where !
j
(x, u) is a weight for estimand or effect j that can be empiri-
cally determined and the limits (a, b) are known in any application.
We can generate LATE as a special case of this formula. When the
model has limited support (regions where MTE can be identified), the
estimator automatically adjusts for it.
25
Bounds for the treatment parameters are presented in
Heckman and Vytlacil (2006b). Different instruments produce differ-
ent weights and these weights are generally not the weights required to
define the standard treatment effects. Our approach is far more gen-
eral than the piecemeal type of analysis of what IV estimates of the
sort presented by Sobel in his comments on the statistical literature.
Each of his special cases drops out from our general analysis. The
MTE
approach presents a nonparametric control function analysis
where the propensity score plays a conceptually distinct role from the
role it plays in matching models (Heckman and Vytlacil 2006b). Our
analysis is not to be found in the statistics literature.
Sobel is clearly a fan of the LATE approach. Therefore, he has
to be a fan of MTE. The Imbens-Angrist (1994) LATE parameter is a
discrete version of the Bjo¨rklund-Moffitt (1987) marginal gain para-
meter introduced into the evaluation literature in a selection model
framework. The Bjo¨rklund-Moffitt parameter is the mean gain to
participants induced into the program by an instrument. They identify
the parameter in a selection framework. Imbens-Angrist show how IV
can approximate it. Heckman and Vytlacil (1999) show how local
instrumental variables (LIV) identify it. Heckman, Urzua, and
Vytlacil (2006) and Heckman and Vytlacil (2006b) show that IV and
selection models are closely related. IV and its extension Local IV
(LIV) estimate the slopes of the models estimated by selection models
in levels.
As pointed out in Heckman, Urzua, and Vytlacil (2006) and
Heckman and Vytlacil (2005), the ‘‘monotonicity’’ assumptions made
in the LATE literature are not innocuous. If, in response to a change
in an instrument, some people go into treatment and others drop out,
instrumental variables do not identify any treatment effect but they
do identify a weighted average of two way flows (Heckman and
25
Software for estimating MTE and generating all of the treatment
parameters is available from Heckman, Urzua, Vytlacil (2006). See the website
http://jenni.uchicago.edu/underiv.
REJOINDER: RESPONSE TO SOBEL
153
Vytlacil 2005, 2006b; Heckman, Urzua, and Vytlacil 2006). The recent
IV
literature is asymmetric. Outcomes are permitted to be hetero-
geneous among persons in a general way. Choices of treatment are
not permitted to be heterogeneous in a general way.
7. POLICY RELEVANT TREATMENT PARAMETERS
The Policy Relevant Treatment Effect (PRTE) introduced in
Heckman and Vytlacil (2001a) and elaborated in Heckman and
Vytlacil (2005, 2006a,b) is a good example of the benefits of the
econometric approach. It is defined by stating a policy problem—
estimating the effect of a policy on mean outcomes—and showing
that this treatment effect can be generated as a weighted average of
the marginal treatment effect with known weights using formula (1)
presented in the preceding section. Standard IV and matching estima-
tors do not, in general, identify this parameter.
Policy problems dictate the identification and estimation strat-
egy in our approach. As shown in Heckman (2001), Heckman and
Vytlacil (2001a, 2005, 2006b), Heckman, Urzua, and Vytlacil (2006)
and Carneiro, Heckman, and Vytlacil (2005), the weights on MTE
required to form the PRTE parameter are generally not the same as
the weights for OLS, matching or IV, although an IV estimator can be
devised to identify the PRTE.
Heckman and Vytlacil (2005) develop an algorithm for defin-
ing causal effects that answer specific policy problems from a general
list of possible problems rather than relying exclusively on the stan-
dard set of causal effects discussed by Sobel in his Section 3 that
answer only a few narrowly selected policy problems.
26
Sobel ignores
parameters like the PRTE and fails to recognize that the standard
treatment estimators do not identify this parameter. Heckman and
Vytlacil develop estimators for specific well-posed policy problems
26
In his defense of ACE, Sobel makes a familiar error. In defending
ACE
as estimating the effect of a policy with universal coverage compared to the
effect with no coverage, he fails to account for the effects of large scale programs
on potential outcomes—what economists call ‘‘general equilibrium’’ effects.
Heckman, Lochner, and Taber (1998a,b) show that these are empirically impor-
tant in the analysis of education policies. These effects violate the SUTVA
assumption of Holland (1986) or the invariance assumption of Hurwicz (1962).
154
HECKMAN
rather than hope that a favored estimator just happens to hit the
selected target. This is a large advance over the existing literature in
statistics. Just compare Sobel’s discussion of IV with our own.
8. ESTIMATING THE PROPORTION OF PEOPLE WHO
BENEFIT FROM A PROGRAM
Sobel’s discussion of the benefits of randomization illustrates all of
the problems with the ad hoc statistical approach he favors.
Randomized trials cannot identify Pr(Y
1
> Y
0
). In a large sample,
this is the proportion of the population that benefits from a pro-
gram.
27
See Heckman (1992). This is because randomized trials pro-
duce Y
1
or Y
0
but not both for each person. The parameter
Pr(Y
1
> Y
0
) is not even contemplated in the Neyman (1923)–Rubin
(1978) setup. Using the Roy model (Heckman and Honore´ 1989) or
more general models (Carneiro, Hansen, and Heckman 2001, 2003;
Cunha, Heckman, and Navarro 2005, 2006; Cunha and Heckman
2006a,b; Heckman, Lochner, and Todd 2006) it is possible to estimate
this proportion. Modeling the unobservables and their relationship with
the treatment selection rule and any related measurement equations
plays an important role in their analysis. The statistical treatment effect
literature is silent on this crucial parameter. Modeling the dependence
among the unobservables in choice, outcome and auxiliary measure-
ment equations, is the key to identifying this proportion.
Sobel says that ‘‘much stronger assumptions’’ are required to
estimate this parameter. In any specific case, this claim is not true.
The assumptions required to justify randomization (no randomization
bias; no contamination or crossover effects; see Heckman, LaLonde,
and Smith 1999) are different and not weaker or stronger than
the assumptions used to identify the Roy model and its extensions.
Indeed when randomization breaks down, Roy models and their
generalizations can exploit the attrition and self selection information
to identify Pr(Y
1
> Y
0
). See Heckman (1992) and Heckman and
Vytlacil (2006a,b).
27
I keep the conditioning on covariates implicit. I assume a heteroge-
neous response model.
REJOINDER: RESPONSE TO SOBEL
155
Dostları ilə paylaş: |