Direct and Indirect Effects

• Activity Number: 79
• Type: Invited
• Date/Time: Monday, August 8, 2005 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Judea Pearl*+
• Companies: University of California, Los Angeles
• Keywords: causal inference

Abstract

The direct effect of one variable on another usually is defined (and measured) as the change produced in the latter variable after holding constant all intermediate variables between the two. Indirect effects present conceptual and practical difficulties (in nonlinear models) because they cannot be isolated by holding certain variables constant, and cannot therefore be defined---let alone identified---in experimental studies. This talk formalizes an alternative, more general definition of partial effects transmitted through restricted sets of paths. The new definition, expressed mathematically in the language of nested counterfactuals, permits the assessment of direct and indirect effects in both linear and nonlinear models,and provides policy-related interpretations for these effects. The talk establishes formal conditions under which direct and indirect effects can be estimated consistently from experimental and nonexperimental studies, thus extending effect-decomposition to nonlinear and nonparametric models.