Abstract
The idea that mechanisms are crucially important to differentiate between genuine and spurious causal relations is ubiquitous both in the philosophical and in the social scientific literature. Yet philosophers of the social sciences have seldom attempted to spell out systematically the way in which mechanistic reasoning or evidence are concretely used to deal with spurious association and the problem of confounders in the social sciences. In this paper, we analyze two recent such accounts, proposed by Harold Kincaid and Daniel Steel. We show how these two accounts radically differ in their notion of mechanism (a process account, and a complex system account, respectively), and how this ultimately impacts in the way in which they understand the inferential role of mechanisms in the social sciences. We then confront both accounts with the details of a well-known controversy around the purportedly causal association between the legalization of abortion and the subsequent fall in criminality in the United States. We show the limitations of both accounts in representing accurately the role of mechanistic evidence and hypotheses in practice.
1. Introduction
Although ubiquitous in the social sciences and often difficult to solve, the problem of potential confounders is easily defined: we can say that X and Y are confounded when we have measured their association without regard to a third variable, Z, that influences both X and Y. In such cases, Z is called a confounder of X and Y, because it confounds, or biases, our reading of the association between X and Y. In the most extreme case, if Z (being a common cause to both X and Y) is fully responsible for the association between these variables, then we can say the association between them is not only biased, but spurious, and therefore, not causal.
The Problem of potential confounder
Many of the important philosophers and social scientists defending the centrality of mechanisms in the social sciences have argued that searching for mechanisms is a good strategy to confront the problem of potential confounders.
1
The argument is that knowledge of mechanisms connecting social phenomena improves our capacity to distinguish spurious associations from correlations that represent genuine causal relations. The following quotes are clear examples: The search for generative mechanisms consequently helps us distinguish between genuine causality and coincidental association. (Hedström and Swedberg 1999, 9) [T]o identify a causal relation between two kinds of events or conditions, we need to identify the typical causal mechanisms through which the first kind brings about the second. (Little 1995, 34) The relevance of mechanisms is not limited to explanation. Especially in non-experimental contexts, mechanisms often have a crucial role to play in distinguishing true causal relations from spurious correlations. Mechanisms help in causal inference in two ways. The knowledge that there is a mechanism through which X influences Y supports the inference that X is a cause of Y. In addition, the absence of a plausible mechanism linking X to Y gives us a good reason to be suspicious of the relation being a causal one. (Hedström and Ylikoski 2010, 54)
While statements like these abound, there have been, to date, few attempts to systematically account for the way in which mechanisms concretely contribute to solving the problem of potential confounds. In the pages that follow, we try to contribute to the understanding of the role of mechanisms in the context of social scientific causal inference, by tracing and contrasting two different accounts, proposed by Daniel Steel (2004b, 2007, 2011) and Harold Kincaid (2011, 2012), of how mechanisms help us confronting confounding situations.
In a first part of the paper, we introduce Steel’s and Kincaid’s accounts, showing that they are built on different understandings of what mechanisms are in the social sciences. In the second part, we put both accounts to the test of social scientific practice, by analyzing a well-known case study: the controversy regarding the causal status of the association between the rise in abortion and the decrease in crime in the United States at the end of the 20th century, as first proposed by Donohue and Levitt’s (2001) paper, and analyzed extensively by Daniel Steel (2013). We conclude by showing the problems that both accounts faced when contrasted with the actual scientific practice.
Before we proceed, two considerations are in order. In the comments quoted before, Kincaid and Steel have found two relevant distinctions. First, while some authors seem to be committed to the strong claim according to which mechanisms are necessary to distinguish true causal relations from spurious correlations, others restrict their defense to a weak claim according to which mechanisms can be useful for this task, leaving open the possibility for other methodological tools to do the work (Kincaid 2012). Second, the idea that mechanisms have a role to play in dealing with the problem of potential confounders has a positive and negative side (Steel 2004b, 2007): on one hand, one could use positive evidence in favor of the existence of a mechanism connecting two social phenomena, to confirm a causal relation between them, and on the other, one could use the impossibility to conceive any plausible mechanism connecting two social phenomena, to conclude that a correlation between the variables that represent them is spurious.
Kincaid 2 has argued that mechanisms are not necessary to deal with the problem of potential confounders. According to the author, identifying and controlling potential common causes, as well as randomizing, are alternative methodological strategies to distinguish true causal relations from spurious correlations, that do not presuppose knowledge about mechanisms, and that are also at hand in social research. In turn, Daniel Steel 3 shows a pragmatic limitation of the negative side of the idea that mechanisms can help us to confront potential confounding situations. The author argues that, in the social sciences, there are too few situations where we are not capable of conceiving a mechanism connecting to social phenomena. To the contrary, the more common problem faced by social scientists is the overabundance of potential mechanisms.
