opr196VS
-
Upload
natiatsiqvadze -
Category
Documents
-
view
214 -
download
0
Transcript of opr196VS
-
7/31/2019 opr196VS
1/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantiles regression
Pauline Givord
CREST, INSEE
2011/2012
Pauline Givord Evaluation of Public Policies
http://find/http://goback/ -
7/31/2019 opr196VS
2/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantiles regression
Pauline Givord
CREST, INSEE
2011/2012
Pauline Givord Evaluation of Public Policies
http://find/ -
7/31/2019 opr196VS
3/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Introduction
95% of applied econometrics is concerned with averagesbut many variables (earnings, test scores...) have continuousdistributions: they can change in a way not revealed by anexamination of averages
Pauline Givord Evaluation of Public Policies
http://find/ -
7/31/2019 opr196VS
4/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Moving Beyond Average
Growing interest on distributional outcomes beyond simple averages(what is happening to the entire distribution) :
inequalities analysis
ex: at average real wages in US since the 80s, but upperearnings quantiles have been increasing while lower quantileshave been falling (Buchinsky, 1998,...)policy maker might be interested in the effect of treatment ondispersion of the outcome, or its effect on lower tail of theoutcome distribution:
Heckman, Smith and Clements (1997): many persons would judgeprograms to be successful if (...) enough of the right kinds of persons, reaped benets from them even if the average participant
did not.Pauline Givord Evaluation of Public Policies
http://find/ -
7/31/2019 opr196VS
5/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile Regression
rapidly expanding empirical (and theoretical) quantile regression
literature in economicsquantile regression:provides a convenient linear framework for examining how thequantiles of a dependent variables change in response to a setof independent variablesallows the estimation of linear conditional quantile functions
Pauline Givord Evaluation of Public Policies
http://find/ -
7/31/2019 opr196VS
6/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Course Overview
This course :1. (briey) introduces quantile regression
for a more detailed presentation see Buchinsky (1998) andKoenker et Hallock (2002)
2. and discuss some major application(s) to the evaluation
framework : quantile treatment effect
Pauline Givord Evaluation of Public Policies
http://find/ -
7/31/2019 opr196VS
7/45
Denition and Notation
-
7/31/2019 opr196VS
8/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Denition and NotationQuantile RegressionEstimationInterpretationDiscussion
Minimization
Helpful to think about quantile as the solution to a minimizationproblem.
simplest case : Medianit solves argmin b i |Y i b | : Least Absolute Deviation (LAD)
Pauline Givord Evaluation of Public Policies
Denition and Notation
http://find/ -
7/31/2019 opr196VS
9/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Denition and NotationQuantile RegressionEstimationInterpretationDiscussion
Intuition
First order conditions
S ( y , b )
b =
i
(1(u i > 0)(+ 1) + 1(u i < 0)( 1))
with u i = y i b .minimum for b such as we have as many negative and positiveresiduals.i.e. we have as many y i higher than b as we have y i smallerthan b : denition of median.
Note that the location of the median depends only on the signs of the residuals ( y i b ), so robust to outliers
Pauline Givord Evaluation of Public Policies
Denition and Notation
http://find/http://goback/ -
7/31/2019 opr196VS
10/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Denition and NotationQuantile RegressionEstimationInterpretationDiscussion
Check Function
This formula can be generalized to any other quantile :the th sample quantile solves
argmin b i :Y i b
|Y i b | + i :Y i < b (1 )|Y i b |
or : argmin b i (Y i b )solution to a problem that minimizes the weighted sum of the
absolute value of the residualsthe weight function is called the check function:
(u ) = u ( 1(u < 0))
Pauline Givord Evaluation of Public Policies
Denition and Notation
http://find/http://goback/ -
7/31/2019 opr196VS
11/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Denition and NotationQuantile RegressionEstimationInterpretationDiscussion
Quantile Regression
We are interested in the conditional distribution according toobservable X.e.g. : Quetelets growth chart (distribution of weight/heightconditionally on age)
Pauline Givord Evaluation of Public Policies
Denition and Notation
http://find/ -
7/31/2019 opr196VS
12/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
Quantile Regression
Motivation : observables X affect the entire shape of thedistributionthe impact on tail quantile can be very different than the impact on
the central quantilethe quantile regression assumes a linear dependence of thequantile in these observables.quantile regression model replaces the b in the precedingprogram by a linear function of the covariateswe estimate :
= arg min (Y i X i )
Pauline Givord Evaluation of Public Policies
Denition and Notation
http://find/ -
7/31/2019 opr196VS
13/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
Estimation
the check function is not differentiable, so common gradientprocedure cannot be usedcan be written as the solution of linear programming model(Koenker and Bassett, 1978).implementation by stata (qreg, sqreg) or R (rq), sas
(quantreg).
