Mechanisms and Extrapolation in the Abortion-Crime Controversy - Mechanism and Causality in Biology and Economics - page 193

the key factor of being born unwanted is unmeasured in USA data. Similarly, many
uncertainties would be involved in the details of the scale-up model through which
this extrapolated result is integrated with other information to produce the estimate
of the impact of legalized abortion in the 1970s on crime rates in the 1990s.
Estimates of the impact of legalized abortion on crime based on national-level
statistical data also confront a variety of uncertainties, for instance, concerning the
proper modeling approach and how to account for other factors—such as the crack
epidemic of the late 1980s—that might affect the results. Indeed, the discussions
between Donohue and Levitt and their critics have predominantly focused on such
issues (see Foote and Goetz 2008 ; Donohue and Levitt 2004 , 2008 ; Joyce 2003 ).
However, I should emphasize that the point here is definitely not to insist upon the
infirmity of causal inferences grounded in extrapolation and observational data.
Uncertainties frequently arise in experiments too, especially those involving human
subjects (for instance, due to noncompliance, i.e., the failure of some subjects in the
experiment to follow the experimental protocol). Such uncertainties are inherent in
any attempts to learn about causation in large complex systems wherein numerous
practical and ethical concerns restrict the types of studies that are possible. Conse-
quently, scientific inference in such situations usually must build a cumulative case
from a variety of lines of evidence none of which is decisive in isolation.
Although that may seem a rather obvious point, it does seem to get overlooked in
some critical discussions of extrapolation. For instance, LaFollette and Shanks
( 1996 ) argue that results from animal experiments can never be extrapolated across
species boundaries (e.g., from rats to humans) because causally relevant differences
between populations are always present. The discussion of extrapolation in Sect. 3.2
has already illustrated several shortcomings with this line of argument. Extrapolation
need not be direct, and it may be possible to adjust for relevant differences between
the model and target. Moreover, some types of causal claims—such as claims about
positive causal relevance—can be directly extrapolated even when considerable
differences exist. Nevertheless, it is true that extrapolation is often haunted by the
possibility that relevant differences between model and target have not been ade-
quately accounted for. But this is only to say that there is often an unavoidable
element of uncertainty inherent in extrapolations, just as there is in any other method
for learning about causation in very complex systems. That in no way precludes
extrapolations from being one useful line of evidence among others.
However, one might object that extrapolation can never be more than very weak
evidence, useful only when information concerning the target population is grossly
incomplete. Suppose that in initial stages of the investigation, studies performed on
the target alone provide only rather uncertain evidence for the causal relationship
and that in this context the extrapolation strengthens the overall case. Moreover,
suppose that subsequent studies of the target population are able to make a
compelling argument for causal claim in a way that does not require any reliance
on the model. 10 This type of situation shows that the importance of extrapolation in
10 Reiss ( 2010 ) suggests that an example discussed in (Steel 2008 , chapter 5) follows this plot line.