Given these distinctions, here, we focus on a more restricted claim: that positive evidence in favor of a mechanism can be useful to deal with the problem of potential confounders.
2. Mechanisms and the Problem of Potential Confounders
2.1. Two Notions of Mechanism
As we have mentioned, underlying Steel’s and Kincaid’s accounts about the role of mechanisms in the problem of potential confounders, it is possible to identify two different notions of mechanism. According to the first notion, mechanisms are complex systems: sets of interactive components that are organized in a way that they produce a causal regularity. 4 In the second notion, mechanisms are considered to be causal processes: sequences of intervenient causal steps that connect a putative cause with its effect. 5
According to Glennan, a toilet is an example of a complex system mechanism: the structured interaction of its parts—the float boat, the fill valve, the flapper, and so on—produce a causal regularity—pushing the handle is followed by the flushing of the water. And the chain of events that lead to the breaking of a window is an example of a causal process mechanism: in an improvised baseball field, a pitcher throws a ball to the home plate, then the hitter hits the ball changing its trajectory, and the ball ends up impacting the window (Glennan 2009, 291).
Steel’s account is based on a complex system conception of mechanisms. The author defines social mechanisms as “complexes of interacting individuals, usually classified into specific social categories that generate causal relationships between aggregate-level variables. A mechanism will be said to be from the variable X to the variable Y if it is a mechanism through which X influences Y” (Steel 2004b, 59). As an illustration of social scientific mechanism, Steel (2004b, 2007) develops an example from Malinowski’s research on Trobianders’ polygamy (1935) and analyzes one of its claims qua causal hypothesis: the wealth of Trobriand chiefs is causally explained by the number of wives they have. The mechanism that connects the number of wives with wealth is composed of the individual Trobrianders—categorized in social roles like chief, wife, brother-in-law, and so on, interacting according to two social practices. First, it is a custom that men contribute with gift of yams to the household of their married sister, where the gifts are larger when the sister marries a chief. Second, the yams are used by Trobrianders as the primary source for financing political endeavors and public projects.
In turn, Kincaid’s account is based on a causal process conception of mechanisms. The author distinguishes between two kinds of social mechanisms, vertical and horizontal, and argues that only the latter have a role to play in dealing with the problem of potential confounders. According to Kincaid (2012, 50), vertical mechanisms are the underlying component parts that “give C its capacities to cause E and explain the changes in E.” An example of vertical mechanism are the budget constraints and utility functions of individuals that give the aggregate demand its capacity to produce changes on prices.
On the other hand, horizontal mechanisms are the intervening causes that make for a continuous process, that is, the mechanisms that we would want to know about when “asking for the steps that led from C to E.” Kincaid (2012, 49; 2014, 145) refers to them as “horizontal” because they are mechanisms at the same level as what the mechanisms relate. Thus, for example, when we explain how changes in interest rates cause changes in employment, mediated by changes in aggregate investment, we are referring to a horizontal mechanism.
2.2. How Do Mechanisms Help Us Confront the Problem of Potential Confounders?
2.2.1. Steel’s account
Steel (2004b, 66) argues that the main challenge in further developing the idea that mechanisms are useful to confront potential confounding situations is that it demands an explanation of how it is possible to learn about mechanisms “in a way that does not run directly into the problem of confounders, which is the problem that mechanisms are supposed to help us overcome.”
To understand Steel’s answer to this challenge, let us start by supposing we are studying a social system S, such as an indigenous society (as in the Trobriand example), a transnational entity, such as the European Union, or a domain-specific organization, such as a country’s economy. According to Steel, in every social system, it is possible to distinguish two different sets: a set
According to Steel, if we want to learn the causal relations among the macrofeatures of our social system, we could follow one of two strategies:
We could directly infer the causal relations based on the marginal and conditional associations among the variables in
We could indirectly infer the causal relations finding the configurations of the components in
In Steel’s account, mechanisms are useful to confront confounding situations at the level of the macrofeatures of the system, in cases where indirect causal inference is more easily attainable than direct causal inference. Although the author does not offer an explicit list of conditions where this occurs, it is possible to identify in his work three cases where we are in a better position to confront the problem of potential confounders at the level of the components than at the level of the macrofeatures of the system. 6
Regarding the first two cases, Steel (2007, 188-89) says, One way this could be is if it is possible to perform experiments on the components, but not the system as a whole. For example, experimental economists can perform randomized experiments involving individuals or small groups but not entire economies. Similarly, ethical considerations prohibit an experiment in which persons are exposed to aflatoxin B1, yet it is possible to experimentally study, say, the metabolism of that compound in vitro by means of cell cultures. Even when experiments cannot be performed on the component parts of the system, there may be better observational data with regard to the relevant features of the components than for the system as a whole. Or it may be that the possible confounders have been more exhaustively listed and measured with regard to the components than for the macrofeatures of the system. In short, there may be a variety of practical reasons why the causal relationships among the variables in C can be more directly ascertained than among those in V.
