Causation

Lessons from John Snow Example• Form a hypothesis of the exposure-disease

relationship

• Identify appropriate study population, study area, and study period to obtain:- desired exposure contrast

- exposure groups comparable on other risk factors (minimize confounding)

- best available data on exposure and disease (and other risk factors)

Lessons from John Snow Example• Minimize selection, information and confounding

biases through:- design of the study- “shoe-leather” epidemiology

• Obtain a measure of the exposure-disease (dose-response) relationship

• Rule out alternative hypotheses and make causal inference- Snow’s tables and map- Examples of those in cluster area who did not use the

Broad Street Pump and did not get the disease- Two cases living outside the area whose only contact with

the area was drinking from the pump.

Checklist for epidemiological and statistical reviewsFrom: Rushton L. Reporting of occupational and environmental research: Use and misuse of statistical and epidemiological methods. Occup Environ Med 2000;57:1–9

Design:

Is there a clear statement of the study objectives?Is there a clear description of the study design?Is the study design appropriate?In there a clear description of the study populations?Is there a clear description of how the main variables were measured?Is there an assessment of the adequacy of the sample size?

Presentation and analysis:

Are the statistical procedures adequately described or referenced?Are the statistical procedures appropriate?Are the response rates reported?Are the data and the results adequately described?Are confidence intervals and significant levels given where appropriate?

Checklist for epidemiological and statistical reviews

Bias and confounding:

Are the response or tracing rates satisfactory?

Are you reasonably satisfied that the results are not explained by bias in subject selection or measurement?

As far as you can tell, have confounding variables been taken into account appropriately—either in the analysis or in the discussion?

Serious errors and omissions occurring in occupational epidemiology papers

From: Rushton L. Reporting of occupational and environmental research: use and misuse of statistical and epidemiological methods. Occup Environ Med 2000;57:1–9

Design:Adequacy of sample size not consideredBias in selection of subjects

Execution:Data collection problems and missing data not adequately reportedNon-respondents not investigatedSample selection and exclusions inadequately justified

Serious errors and omissions occurring in occupational epidemiology papers

Analysis:Inappropriate or incorrect analysis of data (e.g., matched case-control data)Modeling incorrect — e.g., inadequate adjustment for confounders

Presentation:Inadequate description of the methodology and statistical proceduresNo presentation of risk estimates and their confidence intervals

Interpretation:Potential bias due to sample selection, no or poor response, missing values, exclusionsLack of statistical power not consideredMisunderstanding and misinterpretation of results from models

What is Causation?

• “We may define a cause to be an object followed by another…where if the first object had not been, the second never had existed” (David Hume)- Hume defined 3 properties of a cause:

- Association (cause and effect occur together)- Time order (cause precedes effects)- Connection or direction (predictable link between

cause and effect)

What is Causation?

• Pragmatic definition: A cause is something that makes a difference

• “The cause of any effect consists of a constellation of components that act in concert.” (JS Mill)

• Factors in causation:- Predisposing factors creating enhanced susceptibility to the exposure

- Age, sex, previous illness- Enabling factors favoring disease development

- SES, poor nutrition, inadequate access to care- Precipitating factors: exposure to the agent - Reinforcing factors:

- repeated exposure, exposure to other agents, stress

What is Causation?

• Factors in causation:- Predisposing factors creating enhanced

susceptibility to the exposure- Age, sex, previous illness

- Enabling factors favoring disease development

- SES, poor nutrition, inadequate access to care

- Precipitating factors: exposure to the agent - Reinforcing factors:

- repeated exposure, exposure to other agents, stress

Causal Pie Model

• Multi-factorial model of causation- Each pie is a causal mechanism (“sufficient cause”) for the disease.

- A disease has multiple causal mechanisms.- Each letter represents a causal factor that is a component of a causal

mechanism. - A given causal mechanism requires the joint action of several

component factors- “A” = necessary cause in this example since it causes disease

through all three causal mechanisms.

