Residency · Residency · Preventive Medicine
Causal Inference Frameworks: Bradford Hill and Beyond
Overview
Establishing causality is the ultimate goal of much epidemiologic research, yet correlation does not equal causation. Systematic frameworks have been developed to help evaluate whether an observed association is truly causal. Modern causal inference has moved well beyond Hill's original criteria to include formal mathematical frameworks that provide rigorous tools for causal reasoning.
Bradford Hill's Criteria (1965)
The Nine Viewpoints
In 1965, Sir Austin Bradford Hill proposed nine viewpoints for evaluating whether an observed association might be causal. Strength of association holds that larger effect sizes are more likely to be causal, though weak associations can still reflect true causation. Consistency means the association has been observed repeatedly across different populations, settings, and study designs. Specificity suggests that one cause leads to one effect, though this is considered the weakest criterion since many exposures cause multiple outcomes. Temporality — that the exposure must precede the outcome — is the only criterion considered absolutely necessary. Biological gradient (dose-response) means that increasing exposure leads to increasing risk, though threshold effects can complicate this pattern. Plausibility requires that a biologically plausible mechanism exists, though this is limited by current scientific knowledge. Coherence means the association does not conflict with what is known about the disease's natural history and biology. Experiment indicates that removal of the exposure reduces disease incidence, representing some of the strongest evidence. Analogy holds that similar exposures causing similar outcomes lends support.
| Criterion | Definition | Example (Smoking & Lung Cancer) | Necessity |
|---|---|---|---|
| Strength | Large effect size | RR of 10-30 for heavy smokers | Not necessary (weak associations can be causal) |
| Consistency | Replicated across settings | Observed in dozens of countries and study designs | Not necessary but strengthens inference |
| Specificity | One cause → one effect | Weakest criterion; smoking causes many diseases | Not necessary |
| Temporality | Exposure precedes outcome | Smoking initiation precedes cancer by decades | Absolutely necessary |
| Biological gradient | Dose-response relationship | More pack-years → higher risk | Not necessary (threshold effects exist) |
| Plausibility | Biologically plausible mechanism | Carcinogens in tar damage bronchial epithelium | Not necessary (limited by current knowledge) |
| Coherence | Consistent with known biology | Lab animal experiments, histopathology findings | Not necessary |
| Experiment | Removal of exposure reduces risk | Cessation reduces lung cancer risk over time | Not necessary but strong evidence |
| Analogy | Similar exposures cause similar outcomes | Other inhaled carcinogens cause cancer | Not necessary |
Important Caveats
Hill explicitly stated that these are "viewpoints," not rigid criteria or a scored checklist. No single criterion except temporality is either necessary or sufficient for establishing causation. They were never intended to be tallied up as a score. They guide scientific judgment rather than providing an algorithmic decision process. Hill developed these viewpoints in the context of evaluating the evidence linking smoking to lung cancer.
Counterfactual Framework (Potential Outcomes)
Fundamental Concept
The counterfactual framework defines a causal effect as the difference between what actually happens under an exposure and what would have happened in the absence of that exposure — the counterfactual scenario. For any individual, only one potential outcome can ever be observed, which constitutes the fundamental problem of causal inference. At the population level, average causal effects are estimated by comparing groups that differ in their exposure status.
Key Definitions
The Average Treatment Effect (ATE) is the mean difference in outcomes if everyone in a population were exposed versus if no one were exposed. The Average Treatment Effect on the Treated (ATT) quantifies the effect specifically among those who actually received the exposure. Exchangeability means that the exposed and unexposed groups would have had the same outcome distribution if their exposure status were reversed. Positivity requires that every subgroup defined by covariates has some probability of receiving each exposure level. Consistency demands that the exposure be sufficiently well-defined so that potential outcomes are unambiguous.
Assumptions for Causal Inference
Valid causal inference requires no unmeasured confounding (exchangeability conditional on measured covariates). It also requires the Stable Unit Treatment Value Assumption (SUTVA): one person's treatment does not affect another person's outcome, and there is only one version of each treatment level.
Directed Acyclic Graphs (DAGs)
Formal Framework
DAGs provide a graphical representation of causal assumptions using nodes (representing variables) and directed edges (representing causal effects). Developed primarily by Judea Pearl and applied to epidemiology by Robins, Hernan, and others, they allow formal identification of confounders, mediators, and colliders within a hypothesized causal structure.
Causal Identification Rules
The concept of d-separation states that two variables are conditionally independent if all paths between them are blocked. The backdoor criterion identifies the set of variables that must be conditioned upon to block all non-causal (backdoor) paths from exposure to outcome. The front-door criterion provides an alternative strategy when backdoor adjustment is impossible, using mediators to estimate the causal effect.
Common DAG Structures
The fork structure (X <- C -> Y) represents a common cause or confounder, and adjustment for C is appropriate. The chain structure (X -> M -> Y) represents a mediator, and adjusting for M should be avoided when estimating the total effect. The collider structure (X -> C <- Y) means that C should not be conditioned upon, because doing so opens a biasing path. M-bias is a specific collider structure where adjusting for a variable that superficially resembles a confounder actually introduces collider bias.
Target Trial Emulation
Concept
Target trial emulation is a framework proposed by Hernan and Robins for designing observational studies with causal ambitions. The investigator explicitly specifies the hypothetical randomized controlled trial that would ideally answer the causal question, then designs the observational analysis to emulate that trial as closely as possible using available data.