Pauline Givord Evaluation of Public Policies
Denition and Notation
http://find/ -
7/31/2019 opr196VS
14/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
Results
we could run a quantile regression for each different of :
Q Y |X ( ) = X
the impact is summarized by the function coefficient:
i.e. we will get a different coefficient vector for every value of presentation of results: usually graph of the coefficientestimates as a function of the quantiles
Pauline Givord Evaluation of Public Policies
Denition and Notation
http://find/ -
7/31/2019 opr196VS
15/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
Example : Koenker and Hallock
Birthweight
Pauline Givord Evaluation of Public Policies
I d iDenition and NotationQ il R i
http://find/ -
7/31/2019 opr196VS
16/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
Interpretation
we measure how a quantile of the conditional distributionchanges with change in the covariates
parameter of interest: EQ Y | X ( ) X j i.e. marginal change in the th conditional quantile after amarginal change in X j .if x j is entered linearly, it is just j not individual interpretation: does not imply that a subject inthe th quantile of one conditional distribution would still ndhimself there with different value of X
Pauline Givord Evaluation of Public Policies
I t d tiDenition and NotationQ til R g i
http://find/ -
7/31/2019 opr196VS
17/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
Conditional versus Unconditional Quantile
we obtained estimates of the impact on a covariate on theconditional quantilein average estimation framework (OLS), it is sufficient toobtain consistent estimates of the impact of X on thepopulation unconditional mean of the outcome Y :a linear model for conditional meansE [Y |X ] = X impliesthat E [Y ] = E [X ] (law of iterated expectations)BUT conditional quantiles do not average up to their
population counterparts:q Y ( ) = E [q Y |X ( )]
estimates cannot be used to estimate the impact of X on thecorresponding unconditional quantile
Pauline Givord Evaluation of Public Policies
IntroductionDenition and NotationQuantile Regression
http://find/ -
7/31/2019 opr196VS
18/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
Conditional versus Unconditional Quantile : Example
Firpo, Fortin and Lemieux (2009) Unconditional quantileregressions , Econometrica
empirical question:what is the impact on median earnings of increasingproportion of unionized workers, holding everything elseconstant?estimates obtained by running a quantile regression could not
answer this simple questionFFL propose an estimation of the unconditional quantiletreatment effect (under exogeneity assumption for T )
Pauline Givord Evaluation of Public Policies
IntroductionDenition and NotationQuantile Regression
http://find/ -
7/31/2019 opr196VS
19/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
cond. versus uncond. quantile impact of union status onearnings
Pauline Givord Evaluation of Public Policies
IntroductionDenition and NotationQuantile Regression
http://find/ -
7/31/2019 opr196VS
20/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
cond. versus uncond. quantile impact of union status onearnings
unions reduce within-group dispersion, where groupconsists of workers with sameX and increase conditional mean of wages of union workers (i.e.between group inequalities)so tend to increase wages for low unconditional quantiles (both
effects go in same direction)but can decrease wages for high unconditional quantiles(effects go in opposite directions)
Pauline Givord Evaluation of Public Policies
IntroductionDenition and NotationQuantile Regression
http://find/ -
7/31/2019 opr196VS
21/45
IntroductionQuantile Regressions
Quantile Treatment Effect
Quantile RegressionEstimationInterpretationDiscussion
Advantages on OLS
estimate is less sensitive to outliers values of the outcomeequivariance to monotone transformation :if h is a monotone function, P (T < t |X ) = P (h(T ) < h(t )|X )then Q h(Y ) |X ( ) = h(Q Y |X ( ))The mean does not share this property:
E (h(Y )|X ) = h(E (Y |X ))
useful for censored data (cf Buchinsky, 1998), as thequantiles of the censored conditional quantiles functioncorrespond to the quantiles of the uncensored conditionalquantile function
Pauline Givord Evaluation of Public Policies
Introduction Denition
http://find/ -
7/31/2019 opr196VS
22/45
Quantile RegressionsQuantile Treatment Effect
Identication : CIAIdentication: IV
Denition
Evaluation framework : we are interested in the effect of a binarytreatment T on an outcome Y .
