causality - university of cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/mlss09_dawid_1.pdf ·...
TRANSCRIPT
![Page 1: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/1.jpg)
Statistical Causality
Philip DawidStatistical Laboratory
University of Cambridge
![Page 2: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/2.jpg)
Statistical Causality
1. The Problems of Causal Inference2. Formal Frameworks for Statistical Causality3. Graphical Representations and Applications4. Causal Discovery
![Page 3: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/3.jpg)
1. The Problems of Causal Inference
![Page 4: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/4.jpg)
Conceptions of Causality
• Constant conjunction– Deterministic
• Mechanisms– “Physical” causality
Agency– Effects of actions/interventions
Contrast– Variation of effect with changes to cause
![Page 5: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/5.jpg)
Causal Queries
• If I had taken aspirin half an hour ago, would my headache would have gone by now? – “Causes of Effects”, CoE– Counterfactual– LAW
• If I take aspirin now, will my headache be gone within half an hour?– “Effects of Causes”, EoC– Hypothetical– SCIENCE, STATISTICS
![Page 6: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/6.jpg)
Causal Enquiry
• Experimentation (“doing”)– To find out what happens to a system when you
interfere with it you have to interfere with it (not just passively observe it) – George Box
• Observation (“seeing”)– Cost– Ethics– Practicality
• No necessary connexion!
![Page 7: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/7.jpg)
Problems of observational studies
An observed association between a “cause” and an “effect” may be spurious:
– Reverse causation– Regression to mean– Confounding
• common cause• differential selection
– …
![Page 8: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/8.jpg)
Problems of observational studiesThe facts about fuel (Which?, August 2007)Mr Holloway said that a colleague of his used to drive from London to Leeds and back, using Shell petrol to go up there and BP fuel to drive back. He was convinced the BP petrol gave better fuel economy, but Ray had another explanation: ‘I pointed out that Leeds is at a higher altitude that London: he was going uphill one way and downhill the other!’
![Page 9: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/9.jpg)
Problems of observational studiesHRT and coronary artery diseaseObservational research on women taking post-menopausal hormone therapy suggested a 40–50% reduction in coronary heart disease. A large clinical trial found an elevated incidence. Hazard ratio estimates in the observational study were up to 48% lower than in the clinical trial.
![Page 10: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/10.jpg)
Problems of observational studiesVitamin supplements and mortality
Many observational studies appeared to indicate that antioxidant supplements (vitamins A and E, β-carotene) reduce the risk of disease.
Randomized controlled trials showed that they increase mortality.
![Page 11: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/11.jpg)
Problems of observational studiesCalcium channel blockers
Non-experimental studies suggested an increased risk of myocardial infarction associated with the short-acting calcium channel blocker (CCB) nifedapine. It took almost a decade to obtain RCT evidence, which showed that long-acting nifedapine is safe.
![Page 12: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/12.jpg)
Simpson’s Paradox
![Page 13: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/13.jpg)
Causal Inference
• Association is not causation!• Traditionally, Statistics dealt with association
– Theory of Statistical Experimental Design and Analysis does address causal issues
• but of no real use for observational studies
• How to make inferences about causation?– “bold induction”, to a novel context
• Do we need a new formal framework?
![Page 14: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/14.jpg)
2. Formal Frameworks for Statistical Causality
![Page 15: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/15.jpg)
Some Formal Frameworks
• Structural equations• Path diagramsDirected acyclic graphs• …
Probability distributions• Potential responses• Functional relationshipsExtended conditional independence• …
![Page 16: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/16.jpg)
A SIMPLE (??) PROBLEM
• Randomised experiment
• Binary (0/1) treatment decision variable T
• Response variable Y
Define/measure “the effect of treatment”
![Page 17: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/17.jpg)
Probability Model
• Measure effect of treatment by change in the distributionof Y: compare P0 and P1
– e.g. by change in expected response:δ = µ1 − µ0 (average causal effect, ACE)
• Specify/estimate conditional distributionsPt for Y given T = t (t = 0, 1)
• Probability model all we need for decision theory– choose t to minimise expected loss EY » Pt
{L(Y)}
(Fisher)
[e.g. N(µt, σ2) ]
![Page 18: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/18.jpg)
T
Decision Tree
0
1
y
y
Y
Y
Y~P0
Y~P1
L(y)
L(y)
Influence Diagram
T YY | T=t ~ Ptt
LL(y)
![Page 19: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/19.jpg)
• Then value of Y would change by exactly δ = µ1 − µ0
Structural Model
• Deterministic relationship• Value of E for any unit supposed the
same if we were to change T from 0 to 1
– individual causal effect (ICE)
(E = “error”, “effect of omitted variables”,…)
Y = µT + E[e.g., E » N(0, σ2)]
![Page 20: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/20.jpg)
• Split Y in two:Y0 : potential response to T = 0Y1 : potential response to T = 1
Potential Response Model (Rubin)
– necessarily unobservable!
