Dario Debowicz explains the methodology behind his analysis of the impact of Bolsa Familia transfers within different municipalities in Brazil. Read the full research at: http://bit.ly/1plN4yI
Methodology: Quantile regression panel data for Bolsa Familia
1. METHODOLOGICAL
PRESENTATION
QUANTILE REGRESSION PANEL
DATA FOR BOLSA FAMILIA SOCIAL
TRANSFERS
Armando Barrientos & Dario Debowicz
BWPI, University of Manchester
2. CURRENT RESEARCH
The effects of Bolsa Familia social transfers in Brazil on labour supply
and school attendance (DFID IRIBA Phase I).
Goal is to analyse the distribution of program outcomes across
municipalities in Brazil using quantile regressions, which have not been
used in this context...
taking into account the indirect effects of the program on
beneficiaries and non-beneficiaries.
controlling for time-invariant conditions at the municipality level.
(partially) addressing the issue of endogeneity in program
assignment.
Bolsa Familia is Brazil’s flagship social protection program. Its national
budget (0.5% GDP) is allocated among municipalities, mainly as a
function of the distribution of pre-program poverty, based on household
data. Municipalities then allocate the transfers at the household level.
The transfers are conditional on children’s school attendance and
health check-ups.
3. METHODOLOGY AND DATA
Quantile regressions panel data for short period following
Abrevaya & Dahl (2008), on municipality-level data.
Our database is constructed using an annual cross-section
of a nationally representative household-level survey
(PNAD), focusing on the municipalities for which PNAD is
representative (273).
We use seven waves to build our panel, from the start of
the program (2003) to the last year when the set of surveyed
municipalities remains unchanged (2009).
Among the covariates, we include pre-program poverty data
used by the Brazilian government to allocate the program
quotas among municipalities (PNAD 2001).
Households benefiting from the program are identified via a
specific question in the survey (when possible) or via typical
transfer values (otherwise).
5. PRECISION OF PARAMETER ESTIMATES
Usual bootstrap procedures to calculate standard
errors imbedded in Stata are not consistent for this
method (Abrevaya and Dahl, 2008).
Starting with Abrevaya and Dahl’s bootstrap
estimation procedure, we extend it to include more
than two years and time-invariant regressors, such
as pre-program poverty at municipality level, using
Stata.
8. FINDINGS (1 OF 2)
The estimated effect of the municipal coverage of Bolsa Família on the
distribution of female school attendance among conditional quantiles of
outcome indicates significant heterogeneity, in contrast to OLS results.
While point estimates are positive for all conditional quantiles analysed, the
effect is statistically significant (with a ten percent significance level) from
quantile .10 to quantile .40 of the outcome distribution. The maximum is
achieved in quantile .10, where a less than 8 (0.1268-1) percentage point
(p.p.) increase in program incidence is needed to “buy” a p.p. of female
school attendance. This contrasts sharply with the median municipality,
where more than 28 (0.0349-1) p.p. of program incidence are needed for the
same target, or with municipalities in the top of the distribution, where the
program leads to practically no school attendance increases.
9. FINDINGS (2 OF 2)
A formal test of the uniformity of Bolsa Família effects on female school
attendance among quantiles, which we conduct by extending the analysis of
AD, rejects the null of uniformity even at one percent level of significance.
At the same significance level, a test of significance of the correlated random
effects rejects the null of insignificance. This suggests that unobservables
affecting female school attendance may be captured, in part, by repeated
observations on the program incidence, and that results not accounting for
them could lead to a significant bias in the estimation of program effects.
A test of overall significance of the regression does not reject the null of lack
of significance, suggesting significant variation of school attendance was left
unexplained.
With the notable exception of the bottom quantile analysed, the program has
worked to reduce differential outcomes across the conditional schooling
distribution of municipalities. This adds to the evidence in support of Bolsa
Família’s contribution to inclusive growth.
10. FOLLOWING STEPS AND FUTURE
RESEARCH
Try to include 2000 census poverty, yearly Bolsa Familia benefits, and
other municipal-level information in our database, for which we are trying
to get the names of the municipalities in PNAD from the Brazilian statistical
office (IBGE).
Future research could look into other program outcomes of interest.
11. RESULTS: LABOR SUPPLY (%)
.1 .2
-.1
0
QRPD
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
Quantiles
OLS
Source: Authors’ quantile regression panel data and OLS regressions. For
quantile regressions, point estimates and 90% and 95% confidence
intervals are included. Municipal cluster standard errors are reported in
parenthesis. All specifications include all regressors as listed in Table A1 in
the Appendix. The sample size is 1,911 municipal-level cases.