Causal Pie Model• Each “pie” represents a “sufficient cause”

- A set of minimal conditions and events that inevitably produce the disease

- A complete causal mechanism consisting of a constellation of component causes

- Requires the joint action of its component causes- Each component cause plays a necessary role

• A necessary cause is a component cause in every causal mechanism- A component cause in every “pie” - If absent, then the disease does not occur

Examples of component causes in a causal

mechanismSocial

pressuresAir pollution

Lack of exercise

Cigarettesmoking

Workexposures

Hereditaryfactors

Pre-existingmedical

condition

diet

Implications of the Causal Pie Model

• The “strength” of a component cause depends on the prevalence in the population of the other component causes in the causal mechanism

– A factor will appear to have a strong effect if the other component causes are common in the population

– A factor will appear to have a weak effect if the other component causes are rare in the population

Implications of the Causal Pie Model

• No component cause acts alone to produce disease

• Every case of a disease is the result of multiple component causes acting jointly in a causal mechanism (“pie”).

• Therefore,• each case of a disease could be attributed to each of the separate

component causes that make up the causal mechanism (“pie”)

• the sum of the fractions of a disease attributable to each of its causes does not equal 100% but instead has no upper limit

An example of the naïve view that every case of a disease has a single cause, and that two or more causes cannot contribute to the same case.

Rushton L. How much does the environment contribute to cancer? Occupational and Environmental Medicine 2003;60:150–156

Causal Inference

Two questions:

• Is the exposure associated with the disease in our study?- Statistical inference

- Obtain a point estimate of effect (e.g., RR)- Evaluate the precision of the estimate of effect- Assess how likely the result is compatible with no

exposure effect

Causal Inference

Two questions:

• Is the observed association a causal one?- Causal inference

- Weighing the evidence and ruling out chance and bias as causes

Inference in survey research• Examples:

- Opinion polls- Inferring exposure prevalence or disease rate in a

community based on a random sample of that community

• Inference is from the sample to the source population from which sample was taken- Requires a random or “representative” sample be taken

from the source population- Inference is specific to time and place of the sampling- A key assumption underlying the meaning of P-values and

confidence intervals (random sampling) is satisfied

Inference in observational studies• The study population usually is not a random or

representative sample from a source population.

• Implications:- Inference is not from a sample to a source population from which the

sample was taken- Inference is from the study population to a population of interest- Inference is based on biological plausibility and other causal

criteria

• Examples of types of inferences:- Effect of an occupational exposure ⇒ effect in general population- Effect of exposure in one plastics plant ⇒ Effect of exposure in all plastics

plants- Effect of high occupational exposure ⇒ effect of low community exposure- Effects in animal study ⇒ effects in humans

Inference in observational studies• The purpose of statistical inference:

• compare observed result with what would have happened in a hypothetical situation if exposures had no effect and were randomly assigned to the study population over a large number of repetitions.

• Problem:• In observational studies, exposures are not randomly assigned

• Implications:- Key assumption underlying the meaning of the p-values and confidence

intervals (random allocation of exposure) is not satisfied.- Confidence intervals and p-values become indirect, descriptive

measures of precision and the compatibility of the observed results with the null and alternative hypotheses

Statistical Inference in observational studies

• The point estimate (e.g., risk ratio or odds ratio) provides the primary evidence for determining whether an observed exposure-disease association exists in our study

• Confidence interval can be used in a descriptive fashion as indirect measure of the precision of the point estimate as well as the compatibility between the observed result and a range of hypothetical results including a result of no effect (i.e., the null hypothesis)

• P-value can be used in a descriptive fashion as indirect measure of the extent of compatibility between the observed result and the hypothetical result of no exposure effect (i.e., the null hypothesis)

• The confidence interval and p-value provide supplementary information that can help in determining whether an observed exposure-disease association exists in our study

Causal Inference• After deciding that an exposure-disease association

has occurred in a study, the next question is whether this association is a causal one.

• Purpose of causal inference: • to weigh the evidence for a claim that the observed

exposure-disease association (also called a “statistical association”) constitutes a causal association

• reasoning, using causal “criteria”, to weigh the evidence• rule out chance and biases as causes

Causal Inference

• Causal criteria are used to weigh the evidence for a causal association• Although the results of an individual

epidemiological study can be evaluated, often the results of several studies are evaluated in a qualitative review or a quantitative meta-analysis.

