Causal pathway diagram. Prior to finalizing the inclusion/exclusion criteria and hypotheses for this study, a causal diagram was developed to visually encode a priori assumptions about the relation between exposure (previously undiagnosed versus previously diagnosed diabetes), outcomes, and covariates, taking into account the strengths and limitations of the Surveillance, Epidemiology, and End Results (SEER) - Medicare database. The diagram depicts directed paths (a head-to-tail sequence of arrows, or a “one-way street”) between previously undiagnosed diabetes and both cancer stage at diagnosis and mortality. One of the directed paths between previously undiagnosed diabetes and mortality contains cancer stage as an intermediate variable. In other words, the impact of previously undiagnosed diabetes on mortality is partially explained by its intermediate impact on cancer stage. The other directed path contains no intermediate variables. In addition, the diagram depicts undirected paths (paths in which the arrows are not all head-to-tail) between previously undiagnosed diabetes and both cancer stage at diagnosis and mortality, which “flow through” patient demographic, socioeconomic, and clinical characteristics. In this instance, the undirected paths between previously undiagnosed diabetes, cancer stage, and mortality are biasing paths (and the variables on those paths are potential confounders) for the association between exposure and outcomes because they do not represent effects of previously undiagnosed diabetes on the outcomes, yet can contribute to (confound) the association between previously undiagnosed diabetes and outcomes. These should be “blocked” either by study design, including patient selection, or by adjustment in the analyses, to maximize the likelihood that the observed residual associations between exposure and outcomes are unbiased.
Griffiths et al. BMC Cancer 2012 12:613 doi:10.1186/1471-2407-12-613