Thus, the first case describes situations where we can experiment with the components of the system but not with the system as a whole, and the second case refers to situations where it is easier to identify and control the potential confounders at the level of the components than at the level of the macrofeatures.
The third case takes place when indirect causal inference identifies mechanisms that are built on simple generalizations about human cognitive and psychological tendencies for which we do not take seriously the possibility of an unmeasured confounder. Steel argues that, in this case, once we have identified the relevant elements of the mechanisms—the components and its configuration—the causal relation between the macrofeatures connected by the mechanism often follows straightforwardly.
Malinowski’s example of Trobiand polygamy would, according to Steel, illustrate this third case. Recall that the mechanism proposed by Malinowski is composed by the individual Trobianders, categorized in relevant social roles, acting according to two social practices: men must contribute with a gift of yams to the household of their married sister, and yams are used as a source for financing political endeavors and public projects.
According to Steel (2007, 189), this mechanism carries, or is also composed of, a series of causal generalizations concerning “human aspiration for wealth and social status.” 7 Steel considers that these type of generalizations can be plausibly regarded as a part of one’s background causal knowledge, and thus, they can be considered to be guarded against the problem of potential confounders. Therefore, once we have identified the components of the system relevant for the mechanism—the individual Trobrianders described by their social roles- and its configuration—the social practices that rule their behavior—we can safely infer that the number of wives has a positive effect on the wealth among the trobrianders. 8
In sum, for Steel, mechanisms can help us confront the problem of potential confounders in situations where the level of the components of a social system display a form of privileged epistemic access to this problem, relative to the level of macrofeatures. The author further argues that these situations occur when we are in better position to experiment or control at the level of the components of a system than at the level of its macrofeatures, and when researchers have a greater background knowledge at the level of the components.
2.2.2. Kincaid’s account
In Kincaid’s (2012, 52-53) account, mechanisms can help us confront confounding situations by way of making the tests of potential confounders more stringent. The argument goes as follows:
Suppose that we find a correlation between two variables X and Y, and we suspect that the correlation could be the spurious effect of a confounder Z, not identified in the data, which is the common cause of X and Y.
Suppose, also, that we later find correlational evidence of a mechanism M mediating between X and Y. 9
What does the mechanism do for our suspicions of an unmeasured confounder? It demands for any potential confounder Z to do more work in order to show that the X-Y correlation is spurious. If the X-Y is actually a spurious correlation produced by a common cause Z, and we have evidence of a mechanism M mediating between X and Y, then Z has to explain, not only the correlations between X and Y, but also the correlations between X and M, and between M and Y. This means that conditioning on the potential confounder must show that X, M, and Y are independent of each other.
Naturally, the number of correlations that a potential confounder would have to explain will increase with the number of intervening causes we add to the model. Thus, the more complex the mechanism—in terms of the number of steps that connect the cause with the effect—the more stringent the test of potential confounders will be.
2.3. Complex Systems, Causal Processes, and Reduction
As we saw, Steel’s and Kincaid’s accounts presuppose different definitions of mechanisms and, therefore, develop distinct explanations of the role of mechanisms in the problem of potential confounders. The main difference between both accounts is related to where mechanisms are placed in relation to the potentially confounded correlation: Steel’s account distinguishes between the level of the macrofeatures of a social system, where the potentially confounded associations are found, and the level of its components, where mechanisms lie. In contrast, in Kincaid’s account, mechanisms are labeled as “horizontal” because they are placed at the same level of the potentially confounded correlation. In this sense, it is possible to define Steel’s and Kincaid’s accounts, respectively, as reductive and nonreductive methodological strategies to face the problem of potential confounders.
