Social epidemiologists study effects of variables such as education or income on health outcomes. Because other factors may influence both the exposure and the outcome, adjustments are commonly made in an effort to estimate the “independent” effect of exposure. The validity of common adjustment strategies when estimating the outcome distribution under hypothetical interventions of the exposure is potentially compromised by structured relations between covariates, observed and unobserved. These considerations of covariate structure may be particularly important for the study of “distal” socioeconomic factors that affect health through specified intermediates, therefore making standard adjustments in social epidemiology potentially problematic. Two related approaches have been proposed for defining and estimating causal effects in light of covariate structure: Robins’ g-computation algorithm and Pearl’s non-parametric structural equations. We review the conceptual foundation for these techniques, and provide a heuristic example using data from the National Longitudinal Mortality Study (NLMS) to demonstrate the extent to which selected causal effects (contrasts between hypothetical intervention regimens) are sensitive to structured relations among measured and unmeasured covariates, even in very simple systems.