Let Y 0 and Y 1 the potential outcomes with and withouttreatmentF Y 0 and F Y 1 the corresponding distributions.
Pauline Givord Evaluation of Public Policies
Introduction Denition
http://find/ -
7/31/2019 opr196VS
23/45
Quantile RegressionsQuantile Treatment Effect
Identication : CIAIdentication: IV
Quantile Treatment Effect
we dene the th quantile treatment effect (QTE):
= F 1Y 1
( ) F 1Y 0
( )
Horizontal distance between both distributions (Lehmann,1974 and Doksum, 1974).Similarly, we could dene the restriction to the treated
(QTET): |T = 1 = F
1Y 1 |T = 1 ( ) F
1Y 0 |T = 1 ( )
Pauline Givord Evaluation of Public Policies
Introduction Denition
http://find/ -
7/31/2019 opr196VS
24/45
Quantile RegressionsQuantile Treatment Effect
Identication : CIAIdentication: IV
Interpretation
without further assumption on the joint distribution of potential outcomes, we estimate the difference of the quantilesand not the quantile of the difference (i.e. the treatmenteffect) Y 1 Y 0 no individual interpretation : a nding of a treatment effectof at the th quantile says nothing about the treatmenteffect for the person at the th quantile of the untreatedoutcome distribution
In many applications, this is sufficient to answer economicallymeaningful questionschanges in the median, in the lower tails of the distribution, of the Gini coefficient...otherwise, need to make assumption on joint distributions
Pauline Givord Evaluation of Public Policies
Introduction Denition
http://find/ -
7/31/2019 opr196VS
25/45
Quantile RegressionsQuantile Treatment Effect
Identication : CIAIdentication: IV
Rank Invariance Assumption
rank invariance: implies that the treatment does not alter the
ranking of the units:If i as Y 0 i < Q Y 0 ( ), then Y 1 i < Q Y 1 ( )when it is not likely to be satised for all observations,Heckman, Smith and Clements (1997) propose bounds for thequantile treatment effect under several assumptions on theranking
Pauline Givord Evaluation of Public Policies
IntroductionQ il R i
DenitionId i i CIA
http://find/ -
7/31/2019 opr196VS
26/45
Quantile RegressionsQuantile Treatment Effect
Identication : CIAIdentication: IV
Identication
preceding quantile regression methods could be used (we justemphasize the impact of a particular explanatory variable T )
but so far we have (implicitly) ignored potentially endogeneityexcept in case of experimental data, we have to deal with thesame selection effects as usual
Extension of the classical identication methods toestimate counterfactual distributions.still ongoing research eld...
