Last week we discussed the idea of a structural, as opposed to an
individual-level explanation of a phenomenon.1 Haslanger (2015), What Is a (Social) Structural
Explanation? Philosophical Studies. We also talked about
structural discrimination and oppression. See Haslanger (2012), Chapter
11 - Oppressions: Racial and Other, in Haslanger (2012), Resisting
Reality: Social Construction and Social Critique, Oxford University
Press. This week we continue
to talk about structural explanation in the social sciences, and see to
what extent the manipulability account of causation can adequately model
it.2 The main readings for today are: Ross (2023), What Is
Social Structural Explanation? A Causal Account, Nous; Malinsky
(2018), Intervening on Structure, Synthese 195:2295-2312.
We have come full circle: we started with the manipulability account
of causation and today we return to it.3 James Woodward (2003), Making Things Happen: A
Theory of Causal Explanation, 2003, Oxford University Press.
Ross on Social Structural Causes
These three examples of structural causality are helpful to fix
ideas. First example:
Jason has a factory job that starts at 6am and he commutes via a city
bus. He takes the first available morning bus and manages the 45-minute
commute to arrive on time. As Jason is poor, he lacks financial and
other resources that would allow for alternative travel options. After
financial changes, the city implements cutbacks which eliminate Jason’s
bus-route. The early route he usually takes is discontinued and there
are no other routes to get him to work on time. After the manager states
that no other shifts are available, Jason cannot arrive on time, and he
loses his job. (section 2.1)
Second example:
Some studies report a higher prevalence of less “healthy” diets in
individuals from minority and lower socioeconomic groups and there is
interest in explaining why this is the case … some impoverished US
neighborhoods are considered “food deserts” in the sense that there are
no close grocery stores selling fresh produce. These locations often
lack bus routes to nearby stores and they contain barriers for walking
options … For individuals living in these areas, the lack of resources
(including time and finances) makes “choosing” such a diet extremely
difficult, if not impossible. (section 2.1)
Third example:
A third example concerns explanations of the Black-white wealth gap
among Americans. A common social structural explanation of this racial
wealth gap appeals to “historical and contemporary structural factors” …
Financial services were denied to Blacks through redlining practices and
discriminatory covenants prevented them from owning, occupying, or
leasing property, which excluded them from receiving Federal Housing
Administration loans … The systemic denial of resources to Blacks made
home ownership–and the ensuing accumulation and transmission of
wealth–exceedingly difficult, if not outright impossible. (section
2.1)
Contrary to what many think, there is nothing mysterious about
structures acting as causes. Social structures are causes simply because
they meet the counterfactual test of causation in the manipulability
account.4 A different question—which Ross does not address—is, how
does the manipulability account model the distinction between
structuring (or constraining) causes and triggering causes? How would
you answer this question? See the other reading for today by Malinsky.
Consider the first example mentioned earlier:
In this case, the relevant structure is bus transit availability (T),
which is either available (1) or not (0) in a given location. Within the
interventionist framework, it makes perfect sense to say that this
structure causes and explains the job loss (L) outcome. Given that Jason
is willing to attend work, the absence of this resource prevents him
from going, which causes the job loss. The operative counterfactual here
is that if this resource were available, then the job loss would not
have occurred.(section 3.1)
How should structures as causes be understood? Haslanger (as we saw
last week) suggests that we understand structural explanations of social
phenomena in terms of a part/whole relationship.5 Recall the examples from last week: the treat in the
ball; the grading curve; standing up when the Queen enters; the
invisible foot. But, while some
part/whole relationships are explanatory and causal, others are not. How
are we to distinguish between the two?
Jason is “part” of his family, church community, individuals on
planet Earth, and so on, but none of these explain his job loss. Of
course, this doesn’t demonstrate that part-whole relations cannot be or
are never explanatory, just that we need something more to capture what
is. If some factor is explanatory in virtue of standing in a part-whole
relationship to the outcome, yet many irrelevant factors also have this
feature, what distinguishes the relevant factors from the irrelevant
ones? (section 2.3)
Instead of thinking in part/whole terms, Ross prefers to think of
structural causes as constraints. Here is an example from
Dretske:6 Dretske (1988), Explaining behavior: Reasons in a
world of causes, MIT Press.
a switch is electrically wired to either a light that shines or a
bell that rings. Suppose you want to explain the behavior of this
system–what explains why the light shines or the bell rings? In this
case, there are two main causal factors: the (1) switch that is on/off
and the (2) wire which determines what downstream system the switch is
connected to. These factors are both causes and they interact to produce
the system’s behavior–both need to be in a particular state for the
system to exhibit a given behavior. However, while both causes are
involved they play different causal and explanatory roles … . The wire
is a structuring cause because it shapes, guides, and constrains the
behavior of the system, namely, whether it is the light or bell that
turns on. On the other hand, the switch is a triggering cause because it
controls when the system’s behavior is produced. (section 3)
Dretske’s picture can be applied to the social sciences and can
model the relationships between individuals and social structures. So
here is the core of Ross’s proposal: structures as causes should be
understood as constraints on the available choices individuals
can make.
