Using structural equations to estimate effects of behavioral interventions

Abstract

When evaluating intervention programs designed to influence complex behavioral outcomes, researchers are often faced with confounding background variables that need to be controlled, as well as several intervening or mediating variables such as self‐efficacy or locus of control. However, because intervening variables are assumed to have a causal effect on program outcomes, their impact should be estimated and not controlled. We compare the difference between means using t tests, single‐equation multiple regression, and structural equation models for their ability to characterize accurately the causal pathways considered in behavioral interventions. We find that the mean difference accurately estimates the total effect of the treatment on the outcome of interest but does not allow for decomposition of treatment effects attributable to intervening variables. Multiple regression analysis defines the treatment effect as a unique component that does not consider relations among intervening variables. In contrast, structural equation modeling not only provides an accurate estimate of the total treatment effect, but it also allows for decomposition of effects into those directly related to the treatment and those operating through theoretically specified intervening causal variables. Because many behavioral theories suggest that treatment effects of social interventions result from effects of intervening variables, treatment effects cannot be adequately represented conceptually or statistically by single‐equation analyses.