REJOINDER: RESPONSE TO
SOBEL*
James J. Heckman
†
‘‘The term ‘cause’ is highly unspecific. It commits us to nothing
about the kind of causality involved nor about how the causes
operate. Recognizing this should make us more cautious about
investing in the quest for universal methods for causal inference.’’
Cartwright, 1999, Chapter 5
.
Sobel claims to disagree with many of the points made in my
paper. He also claims that much if not all of what I say is already in
the statistical treatment effect literature. He treats my Section 4 as a
literature review rather than an illustration of the basic principles
made in Sections 1–3 of the paper, as I intended it to be. In joint
work with Edward Vytlacil, I present a comprehensive literature
review (Heckman and Vytlacil 2006a,b).
The primary objective of my paper is to present a general and
coherent view of causality as it applies to social science. As part of my
analysis, I address the approach to causality popularized in statistics
by Donald Rubin, Paul Holland and other statisticians. This is an
approach to which Sobel subscribes. As my essay documents, the
statistical approach suffers from many limitations and in many funda-
mental respects is a recapitulation of older approaches in econo-
metrics, well understood by economists, that have been enhanced
and developed further by contemporary econometricians. I am
disappointed that, rather than addressing my arguments, Sobel
restates misleading arguments made in the statistics literature. In
*This research was supported by NSF-SES-0241858 and NIH-R01-
HDO43411. I thank Jennifer Boobar, Steve Durlauf and Hanna Lee for com-
ments on this rejoinder.
†
University of Chicago and the American Bar Foundation
135
responding to Sobel, I am in essence responding to Rubin, Holland
and other statisticians whose views are reiterated by Sobel in his
commentary.
1. WHAT MY PAPER IS ABOUT
My paper moves discussions of causal inference away from vague
philosophical discussions about what ‘‘really’’ constitutes causality to
a precise discussion of three prototypical policy problems.
1
In my
interpretation, causal models are tools for policy analysis. Different
policy problems place different demands on models and data. I
articulate the econometric approach that (a) defines the problems of
interest precisely; (b) describes the environments, outcomes and
choices of the agents being studied precisely and (c) presents condi-
tions on data and models under which the policy problems can be
solved. The objective of my paper is not to attack statistics but rather
to attack serious policy problems. Sobel attacks the explicit approach
developed in econometrics and confuses clearly formulated abstract
models for outcomes and selection of outcomes with assumptions
made within the context of the explicit models that are maintained
in particular applications of the models.
Building on my previous analysis (Heckman 2001), I reconcile
the statistical treatment effect literature and the econometrics litera-
ture by noting that the wider set of questions addressed by the latter
entails considering more ambitious models. Whether the particular
assumptions required for identifying a parameter are satisfied is a
different problem than the problem of determining conditions
under which a question can in principle be answered. I draw on
Marschak (1953) and later economists to note that for certain nar-
rowly focused policy questions, it is often possible to get by with
much weaker assumptions and data requirements when crafting
acceptable answers.
1
Cartwright (2005) provides an illuminating discussion of alternative
and often inconsistent uses of the term ‘‘causality.’’
136
HECKMAN
My essay is about all three policy questions, P1–P3, and
not solely about P1. Sobel, however, largely focuses on P1. He briefly
touches on P2 and considers a special case of an exogenous condition-
ing set. His discussion of problem P3 is about extrapolation
from factorial experiments instead of a careful discussion of how
to forecast new programs with new characteristics as discussed
in my paper and in Heckman and Vytlacil (2005, 2006b). His discus-
sion is defensive and does not grapple with the larger aims of my
paper.
The careful reader of my paper, Sobel’s discussion and the
recent literature on causal inference in econometrics and statistics
will recognize that Sobel ignores major points that are developed in
the econometrics literature and are absent in the statistical treatment
effect literature. These are:
1.
Development of an explicit framework for outcomes, measure-
ments and the choice of outcomes where the role of unobservables
(‘‘missing variables’’) in creating selection problems and justifying
estimators is developed.
2.
The analysis of subjective evaluations of outcomes and the use of
choice data to infer them.
3.
The analysis of ex ante and ex post realizations and evaluations of
treatments. This analysis enables analysts to model and
identify regret and anticipation by agents. Developments 2 and
3 introduce human decision making into the treatment effect
literature.
4.
Development of models for identifying entire distributions of
treatment effects (ex ante and ex post) rather than just the tradi-
tional mean parameters focused on by statisticians. These distri-
butions enable analysts to determine the proportion of people
who benefit from treatment, something not attempted in the
literature Sobel draws on.
5.
Development and identification of distributional criteria allowing
for analysis of alternative social welfare functions for outcome
distributions comparing different treatment states.
6.
Models for simultaneous causality relaxing the recursive frame-
works adopted by Rubin (1978) and Holland (1986).
7.
Definitions of parameters made without appeals to hypothetical
experimental manipulations.
REJOINDER: RESPONSE TO SOBEL
137
Dostları ilə paylaş: |