Causal inference

1
Causal inference
Doris Radun
15th EPIET Introductory Course, Lazareto,
Menorca 2009
2
What is a cause?- Definition -
Antecedent event, condition, or
characteristic that was necessary for the
occurrence of the disease event and without the
disease event either would not have occurred at
all or until some time later.
Source Rothman KJ, Greenland Causation and
Causal Inference in Epidemiology,
(Am J PH, 2005)
3
Henle-Koch-Postulates 1809-1885 1843-1910
4
Henle-Koch-Postulates 1809-1885 1843-1910

• pathogen can be identified in ill person

5
Henle-Koch-Postulates 1809-1885 1843-1910

• pathogen can be identified in ill person

• pathogen is culturable

6
Henle-Koch-Postulates 1809-1885 1843-1910

• pathogen can be identified in ill person

• pathogen is culturable

• cultured pathogen causes
• same illness in test animal

7
Henle-Koch-Postulates 1809-1885 1843-1910

• pathogen can be identified in ill person

• pathogen is culturable

• cultured pathogen causes
• same illness in test animal

• and can be isolated

8
Henle-Koch-Postulates 1809-1885 1843-1910

• pathogen can be identified in ill person

• pathogen is culturable

• cultured pathogen causes
• same illness in test animal

• and can be isolated

9
Causality and Epidemiology

• Search for causes and effects
• Daily work in epidemiology
• e.g.,
• - source of an outbreak?
• - risk factors?
• - predictors of disease?
• in order to give recommendations,
• implement prevention measures

10
Analytical studies in epidemiology
Comparison of groups

In determining differences in groups
compared, hypotheses are derived
? hypotheses are being tested
11
Do statistical associations automatically mean
that theres a causal relationship?
P lt 0.001
Chance Bias Confounding True association
?
as well as faulty study design
12
Impossibility to prove
In empirical science, it is impossible to prove.
(D. Hume)
Hypothesis All swans are white.
(K. Popper)
13
Epidemiology and causality
Examples of epidemiologists who had a difficult
time John Snow and cholera epidemic, Ignaz
Semmelweis and puerperal fever, .
14
Concept of causal pies (K.
Rothman)

• Is a factor (or a group of factors) sufficient
• to cause disease x?
• Is a factor (or a group of factors) necessary
• to cause disease x?

15
Elements for sufficient cause

• - Light switch on ? LIGHT!!!
• - Light switch Electric cable Light bulb ?
  LIGHT
• sufficient cause
• Any of the three components is also a necessary
  cause.

Ls
Ec
Lb
16
Multicausality
One causal mechanism
Single component
Sufficient cause 1 Sufficient cause 2
Sufficient cause 3
17
Multicausality and Interaction

• Different causal factors act together
• and produce an effect.

Alcohol drinking

-

12 4 3 1
Smoking status
-
Cohort study Rates of carcinoma of the tongue
per 100.000 person years
18
Multicausality and Interaction
Alcohol drinking
-

12 4 3 1
Smoking status

-
Rates of carcinoma of the tongue per 100.000
person years
What proportion of cases among smokers, who also
drink, is attributable to the effect of smoking?
19
Multicausality and Interaction
Alcohol drinking
-

12 4 3 1
Smoking status

-
Rates of carcinoma of the tongue per 100.000
person years
What proportion of cases among smokers, who also
drink, is attributable to the effect of smoking?
12 3 9, i.e. 9/12 cases 75
What proportion of cases among smokers, who also
drink, is attributable to the effect of drinking?
12 4 8, i.e. 8/12 cases 67
20
Multicausality and Interaction
Alcohol drinking
-

12 4 3 1
Smoking status

-
Rates of carcinoma of the tongue per 100.000
person years
What proportion of cases among smokers, who also
drink, is attributable to the effect of smoking?
12 3 9, i.e. 9/12 cases 75
What proportion of cases among smokers, who also
drink, is attributable to the effect of drinking?
12 4 8, i.e. 8/12 cases 67
21
Multicausality and Interaction
Alcohol drinking
-

12 4 3 1
Smoking status

-
Rates of carcinoma of the tongue per 100.000
person years
What proportion of cases among smokers, who also
drink, is attributable to the effect of smoking?
12 3 9, i.e. 9/12 cases 75
What proportion of cases among smokers, who also
drink, is attributable to the effect of drinking?