• Consider (for any unit) the pair Y = (Y0 , Y1)– with simultaneous existence and joint distribution
• Treatment “uncovers” pre-existing response:Y = YT (determined by Y and T)
– other PR unobservable, “counterfactual”• Unit-level (individual) [random] causal effect
Y1 − Y0
![Page 21: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/21.jpg)
General Functional Model
• Value of U supposed the same , independent of value of T– and of whether we intervene or just observe
• Formally includes:– Structural model: U = E, Y = µT + E– PR model: U = Y, Y = YT
(U = “unit characteristics”)Y = f(T, U)
![Page 22: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/22.jpg)
• Any functional model Y = f(T, U) generates a PR model: Yt = f(t, U)
• Any PR model generates a probability model:Pt is marginal distribution of Yt (t = 0, 1)
• Distinct PR models can generate the samestatistical model– e.g., correlation between Y0 and Y1 arbitrary
• Cannot be distinguished observationally• Can have different inferential consequence
– can be problematic!
Potential Response Model
![Page 23: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/23.jpg)
Potential Responses: Problems
NB: ρ does not enter! – can never identify ρ– does this matter??
• Corresponding statistical model:
• PR model:
![Page 24: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/24.jpg)
Potential Responses: ProblemsUnder PR model:
We can not identify the variance of the ICE
We can not identify the (counterfactual) ICE, after observing response to treatment
E(Y1/Y0) depends on ρ We can not estimate a “ratio” ICE
var(Y1 − Y0) = 2(1 − ½) ¾2
E(Y1 − Y0 | Y1 = y1) = (1 − ½) y1 + (½µ1 − µ0)
![Page 25: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/25.jpg)
OBSERVATIONAL STUDY
• Treatment decision taken may be associated with patient’s state of health
• What assumptions are required to make causal inferences?
• When/how can such assumptions be justified?
![Page 26: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/26.jpg)
Functional Model
T = treatment receivedU = “unit characteristics”
– value supposed unaffected by treatment or how it is applied
– but could influence choice of treatment T observational dependence between T and U
),( UTfY =
Response to applied treatment t: Yt = f(t, U).Observational distribution of Y, given T = t, same as distribution of Yt if
![Page 27: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/27.jpg)
T Y
U
Functional Model
U = “unit characteristics”– value supposed unaffected by treatment or how it
is applied– but could influence treatment choice
),( UTfY =
UPU ~
![Page 28: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/28.jpg)
T Y
U
Functional Model
“No confounding” (“ignorable treatment assignment”) if
(treatment independent of “unit characteristics”)
TPT ~ ),( UTfY =
UPU ~
![Page 29: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/29.jpg)
T Y
Y
“No confounding” (“ignorable treatment assignment”) if
(treatment independent of potential responses)
TPT ~
PR interpretation (U = Y)
TYY =
YY P~
![Page 30: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/30.jpg)
PR interpretation (U = Y)
• Value of Y = (Y0, Y1) on any unit supposed the same in observational and experimental regimes, as well as for both choices of T
• No confounding: independence of Tfrom PR pair Y
How are we to judge this??
![Page 31: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/31.jpg)
• “Treatment regime indicator” variable FT– intervention variable– non-random, parameter
• Values:FT = 0 : Assign treatment 0 (⇒ T = 0)FT = 1 : Assign treatment 1 (⇒ T = 1)FT = ; : Just observe (T random)
Statistical Decision Model
(Point intervention: can generalize)
![Page 32: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/32.jpg)
• Causal target: comparison of distributions of Ygiven FT = 1 and given FT = 0– e.g., E(Y | FT = 1) − E(Y | FT = 0)
average causal effect, ACE
• Causal inference: assess this (if possible) from properties of observational regime, FT = ;
Statistical Decision Model
![Page 33: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/33.jpg)
True ACE is
E(Y | T = 1, FT = 1) − E(Y | T = 0, FT = 0)Its observational counterpart is:
E(Y | T = 1, FT = ;) − E(Y | T = 0, FT = ;)“No confounding” (ignorable treatment assignment)
when these are equal.
Can strengthen:p(y | T = t, FT = t) = p(y | T = t, FT = ;)
distribution of Y | T the same in observational and experimental regimes
Statistical Decision Model
![Page 34: CAUSALITY - University of Cambridgemlg.eng.cam.ac.uk/mlss09/mlss_slides/MLSS09_Dawid_1.pdf · Statistical Causality 1. The Problems of Causal Inference 2. Formal Frameworks for Statistical](https://reader031.vdocument.in/reader031/viewer/2022031507/5c944c9b09d3f27d4f8c8c94/html5/thumbnails/34.jpg)
Extended Conditional Independence
Distribution of Y | T the same in observational and experimental regimes:
Y | (FT, T) does not depend on value of FT
Can express and manipulate using notation and theory of conditional independence:
(even though FT is not random)