Versión de HTML Básico
Lorena Moreno
114
Analiti a, Revista de análisis estadístico, Vol. 13 (1), 2017
Relying on the latest national Living Standard Measurement Survey, and in the richness
of its questionnaire, I delineated a two stage quasi-experimental FRDD study. The first
stage reproduces the eligibility index RSII, based on a Categorical Principal Component
Analysis. The created 0 to 100 index accounts for 36,5% of the total variance of the selected
variables and has a 62,2% correlation with the monthly per-capita aggregate consumption
(similar to the original index). The official threshold was adopted; therefore, households with
values at least as lower as 28,2 are considered eligible. For the second stage, I exploited the
characteristics of the programme, mainly the continuous RSII and the cutoff, to implement
a RDD on its fuzzy version given that the reception of the bonus showed to be not perfectly
determined by the eligibility status (or instrument). The size of the sample for the evaluation
had to be reduced due to intricate policy and methodological changes happening in the
surveying period, trying to avoid extra noise in the results. Additionally, since the outcome
was only registered for boys and girls up to 59 months old, the valid final sample was
compressed to 6.174 observations.
The identification strategy of the RDD, was tested indirectly via manipulations tests and
covariate balance. Both, empirically support the smoothness of the potential outcomes for
the transfer recipients and non-recipients around the cutoff. Once the possibility of effect
falsification was rejected estimates were drawn from two approaches a parametric two-stage
least square instrumental variable regression analysis and a non-parametric local polynomial
regression with robust bias-corrected estimates and confidence intervals.
For the IV estimates the bandwidth selected was
±
3, for which covariate average values by
eligibility group showed to be not significantly different. Based on three specifications, linear
(1), quadratic (2), and quadratic with controls (3), I found that the first stage accounted
for an increase in the probability of treatment of around 18% to 19%. The effect of the
instrument in the outcome, i.e. the intention-to-treat, showed a significant coefficient only
for specification (3) with a -0,41 s.d. The ratio of these ITT (outcome discontinuity) to the
first stage (treatment discontinuity), is the effect of the BDH for the compliers within the
chosen bandwidth. None of these ratio estimates was found to be significantly different from
0 which could lead to fail to reject the null hypothesis of no effect. Though, even when the
minimal RDD restrictions are likely to be met, under the fuzziness of the present design the
last conclusion might seem a bit extreme.
Complementarily, the findings from the non-parametric procedure based on the work
by Calonico, Cattaneo, Titiunik and others, account for estimates very similar to those
from the IV method. With the same specifications as my parametric, but relying on data-
driven bandwidths, the local polynomial robust estimates showed no statistically significant
effect of the BDH over stunting z-scores. Though, as previously mentioned we ought to be
cautious when concluding about the policy. A finding that could be emphasized is that the
10% level significant conventional non-parametric effect estimate for specification 3 (
α
= -
2,032, p-value = 0,082) changed to non-significant when robust bias corrected (
α
= -2,042, p-
value = 0,122). Consequently, if the present research would have only relied on a conventional
32