Causal Inference

Causal Inference: Key Concepts

• Exposure and Outcome

  • Exposure is any factor that could be linked to the outcome of interest. The relationship between exposure and outcome can be:

  • Associated

  • Causal

  • Non-causal due to bias or chance

  • The core challenge in epidemiology: distinguishing association from causation.

Fundamental Question

• The fundamental question of epidemiology in these slides:[Exposure] [Outcome] are often:

  • Exposure Associated with Outcome

  • Exposure Causes Outcome

• Epidemiology is viewed as an art: how to distinguish associational from causal relationships and assess whether the observed result is valid (true).

• Core questions: Do both exposure and outcome co-exist? Does exposure cause the outcome? Are there other explanations?

Explanations for an Association

• Causal: The exposure is part of the causal pie leading to disease.

• Reverse Causation: The disease causes the exposure.

• Chance: There is no true association despite an observed link.

• Bias: Systematic errors in study design, data collection, or analysis distort the observed association.

Models of Causation (Overview)

• Web of Causation / Chain of Causality (1960s): causation results from a complex set of interconnected factors—not a single cause.

• Hill’s Criteria (1964): heuristic guidelines to assess causality, not rigid rules.

• Modified Determinism (1970s): causal pies; specific combinations of factors that are sufficient, component, or necessary.

• Pragmatic Epidemiology Causal Definition (1991): a practical framework focusing on essential properties of a cause.

Web of Causation / Chain of Causality

• Causation arises from a complex, interconnected system of factors, not a single factor alone.

• Includes both individual and environmental factors, e.g., social factors, environmental factors, work, family, friends, eating habits, genetics, job and working conditions, access to health care, culture.

• Shift toward considering multiple causes of disease.

• Public health relevance: prevention efforts can be amplified by addressing multiple risk factors.

• Provides a way to link social determinants with biomedical etiologic factors.

• Precipitating example: Lead poisoning Web of Causation (lead exposure network).

Lead Poisoning Web of Causation (Illustrative Example)

• Lead exposure pathways include lead paint in old housing, plumbing with lead, lead in soil and dust, water from lead pipes, and exposures from various sources (soil, dust, water, food).

• Precursors and contextual factors include low education, lack of daycare, lack of supervision, mouthing behavior and pica in children, low income or unemployment, fewer parents at home, working parents, and other sources like hobbies using lead-based materials, soil and dust, and high-risk living environments.

• Exposure pathways: ingestion and inhalation of lead lead to increased lead absorption and ultimately lead poisoning.

• The figure illustrates a web where precipitating conditions (e.g., poor housing, pica) interact with environmental and social factors to produce lead poisoning.

Criticisms of the Web of Causation

• Nancy Krieger (1994): Epidemiology and the Web of Causation: Has Anyone Seen the Spider?

  • The field has not clearly identified a specific “spider” or central causal agent weaving together factors.

• Fundamental criticisms:

  • Lack of explicit causal mechanisms: addressing one factor can still confer some benefit, so the mechanism is not always needed to justify intervention.

  • Little discussion of the origins of causes vs. interactions among causes.

  • No underlying theory or process for identifying elements of the web.

  • Focus on proximal factors (usually at the individual level) rather than population determinants.

  • No clear differentiation between individual determinants and population-level determinants.

Biomedical Individualism vs Public Health Perspective

• Biomedical Individualism criticisms (Slides 27–28):

  • Biomedical model focuses on biological determinants addressable by healthcare.

  • Social determinants are secondary to biological determinants.

  • Populations are treated as the sum of individuals; population patterns reflect individual cases.

  • This view is not fully aligned with public health, which emphasizes population-level determinants and prevention.

Hill’s Causal Criteria (Bradford Hill)

• Hill’s Guidelines were never intended as rigid criteria; they are viewpoints to help assess causality:

  • Experimental evidence

  • Temporal relationship

  • Strength of the association

  • Dose-response

  • Biological plausibility

  • Consistency

  • Analogy

  • Specificity

  • Coherence

• Quote (Hill): None of the nine viewpoints provide indisputable evidence; they help judge whether there is any other explanation more likely than cause and effect.

  • Source: Hill AB. The Environment and Disease: Association or Causation? Proc R Soc Med 1965; 58:295–300.

Hill’s Criteria: Detailed View

• Temporal relationship: the exposure precedes the development of the outcome.

• Strength of the association: stronger associations are more likely to be causal.

• Dose-response: greater exposure is associated with more outcome.

• Consistency: association observed in different contexts with same results.

• Biological plausibility: a reasonable biological mechanism linking exposure to outcome.

• Coherence: findings are coherent with existing theory and data; laboratory and epidemiological results align.

• Analogy: similar to other established causal relationships.

• Specificity: a given exposure leads to a specific outcome (in practice, often weaker in complex diseases).

• Experimental evidence: there is scientific evidence of an association (e.g., randomized trials, natural experiments).

Modified Determinism (Ken Rothman): Causal Pies

• Concept: Causal pies are composed of component causes that together form a sufficient cause for disease.

• Key ideas:

  • Sufficient cause: a complete set of causes that inevitably produces the disease when present.