The reductive and nonreductive character of these accounts raises different philosophical challenges. On one hand, Steel needs to show that his proposal is not undermined by the common argument posed against reductionism: the possibility that a generalization at the level of the macrofeatures of a system could be realized by multiple different mechanisms at the level of the components. On the other hand, Kincaid needs to justify the possibility of causal relations at the level of social aggregates.
Although our main task is to critically evaluate the extent to which Steel’s and Kincaid’s accounts can be used as frameworks to understand actual scientific practice, a brief introduction to how the authors answer to this challenge will give us a better understanding of the differences between both accounts.
Steel (2004a, 2007) offers some insights about the possibilities and limitations of reductionism that can be put in tandem with his account of the role of mechanisms in the problem of potential confounders. The author defines reduction as an explanatory strategy that can be pursued to achieve different goals. Thus, for example, reduction can be used for unification purposes: to show that a number of different generalizations at the level of the macrofeatures can be explained by a lesser number of mechanisms. But it also can be pursued to achieve decomposition: to demonstrate that a mechanism, shared by a set of systems, is sufficient to explain a generalization displayed by those systems.
According to Steel, when reduction is used for the purpose of unification, the fact that a generalization at the level of the macrofeatures can be multiple, realizable by different mechanisms at the level of the components, is a real problem for reduction. But if reduction is put at the service of decomposition, then multiple realizability is not a challenge: the fact that a generalization is realized by a particular mechanism is compatible with the possibility that the generalization could have been produced by a different mechanism.
Steel’s account can be interpreted as a decomposition-seeking reduction: mechanisms can be used to rule out a potential confounder hypothesis because they show us that the regularity displayed by a social system is actually realized by a particular mechanism. This way, Steel can follow a reductive strategy to solve the problem of potential confounders without facing the problem of multiple realizability.
For his part, Kincaid (1990, 1996, 2009) raises several arguments in favor of causation at the social aggregate level. According to the author, the fact that social entities, and its causal capacities, are composed of, and supervene on, individuals (as described by the vertical arrows in Figure 2), does not rule out causation at social aggregate level (the horizontal arrows in Figures 1 and 2). Kincaid argues that social level causal claims can identify causal patterns that are not identifiable at the individual level. This is the case, because, although the causal capacities of any particular social entity are token identical to the sum of individuals that compose it, social entities are multiple realizable at the individual level, and, therefore, not definable in individualist terms.
Vertical mechanisms (Kincaid 2012, 49).
Horizontal mechanisms (Kincaid 2012, 49).
Kincaid, also argues that the relations between social aggregates can be described using the common markers of causation. The author distinguishes between three kinds of social aggregate: logical aggregates, synthetic aggregates, and complex wholes. Logical aggregates are simple sums or averages of individual characteristics, like the totality of employed individuals in a city. Synthetic aggregates are more complex constructions like the GDP or the interest rate of a country. Finally, complex wholes are social entities with internal structure that exist in space and time in a way that logical and synthetic aggregates do not, for example, a company or a political party. According to Kincaid, these three kinds of social aggregates can be candidates for causal relata of relations described by regularity, counterfactual, or probabilistic markers of causation.
3. The Test of Social Scientific Praxis
We are now in a position to contrast Kincaid’s and Steel’s accounts regarding the role of mechanisms in causal inference. In what follows, we use a case study to put each of the accounts presented under the scrutiny of current social scientific representative practices (and arguably, and broadly conceived, best ones). As we will see, this exercise shows some important limitations in both accounts.
3.1. The Abortion-Crime Controversy
As a case study, we use the abortion-crime controversy spurred by John Donohue and Steven Levitt’s 2001 paper, “The impact of legalized abortion on crime.” This paper offers several advantages as a case study: first, it is a well-known case of a controversial causal statement in the social sciences. Second, it has been studied extensively by Steel (2013), so that we can directly know how Steel thinks that it can support his thesis regarding the problem of confounders, thus avoiding problems of incorrectly interpreting his account when applied to a specific case. Third, the central bulk of evidence supporting the claim directly corresponds to Kincaid’s straightforward approach to the problem in terms of aggregate level variables.
In 2001, Donohue and Levitt argued that the increase in the number of abortions in the United States following the 1973 Roe vs. Wade decision was the main cause of the crime reduction during the 1990s. The context to the claim was one in which, contrary to the predictions of politicians and experts, the United States had just experienced a marked decrease in its criminality at the end of the 20th century, where homicide rates decreased by more than 40%, while property and violent crimes decreased by roughly 30% (Donohue and Levitt 2001, 379). While several causal explanations of the phenomenon were on offer (such as an improvement in police strategies, economic growth, the increase in the number of incarcerations, etc.), Donohue and Levitt’s paper, with its econometric display and Levitt’s impressive pedigree as an economist, was definitely among the better known and controversial ones.
