Causal inference – some aspects
[email protected] Department of Public Health and Caring Sciences Uppsala University, Sweden
The most significant purpose of epidemiology is to acquire knowledge of causal mechanisms that form a basis for preventive measures against diseases not currently preventable MacMahon&Pugh, 1970
Causal inference = conclude whether the association is causal or not
Judea Pearl. Causality. Models reasoning, and inference. Cambridge University Press 2000, 2nd edition 2009
Sander Greenland, James Robins, Miguel Hernan, Charlie Pool, Timothy Lash, Ken Rothman, Jan Vandenbroucke, Jay Kaufman, Malcolm Maclure Tyler VanderWeele, Danella Hafeman, … Modern Epidemiology: Harvard, UCLA, Chapel Hill, …
Epiloque: historical views on causality • The Bible – facts, predetermined – Explanations: merely for passing on responsibility – Final causes (Aristotle)
• Engineering – even objects could have causal character
• Galileo (1600) – First how then why. Algebra the language
• Hume (1700) – the final why is superfluous, causal connections the product of observations
How: Natural laws, symmetrical Russel: causality obselete, like the monarchy
F = ma , am = F , F/a = m But engineering: directionality Galton, Pearson - associations, contingeny tables Fischer - randomisation AND intervention Pearl: Artifical intelligence and computor science Graphs to handle causality in complex observational data
Philosophy of science • Epistemology: how to know • How to get knowledge – how to be a scientist • Simplistic: inductivion vs falsification • Rothman KJ (ed). Causal Inference. Chestnut Hill, ERI, 1988
Induction • Compare with deduction • Francis Bacon, the father of empiricism • The more times you observe an association (and the better the quality of the observation) the more probable it is that the association is causal • ”with empirical bricks you build the castle of knowledge” • Criteria: Hill’s guidelines • Methods: better studies, concensus meetings, systematic reviews, metaanalysis, Cochrane
Refutationism (falsification) • (David Hume) Karl Popper – ”modern epidemiology”
• The truth is not eternal, only valid until disapproved • Skepticism, identify the alternative hypotheses and the critical test
Hill’s guidelines (1965) from Doll R (2002)
Strength of the association Consistency Dose response Temporal relationship Biological plausibility Coherence of evidence Experiment Analogy Specificity
Causal inference – two major questions according to Pearl • What empirical evidence is required for legitimite inference of causal effect relationships? • Given that we are willing to to accept causal information about a phenomenon, what inferences can be drawn (and how)?
Causal models: Conditional counterfactuals -potential outcome models Sufficient-component cause model Structural equation modelling Graph models: Directed acyclic graphs DAG
Causal parameters
But first … An important feature of the object of most
epidemiological investigations…
Multicausality
246 suggested risk factors of acute myocardial infarction (Hopkins & Williams. Atherosclerosis 1981)
Koch-Henle principles the cause should be found in all cases (necessary) cultivation of cause outside the body the cultivated cause should reproduce disease (sufficient)
Multicausality all in a row as a single causal chain , all necessary and sufficient or another model?
Multicausality • Most (every) diseases have many causes. • Ex: All smokers don’t get lung cancer and all cases of lung cancer have not smoked. (not even smoking is a sufficient or a necessary cause)
• There must be a third type of cause – Sufficient cause – Necessary cause – Contributing cause (=component cause)
• The analytical strategies in epidemiology are simplistic • Separate theoretical subject matter questions from methodological issues • Mostly one occurrence relation at a time (or 2) in the empirical analysis • The theoretical model is build in theory and empirically supported piece by piece • Like cutting ”white holes” of epidemiology in the ”black curtain”
S
U
T O
N
Q
P
H
I
E
V R
J
F
K
L
M
G A
B
C
D = disease
Some components of a web of causation. Modified from fig 2 in chapter 2 in MacMahon and Pugh (1970). D=Disease. A-C and E-V are component causes. The sequence U-R-K-G-A-D is an example of a causal chain. I is interacting with K in producing G.
K
?
D = disease
I
?
K
?
D = disease
S
U
T O
N
Q
P
H
I
E
V R
J
F
K
L
M
G A
B
C
D = disease
Some components of a web of causation. Modified from fig 2 in chapter 2 in MacMahon and Pugh (1970). D=Disease. A-C and E-V are component causes. The sequence U-R-K-G-A-D is an exaple of a causal chain. I is interacting with K in producing G.
By moving the white holes like looking glasses
we may zip from the black box in our theories
First question: The quality of the observations
used for making causal inference
Conditional counterfactuals
Question: is smoking causing lung cancer? • Analytical strategy – Follow up smokers: do they get lung cancer? – A coincidence? Which control information?
