Causal Inference

Causal inference is the process of drawing conclusions about cause and effect relationships from data, using study design features, statistical methods, and theoretical reasoning to distinguish genuine causation from mere correlation or confounded association. It addresses the fundamental question: does the observed relationship between two variables reflect a true causal effect, or is it an artifact of confounding, bias, or chance?

Establishing causation is the ultimate goal of much scientific research, yet it is also one of the most challenging tasks. Clinicians need to know whether a drug truly causes symptom improvement before prescribing it. Policy makers need to know whether a program actually reduces poverty before scaling it. Public health officials need to know whether an environmental exposure causes cancer before regulating it. Incorrect causal conclusions can lead to ineffective treatments, harmful policies, and wasted resources. The methods of causal inference provide the intellectual and methodological tools for distinguishing genuine causes from spurious associations.

Causal inference rests on several complementary frameworks. The counterfactual framework (also called the potential outcomes framework, developed by Rubin and Neyman) defines a causal effect as the difference between what would happen to the same individual under different treatment conditions.^[4][5] Because only one condition can be observed for each individual (the fundamental problem of causal inference), randomization is used to create groups that are comparable on average, allowing the average causal effect to be estimated.^[4][5]

Randomized controlled trials are considered the gold standard for causal inference because randomization ensures that the treatment and control groups differ only by chance, removing systematic confounding; chance imbalances can still occur, especially in small trials. When randomization is not feasible, quasi experimental methods and observational study designs can provide evidence for causal relationships, though they are more susceptible to confounding and bias.

The Bradford Hill criteria, proposed in 1965 (Hill himself called them "viewpoints" rather than criteria), provide a set of considerations for evaluating causal evidence from observational studies: strength of association, consistency across studies, specificity, temporality (cause precedes effect), biological gradient (dose response relationship), plausibility, coherence, experimental evidence, and analogy.^[3] No single criterion is sufficient to establish causation, and they are used as guidelines rather than rigid rules.^[3]

Modern causal inference methods include instrumental variable analysis (which uses a variable that affects the outcome only through its effect on the exposure), regression discontinuity designs (which exploit threshold based treatment assignment rules), difference in differences analysis (which compares changes over time between treated and untreated groups), propensity score methods (which balance observed confounders between groups), and directed acyclic graphs (which formalize causal assumptions and guide the selection of adjustment variables).

Causal inference methods are relevant whenever the research question asks whether one variable causes a change in another. Randomized experiments are the first choice when feasible and ethical. When randomization is not possible, researchers should select the quasi experimental or observational method that best addresses the specific sources of confounding in their study. Causal inference is particularly important in clinical research, epidemiology, economics, and policy evaluation, where the consequences of incorrect causal claims are significant.

An epidemiologist wants to determine whether a community water fluoridation program causes a reduction in childhood dental caries. A randomized trial is impractical because fluoridation is decided at the community level. Instead, the researcher uses a difference in differences design, comparing changes in dental caries rates over 10 years in 30 communities that adopted fluoridation with 30 matched communities that did not. Before adoption, caries rates had followed similar trends in both sets of communities, supporting the parallel trends assumption on which the design relies. After adjusting for socioeconomic factors, access to dental care, and baseline caries rates, the analysis finds that communities that adopted fluoridation experienced a 2.1 additional point decline in caries per 100 children (95% CI: 1.2 to 3.0) compared to the control communities, supporting a causal interpretation.