  • Component causes: the individual factors that contribute to a sufficient cause and may be needed in some cases.

  • Necessary cause: a factor that is present in all sufficient causes for a given disease (if present in all pies).

  • A pie represents one pathway to disease; multiple pies can exist for the same disease.

Sufficient, Component, and Necessary Causes (Diagrams Abstracted)

• Sufficient causes: sets of factors that, when combined, will inevitably produce the outcome (there may be multiple different sufficient sets).

• Component causes: factors that appear in some sufficient causes but not in all.

• Necessary causes: factors that must be present in every sufficient cause for the disease.

• The idea is that disease can result from multiple different causal pathways (pies).

Real-World Example: Active Tuberculosis (Aschengrau & Seage)

• Sufficient cause for Active TB could be composed of the following exposures/conditions:

  • Exposure to TB bacteria

  • Poor nutrition

  • Crowding

  • AIDS

  • Absence of BCG vaccination

  • Poor ventilation

  • Absence of BCG (again listed as a factor in some pies)

• These components can combine in different ways to form sufficient causes for active TB.

• Real-world exercise: Which causes are sufficient? (List shows several potential pies.)

• Real-world example question: Why does Pie #3 need poor nutrition in addition to everything in Pie #2 to be sufficient?

  • Illustrates that different combinations of factors can be needed to reach a sufficient cause in a given pie, reflecting interactions among risk factors.

Practical Assessment of Pies in Practice

• Sufficient Causes: which combinations are enough to produce the disease?

• Component Causes: which factors are required in some contexts?

• Necessary Causes: which factors appear in all causal pies?

• Example questions illustrate how specific exposures interact (e.g., nutrition, crowding, AIDS) to create a sufficient set.

Does Modified Determinism Work in Practice?

• Pros:

  • Conceptually sensible; helps explain biological mechanisms leading to outcomes.

  • Explains patterning of risk factors among those who develop the disease.

  • Flexible; pies can be reformulated as new information emerges.

• Cons:

  • Difficult to operationalize because we often don’t know all causes.

  • In practice, identifying a complete set of sufficient causes is rare.

  • Does not necessarily address the quantitative aspect of risk (how much each factor contributes).

Why the Conundrum? Smoking and Lung Cancer

• Example question: We know smoking causes lung cancer, but what specifically about tobacco smoke triggers the disease process?

• Questions raised:

  • Why don’t all smokers get lung cancer, and why do some non-smokers get it?

• This reflects the complexity of interactions among multiple causes and the presence of multiple causal pies.

Reductionism vs Upstream Causes

• Reductionism: the idea that complex causal factors can be explained by identifying the simplest combinations.

• Alternative: consider more upstream causes, including distal, intermediate, and proximal factors, for a fuller causal account.

How to Spot a “True Cause” (Susser, 1991)

• Susser introduces a grammar for pragmatic epidemiology focusing on essential properties of a cause rather than checking properties in a single case.

• Criteria (three core):

  • Association: Factor X must occur with Y; a statistical relationship between X and Y; if no relationship, not causal.

  • Time order: X must precede Y.

  • Direction: X causes Y; Y does not cause X.

Pragmatic Causal Criteria in Practice

• If there is an association, evaluate time order using the study design.

• If there is association and time order, evaluate causal direction using factors such as:

  • Consistency

  • Strength

  • Specificity

  • Predictive performance

  • Coherence

Summary and Take-Home Messages

• The Fundamental Question: Exposure associated with outcome, exposure causes outcome, or both exist?

• Causation is the Holy Grail of epidemiology but difficult to prove definitively.

• Throughout this course, we will learn strategies to test hypotheses of causal relationships.

• Often, it is more important to make a case and rule out other explanations than to obtain definitive proof of causation.

Fundamental Question of Epidemiology (Recap)

• Exposure Associated with Outcome

• Exposure Causes Outcome

• The two options are not mutually exclusive in practice; many causal inferences rely on accumulating evidence across criteria and study designs.

Closing Note

• Real-world epidemiology relies on synthesizing multiple lines of evidence, using frameworks like Hill’s Criteria, causal pies, and pragmatic Susser criteria to build a coherent argument for causality rather than relying on a single definitive test.

ext{Dose-response: greater exposure} \Rightarrow ext{more outcome}
ext{X precedes Y (time order)}
ext{Causal pie: sufficient set of factors that inevitably causes disease}

Appendices (as referenced in slides, for quick recall)

• Hill’s Criteria: Experimental, Temporal, Strength, Dose-response, Consistency, Biological plausibility, Coherence, Analogy, Specificity

• Susser Criteria: Association, Time order, Direction; followed by evaluation via Consistency, Strength, Specificity, Predictive performance, Coherence

• Web of Causation: Multi-factorial networks linking social and biomedical determinants

• Modified Determinism: Pie diagrams showing sufficient, component, and necessary causes

• Real-world TB example: Presents multiple pies with exposures like TB bacteria, nutrition, crowding, AIDS, BCG absence, ventilation

• Lead poisoning web: Example of a web of precursors and precipitating conditions including socioeconomic and environmental factors