• The ideal: the counter factual situation – To follow up the same persons during the same time but without smoking. But impossible.
• The practical solution: – To compare with non-smokers in the study base
Comparability! • Restriction • Stratified analyses and statistical adjustment • Forward or backward selection or significance testing, change in estimate • Instrumental variables, propensity score matching • Case only designs • DAG and overadjustment
Instrument variables To simulate the randomisation of exposure in RCT
U (unknown condounders) Z (instrument)
E (exposure
Three criteria 1. Z has a causal effect on E 2. Z effects D only through E 3. U may not cause Z
Examples a. The price of tobacco b. ”natural experiments” c. Randomization in itself d. Mendelian randomization
D (disease)
The Janus face of statistical adjustment • Confounders versus colliders • Janszky I, Ahlbom A, Svensson AC. Eur J Epidemiol 2010;25:361-363
• DAG: Directed acyclic graphs • Shrier I, Platt RW. Reducing bias through directed acyclic graphs. BMC Medical Research Methodology. 2008;8:70
Causal parameters, causal contrast, effect measure • An intuitive understanding of the characteristics of the causal relation • Contrast of a target under two exposure distributions, the counterfactual approach • Not correlation coefficients, percent of variance explained, p-values, chi-2 statstics, and standardized regression coefficients
The sufficient-component cuase model helps with • Understanding the meaning of the strenght of a risk factor • Understanding measures like attributable fraction • How to chose a study base • Understanding how the synergy (interaction) between risk faktors shall be examined • Identifying interventiv alternatives
The sufficient-component cause model Abstract example: a disease with 3 sufficient causes
A
E D
A
H B
G
C
B
I
C
F
F
Empirical example: causes of tuberculosis
B
A E
C
A F
A
J
D
A G
H
A
A= tbb B= nutrition C= housing D= measles E-H= other
Simplistic:
A sufficient cause with E
C
E = exposure C = the causal complement of E
First, the epidemiological task of finding empirical evidence supporting that a specific exposure (E) is causal: persons
I1 E
E U1
U2
time The ideal: the counterfactual situation
The usual cohort study:
persons
I1 E
E U1
U2
I0 E
U3 time
E U1
U3
U2
E
U1
I1 – I0 = RD
(RD > 0)
(I1 – I0) / I0 = I1 / I0 – 1 = RR – 1
(RR > 1)
The strength of a risk factor E
C’ C’= the component causes constituting the causal complement of E.
RR = C
C = the class of all other sufficient causes of the Outcome.
RR (or RD) is not a biological characteristic of a risk factor. The ”strength” depends on the availability of other component causes of the outcome in the population. Ken Rothman about strenght: ”if it plays a causal role in large proportion of cases”
Now, the combined action of two exposures: persons
IAB AB
AB
A
B U
A
A
U
B
U
U
IAB U
U
IAB AB
B
U
U
IAB AB
U
time
A
The empirical criterion:
A
B
>0
U
B U
A A
B U
U
A
U
-
= B
U
B
U
U
+
U
U
U
IAB – IAB - IAB + IAB = (IAB – IAB) – (IAB - IAB) > 0 Risk difference (RD) over strata of the other variable, or Departure from additivity of absolute risks
Why is the sufficient-component model helpful? With the help of the model we can define interaction = two causes acting together in the same sufficient cause Further an empirical criterion of interaction can be derived through the model (departure from additivity of absolute risks)
No interaction = no synergy = independent effects (RERI=0 or AP=0
or SI=1)
Means: - The two component causes are not part of the same sufficient cause - Risk evaluaton could be done separately for each risk factor (witout considering the effect of the other) as they don’t increase each other’s effect.
Interaction between causes - Combined (or joint) action of two causes - Synergy (or antagonism) - Conditional causation - Specific susceptibility
• The lay perspective on conditional causation as an important feature of multicausation • Let there be light!
The advantage of interaction, synergy or conditional causation in a multicausal structure is for example -It provides intervention alternatives
-Everything may be explained several times -There is never only a certain fraction left to explain
-Unavoidable risk factors may avoidable effects
It is a phenomenon of the real world. Does it apply to disease causation? - smoking + exposure to asbest - mutation + diet
lung cancer
PKU (gene-environment interactions)
- psychological demands + decision latitude (job strain) - atherosclerosis + thrombosis
CHD
- low birth weight + later obesity
CHD
Suggested: most causes work together with other causes - and this is something that we would like to discover
Causal models • The sufficient-component cause model (causal fields) Mackie, Rothman
• The counterfactual approach (respons types)
• Graphical models (DAG: directed acyclic graphs)
