of nonrandomized selection rules in the fashion pioneered and devel-
oped in econometrics.
7
The act of defining a model is a purely mental activity. It may
draw on preexisting theory (which is itself derived from earlier mental
acts), interpretations of data (which involve a mental act using models
for the phenomenon being studied and models of statistical inference)
and the rules of logic. There is no purely empirical process for
discovering or defining causality. All causal knowledge is conditional
on maintained assumptions.
8
Sobel (p. 106) dismisses the conditional nature of causality,
writing that ‘‘I do not believe that most scientists (or philosophers)
would subscribe to this view and were they to do so they would
presumably have little further interest in causality.’’ This claim runs
contrary to a large body of thought in philosophy associated with
Kant and Hume, among others.
Indeed, ‘‘causality’’ is not a central issue in fields with well-
formulated models where it usually emerges as an automatic by-
product and not as the main feature of a scientific investigation.
Moreover, intuitive notions about causality have been dropped in
pursuit of a rigorous physical theory. As I note in my essay with
Abbring (2006), Richard Feyman in his work on quantum electro-
dynamics allowed the future to cause the past in pursuit of a scienti-
fically rigorous model even though it violated ‘‘common sense’’ causal
principles. The less clearly developed is a field of inquiry, the more
likely is it to rely on vague notions like causality rather than explicitly
formulated models.
Most scholars intend to describe the real world with their
models. Certainly, in addressing P1–P3, I am referring to real world
problems. However, any empirical or theoretical analysis rests on
assumptions. The clearer analysts are about these assumptions and
the more they are able to test them, the more clearly stated are the
sources of agreement or disagreement among analysts. Such clarity
determines the next steps in the scientific process of constructing
7
When Sobel writes that TT
¼ ACE ¼ TUT when treatment decisions
do not depend on potential outcomes, he presents a garbled version of a precise
result established in my 1974 and 1976 papers.
8
Indeed all knowledge is conditional in this sense. For example, much of
modern mathematics is predicated on the Axiom of Choice.
144
HECKMAN
better models and better data to narrow down the zones of disagree-
ment among analysts.
Although Sobel objects to my assertion about the importance
of maintained assumptions and a priori beliefs in the causality enter-
prise, he proceeds, like many statisticians, to pass judgements about
what is and is not good practice. Unlike the econometric approach,
Sobel adopts the statistical approach which often hides key assump-
tions by invoking slogans instead of science.
Each of Sobel’s judgements is based on maintained assump-
tions and beliefs ‘‘in his mind.’’ They involve his implicit assumptions
and value judgements. Statisticians like him who are not explicit do
not convey their private thoughts in an objective, publicly interpre-
table way. Like many statisticians working in the field of causal infer-
ence, he is not clear about many of his crucial implicit assumptions.
Implicit assumptions are entailed in writing down the arguments of
causal relationship (1) (his notation) and characterizing its properties.
Sobel writes ‘‘the focus of the statistical literature is primarily
on obtaining the best possible estimate of the causal parameter of
interest’’ without defining ‘‘best possible,’’ the basis for the choice of
the ‘‘parameter of interest’’ or the question being addressed.
9
In his
defense of matching, he appeals to conditional independence assump-
tion (11) (in the notation of his paper) as ‘‘ straightforward’’ and that
it ‘‘readily lends itself to use by empirical investigators.’’ Some statis-
ticians often use the phrase ‘‘it works well in practice’’ without defin-
ing ‘‘works well.’’ Another common slogan used to justify matching is
that ‘‘my clients understand it.’’ In these and numerous other
instances, Sobel and a large statistical community implicitly appeal
to a variety of conventions rather than presenting explicit rigorous
models and assumptions. The credo ‘‘let sleeping dogs lie’’ is good for
sales, but it is bad for science. Instead of invoking slogans as a
solution to problems, the structural approach emphasizes understand-
ing the underlying mechanisms producing outcomes and selection
rules.
As an application of the scientific approach, my discussion of
matching develops the point that (i) there is no rigorous basis for
picking the set or sets of conditioning variables that make the method
9
The ‘‘best possible estimate’’ is defined precisely in Bayesian and Wald
decision theories. See, e.g., the discussion in Manski (2000).
REJOINDER: RESPONSE TO SOBEL
145
‘‘work.’’ Heckman and Navarro (2004) show how the conventional
model selection rules for picking the conditioning variables W in a set
of data can produce badly biased estimates of the average causal
effect and other parameters. (ii) The method assumes that the analyst
has as much relevant information as the agent being studied (see, e.g.,
Heckman and Navarro 2004, for a precise definition of ‘‘relevant’’
information). If the agent knows more than the analyst and acts on it,
matching breaks down. (iii) The analyst assumes that people at the
margin of being attracted into a program are the same (have same
outcomes on average) as average participants. Matching is just non-
parametric regression analysis. It is more careful than Ordinary Least
Squares (OLS) in accounting for empirical support problems but it
assumes that the conditioning variables that the analyst has at his
disposal fortuitously solve selection problems.
10
To take another point, like many statisticians, Sobel resolutely
defends randomization. The Rubin (1978)–Holland (1986) papers
take as their benchmark randomized trials where treatments are
selected by a hypothetical randomization. As I point out in my
essay, even under ideal conditions, unaided randomization cannot
answer some very basic questions such as what fraction of a popula-
tion benefits from a program.
11
And in practice, contamination and
cross over effects make randomization a far from sure-fire solution
even for constructing ATE or ACE (see the evidence on disruption
bias and contamination bias arising in randomized trials that is pre-
sented in Heckman, LaLonde, and Smith 1999; Heckman, Hohmann,
Smith, and Khoo 2000). Sobel makes a series of implicit assumptions
about what questions should be answered, the effects of randomiza-
tion on participants and the like.
Sobel also disagrees with my claim that statisticians conflate
the three tasks shown in my Table 1. The analysis of Holland (1986,
1988) is a good illustration of my point. It also illustrates the central
10
In one of his many attributional errors, he echoes Imbens (2004) and
credits Barnow, Cain, and Goldberger (1980) with the phrase ‘‘selection on
observables’’ to describe matching. The term originates in Heckman and Robb
(1985) and is not to be found in Barnow, Cain, and Goldberger.
11
See Carneiro, Hansen, and Heckman (2001, 2003), where this para-
meter is identified using choice data and/or supplementary proxy measures. See
also Cunha and Heckman (2006a,b) and Cunha, Heckman, and Navarro (2005,
2006).
146
HECKMAN
Dostları ilə paylaş: |