Transcript Who is in charge of causation?

WHO IS IN CHARGE OF
CAUSATION?
Olaf Dekkers/20-11-2014
Dept Clinical Epidemiology Aarhus and Leiden
Le Malade Imaginere (1673), Act III
Basic problem I


We don’t usually see that x causes y
Causal judgement is based on inference
Basic problem II


Inferences can be wrong
Statistics do not differentiate between association
and cuasation
Basic problem III

We are taught, over and over, to be sceptical
 Why
most research findings are false. Ioannidis 2005
 Am J Epidemiology does not allow the use of the word
‘effect(s)’ to denote association(s)
 It
may promise more than an observational study can deliver
 Only studies in which there is an intervention to change
exposure can study causes
 It misleads the uncritical reader
So…


Making causal inferences in clinical research is not
straightforward (old news)
But we should not refrain from searching for causes,
as clinical research aims to improve health
Where do we stand today?
Theories of causation
- Regularity theories (Hume)
- Probabilistic theories
- Counterfactual theories
- Interventionism
Counterfactual theory of causation
Counterfactual theory
Intuitive interpretation:
If it is the case that Y occurs after X and that Y does
not occur in the absence of X, than X is a cause of Y
More formally:
“X causes Y because the counterfactual ‘if not X then
not Y’ is true.”
cf Paul LA Counterfactual theories
“X causes Y”

Counterfactual thinking is intuitively appealing
 Italy
lost the game (Y), because the referee was not
objective (X)
 She was late (Y), because of the snow (X)
 She/he fell deeply in love with him/her (Y), because of
his beautiful eyes (X)
 Attacking Pearl Harbor (X), caused the USA to enter
WWII (Y)
Why do we need causation theories?

What is a philosophical theory of causation giving
an account for?
 The
theory facilitates causal inference, and gives an
answer to the question how we can know whether X is a
cause of Y (epistemic notion)
 In
 The
epidemiology the epistemic notion is more relevant
theory might facilitate our understanding of what
causation (in general) stands for (semantic/ontologic
notion)
Epistemic notion of causation

Epistemic notion
 Obesity
(x) is a cause of mortality (y) because the
counterfactual ‘if not X then not Y’ is true (in a least one
person) (Counterfactual theory)
 Obesity (x) is a cause of mortality (y) because being
obese increases mortality probability (Probability
theory)
Ontologic notion of causation

Semantic/ontologic notion
 There
are no working forces, there are only
probabilities (Probability theory)
 This notion is in a counterfactual framework difficult as
the theory refers to non-occurring situations (or even
worse: non-existing worlds)
Counterfactuals: philosophical
problems
 Pre-emption
(see next slide)
 Counterfactual
dependence is not a necessary condition for
causation
 Prevented
causes
 Think
of a new extremely powerful weapon (X) causing
world-destruction (Y). Causation cannot be judged in a
counterfactual framework
 Many
irrelevant causes
 Counterfactual
causation
dependence is not a sufficient condition for
Pre-emption
Smoking
Hypertension
AMI
An unstable christmas tree
Counterfactual theory needs many auxillary theories
to keep it upright
Counterfactuals in epidemiology



What makes the theory so appealing in epidemiology?
Counterfactuals point towards non-existing situations: ‘If
not X, etc’
In medical research counterfactuals do not exist



Even not in cross-over studies
But: we are used to control groups
Control group as approximation for the counterfactual

Counterfactual thinking resembles the way we do studies
Counterfactual in epidemiology
BUT

No practical decision rule to infer causality


Refers to non-occurring situations
There is however a set of basic assumptions for causal
inference in the counterfactual framework (cf Hernan):
‘Exchangeability’
 ‘Positivity’
 X=1 and X=0 (exposure contrast) should be well-defined

Counterfactuals and exchangeability
Patients (X=1)
Outcome Y=y|X=1
Patients (X=0)
Outcome Y=y|X=0
Counterfactuals and exchangeability
Patients (X=1)
Outcome Y=y|X=1
Patients (X=0)
Outcome Y=y|X=0
Counterfactual definition of causation:
If it is the case that Y occurs after X
and that Y does not occur in the absence of X, than X is a cause of Y
This can be inferred from a comparison under exchangeability:
Two groups would have had an identical outcome (Y=y)
if the exposure (X=1 or X=0) would have been the same
Exchangeability
 Two
groups would have had an identical outcome (Y=y)
if the exposure (X=1 or X=0) would be the same
 If
the baseline prognosis is the same, then a difference
in outcome between groups has a causal interpretation
 This
is threatened by confounding in observational
studies
Confounding and selection-bias



Causal question Z in an observational study
Epidemiological answer: bias and confounding
unlikely
Then we are (philosophically speaking) referring to
a counterfactual theory of causation
RCTs: a counterfactual paradigm
A RCT
DREAM
RCTs

The randomized design fits the counterfactual
framework
 The
control group approaches the counterfactual
situation by design (randomization)
 It is deprived from theory
RCTs


The randomized design fits the counterfactual
framework
It gives decision rule for causation:
 if
a treatment x is randomized, then a difference in
outcome Y between two groups has logically a causal
interpretation because the counterfactual statement
holds
RCT: a counterfactual paradigm


The counterfactual
model and the RCT fit
well
Stop here: don’t move
philosophically further
Epistemic reduction?



The counterfactual model and the RCT fit well
That does not mean that we can only have causal
inference in case of randomization
Reason: the counterfactual model is (as we have
seen) neither sufficient nor necessary to infer
causality
Ontologic reduction?

The counterfactual model and the RCT fit well
A
reason for some to accept only interventions as true
causes (interventionism)

This does not mean that there is only causation in
case of a (randomized) intervention
 Things
you can hardly intervene like SES
Interventionism


Makes causation human-like
Many phenomenon difficult to capture in an
interventionist framework
 Big
Bang
 Volcano Eruptions
 Postulation
of hypothetical interventions (which do not
differ from counterfactuals)
Who is in charge of causation?
Where do we stand today?
What do we believe?


Option 1: Roziglitazone reduces blood glucose level
because the counterfactual (if not X then not Y) is
true?
Option 2: Roziglitazone reduces blood glucose level
because of an intrinsic drug capacity to do so?
Counterfactual theory presupposes
causation

If we adopt option 2:
 The
counterfactual notion is true because x is a cause of
y, and not the other way round
 Counterfactual
thinking can still be used to judge
causality as it is be derived from causality
 But:
it can not be the only argument for causation
How foolish are we today?


This was not shown in a
RCT
We have no ultimate
counterfactual proof
How foolish are we today?

RR roken – myocardinfarct = 2.0

Risico normaliseert niet na weghalen risicofactor
JAMA 1982
We (researchers) have to do more


There are reasons to think that risk factors and drugs
have some inherent effects
But we have to do more for causal inference
Experiments (thinking along the lines of counterfactual
theories is helpful)
 Explaining why x causes y
 Stating the conditions under which x causes y


This means that the philosopher is not in charge of causation
In summary



Counterfactual thinking fits epidemiology, probably
because we use preferably a controlled study
design
Counterfactuals can be helpful when thinking how to
design/perform/analyze a study
Causality can not be reduced to counterfactual
dependence, and causal inference requires more
than counterfactual dependence