Donohue and Levitt’s paper offers several types of aggregate evidence in favor of their hypotheses. For example, they show how the beginning of the crime reduction trend coincides with the time when the first cohort born in the context of legalized abortion reaches the age of greatest propensity to commit crimes (18-24 years) and that the states that legalized abortion before the Roe vs. Wade decision (Alaska, California, Hawaii, New York, and Washington) experienced a decrease in their crime rates before the rest of the states. In their statistical analysis, an increase in the effective abortion rate of 100 × 1000 births is associated with a reduction of 12% in murders, 13% in violent crimes, and 9% in property crimes, net of controlling for the effect of variables such as levels of poverty, alcohol consumption, and numbers in police forces (among others).
The authors also propose the mechanism that would explain how abortion legalization lowers criminality. The mechanism can be divided in three parts:
the abortion legalization has a selective effect in the population: both women who are less able to raise a child (because of a disadvantageous social, economic, and family environment) and women who are less willing to provide a nurturing home (because of an unwanted pregnancy) are the groups of the population that experience the highest increase in abortion after legalization.
Being raised in a disadvantaged environment (poorer neighborhood and school, teenage and single mothers) increases the probability of becoming a criminal in adulthood.
Unwanted pregnancies are more likely to translate into unfavorable parental behavior (maternal rejection, erratic/harsh practices) that also increases the probability of delinquency (Donohue and Levitt 2001: 388).
The abortion-crime mechanism
As part of their evidential arsenal, Donohue and Levitt offer evidence in favor of each of the parts of the mechanism:
With respect to the selective effect of the abortion legalization, they present evidence whereby the increase in abortion “was roughly twice as great for teenagers and non-white mothers as it was for non-teen and white population” (Donohue and Levitt 2001, 387). 10
In regard to the relation between social disadvantage and crime, Levitt and Donohue report, for example, data according to which prison inmates in the United States have a significantly higher probability (than the average population) of having been raised by single parents, by parents who had substance abuse problems, or by parents who were, themselves, abusive (Donohue and Levitt 2001, 388-89).
Finally, they also present evidence of the association between being an unwanted child and engaging in criminal activity later in life, notably, a set of matching or quasi experimental studies following the trajectories of children raised by mothers who requested, but were legally denied, access to an abortion. The studies, carried out in several European countries, followed the lives of these families and matched them to other families of similar socioeconomic characteristics. Their results show that when these (unwanted) children reached adulthood, they were disproportionally more likely to participate in criminal activities that other children raised in similar socioeconomic backgrounds.
3.2. Steel’s Account: Mechanisms as Complex Systems and the Evidential Accessibility of Their Component Parts
Recall that, in our interpretation of Steel’s account, mechanisms help with the problem of confounders under three circumstances. According to Steel, in these three cases, the problem of potential confounders is more treatable at the level of the components than at the level of the macrofeatures of a system. The first two cases describe, respectively, situations where experimenting and controlling for potential confounders is more feasible at the level of the components than at the level of the macrofeatures. The third case refers to situations where the mechanisms at the level of the components are built on causal generalizations that can be considered exempt from the problem of potential confounders. Here, for reasons that have to do with the applicability to our case study, we leave aside the third case and focus on the first two. Our conclusion, as will become clear through the analysis, applies to all three cases.
3.2.1. Experimenting at the level of components in the abortion-crime controversy
Steel considers the abortion-crime hypothesis an example of the case where we are not able to manipulate the macrofeatures of the system, but we can, in fact, experiment with its components. In Steel’s categories, the rates of effective abortion and criminality (the putative cause and effect) are the macrofeatures of the system, while the components of the system are constituted by individuals of two types: first, pregnant women who, if given the legal possibility, could decide whether or not to terminate their pregnancies, and second, children categorized by their socioeconomic and parental conditions.
While one cannot, rather obviously, experimentally manipulate the rate of abortions to observe the effect in the rate of criminality, Donohue and Levitt do use quasi-experimental evidence regarding the effect of being raised as an unwanted child on the propensity of becoming a criminal later in life. This quasi-experimental evidence about the causal relation between unwantedness and criminality, in conjunction with some assumptions about the effect of legalization on the effective number of abortions, would help us gain access (to put it in purposely ambiguous terms) to the causal relation between abortion and crime.
