Path analysis in genetic epidemiology: A critique

Path analysis, a form of general linear structural equation models, is used in studies of human genetics data to discern genetic, environmental, and cultural factors contributing to familial resemblance. It postulates a set of linear and additive parametric relationships between phenotypes and genetic and cultural variables and then essentially uses the assumption of multivariate normality to estimate and perform tests of hypothesis on parameters. Such an approach has been advocated for the analysis of genetic epidemiological data by D.C. Rao, N. Morton, C.R. Cloninger, L.J. Eaves, and W.E. Nance, among others. This paper reviews and evaluates the formulations, assumptions, methodological procedures, interpretations, and applications of path analysis. To give perspective, we begin with a discussion of path analysis as it occurs in the form of general linear causal models in several disciplines of the social sciences. Several specific path analysis models applied to lipoprotein concentrations, IQ, and twin data are then reviewed to keep the presentation self-contained. The bulk of the critical discussion that follows is directed toward the following four facets of path analysis: (1) coherence of model specification and applicability to data; (2) plausibility of modeling assumptions; (3) interpretability and utility of the model; and (4) validity of statistical and computational procedures. In the concluding section, a brief discussion of the problem of appropriate model selection is presented, followed by a number of suggestions of essentially model-free alternative methods of use in the treatment of complex structured data such as occurs in genetic epidemiology.