12. WORKING PAPER:
Antipoverty transfers and inclusive growth in Brazil:
http://bit.ly/1plN4yI
Other working papers from the International Research Initiative on Brazil and
Africa (IRIBA) are available at:
http://www.brazil4africa.org/publications/
Editor's Notes
PNAD=Pesquisa Nacional por Amostra de Domicilios, National Survey of Residences
𝑄 𝜏 𝑦 𝑚𝑡 𝑥 𝑚 are the conditional quantiles of the response variable 𝑦 𝑚𝑡 , 𝑥 𝑚𝑡 ′ is a row vector of covariates of municipalities 𝑚 at time 𝑡. 𝜓 𝜏 𝑡 , 𝑦𝑒𝑎𝑟 𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐 𝑖𝑛𝑡𝑒𝑟𝑐𝑒𝑝𝑡𝑠. Linear function.
𝛽 𝜏 denotes a time-invariant effect column vector by which the covariates effect the conditional quantiles of the observables above and beyond the effects that work through the unobservables, the 𝛽 𝜏 corresponding to 𝐵𝐹 being our main focus of interest. 𝜓 𝜏 𝑡 is a location shift in the conditional quantiles and the last generic term 𝑥 𝑚𝑡 ′ λ 𝜏 𝑡 captures the effects of the unobservables into the conditional quantiles at the different times 𝑡, with the unobservables being a linear projection onto the observables.
𝐿𝐴 stands for adult labour participation rate in the working-age population (18-64 years olds);
𝐿𝐴𝐼 𝑚,𝑡 denotes the share of informal employment in the population of working age, where informality includes all those reporting having worked in the week previous to the survey but no in formal employment (contribution to social security).
We report on the estimated effects of municipality coverage of Bolsa Familia on the distribution of labour outcome variables for the working-age population, the 𝛽 𝜏 associated with Bolsa Familia coverage. Figure shows the estimates for the effects of Bolsa Família on adult labour force participation rates. To facilitate comparison of our results with those in Ribas and Soares (2011), we can look at the median in the distribution, both are measures of central tendency. In contrast to the non-statistically significant effects of Bolsa Família on adult labour supply found by Ribas and Soares, we observe a positive and statistically significant effect that for the municipality in the median of adult labour supply participation rate. The estimate suggests that adult labour supply increases by 0.09 percentage point for each percentage point increase in programme coverage. We also observe that above-median municipalities have lower effects. Overall, we find that Bolsa Familia tends to reduce the spread of labour participation rates among municipalities We plan to test this hypothesis formally.
We report on the estimated effects of municipality coverage of Bolsa Familia on the distribution of labour outcome variables for the working-age population, the 𝛽 𝜏 associated with Bolsa Familia coverage. Figure shows the estimates for the effects of Bolsa Família on adult labour force participation rates. To facilitate comparison of our results with those in Ribas and Soares (2011), we can look at the median in the distribution, both are measures of central tendency. In contrast to the non-statistically significant effects of Bolsa Família on adult labour supply found by Ribas and Soares, we observe a positive and statistically significant effect that for the municipality in the median of adult labour supply participation rate. The estimate suggests that adult labour supply increases by 0.09 percentage point for each percentage point increase in programme coverage. We also observe that above-median municipalities have lower effects. Overall, we find that Bolsa Familia tends to reduce the spread of labour participation rates among municipalities We plan to test this hypothesis formally.
We report on the estimated effects of municipality coverage of Bolsa Familia on the distribution of labour outcome variables for the working-age population, the 𝛽 𝜏 associated with Bolsa Familia coverage. Figure shows the estimates for the effects of Bolsa Família on adult labour force participation rates. To facilitate comparison of our results with those in Ribas and Soares (2011), we can look at the median in the distribution, both are measures of central tendency. In contrast to the non-statistically significant effects of Bolsa Família on adult labour supply found by Ribas and Soares, we observe a positive and statistically significant effect that for the municipality in the median of adult labour supply participation rate. The estimate suggests that adult labour supply increases by 0.09 percentage point for each percentage point increase in programme coverage. We also observe that above-median municipalities have lower effects. Overall, we find that Bolsa Familia tends to reduce the spread of labour participation rates among municipalities We plan to test this hypothesis formally.