This kind of argumentation would in fact seem satisfactory if the causal connection between abortion and crime were to be used in a how-possible explanation. The quasi-experimental evidence just quoted could be relevant if our aim is to persuade someone of the abortion-crime connection who wonders how it is possible that two seemingly distant phenomena (legal reform on reproductive rights in the 70s and crime rates in the 90s) are connected causally. But let us be reminded that Steel’s attempt (and our focus here) is that of accounting for how mechanisms can help in the discarding of causal confounders. This is a very different business from that of providing plausible explanations, for often, the most plausible explanations are held as true because of their plausibility, even if eventually, confounders that are hidden for some time end up showing their face, and consequently show that prima facie plausibility is often not a good argument if one is interested in causal inference. So, the question that remains is, “Does the type of mechanistic evidence that Levitt and Donohue present help to solve the problem of confounders in the way described by Steel?”
The problem with Steel’s interpretation of Levitt and Donohue’s case is that, contrary to Steel’s account, there is no relevant sense in which the level of the components of the system (the level of individuals) is epistemically more accessible than the level of the macrofeatures of the system. The case illustrates an instance in which the relation between the components and the systemic features is one of simple aggregation: the macrofeatures are nothing but the sum of the components. In this respect, the abortion-crime association is rather representative of a vast proportion of quantitative research in which there are no interesting features at the macro level that are not mirrored at the microlevel.
In this sense, experimenting at the macro level would not so much be unfeasible or unethical, but rather, not even possible from a theoretical point of view (i.e., it is not even possible in a Woodwardian sense in which we can conceive experiments as idealized, hypothetical exercises; Woodward 2003). Because the rate of abortion, as a macro-phenomenon, is nothing more than the aggregation of individual actions, whatever experimental intervention we would wish to perform (in the idealized absence of any constraints) would anyways need to be done at the individual level.
3.2.2. Controlling for potential confounders at the level of the components
Steel’s account presupposes that there are situations where confounders that operate at the level of the components are different to those at the macro-level. As we have just argued, this is probably not a common case in the social science where, more often, and especially when dealing with statistical evidence, confounders at the individual and aggregate level cannot really differ (e.g., the rate of unemployment is rather straightforwardly based on an estimate of the aggregation of unemployed individuals at a given time). Yet, we could still argue that, even if the applicability of the account is thus reduced, the account may still be relevant in those cases in which confounders at the level of the components are in fact different from those at the level of the macrofeatures of the system.
Two problems arise:
First, it is indeed difficult to conceive of systems where possible confounders (to a given causal relation) are different at different levels of aggregation. As we say, the abortion-crime case is not one of those cases, despite Steel’s suggestions to the opposite: the fact that there is quasi-experimental evidence to be found at the individual level is not proof to the contrary and does not support the general notion of confounders (to the same causal claim) being different for different ontological levels.
Suppose, however, that they were indeed different. The problem then would be that no amount of water-tight proof that the purported mechanism at the component level was a genuine one would contribute to increase our confidence in the causal claim at the macro level. If confounders are truly different at different levels of aggregation: how can controlling them at one level help us ensure that they are controlled at other levels? How can we know that the given causal effect we are interested in is the result of the workings of a particular mechanism, and not due, instead, to some other (perhaps unknown) confounder operating at a different level?
Take again the abortion-crime association: the mechanistic evidence at the level of individuals makes the purported mechanism plausible, but it does not provide grounds to rule out that most (or even all) the macro level associations between abortion legalization and criminality are due to some latent confounder. The work by Reyes (2007), also reviewed by Steel, shows that we have good reason to suspect that an important contribution to the crime reduction of the 1990s comes from the change in regulation regarding the lead content of gasoline and other industrial products, where lead exposure during childhood has been linked to a number of behavioral problems and learning difficulties. The fact that, in many states, the end of exposure to lead coincides, in time, with the legalization of abortion is an unfortunate accident. It surely makes life harder for those researchers interested in the determinants of the 1990s crime decline. Understanding how unwanted pregnancies can, in general, lead to a future life of crime has, alas, a very limited role in discerning the historical causal impact of the end of lead exposure.
In sum, Steel’s account needs to be confronted with either one of the following possibilities:
If the confounders that threaten the putative causal relationship between two factors are the same across the macro and the micro level, there seems to be no reason to impute a privileged epistemic access to the micro level.
If the confounders are different in each of the levels, then no amount of assuredness in the genuine character of the micro associations will assure us that some unknown, macro level confounder is the actual cause of the macro association.
