A concept of causation in epidemiology
A concept of causation in epidemiology
Causation – Seminar
20 November, 2014
Causes – the C-word in epidemiology
• We study associations but we estimate effects
because we are interested in causes
1. (Some) diseases have causes
2. Some of these causes are avoidable
3. Therefore some diseases are preventable
• But what is causation?
• How do causes operate?
• And how can they be detected?
• First the history: Hume – Mackie - Rothman
Hume – one object E followed by another object D
rarely if ever seen
the necessary part
- circular if the exposure is
part of the definition of the
the sufficient part
seen but not often
• And monocausal effects are rare, even a
simple – press the switch, the light is on - has
a sequence of causes from distal to proximal
• Causes act in a probalistic manner, why?
• Causes usually have delayed effects, why?
• Mackie – Rothman causal model explains why
causes act probalistic.
The probability of getting L.C. if you smoke is a
function of the frequency of other component
causes Asb, G1 and G2.
The induction time is the time it takes to
complete a causal field and latency time is the
time it takes for the disease to surface to clinical
Is the model true? More important is – does it
work? Inspire to better understanding, more
research. Shows why etiologic fractions sum up
to more than 100%
• Gwas studies indicate a very complex genetic
set of causes perhaps that is why we see so
few gene-environmental interactions.
• Problems in meta-analyses. No reason to
expect simple additive or multiplicative effects
Hume’s second definition.
Or, in other words, where the first object had not been
the second would never exist.
Leads to counterfactual analysis of causation and a
possible definition of causation. Leads also to the RCT
and other designs that can provide the expected
disease occurence among exposed had they not been
Current definition of causation
• E is causing D if E increases the probability of
• Or like
• E is causing D if E blocks the estrogen receptor
leading to ... D in the presence of E
• Epidemiology versus toxicology in public
If E is a cause of D in a population it must be true
for at least one person that he or she would not
have had this disease at this time if he or she had
not been exposed.
Counterfactual reasoning implies we can
imagine a world where the exposure is not
Counterfactual reasoning also tells us why we
do not compare cases and controls in a casecontrol study.
We try to reconstruct the IR among exposed
and not exposed.
We try to imagine what would be the disease
occurrence among the exposed had they not
been exposed and we use not exposed for
that well knowing that they are surrogates for
counterfactual observation time.
Causes are upstream or downstream.
Many sexual contacts cause cervical cancer?
Cervical cancer is caused by a specific variant of
Both levels of causation can be used in prevention.
Down stream causes may become rather trivial –
accident research for example
Causes may be presented graphically – as done
by using DAGs.
Put simply, causality has been mathematized.
Judea Pearl: Causality. Cambridge University
DAGs also illustrate situations where causal
inference is justified.
A randomizes trial, incl. Mendelian Randomiza-tion.
Nothing feeds into R and R only affects E (incl. a
gene?). No backdoor path – closed at E. If R is
associated with D there is a causal link between E
People without G1,G 2,G3 (2 out of 3)
People with G 2,and G3
People with G1
Quantification of causes
• Even if diseases have many causes – OK to
focus on one.
• The causes of interest are those that can be
manipulated – Følling’s Disease.
Do we need a concept of causation?
The surgeons point of view
• If someone kicks me in the back and it hurts –
I know what causes the pain and I want
revenge - not a lecture in philosophy
• Use causal concepts but leave the nature of
causation to philosophers?
• Perhaps but leaves much of the basic theory
in the dark.
Causal interpretation problem
– Alex Broadbent
• If nobody in the population smoked the risk of
L.C. in that population would be 50% less
• - But estimate is too high what would replace
• Counterfactual; a cause makes a
difference: Had it been different or
absent, then the effect would have
been different or absent (AB).
“A measure of causal strength is a measure of
the net difference in outcome explained by the
To explain a difference in outcome such that the
outcome in group A is greater than the outcome
in group B by degree, u , we must have a
difference in exposure which could at least
explain this degree, u.” (AB).
Strength of association is a causal criterium
• A result, claim, theory, inference, or other
scientific output is stable if and only if
a) It is not known to be contradicted by good
b) Given best current scientific knowledge, it
would probably not be seen contradicted by
good scientific evidence, if good research
were to be done on the topic (AB)
• The ‘What could possibly go wrong?’
• When you have eliminated all other
explanations what remains must be the truth.
• Sherlock Holmes
• The Mackie/Rothman causal model has been
useful in providing an understanding (true or
false) of many of the concepts we use. It
inspires to new theories and hypotheses.
• DAGs has been useful in understanding how
causes operate and in helping separating out
causal paths of interest from causal paths of
• Counterfactuals has focused our attention on
what we aim at in our designs and analyses.
• Cross over design, case-only designs, sib
studies etc. – not only RCTs.