22
Multicausality and Interaction
Alcohol drinking
-

12 4 3 1
Smoking status

-
Rates of carcinoma of the tongue per 100.000
person years
What proportion of cases among smokers, who also
drink, is attributable to the effect of smoking?
12 3 9, i.e. 9/12 cases 75
What proportion of cases among smokers, who also
drink, is attributable to the effect of drinking?
12 4 8, i.e. 8/12 cases 67
23
Multicausality and Interaction
Alcohol drinking
-

12 4 3 1
Smoking status

-
Rates of carcinoma of the tongue per 100.000
person years
What proportion of cases among smokers, who also
drink, is attributable to the effect of smoking?
12 3 9, i.e. 9/12 cases 75
What proportion of cases among smokers, who also
drink, is attributable to the effect of drinking?
12 4 8, i.e. 8/12 cases 67
24
Conclusions from causal pies

• One effect often has several causes/levels of
  causes
• Components interact
• Attributable fractions of a disease regarding
• different component causes can exceed 100

25

• Sir Austin Bradford Hill
• 1897 1991
• British epidemiologist
• 50 ies Doll Hill
• Smoking and Lung Cancer
• developed viewpoints
• to draw conclusions
• (Bradford Hill viewpoints)
• decision making on basis
• of epidemiological data

AB Hill The Environment and DiseaseAssociation
or Causation?Proc Royal Soc Med 196558295-300
26
Nine viewpoints regarding causalityHill (1965)

• 1. Strength of Association
• 2. Consistency
• 3. Specificity
• 4. Temporality
• 5. Biological gradient (dose response)
• 6. Plausibility
• 7. Coherence
• 8. Experimental Evidence
• 9. Analogy

27
Strength of Association

• Strong associations are more likely being causal
  
• than weak ones.
• Smoking gt 20 cigarettes/day ? laryngeal
  carcinoma (RR 20)
• But beware not all strong associations are
  causal
• (Downs syndrome and birth order ? confounding)

Addendum proportion of cases explained by
exposure!
Weak associations do not rule out
causality (e.g., passive smoking and lung cancer
(RR 1.4))
28
Consistency

• If an association is repeatedly observed in
  different populations under different
  circumstances.
• e.g., smoking and lung cancer
• gt 100 studies during past 30 years revealed
• an elevated risk

Absence of consistency does not rule out a causal
relationship
29
Specificity

• One cause has one effect.
• Main argument for those who attempt to argue that
  smoking has not much to do with lung cancer

Specificity strengthens the body of evidence for
causality, however absence of specificity does
not rule out causality
30
Temporality

• Exposure precedes disease.
• The only fundamental criterion when suspecting
• a causal association!

Reversibility?
In an intervention study, reversibility is
another hint pointing to a causal relationship
31
Biological gradient (dose response)

• Risk increases with more intense/more frequent
  exposure.
• Paralleling association implies causality
• The more cigarettes are smoked, the greater the
  risk of lung cancer.

Paralleling observations are not always causal
(e.g., Downs syndrome and birth order)
32
Plausibility

• Being in line with current biologic/common
  knowledge.
• Is based on previous knowledge
• John Snow and the Cholera epidemic in London
• (Vibrio cholerae or other bacteria were not
  detected until then)

33
Coherence

• Interpretation of a causal association does not
  conflict with pathomechanisms of the disease,
• or its natural history.
• Smoking and lung cancer
• Histopathology of respiratory epithelium in
  smokers

Absence of coherence cannot been taken as
evidence against a causal relationship
34
Experimental Evidence

• Human experiments, experiments with animals
• Does the withdrawal of an incriminated exposure
• cause a decrease in incidence?

35
Analogy

• Presence of other cause-effect-relationships
• in analogy to those under investigation
• If a certain drug causes cancer, related drugs
  from the same pharmacological group may cause
  cancer, too.

Weak criterion for causality
36
A few points to finish

• beware of checklists!
• (they also make you stop thinking)
• beware of plausibility
• always aim for better evidence
• keep an open mind
• remain critical ( especially own studies)

37
Summary of the Bradford Hill viewpoints

• None of my nine viewpoints can bring
  indisputable evidence for or against the
  cause-and effect hypothesis ...

(Sir Austin Bradford Hill)