Path Analysis: a Critical Evaluation Using Long-term Experimental Data

-We evaluated the ability of path analysis to characterize direct and indirect interactions in a natural ecological system. Based on previous experimental work, we constructed a path diagram reflecting presumed relationships between kangaroo rats (Dipodomys spp.) and two other small rodent species. The technique was applied to examine the direct effects of competition and indirect effects mediated through vegetation. Path analysis gave varying and sometimes uninterpretable results when applied to data collected from unmanipulated and manipulated systems. The variation most likely can be attributed to some combination of inadequate specification of the path diagram, differences between unmanipulated and manipulated systems, varying responses to experiments of different duration, and natural temporal variation. Because these issues are applicable to many kinds of complex systems, we suggest caution before applying and when interpreting the results of path analysis. A major problem of contemporary science is to understand the structure and dynamics of complex systems. Many physical, biological, and social systems are composed of multiple interacting components. Examples include chemical reactions involving multiple reagents, neural networks (both real and artificial), and animal and human societies. Of particular interest to this audience are interactions among species in ecological communities, among genes in developmental pathways, and among traits in determining fitness. In such complex systems, it can be difficult to isolate causes and effects because each component potentially can influence many others through a network of direct and indirect interactions. Path analysis is one technique that increasingly has been used to quantify causal pathways in networks of interactions. It was originally developed to deal with variables that are correlated because of "a complex of interacting uncontrollable, and often obscure causes" (Wright 1921, p. 557; see also Wright 1934, 1960; Li 1975; Pedhazur 1982). Basically, the technique decomposes the overall correlation between two variables into the direct effects of one on the other, indirect effects mediated by other variables, and spurious effects due to common causes. The computed path coefficients indicate the amount of change expected in the *Author order was determined by a coin toss. Smith is the author to whom correspondence should be addressed; E-mail: [email protected]. ?Present address: Department of Biology, California State University, Northridge, California