3.3. Kincaid’s Account: Mechanisms as Causal Processes and Their Role in the Tests of Potential Confounders
Recall that Kincaid’s account considers that, in qualitative causal claims, the enabling role of mechanisms is that of making the test for potential confounders more stringent, in the sense that if there is reason to suspect that the correlation between our independent and our dependent variable is actually due to a third, uncontrolled factor, mechanisms will provide us additional checks to perform: we will have to test whether, conditional on the suspected confounder, our putative cause and effect are both independent of the mechanism, understood as a mediating (in his terms, horizontal) variable.
Kincaid’s description of the role of mechanisms does not seem, however, relevant to the case in hand. Suppose, for the sake of illustration, that we suspect a potential confounder that is indeed the genuine cause of the abortion-crime negative association: to go further, suppose that in richer societies, there are higher abortion rates and lower crime rates and so the abortion-crime association could be suspected to be explained by GDP growth. According to Kincaid, this income variable would, in the face of a purported mechanism, need to screen off, not only the relationship between abortion and crime but also the relation between abortion and “unwantedness” and the relationship between “unwantedness” and crime. However, the rate of undesired pregnancies or of being “unwanted” is not an observable variable in the population for which Donohue and Levitt are inferring the association between abortion and crime, and therefore, Kincaid’s proposed additional tests cannot be performed. But according to Kincaid, this is the relevant role of mechanisms in helping to adjudicate in problems with potential confounders: short of having a database containing the variable that is supposed to mediate between cause and effect, mechanisms cannot do anything else for you in this regard.
Arguably, this could be considered to represent not so much a problem of Kincaid’s account, as a contingent fact about our particular example. However, it does show that Kincaid’s account limits the usefulness of mechanisms to causal inference to situations in which the available data is of a particular form. Namely, the specific role that Kincaid concedes to mechanism necessitates,
That the mechanistic (horizontal) variable be directly observable for the relevant population and sample.
That the potential confounders are known and measurable in the same population and sample.
And that the available data display enough variation to be able to observe whether, in fact, the potential confounder Z can screen off not only the associations between X and Y but also the associations between X and M, and between M and Y.
The problem, thus, with Kincaid’s account is that it is too restrictive, in the sense that it leaves mechanisms with a too meager role in deciding when causal inference faces the problem of potential confounders. If mechanisms only (or mostly) help in the manner Kincaid describes, then they seldom are of help. And yet, as we see in the abortion-crime controversy, appealing to mechanisms seems to be a crucial part of the establishing of contested causal inferences, even when Kincaid’s restrictive conditions do not hold.
4. Complex Systems, Causal Processes, and the Problem of Confounders
Although important and relevant in their attempts to systematize the role of mechanisms in social scientific inference, both Kincaid and Steel face difficulties when confronted with some of the relevant actual practices of social scientists. The role of mechanisms in helping with causal inferential problems seems to be both ampler and more liberal than these accounts can suggest at times and, possibly, only indirectly related to the problem of confounders.
Kincaid’s account is rooted in a conception of mechanisms understood as causal processes, and yet, his account does not offer much room to the causal tracing of those processes. When helping with the problem of confounders, mechanisms are represented only by a marker in the form of an intervening variable. Absent that variable in the researcher’s database (as it is often the case in the relevant, real cases), mechanisms no longer have a role in helping us distinguish between spurious and genuine causal associations.
Steel’s account instead conceives of mechanisms as complex systems. His ontological commitments are imported from Machamer, Darden, and Craver (2000) definition (devised for the realm of biology, and in particular, in the realm of molecular biology) and adapted to the social sciences, but this seems to come at high cost, as shown in the fact that the details in Steel’s distinction between macrofeatures and components does not sit well with many standard social scientific examples, or at least not with the type of research where the problem of confounders arises as an empirically tractable issue.
While the notion of system, with all of the conceptual challenges that it poses, seems to be adequate to at least some subfields in biology, it has a far less prominent standing in the context of social scientific causal inference. To be sure, beyond its metaphorical value, there are social scientific entities having clear “systemic” features (think of the judiciary system, the system of vouchers to allocate children to different schools, or systems such as social security understood as a public insurance scheme against risk). However, and in the realm of causal inference, these systemic properties are often not relevant to the types of statistical associations that can take place among variables. There is not (other than in some very vaguely metaphorical, and perhaps politically loaded sense) an abortion-crime system, but if there were one, its underlying structure would need to produce a regularity that was stable over time, at least under some conditions. Given the characteristics of this particular case-study, it is unlikely that Steel (2013) would be committed to such an assertion, despite the fact that in his own analysis, he suggests that this case behaves in accordance to his general framework.
