Loading field map
Causal inference in epidemiology emerged as a distinct subfield to address the fundamental challenge of deriving cause-effect relationships from observational data, where randomized experiments are often impractical or unethical. Early epidemiological work relied on heuristic criteria, such as the Bradford Hill viewpoints, to assess causality, but these lacked formal mathematical underpinnings. By the late 20th century, the need for rigorous frameworks to handle confounding, selection bias, and complex dependencies spurred the development of several rival methodological schools, each proposing distinct models of evidence and intervention.
The counterfactual framework, rooted in potential outcomes and associated with Donald Rubin, became a dominant paradigm by defining causal effects as contrasts between observed and hypothetical unobserved states. This approach emphasized design-based thinking and randomization-based inference, shaping how epidemiologists conceptualize treatment effects. In rival fashion, Judea Pearl's structural causal models, using directed acyclic graphs and do-calculus, offered a graphical representation of causal mechanisms, emphasizing identification through back-door and front-door criteria. These two frameworks ignited debates about the primacy of notation versus intuition, with proponents of structural models arguing for clearer causal pathways and counterfactual advocates stressing empirical transparency.
Parallel methodological schools arose to operationalize these theories. Instrumental variable methods, long used in economics, were adapted to epidemiology to mimic randomization when unmeasured confounding exists, relying on variables that affect exposure but not outcome directly. Propensity score methods, developed by Paul Rosenbaum and Rubin, provided a design-based strategy to balance covariates through matching, weighting, or stratification, embedding counterfactual logic into observational study design. While often complementary, these techniques encoded rival assumptions about ignorability and exclusion restrictions, leading to distinct diagnostic and evidential practices.
For longitudinal data with time-varying confounding, g-methods—including g-computation, inverse probability weighting of marginal structural models, and targeted maximum likelihood estimation—emerged as a framework extending counterfactual reasoning to dynamic treatments. Meanwhile, Mendelian randomization solidified as an application of instrumental variables using genetic variants, fostering a school centered on genetic epidemiology's unique causal claims. Today, the subfield navigates tensions between these paradigms, with ongoing synthesis in machine learning integrations, yet the core rivalry between counterfactual and structural causal models continues to organize theoretical discourse and pedagogical approaches in epidemiological causality.