*6.3. Mediation*

A mediator is a variable that lies between other variables within a direct causal pathway to an outcome variable. A directed acyclic graph (DAG) may be used to visually represent direct causal pathways between variables [73]. Acyclic means that a variable's causal pathway does not cycle back directly onto itself. Figure 3 is a DAG that shows a simple mediator causation pathway, based on Baron and Kenny, 1986 [74]. Note that confounders and effect modifiers lie outside the causation pathway in this model. For example, the independent variable may be replaced with an associated independent variable, a confounder, which causes the same effect on the outcome variable. An effect modifier may also change the outcome variable at the end of the causation pathway, as in the modifying effect of age and gender in the association of a risk factor with a disease. When inferring causation during a synthesis, possessing expert knowledge of the subject matter under investigation enables the identification of potential confounding factors [75], effect modifiers, and mediators. Research designs that include participant randomization and stratification of results can also assist in controlling the effects of confounders and effect modifiers.

**Figure 3.** Mediation causal pathway. Causal path C runs from the independent variable to the outcome variable. Causal paths A and B run to and from the mediator, respectively, linking the independent and outcome variables. The absence of the mediator weakens path C, and if the path disappears altogether, the variables are linked indirectly through the mediator.

As causal diagrams have developed, indirect links between variables may be represented by a dotted line, and double-headed solid arrows may represent linked variables with an unspecified common cause [76]. I have also combined double-headed arrows with a dotted line ( ) to represent variables linked indirectly with an unspecified common cause. When conducting a synthesis, the researcher may infer the mediating common cause that indirectly links two variables. To illustrate, low vitamin D levels in patients have been associated with a higher risk of cancer incidence [77]. Based on this association, some researchers have proposed that taking vitamin D supplements may prevent cancer, but recently published clinical trials of vitamin D supplements and cancer prevention do not support this causal inference [78–80]. Having coauthored a textbook chapter on the endocrine regulation of phosphate homeostasis [81], I have background knowledge of vitamin D's role in regulating intestinal absorption of dietary phosphorus—i.e., vitamin D levels are lowered if phosphorus serum levels rise too high, as in clinical and subclinical hyperphosphatemia. Synthesizing the link between lowered vitamin D and hyperphosphatemia with the link between hyperphosphatemia and tumorigenesis [61], I proposed that hyperphosphatemia is a common cause that mediates an indirect association between lowered vitamin D levels and increased cancer risk [69].

When selecting information during knowledge synthesis, conflicting material helps identify areas requiring further in-depth investigation. As demonstrated in the above example of vitamin D supplementation and cancer prevention, there may be additional factors that are missing which thwart the synthesis of a truer overall picture. To illustrate, in the allegory of six blind-men and the elephant, each blind man examined a di fferent part of the elephant by touch: the tail, trunk, tusk, ear, leg, and side, and each man inferred a di fferent description of the nature of an elephant as being like a rope, snake, spear, fan, tree, and wall, respectively. Although their tactile observations were accurate, the men were unable to discover the true overall nature of an elephant because they did not synthesize their findings into new knowledge.

Mediation is also used in literature-based discovery, a synthesis method in which implicit knowledge is discovered from linking together separate bodies of literature [82]. For example, if concept A is related to concept B in one body of literature, and a separate body of literature relates the same concept B to concept C, transitive inference relates A to C, as shown in Figure 4. In this example, B acts as a potential mediator that causatively links two separate bodies of literature in a novel way to infer new knowledge.

**Figure 4.** Transitive inference. A causes B in current knowledge domain 1, and B causes C in current knowledge domain 2. New knowledge is discovered when synthesizing domains 1 and 2 through transitive inference; i.e., A causes C, with B acting as a potential mediator.

I used transitive inference to propose an explanatory theory of how cholesterol oxidation products (COPs) are causatively linked to atherosclerosis [83]. I synthesized concepts from one body of research relating COPs (A) to defects in arterial cell membranes (B) with another body of research relating defects in arterial cell membranes (B) to atherosclerosis (C). In this case, defects in arterial cell membranes (B) acted as a mediator that linked COPs with atherosclerosis. This synthesis helped fill in some theoretical knowledge gaps in the potential cause and mechanism of atherosclerosis and strengthened the evidence for dietary prevention of atherosclerosis by avoiding COPs in thermally treated and processed animal-based foods that contain cholesterol.
