Causal Inference in Epidemiology

Ahmed Mandil, MBChB, DrPHProf of Epidemiology

High Institute of Public HealthUniversity of Alexandria

Headlines Levels of causality Definitions Koch's postulates (1877) Hill's criteria (1965) Susser's criteria (1988, 1991)

Relating Exposures: causes, risk factors,

independent variables to Outcomes: effects, diseases, injuries,

disabilities, deaths, dependent variables

Statistical association versus biological causation: cause-effect relationship

Levels / Types of causality Molecular / Physiological Personal / Social Deterministic / probabilistic What aspect of “environment”

(broadly defined) if removed / reduced / controlled would reduce outcome / burden of disease

Definitions (I) Deduction: reasoned argument proceeding

from the general to the particular. Induction: any method of logical analysis that

proceeds from the particular to the general. Conceptually bright ideas, breakthroughs and ordinary statistical inference belong to the realm of induction.

Induction period: the period required for a specific cause to produce the disease (health-related outcome). Usually longer with NCDs

Definitions (II) Association (relationship): statistical dependence

between two or more events, characteristics or other variables. Positive association implies a direct relationship, while negative association implies an inverse one. The presence of a statistical association alone does not necessarily imply a causal relationship.

Causality (causation / cause-effect relationship): relating causes to the effects they produce.

Cause: an event, condition, characteristic (or a combination) which plays an important role / regular / predicable change in occurrence of the outcome (e.g. smoking and lung cancer)

Causes may be “genetic” and / or “environmental” (e.g. many NCDs including: diabetes, cancers, COPD, etc)

Definitions (III) Deterministic causality: cause closely

related to effect, as in “necessary” / “sufficient” causes

Necessary cause: must always PRECEDE the effect. This effect need not be the sole result of the one cause

Sufficient cause: inevitably initiates or produces an effect, includes “component causes”

Any given cause may be necessary, sufficient, both, neither (examples)

Definitions (IV) Component causes: together they

constitute a sufficient cause for the outcome in question. In CDs, this may include the biological agent as well as environmental conditions (e.g. TB, measles, ARF/RHD). In NCDs, this may include a whole range of genetic, environmental as well as personal / psychosocial / behavioral characteristics (e.g. diabetes, cancers, IHD)

Definitions (V) Probabilistic Causality: in epidemiology,

most associations are rather “weak” (e.g. relationship between high serum cholesterol and IHD), which is neither necessary nor sufficient

Multiple causes result in what is known as “web of causation”or “chain of causation”which is very common for noncommunicable / chronic diseases

Effect Measures / Impact Fractions

Effect measures (e.g. odds ratio, risk ratio) and impact fractions (e.g. population attributable risk) are closely related to the strength of association

The higher effect measures (away from unity) and population attributable risk (closer to 100 %) the more the exposure is predictive of the outcome in question

E.g. PAR of 100 % means that a factor is “necessary”

Deterministic causality (I)

Deterministic causality (II)

Deterministic causality (III)

Deterministic causality (IV)

Deterministic causality (V)

Deterministic causality (VI)

Definitions (IV) Predisposing factors: factors that prepare, sensitize,

condition or otherwise create a situation (such as level of immunity or state of susceptibility) so that the host tends to react in a specific fashion to a disease agent, personal interaction, environmental stimulus or specific incentive. Examples: age, sex, marital status, family size, education, etc. (necessary, rarely sufficient).

Precipitating factors: those associated with the definitive onset of a disease, illness, accident, behavioral response, or course of action. Examples: exposure to specific disease, amount or level of an infectious agent, drug, physical trauma, personal interaction, occupational stimulus, etc. (usually necessary).

Weighing Evidence At individual level: clinical judgment (which

management scheme) At population level: epidemiological judgment

(which intervention) When weighing evidence from

epidemiological studies, we use “causal criteria” (usually applied to a group of articles, to deal with confounding) e.g. Hill’s / Susser’s criteria, which were preceded by Koch’s postulates (on infectious diseases)

Henle-Koch's postulates (1877,1882)

Koch stated that four postulates should be met before a causal relationship can be accepted between a particular bacterial parasite (or disease agent) and the disease in question. These are:1. The agent must be shown to be present in every case of the disease by isolation in pure culture.2. The agent must not be found in cases of other disease.3. Once isolated, the agent must be capable of reproducing the disease in experimental animals.4. The agent must be recovered from the experimental disease produced.

Hill's Criteria (1897 - 1991)

The first complete statement of the epidemiologic criteria of a causality is attributed to Austin Hill (1897 - 1991). They are:

1. Consistency (on replication)2. Strength (of association)3. Specificity 4. Dose response relationship5. Temporal relationship (directionality) 6. Biological plausibility (evidence)7. Coherence8. Experiment

Consistency (I)

Consistency (II) Meta-analysis is an good method for

testing consistency. It summarizes odds ratios from various studies, excludes bias

Consistency could either mean: Exact replication (as in lab sciences,

impossible in epidemiological studies) Replication under similar circumstances

(possible)

Strength of Association

Expressions of Strength of Association

Quantitatively: Effect measure (OR, RR): away from unity (the

higher, the stronger the association) P-value (at 95% confidence level): less than

0.05 (the smaller, the stronger the association) Qualitatively:

Accept alternative hypothesis: an association between the studied exposure and outcome exists

Reject null hypothesis: no association exists

Dose-response relationship (I)

Dose-response relationship (II)

Time-order (temporality, directionality)

Time order

Specificity of Outcome

Specificity of Exposure

Coherence Theoretical: compatible with pre-existing

theory Factual: compatible with pre-existing

knowledge Biological: compatible with current

biological knowledge from other species or other levels of organization

Statistical: compatible with a reasonable statistical model (e.g. dose-response)

Biological Coherence (I)

Biological Coherence (II)

Susser's criteria (I) Mervyn Susser (1988) used similar

criteria to judge causal relationships.

In agreement with previous authors, he mentioned that two criteria have to be present for any association that has a claim to be causal: i.e. time order (X precedes Y); and direction (X leads to Y).

Susser’s Criteria (II) Rejection of a hypothesis can

accomplished with confidence by only three criteria: time order, consistency, factual incompatibility or incoherence.

Acceptance or affirmation can be achieved by only four, namely: strength, consistency, predictive performance, and statistical coherence in the form of regular exposure/effect relation.

Comparison of Causal Criteria

References1. Porta M. A dictionary of epidemiology. New York,

Oxford: Oxford University Press, 2008. 2. Rothman KJ (editor). Causal inference. Chestnut Hill:

Epidemiology Resources Inc., 1988.3. Hill AB. The environment and disease: Association

or causation. Proceedings of the Royal Society of Medicine 1965; 58: 295-300.

4. Susser MW. What is a cause and how do we know one ? A grammar for pragmatic epidemiology. American Journal of Epidemiology 1991; 133: 635-648.

5. Paneth N. Causal inference. Michigan State University.6. Rothman J, Greenland S. Modern epidemiology.

Second edition. Lippincott - Raven Publishers, 1998.

Thank you for your kind attention