Method
T P
>
T
Conventional
-0,5719
0,5674
Undersmoothed
-0,1293
0,8972
Robust Bias-Corrected -0,0914
0,9272
Note:
With a robust bias-corrected local polynomial of order 2 density esti-
mator of -0,0914 (T) and an associated p-value
>
0,10 (P
>
T) we cannot reject
the null hypothesis of no statistically significant differences of the densities
around the threshold
et al. (2015) polynomial data-driven RD plots, with sample averages within bins and 95%
confidence intervals.
The averages of the covariates around the cutoff are very similar, signalling an as good
as random local assignment of the BDH. Therefore, we can expect for households with index
scores just below and above 28,2 to be similar in all confounders, observed and unobserved.
Also, an important step prior to the estimation is the analysis of the association of the
treatment D and the instrument Z. A cross-tabulation (Table 4), as well as an OLS, allows
us to check the level of compliance that the instrument induces, which is 46%. With an F-
statistic of 1.787,39 and a p-value
<
0,001 we can reject the null hypothesis of no significant
association.
This can also be aided by the graphical representation of the probability of the treatment
given the RSII index. As seen in Figure 6, there is a jump at the threshold, meaning that
households with indexes at least as low as 28,2 are about 14% percentage points more likely
to be in the treatment group. Following estimates restrict the observations to the selected
discontinuity samples.
20