Академический Документы
Профессиональный Документы
Культура Документы
Social Sciences
Much of science is consumed with the study of cause and consequence, particularly the
fit between theories of causation and empirical observations of the world. The social
42
sciences have experienced a great deal of internal controversy about the feasibility of
studying cause and consequence. This controversy has been fueled by the sloppy use of
language to interpret results of empirical analyses, the failure to explicitly identify key
assumptions, and the inappropriate use of specific analytic techniques. We do not review
this controversy here. Instead we review some of the key analytic approaches to the study
of causation in the social sciences and highlight the ways in which mixed method data
assigning subjects to experimental conditions. Instead they must usually draw causal
inference from observing people in social settings. The difficulty in drawing conclusions
about causal effects from observational data is that the exposure of interest is usually
allocated in some systematic way – the exposed group differs in composition from the
unexposed group and these differences in composition are related to the outcome being
studied (Moffitt 2003, 2005; Winship and Morgan 1999). In other words, people are not
randomly assigned to the social condition we want to study. This inevitably leads to the
problem that there are alternative explanations for the associations between outcome and
hypothesized cause that we observe. Put yet another way, some factors we do not observe
allow the assignment of people to experimental conditions in the social world, usually
when some type of special government benefits or programs are to be distributed, and
those involved desire a “fair” system of distribution. But these circumstances are rare.
43
An entire literature is devoted to causal modeling in the social sciences. This
literature proposes various analytic techniques for estimating models of cause and
argues that social scientists should give up altogether the effort to study cause and
consequence (Abbot 1998; Bachrach and McNicoll 2003; Freedman 1991; Fricke 1997;
Heckman 1978, 2000; Marini and Singer 1988; Marsden 1991; Moffitt 2003, 2005;
Raftery 1998; Rubin 1974; Smith 2003; Snijders and Hagenaars 2001; Winship and
Morgan 1999). During the past 50 years, a series of hopeful methodological fads emerged
and receded, each hoping to solve the problem of studying causal relationships in the
approaches to studying social causes always suffer from either untestable assumptions, or
Many innovative research designs hold tremendous promise for addressing causal
associations in the research design itself (Campbell and Stanley 1963; Cook and
Campbell 1979; Rosenbaum 1999, 2001). Studies of twins and other “natural”
because they are generally thought to have the benefit of removing the chance that
44
unobserved factors produce observed associations between the change and subsequent
behavior of the people affected (Card and Krueger 1995; Heckman and Smith 1995;
Rosenzweig and Wolpin 2000). That is, they exploit situations in which the exposure of
interest was allocated as though it were randomly assigned. However, these approaches
also have limitations. First, such situations are rare, and the rarity of these situations
dramatically limits the range of topics and settings that can be studied with this approach
(Moffitt 2003, 2005). Second, often times careful examination of exposures that seem to
still produce the observe empirical associations (Bound and Solon 1999; Heckman 2000;
Thus, none of the approaches currently available offers a panacea for the problem
of causal inference for all subjects, settings, or occasions. In fact, we argue there are no
ultimate solutions to the problem of causal inference in the social sciences. The approach
we advocate for the study of cause and consequence in the social sciences lies between
recognition of the flaws in all available observational approaches and abandonment of the
effort. We advocate taking full advantage of every reasonable approach available to build
We argue that advances in causal reasoning in the social sciences must come from
a constant interplay between theory and empirical evidence. In our view the careful social
scientist should try to address the problem of unobserved factors producing observed
45
includes (1) theorizing and reasoning regarding possible unobserved factors that may be
eliminate key rival hypotheses, and (3) designing measurement strategies to observe as
many of the potential causes as possible. Although we believe no one approach will solve
the problem of causal inference, we believe these steps advance the social scientific study
view, neither social scientific methods nor the data gathered and analyzed by social
scientists can demonstrate “proof” of cause. Rather, scientists hypothesize about cause in
the theoretical arguments they create. Empirical analyses of data about the social world
may be consistent or inconsistent with those arguments, but they cannot prove or
demonstrate causation (Marini and Singer 1988). Based on this perspective, the more
reliable evidence brought to bear on a theory the better. Therefore, we espouse an
approach to the study of causation that uses multiple sources of evidence to test, or
challenge, causal theories – an approach that lies at the foundation of the methods we
advocate.
of two specific aspects of measurement in the social sciences. These are the temporal
ordering of measures is important because most causal reasoning has temporal ordering
46
embedded in it. That is, when we reason that X causes Y, almost always we argue that X
occurs before Y. The most common exception to this temporal order is the argument that
anticipation of X in the future causes Y now. However, this exception simply identifies a
possible to observe Z, the temporal order of our reasoning is still quite clear: Z causes Y
and Z occurs before Y. This view of the relationship between causal reasoning and
and 6, and the idea that anticipation of events may produce cause is fundamental to the
redundant measures from multiple sources can be used to help reduce the chances that
bias associated with the measurement is responsible for observed associations. We argue
that the best measurement strategies take advantage of multiple methods to allow the
strengths of some methods to compensate for the weaknesses of other methods, thereby
reducing the likelihood of replicating bias across measures (Rosenbaum 2001). The effort
approach to the study of a specific causal question may be threatened by bias. Replication
varying both research design and measurement strategy, holds great promise for
47
argument for a causal relationship. Measures of all potential causal factors identified by
the social world. Measures of the potential mechanisms responsible for producing an
such mechanisms is one tool for establishing a causal relationship (Rosenbaum 2001).
Again, mixed method data collection is a particularly useful tool in this endeavor.
Readers may associate empirical study of cause and consequence in the social world with
statistical analyses of survey data. However, just as we argue that no one data collection
method is sufficient for the study of cause and consequence, the leading survey
methodologists make the same argument about survey methods. Consider the following
ext Surveys are rather blunt instruments for information gathering. They are powerful in
producing statistical generalizations to large populations. They are weak for in generating rich
understanding of the intricate mechanisms that affect human thought and behavior. Other
techniques are preferred for that purpose. (Groves et al. 2004, p. 378)
Surveys are powerful for providing evidence of associations, but they are less powerful
for discovering the mechanisms responsible for those associations. Less structured
methods, such as observation and unstructured interviewing, are more powerful for
discovering these mechanisms (Moffitt 2000; Seiber 1973). But once such potential
mechanisms are discovered, again survey methods are a powerful tool for establishing
48
associations between these potential mechanisms and the outcome of interest. As a result,
mixed method techniques that combine survey data collection with less structured
interviewing or observational data collection are extremely powerful for advancing causal
techniques.