Research on causal mechanisms is a growth industry, with the largest percentage of studies falling into the camp of qualitative research. The reason for this is the admonition that correlation is not causation, implying the claim that valid causal inference requires evidence on causal mechanisms. While the need for causal mechanisms is occasionally contested, there is an increasing consensus that mechanisms have their place in empirical research and causal inference.
Although there is no commonly held understanding of what a mechanism is, there is at least broad agreement that mechanisms ensure productive continuity between a macro cause and a macro outcome, which implies that mechanisms are located on a lower level of analysis. Thus, the analysis of mechanisms is considered to be the key feature of qualitative case studies and process tracing. Quantitative research, in contrast, is focused on the estimation of causal effects on the macro level.
To a degree, a recent publication by Imai, Keele, Tingley, and Yamamoto (Imai, Kosuke, Luke Keele, Dustin Tingley and Teppei Yamamoto (2011): Unpacking the Black Box of Causality: Learning About Causal Mechanisms from Experimental and Observational Studies. American Political Science Review 105 (4): 765-789) addresses this division of labor between qualitative and quantitative research. Imai et al. develop further mediation analysis in order to allow for the search for causal mechanisms in quantitative research. The following figure illustrates the most basic setting of mediation analysis. A treatment (cause) T works through a mechanism (M), equivalent to the mediator, which is, in turn, tied to the outcome (O). Moreover, the treatment has a direct effect on the outcome. Mediation analysis allows one to estimate both the direct and indirect effects, and the effect of the treatment on the mechanism and of the mechanism on the outcome.
At first glance, a qualitative researcher might be skeptical that mediation analysis is really about mechanisms. I suppose that one counterargument would be that mediation analysis is still about the estimation of effects and relies on what we currently call data set observations. This contrasts with process tracing, which is closely tied to the idea of causal process observations and the gathering of a rich body of diverse pieces of evidence. Thus, the warning that correlation is not causation would apply in mediation analysis as well, whereas it would now pertain to the link between the treatment and the mediator, and the mediator and the outcome. Moreover, one might believe that the distinction between an indirect and a direct effect does not resemble anything we know from case studies, prompting the conclusion that the quantitative understanding of mechanisms does not overlap its qualitative counterpart.
However, this would all too easily dismiss mediation analysis. Two issues deserve closer scrutiny. First, reference to a ‘direct effect’ does not imply that the effect of the treatment is indeed direct, that is, it functions without a mechanism. The blackboxing of the direct effect simply denotes that one is only interested in the mechanisms that one models via the mediator. Thus, the direct effect also works via mechanisms that are not, however, theoretically relevant. Second, Imai et al.’s definition of mechanisms is similar to some of the definitions found in the qualitative literature. Moreover, the two empirical examples that they present include a mechanism that is located on a lower level than the treatment and the outcome, therefore meeting the common understanding in this respect. Third, one might reject the notion of a mechanism having an effect because the concept of a mechanism often serves as an antipode to the idea of an effect. However, in philosophy of science, one also finds some taking the position that mechanisms have effects (Pearl, Judea (2009): Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge: Cambridge University Press.), meaning that Imai et al. are not exploring entirely new ground here. The reason that we might resist the idea of a mechanism having a certain effect is that we usually do not have the data to scrutinize the mechanism via a large-n analysis, requiring it to perform process tracing. Seen from the other side, however, this would mean that with sufficient data, one could estimate the effect attached to a mechanism (as Imai et al. do). Seen in this light, the link between the concepts of mechanisms and effects might be closer than is apparent at first glance.
Even if one were to accept these arguments, there remains the criticism that a significant effect of a mechanism does not resemble causation because the treatment might not be connected to the mediator, and/or the mediator may not relate to the outcome. Thus, even if one accepts that mechanisms can be analyzed quantitatively, there is, of course, still room for process tracing as one needs to discern productive continuity between treatment, mediator, and outcome (which might give rise to problems of infinite regress). But I presume that qualitative researchers see more problems with the quantitative take on mechanisms, among other reasons because of the neat distinction between effects on the one hand and mechanisms on the other. As Imai’s work is rather new, I expect a great many interesting discussions to be ahead.