causality in statistics is recursive and rules out simultaneous causality.
My Section 2.5 is standard econometrics but new statistics that reveals
the power of the econometric approach over ‘‘frontier’’ methods in
statistics like directed acyclic graphs (see Pearl 2000) that are recur-
sive. Given long-standing interest in social interactions by sociologists
(see also Durlauf and Young 2001) it is unfortunate that Sobel
dismisses out of hand this contribution of my essay that allows
sociologists to define and identify models of social interactions that
are ruled out in the Rubin (1978)–Holland (1986, 1988) approach that
he espouses.
Sobel misses another main contribution of my paper: to make
the literature on causality specific by addressing real problems.
Abstract discussions of causality with appeals to philosophy, ‘‘closest
worlds,’’ ‘‘regularity,’’ and the like sound profound but in fact are
superficial since they have no operational content and do not address
policy problems. Most empirical social scientists are not concerned
with philosophy per se, but instead they want honest answers to
clearly stated problems.
My essay is organized around the theme of addressing three
policy evaluation questions that arise in everyday practice. Problem
P1 is what statisticians focus on. Problems P2 and P3 are new prob-
lems that can be answered using the econometric approach. Sobel’s
attempt to address P2—the extrapolation problem—demonstrates the
ad hoc
nature of current statistical approaches and the power of the
econometric approach. By keeping the unobservables implicit, he
disguises an implicit exogeneity assumption for the conditioning vari-
ables in the approach he advocates. Heckman and Vytlacil (2005,
2006b) consider extrapolation under more general conditions.
An alternative approach discussed in Heckman and Vytlacil (2005,
2006b) is to model the dependence between the conditioning variables and
the unobservables, but since Sobel (like Rubin and Holland) does not like
to make unobservables explicit, this route is denied him. He implicitly
agrees that the literature in statistics does not address P3 and his discus-
sion of it confuses problem P2 with problem P3.
He offers a discussion of factorial experiments as a substitute
for a clear discussion of P3. Since he will not make explicit statements
about unobservables, he begs a central question addressed in the
econometrics literature of how to predict the effects of new programs
never previously observed. His version of P3 is a simple extrapolation
REJOINDER: RESPONSE TO SOBEL
141
exercise and hence a version of P2. He does not offer a general
treatment of P3 as I do in the paper and in Heckman and Vytlacil
(2005, 2006b).
Our approach goes well beyond the interpolation and extrapola-
tion advocated by Sobel to consider predictions of programs with features
never previously observed.
6
Making predictions about new programs and
policies is an essential task of social science. If these predictions are not
made in a cautious, principled, explicit way, plenty of ‘‘causal inference
experts’’ stand waiting to fill the vacuum and provide less credible esti-
mates. By discussing hard problems clearly, analysts can pinpoint limits to
knowledge that raise the standards of evidence.
In failing to present a serious discussion of problem P3, which
is a central focus of my essay, Sobel echoes the conventional statistical
approach that ignores this policy evaluation problem. Had he care-
fully considered my analysis of this problem in this paper and in my
work with Vytlacil, his claims about close ‘‘agreement’’ between the
econometric approach and the statistical approach would vanish.
Sobel misses another major theme of my essay and the entire
econometric evaluation literature: that treatment effects should be defined
relative to the problem being analyzed (Marschak 1953; Heckman 2001;
Heckman and Vytlacil 2001b, 2006a,b). The econometric literature devel-
ops the point that the choice of an estimator cannot be separated from the
choice of the question being addressed by the investigator and the a priori
assumptions made by the investigator.
Thus, although ACE, ITT or TT may be traditional para-
meters, they do not address many specific policy questions. My
work with Vytlacil (2001b, 2005, 2006b) discusses specific policy
questions and devises estimators that address them. Manski (2000,
2004), in his interesting work, derives treatment assignment rules for
particular loss functions.
3. IS CAUSALITY IN THE MIND?
Sobel, evidently influenced by my exchange with Tukey (in Wainer
1986, reprinted 2000), sharply attacks me for claiming that ‘‘causality
6
His approach is based on arbitrary functional form assumptions
whereas we present a general analysis guided by theory.
142
HECKMAN
is in the mind’’ and that causal knowledge is provisional. Then,
throughout the rest of his discussion, he demonstrates the validity of
my point by offering a series of unsupported opinions and assertions
about what is ‘‘reasonable’’ and what is not. His opinions and value
judgements are his expressions of intuitive models in his mind that he
never formalizes or makes explicit, but introduces casually using
rhetorical devices. Sobel is typical of many statisticians who keep
crucial assumptions implicit.
My claim is not intended as a defense of solipsism, post-mod-
ernist relativism or the notion that analysts are free to make up any
crazy model they like. Instead, I am saying that all scientific activity is
predicated on assumptions.
A clearly formulated causal model should (a) define the rules
or theories that generate the counterfactuals being studied, including
specification of the variables known to the agents being studied as
well as the properties of the unobservables of the model where the
unobservables are not known to the analyst but may be partly known
by the agent; (b) define how a particular counterfactual (or potential
outcome) is chosen; (c) make clear the assumptions used to identify
the model (or to address the policy questions being considered); and
(d) justify the properties of estimators under the maintained assump-
tions and under alternative assumptions. These are tasks 1
(corresponding to a and b), 2 (corresponding to c) and 3
(corresponding to d) in my Table 1. These ingredients are the hall-
mark of the selection model as analyzed in Heckman (1974, 1976,
1979), Gronau (1974), Roy (1951), Willis and Rosen (1979) and
numerous
other
papers
in
the
econometrics
literature.
Understanding the relationship between the unobservables generating
choice of treatment and the unobservables generating outcomes is the
key to understanding the properties of various evaluation estimators,
a point first made in Heckman and Robb (1985, 1986, reprinted
2000).
One will look in vain in the papers of Neyman, Cox,
Kempthorne, Rubin or Holland for the specification of precise
treatment assignment rules that have been the hallmark of econo-
metric selection models since 1974. Sobel’s claim that Rubin (1978)
or any other statistician has systematically developed treatment
assignment mechanisms is false. Rubin (1976, 1978) contrasts rando-
mization with nonrandomization and does not develop the structure
REJOINDER: RESPONSE TO SOBEL
143
Dostları ilə paylaş: |