Pauline Givord Evaluation of Public Policies
http://find/ -
7/31/2019 opr196VS
27/45
IntroductionQuantile Regressions
DenitionIdentication : CIA
-
7/31/2019 opr196VS
28/45
Quantile RegressionsQuantile Treatment Effect
Identication : CIAIdentication: IV
Unconditional Quantile Treatment Effect
we face the same problem as before : usual quantile regressionsestimate treatment effect at different conditional quantilesbut as stated before, it cannot be used to estimate the impactof the treatment on corresponding unconditional quantilesFirpo proposes a (two stages) semi-parametric direct
estimation of unconditional quantile treatment effect
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
DenitionIdentication : CIA
http://find/ -
7/31/2019 opr196VS
29/45
Quantile RegressionsQuantile Treatment Effect
Identication : CIAIdentication: IV
Identication
Firpo shows that under preceding assumptions the quantile Q Y 1 ( )
could be expressed as an implicit function of theobserved (Y , T , X ):
= E T
p (X )1(Y Q Y 1 ( ))
in this expression, the onlyconditional function to estimate is thescore p (X ) = P (T = 1|X ).
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
DenitionIdentication : CIA
http://find/ -
7/31/2019 opr196VS
30/45
Quantile RegressionsQuantile Treatment Effect
Identication : CIAIdentication: IV
Conditional Quantiles
Similarly, we have:
= E 1 T
1 p (X )1(Y Q Y 0 ( ))
and for conditional quantiles Q Y 1 ( |T = 1) and Q Y 0 ( |T = 1):
= E T p
1(Y Q Y 1 |T = 1 ( ))
and = E
p (X )1 p (X )
(1 T )p
1(Y Q Y 0 |T = 1 ( ))
with p = P (T = 1)
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
DenitionIdentication : CIA
http://find/ -
7/31/2019 opr196VS
31/45
Q gQuantile Treatment Effect Identication: IV
Estimation
Two-stages procedure :1. (non parametric) estimation of p (X )2. Estimates of Q Y 1 and Q Y 0 are obtained by a reweighted
version of the standard quantile regression procedure:consistent estimators of Q Y t ( ) (t = 0, 1) are given by:argmin b t ,i (Y i b )with 1 ,i = T i N p (X ) (i.e. sample analogue of T / p (X )) and
0 ,i =1 T
i N (1 p (X i )) .QTE = Q Y 1 ( ) Q Y 0 ( )
We could similarly estimate QTET.
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
DenitionIdentication : CIA
http://find/ -
7/31/2019 opr196VS
32/45
Q gQuantile Treatment Effect Identication: IV
Empirical Application
identication of the causal impact of a training program onfuture earningsLalonde data set (1986): use bot experimental data set(National Supported Work Program) and observational dataset (PSID)see also Dehejia and Whaba (1999): less destructive resultsthan Lalonde for non experimental data, when one correctly
corrects for differences in observables variables in treatmentand control groupsestimation of the score that balances covariates betweentreated and control groups
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
DenitionIdentication : CIA
http://find/ -
7/31/2019 opr196VS
33/45
Quantile Treatment Effect Identication: IV
Observed and Counterfactual Distributions of PotentialOutcomes (Treatment Group)
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
DenitionIdentication : CIA
http://find/ -
7/31/2019 opr196VS
34/45
Quantile Treatment Effect Identication: IV
Quantile Treatment Effect for the Treated
Pauline Givord Evaluation of Public Policies
http://find/ -
7/31/2019 opr196VS
35/45
IntroductionQuantile Regressions
Q il T Eff
DenitionIdentication : CIAId i i IV
-
7/31/2019 opr196VS
36/45
Quantile Treatment Effect Identication: IV
Instrumental Variable Estimate of the Quantile TreatmentEffect
Abadie, Angrist et Imbens (2002), Instrumental Variables
Estimates of the Effect of Subsidized training on the quantiles of Trainee Earnings, Econometricaextension of AIR framework to quantile regressionestimation of the QTE with an instrument