When dealing with social scientific causal inference problems, such as the problem of confounders, references to society as a system seem to provide the kind of unnecessary metaphysical baggage that separates our accounts from the practical concerns of the scientists involved: the notion of system can be relevant to describe sets of variables that define entities for which we know well their underlying structure (e.g., a cell, but also such machines as a combustion engine, or a flushing toilet). Yet a system is not the relevant notion when one is concerned with making sure that there are not latent, third variables that could explain the identified associations between two given variables in an open environment in which potentially many other factors can have an effect.
Perhaps the fate of future attempts at bringing out the role of mechanisms in causal inference is linked to the fate of a unified concept of mechanism that can work across the sciences (for a recent, careful proposal, see the deflationary account of Ioannidis and Psillos 2017). Maybe, instead, we may have to admit that different conceptions of mechanisms assist causal inference in different ways, or maybe mechanisms are important in causal inference only as a manifestation of the idea that having evidence that is varied and that comes from multiple independent sources is better (Claveau 2012). Whatever way one goes, our analysis advises that, when dealing with causal inferential problems, our ontological commitments regarding mechanisms be rooted in the practices of researchers, rather than stipulated.
Footnotes
Acknowledgements
The authors would like to acknowledge valuable feedback received during the refereeing process. Additionally, they would like to thank Armando Cíntora and Héctor Cebolla, as well as the audiences in the ENPOSS Roundtable and the Logic Department Seminar at UNED for comments and insights. They would like to also acknowledge CONACyT (México) and the research project MECABIOSOC (FFI2017-89639-P) for institutional support and funding.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
1
Elster (1983 47-48), Little (1991, 24-25), Hedström and Swedberg (1999, 9), 
, 54).
2
Kincaid (2012, 54), , 113).
3
Steel (2007, 197-200), , 64-66).
4
The complex system notion of mechanism is the most common and accepted in contemporary philosophy of social sciences. It is also the notion of mechanism defended by those that have raised (without systematically developing!) the idea that mechanisms help us to distinguish true causal relations from spurious correlations. Nevertheless, as will appear clear in the following pages, this notion is not the only one relevant in the philosophy of social sciences, nor the best suited to account for the role of mechanisms in the problem of potential confounders.
5
We take the complex system/causal process distinction from Stuart Glennan (2002, 2009) Nancy Cartwright (2009, ) also recognizes this distinction, although she uses the notions of “underlying structure” and “nomological machines” instead of “complex system.” In (2014), Cartwright argues that both notions of mechanisms are currently at play in the contemporary philosophical debates about causation in the social sciences.
7
Although the author does not give concrete examples of the causal generalizations underlying the Trobianders mechanisms, it is safe to conclude that he is referring to folk psychology statements like “if Trobiand chiefs want to increase their wealth, and they believe that the best way to do this is to get married, then, they will be more likely to marry more women.” See , 73-74).
8
It is important to note that there is a difference between the first two cases and the third one. In the first two cases, the problem of potential confounders exists both at the level of the macrofeatures of the system and at the level of its components, yet we are in a better position to deal with it -by controlling or experimenting- at the level of the components. In the third case, in contrast, the problem of potential confounders only appears at the level of the macrofeatures of the system. This is not to say that at the level of the components, there are not other problems (e.g., according to , 72], “Malinowski . . . faced the challenge of making an inference about a social practice of which he had no initial inkling from beginning observations of large quantities of yams being moved to and fro”). Yet, these difficulties at the level of the micro components are not the same as the problem of potential confounders, that is, we have enough background knowledge to know that there are no third, unidentified causal variables affecting the relationship between the cause and the effect.
9
It is important to note that correlational evidence is not enough to identify a mechanism. This is the case, because correlations alone do not allow us to distinguish mechanisms (X→M→Y) from common causes (X←Z→Y): in both structures, if we fix the values of M and Z, the correlation between X and Y vanish. Therefore, correlational evidence must be accompanied by background knowledge about the phenomena represented by the variables, that allows us to infer that the correlations we observe are the product of a mechanism rather than a common cause.
10
They also make reference to a study that shows, based on the sociodemographic characteristics of the population groups involved, that the children who were not born as a result of the abortion legalization would have been “60 percent more likely to live in a single parent household, 50 percent more likely to live in poverty, 45 percent more likely to be in a household collecting welfare, and 40 percent more likely to die during the first year of life” (: 387).