In social structural explanations, social structure operates as a
“causal constraint” on the behavior of individuals. Social structure
imposes limitations on which options are available to individuals, while
their agency performs the selection. (section 3.3)
The constraints that a structure imposes on individual choices can
be more or less stringent. The constraining force can be so definitive
that it necessarily determines what individuals will do. More often than
not, however, constraining causes will have an effect on the
probabilities of possible outcomes, choices, and decisions.7 It is instructive to go back to the examples in
Haslanger’s paper on social structural explanation and see if they can
all be modeled using structuring causes. Can they? What about the
feedback loop in the Invisible Foot example?
Social structure can operate as a strong constraint in the sense that
it makes some outcomes much easier, more favorable, or more rational
than others. Consider an individual with limited resources (financial,
time, transportation, etc.) that has to walk either 1, 3, or 5 miles to
the nearest grocery store to purchase fresh produce. As the distance of
the store increases–and other resources diminish–it becomes much more
difficult to make the “healthy” decision, despite the fact that it is
still “possible” in some sense. (section 3.3)
Malinsky on intervening on structures
One question in Ross’s account is left open. If structures can
sometimes act as causes by playing the role of constraints, how do they
fit into the manipulability account of causation? As noted earlier,
structures can meet the counterfactual test, but how can they be
intervened on or manipulated? Usually, single variables are manipulated:
they are turned on or off. But structures do not seem to be just single
variables. What would a manipulation or an intervention on a structure
look like?
Here is a simple structural equation model:
\[Y = \theta_1 X_1 + \theta_2
X_2\]
The model says that variables \(X_1\) and \(X_2\) are (potential) direct causes of the
outcome variable \(Y\).8 Can you think of an example? The strength of the
causal dependency is given by the coefficient \(\theta_1\) and \(\theta_2\). Suppose that \(\theta_1=.4\) and \(\theta_2=-1.3\). The vector \((\theta_1, \theta_2)=(.4, -1.3)\) gives
information about the causal structure. So, intervening on the
structure means—at least in this example—to change the values assigned
to the parameters \(\theta_1\) and
\(\theta_2\), say to \((0, .87)\).
Most abstractly, an intervention on the structure … is a setting of
the functions and parameters which characterize the model to a new set
of functions and a new set of parameter values (p. 2301)
Once the intervention is carried out, the counterfactual (causal)
question can be asked: by manipulating the structure (that is,
manipulating the values of the parameters), how would the value of the
variables change as a result? This is rather abstract, but can be
applied to concrete examples:9 It can be helpful to work through the healthcare example
on pp. 2303.
… “patriarchy,” “capitalism,” and other macro-structural features of
a system may be identified with sets of parameters (and/or functions) in
some model and we can ask, for example, “what would be different about
the distribution of wealth if there were no gendered wage gap (no causal
dependence of salary on gender), and if hiring/promotions were
gender-blind?” (p.2304)
An interesting (and perhaps revealing) limitation of this approach is
that structures with feedback loops cannot be modeled as causes:10 So, looks like the Invisible Foot example by Haslanger
cannot be modeled as a cause. Do you think this is a serious limitation
of Malinsky’s account? How does this observation apply to the causal
role of race and gender? Does modeling their causal role require
postulating feedback loops?
Unfortunately, the story is not so simple for models which exhibit
causal feedback … It is typically assumed that the system is measured at
some kind of equilibrium … Roughly speaking, a model has a well-defined
equilibrium distribution only so long as the functions do not “blow up,”
i.e., create an unstable feedback process. Consequently, some
interventions on sets of structural parameters or functions make
counterfactual prediction impossible … Parameters in such dynamic models
must satisfy certain mathematical constraints in order to be
well-behaved in a statistical sense; otherwise the system does not have
a stable distribution. Thus, in considering candidate interventions on
structure in models with feedback or dynamical processes, we can only
investigate counterfactual settings which respect these mathematical
constraints, or else we are setting ourselves up to consider
counterfactual predictions for an unpredictable system.
(pp. 2307-2307)
Coda: Race and Gender
One question the readings for today do not directly address is, how
does the idea of structures as causes help to understand the causal role
of race and gender? Here are a couple of tentative responses. Following
Ross, perhaps race and gender can be modeled as structures that
constrain the choices of individuals. Following Malinsky, perhaps race
and gender can be modeled as the parameters in a given causal causal
structure.