Key Components to Specify
The framework requires specification of eligibility criteria, treatment strategies being compared, assignment procedures (how treatment was "assigned" in the observational data), follow-up period start (time zero alignment), outcome definition, the causal contrast of interest (intention-to-treat versus per-protocol), and the analysis plan.
Avoiding Common Pitfalls
Immortal time bias arises from misalignment of time zero and treatment assignment, creating a period during which the treated group cannot experience the outcome. Prevalent user bias occurs when patients already on treatment are included rather than restricting to new initiators. Active comparator designs — comparing one treatment to another rather than to no treatment — are generally preferred for reducing confounding by indication.
Mendelian Randomization
Concept
Mendelian randomization uses genetic variants as instrumental variables to estimate causal effects of modifiable exposures on health outcomes. Because genes are "randomly assigned" at conception through the process of meiosis, this approach is analogous to randomization. If a genetic variant affects an exposure (such as an alcohol metabolism gene affecting drinking behavior) and is only associated with the outcome through that exposure, it can serve as a natural experiment.
Assumptions
Three core assumptions must hold. The genetic variant must be associated with the exposure (relevance). The variant must not affect the outcome except through the exposure (the exclusion restriction). The variant must not be associated with confounders of the exposure-outcome relationship (independence or exchangeability).
Limitations
Pleiotropic effects — where a gene affects multiple biological pathways — violate the exclusion restriction. Weak instrument bias occurs when the variant explains only a small fraction of exposure variation. Canalization refers to developmental compensation for lifelong genetic effects that may differ from effects of modifying the exposure in adulthood. Population stratification can also create confounding.
Applications
Mendelian randomization has been applied to investigate causal relationships between alcohol consumption and cardiovascular disease, BMI and cancer risk, and LDL cholesterol and coronary disease. The last of these provided genetic evidence supporting statin development well before definitive trial data were available.
Difference-in-Differences and Synthetic Control Methods
Difference-in-Differences (DiD)
The difference-in-differences approach compares the change in an outcome before and after an intervention in the treated group relative to the same change in a control group. The key assumption is parallel trends: both groups would have followed the same trajectory in the absence of the intervention. This method is widely used for policy evaluation, such as assessing the health effects of Medicaid expansion.
Synthetic Control Method
The synthetic control method constructs an artificial control group from a weighted combination of untreated units that together approximate the pre-intervention trajectory of the treated unit. This is particularly useful when only a single unit (such as one state or country) receives the intervention, making traditional comparison groups inadequate.
Regression Discontinuity Design
Regression discontinuity design exploits a threshold or cutoff that determines treatment assignment. By comparing outcomes for individuals just above and just below the threshold, the design leverages the assumption that people near the cutoff are similar in all relevant characteristics except their treatment status. An example is comparing outcomes for patients just above versus just below an age cutoff for cancer screening eligibility.
<image>A timeline diagram illustrating the evolution of causal inference methods in epidemiology. Starting from the left with Koch's postulates (1890), moving to Bradford Hill's criteria (1965), then to the counterfactual/potential outcomes framework (Rubin, 1974), directed acyclic graphs (Pearl, 1990s-2000s), and modern methods including target trial emulation and Mendelian randomization (2010s-present). Each era is depicted as a milestone on the timeline with key concepts and representative publications noted below. Clean, horizontal timeline with color-coded eras.</image>
<image>A schematic diagram explaining Mendelian randomization. Shows three connected elements: a genetic variant (represented as a DNA helix icon) with an arrow pointing to an exposure variable (e.g., LDL cholesterol), which has an arrow pointing to an outcome (e.g., coronary heart disease). A dashed line with an X through it runs directly from the genetic variant to the outcome, indicating the exclusion restriction assumption. Confounders (lifestyle, socioeconomic factors) are shown connected to both exposure and outcome but NOT to the genetic variant, illustrating independence. Annotations explain each assumption. Medical illustration style with clear labels.</image>
<image>A side-by-side comparison diagram of Bradford Hill's nine criteria. Each criterion is shown as a card or panel with its name, a brief definition, a small illustrative icon (e.g., a dose-response curve for biological gradient, a clock for temporality), and a note on its relative strength or importance. Temporality is highlighted as the only necessary criterion. The overall layout emphasizes that these are guidelines for judgment, not a rigid scoring system. Professional medical education style.</image>
Clinical Pearls
Temporality is the only absolutely necessary criterion for causation — if the exposure does not precede the outcome, causation is excluded. Hill's criteria should be used as a guide for scientific judgment, not as a checklist with a passing score. DAGs are essential for identifying what to adjust for, because adjusting for mediators or colliders can introduce bias rather than remove it. Target trial emulation is the modern standard for designing rigorous observational studies — the first step should always be specifying the hypothetical trial. Mendelian randomization is powerful but requires careful assessment of pleiotropy and instrument strength. The E-value quantifies how robust a causal claim is to unmeasured confounding. For board exams, know Hill's criteria by name and understand that only temporality is essential.
References
- Hill AB. The environment and disease: association or causation? Proc R Soc Med. 1965;58:295-300.
- Hernán MA, Robins JM. Causal Inference: What If. Chapman & Hall/CRC; 2020.
- Pearl J. Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge University Press; 2009.
- Hernán MA, Robins JM. Using big data to emulate a target trial. Am J Epidemiol. 2016;183(8):758-764.
- Davey Smith G, Ebrahim S. Mendelian randomization: prospects, potentials, and limitations. Int J Epidemiol. 2004;33(1):30-42.
- VanderWeele TJ, Ding P. Sensitivity analysis in observational research: introducing the E-value. Ann Intern Med. 2017;167(4):268-274.