• The most commonly used criteria were articulated by the British statistician Austin Bradford Hill in 1965

Causal InferenceCausal “criteria” (“Viewpoints”) from Hill (1965)

Strength of AssociationConsistencySpecificityTemporalityBiological GradientBiological PlausibilityCoherenceExperimentAnalogy

Hill’s Viewpoints

“What I do not believe…is that we can usefully lay down some hard-and-fast rules of evidence that must be obeyed before we accept cause and effect.”

(Hill, 1965)

Hill’s viewpoints are not a checklist. A causal association can be inferred even if only a few of the viewpoints are satisfied.

Hill’s Viewpoints

• Strength of Association- Strong associations are more likely to be causal

than weak associations because they are less likely to be explained by undetected biases

- However, the fact that an association is weak does not rule out causality

- Strength of association is measured by the risk ratio, relative risk, odds ratio, or regression coefficient, NOT the p-value

Hill’s Viewpoints• Consistency

- Repeated observation of an association in different populations and under different circumstances and study methods

- Lack of consistency does not rule out causality- A causal factor requires the joint action (in the proper

sequence) of other factors in the causal mechanism. This may not happen in a specific population.

- May be helpful in ruling out a particular bias as the cause of the association, since it is unlikely that the bias would occur across all studies.

Hill’s Viewpoints

• Specificity- Requires that a cause lead to a single

effect

- The criterion is invalid since a cause (e.g., smoking) can have multiple effects

- The criterion may be relevant for some infectious agents

Hill’s Viewpoints

• Temporality

- The necessity that the cause precede the effect

- This “viewpoint” is the only one that is absolutely necessary for an association to be causal

Hill’s Viewpoints• Biological Gradient

- The presence of a monotonic (unidirectional) dose-response curve

- Increasing trend in disease frequency with increasing dose

- Lack of a monotonic dose-response does not rule out a causal association

- Some causal associations have a threshold dose that must be exceeded before the disease frequency increases

- Confounding can produce the appearance of a monotonic dose-response between the exposure and disease

- Non-differential exposure misclassification bias can distort a dose-response curve so that it is no longer monotonic

Hill’s Viewpoints

• Biological Plausibility- An hypothesized relationship between an

exposure and a health outcome makes sense in the context of current biological knowledge

- Not a necessary criterion since “…the association we observe may be new to science or medicine and we must not dismiss it too light-heartedly as just too odd.” (Hill)

Hill’s Viewpoints

Key issues with Biological Plausibility:• No consensus on the amount of evidence sufficient to

establish plausibility• No consensus on rules for assessing relevance or

weight of evidence• Data gaps and uncertainties in current knowledge• Overemphasis on plausibility impedes acceptance of

new facts, especially those that reveal deficiencies in current knowledge

• Evidence underdetermines hypotheses- Biological evidence can support conflicting hypotheses- Assessment of evidence is open to different interpretations

Hill’s Viewpoints

• Coherence- Interpreting an association as causal should not

conflict with what is known of the natural history and biology of the disease

- Lack of coherence does not rule out a causal association

- Conflicting information may be mistaken or misinterpreted

Hill’s Viewpoints• Experimental Evidence

- Hill used the example of a prevention study:- “…Because of an observed association some preventive

action is taken. Does it in fact prevent?” - Others interpret Hill to mean evidence from laboratory

experiments or human experiments- Experimental evidence is often unavailable therefore

lack of such evidence does not rule out a causal association

- Experimental evidence underdetermines an hypothesis- There are often several alternative explanations for the

outcome of any experiment

Hill’s Viewpoints

• Analogy

“With the effects of thalidomide and rubella before us we would surely be ready to accept slighter but similar evidence with another drug or another viral disease in pregnancy” (Hill)

Hill’s Viewpoints

• Hill explicitly rejected the usefulness of statistical significance testing for causal inference:

“No formal tests of significance can answer those questions. Such tests can, and should, remind us of the effects that the play of chance can create, and they will instruct us in the likely magnitude of those effects. Beyond that they contribute nothing to the “proof” of our hypothesis.”

“And far too often we deduce ‘no difference’ from ‘no significant difference’.”

Proposed Guidelines for Carcinogen Risk Assessment

US Environmental Protection Agency, 1996