as Firpo, AAI show that it could be obtained as a weightedversion of the standard quantile regressionapplication to a randomized experiment evaluation
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
Quantile Treatment Effect
DenitionIdentication : CIAIdentication: IV
http://find/ -
7/31/2019 opr196VS
37/45
Quantile Treatment Effect Identication: IV
Notation
Random affectation to treatment : Z = 0, 1treatment T = 0, 1, depends on instrument (denoted by T 0and T 1 )outcome Y depends on treatment Y t (ie Y 0 and Y 1 )observable characteristics X
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
Quantile Treatment Effect
DenitionIdentication : CIAIdentication: IV
http://find/ -
7/31/2019 opr196VS
38/45
Quantile Treatment Effect Identication: IV
Assumptions
1. independence: (Y 1 , Y 0 , T 1 , T 0 ) indep of Z cond. X 2. non trivial assignment : 0< P (Z = 1|X ) < 1
3. rst stage E [T 1 |X ] = E [T 0 |X ]4. monotonicity : P (T 1 T 0 |X ) = 1 (no deers)
compliers still dened as individuals who change their treatmentstatus with instrument : T 1 > T 0
independence of the potential outcome with treatment forcompliers:(Y 1 , Y 0 ) T |X , T 1 > T 0
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
Quantile Treatment Effect
DenitionIdentication : CIAIdentication: IV
http://find/ -
7/31/2019 opr196VS
39/45
Quantile Treatment Effect Identication: IV
Estimation of QTEc
Identiable parameter:
Q (
Y |X
,T
,T 1
>T 0
) = T
+X
i.e. estimation of the QTE for compliers estimation of
( , ) = argminE ( (Y T X )|T 1 > T 0 )
Pb: population of compliers is not identied
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
Quantile Treatment Effect
DenitionIdentication : CIAIdentication: IV
http://find/ -
7/31/2019 opr196VS
40/45
Quantile Treatment Effect Identication: IV
AAD show that we could use the weight function:
(T , Z , X ) = 1 T (1 Z )
(1 0 (X ))
(1 T )Z 0 (X )
with 0 (X ) = P (Z = 1|X ).note that equals 1 if T = Z (Compliers).AAD show that for all functionh(Y , T , X ):
E [h(Y , T , X )|T 1 > T 0 ] =1
P (T 1 > T 0 ) E [h(Y , T , X )]
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
Quantile Treatment Effect
DenitionIdentication : CIAIdentication: IV
http://find/ -
7/31/2019 opr196VS
41/45
Q
the could estimate using
( , ) = argminE [ (Y T X )]
in practice, pb as could be negative, so they use instead thenonnegative weight = E [ |Y , T , X ] = P (T 1 > T 0 |Y , T , X )
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
Quantile Treatment Effect
DenitionIdentication : CIAIdentication: IV
http://find/ -
7/31/2019 opr196VS
42/45
Q
Application : JTPA
experimental data
Job Training Partnership Act (JTPA) : offer services forindividuals facing barriers to employmentrandom assignment (20 000 individuals), but only about 60%of those offered training actually received JTPAnote that very few individuals in the control group receivedJTPA services: less than 2%
Pauline Givord Evaluation of Public Policies
http://find/ -
7/31/2019 opr196VS
43/45
IntroductionQuantile Regressions
Quantile Treatment Effect
DenitionIdentication : CIAIdentication: IV
-
7/31/2019 opr196VS
44/45
Results
Pauline Givord Evaluation of Public Policies
IntroductionQuantile Regressions
Quantile Treatment Effect
DenitionIdentication : CIAIdentication: IV
http://find/http://goback/ -
7/31/2019 opr196VS
45/45
Extension of classical evaluation methods to quantile treatmenteffect estimation :
Lamarche (2007) : xed effects to deal with endogeneity biasin a evaluation of the vouchers experiment (Milwaukee).
Froelich and Melly (2008) extension to discontinuity regressiondesignAthey et Imbens (2003): application to differences indifferences.
application to duration models : Koenker and Bilias (2002),evaluation of the Bonus Experiment
Pauline Givord Evaluation of Public Policies
http://find/http://goback/