causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and...
TRANSCRIPT
![Page 1: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/1.jpg)
Causal reasoning withthe „do“-operator
Björn MederMax Planck Institute for Human Development
York HagmayerKing‘s College London
Michael WaldmannUniversity of Göttingen
![Page 2: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/2.jpg)
Why causal knowledge?
Causal knowledge helps us to understand, predict, and controll our environment
Causal reasoning involves various types of (probabilistic) conditional inferences Diagnostic inferences Predictive inferences Reasoning about interventions
germs
infection
symptoms course
Treatment:vaccination, antibiotics
![Page 3: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/3.jpg)
Hume‘s riddle of causal induction
“The idea of cause and effect is deriv’d fromexperience, which informs us, that such particular objects, in all past instances, havebeen constantly conjoin’d with each other.” (David Hume, 1739/2000, p.63)
How can you find out whether you are allergic against some kind of food?
Problem: Our senses do not provide us with causal knowledge, i.e. causal relations are not directly observable
?Food Symptoms
![Page 4: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/4.jpg)
Claim: Causal knowledge can be reduced to associative knowledge (e.g., Shanks & Dickinson, 1987; Dickinson, 2001)
Problems Associative knowledge does not allow us to distinguish between causal
and spurious relations Associative models do not represent causal directionality (causes
generate effects) As a consequence, associative models lack the expressive power to
capture normative differences between observations and interventions
Causal learning = associative learning?
V V
Cues(Causes)
Outcomes(Effects)
Associative weights V
![Page 5: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/5.jpg)
Associative models: No causal structure
Cues Outcomes
Headache Associative Weights
Cough
Fever
Chills
Infection
Influenza Virus
Rhino virus
Point of intervention
Point of intervention
![Page 6: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/6.jpg)
Causal model theory
Causal model theory assumes that people represent the world in terms of causal relations that mirror the causal texture of the (physical, biological, social) environment (Waldmann & Holyoak, 1992; Waldmann, 1996; Sloman, 2005; Waldmann, Hagmayer & Blaisdell, 2006; Meder, Hagmayer, & Waldmann, 2008, 2009)
In the spirit of Kant, cause and effect are considered as fundamental categories of thought. Asymmetry of causal relations is preserved in mental representations
Causal structure vs. causal strength: distinction between the structure of a causal model and its parameters (base rates, causal strength estimates)
![Page 7: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/7.jpg)
Formalizing causal models: Bayes Nets(Pearl, 1988, 2000; Spirtes, Glymour & Scheines, 1993; Spohn, 1980, 1983)
Directed acyclic graphs (DAGs) to represent causal structures
Parameters to express the strength of these relations (e.g., marginal and conditional probabilities, causal strengths)
Causal Markov condition: ∏ |
Common Cause Model
CB
A
P(A,B,C) =P(A) P(B|A) P(C|A)
Causal Chain Model
CB
A
P(A,B,C) =P(A) P(B|A) P(C|B)
Common Effect Model
CB
A
P(A,B,C) =P(A) P(B) P(C|A,B)
![Page 8: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/8.jpg)
Normative difference between inferences based on passively observed states of a variable
(“seeing”) inferences based on the very same state resulting from an
external intervention (“doing”)
atmosphericpressure
storm
barometer
Seeing vs. doing: Observational and interventional inferences
Intervention
Observation Intervention
atmosphericpressure
storm
barometer
![Page 9: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/9.jpg)
Obs
erva
tion
(„se
eing
“)In
terv
entio
n(„
doin
g“)
Common Cause Model
CB
A
CB
A
do(B)
Causal Chain Model
CB
A
CB
A
do(B)
Common Effect Model
CB
A
CB
A
do(B)
Reasoning about interventions formalized by Pearl‘s (2000) „do-operator“(cf. Spirtes, Glymour & Scheines, 1993; Dawid, 2002)
Observational inference: IF B is observed, THEN...
Interventional inference: IF B is avtively generated, THEN...
Seeing vs. Doing: Observational and interventional inferences
![Page 10: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/10.jpg)
Are people sensitive to the normative distinction between observations and interventions?
Can people infer the consequences of interventions on causal systems?
Requires to
consider the causal model underlying the observed covariations distinguish observing (“seeing”) from intervening (“doing”)
Seeing vs. doing: Observational and interventional inferences
![Page 11: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/11.jpg)
Common Cause Model
CB
A
Causal Chain Model
CB
A
Seeing vs. doing: Observational and interventional inferences
CB
A
0.5
0.10.1 CB
A
0.5
0.1
0.1
DATA DATA=
Waldmann & Hagmayer, 2005
![Page 12: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/12.jpg)
Common Cause Model
CB
A
0.5
0.10.1
Causal Chain Model
CB
A
0.5
0.1
0.1
Seeing vs. doing: Observational and interventional inferences
Observational inference: If you observe B to be present, how likely is C?
Interventional inference: If you generate B by an means of intervention, how likely is C?
![Page 13: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/13.jpg)
0.0
0.2
0.4
0.6
0.8
1.0
Causal model Empirical
P(C|B
)
Observation Intervention
0.0
0.2
0.4
0.6
0.8
1.0
Causal model Empirical
P(C|B
)
Observation Intervention
Common Cause Model
CB
A
0.5
0.10.1
Causal Chain Model
CB
A
0.5
0.1
0.1
Seeing vs. doing: Observational and interventional inferences
![Page 14: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/14.jpg)
0.0
0.2
0.4
0.6
0.8
1.0
Causal model Empirical
P(C|B
)
Observation Intervention
0.0
0.2
0.4
0.6
0.8
1.0
Causal model Empirical
P(C|B
)
Observation Intervention
Common Cause Model
CB
A
0.5
0.10.1
Causal Chain Model
CB
A
0.5
0.1
0.1
Seeing vs. doing: Observational and interventional inferences
![Page 15: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/15.jpg)
Common Cause Model
CB
A
CB
A
Causal Chain Model
CB
A
CB
A
D
CB
A
Confounder Model
D
CB
A
Common Effect Model
CB
A
CB
A
Obs
erva
tion
Inte
rven
tion
do(B) do(B) do(B) do(B)
Observations and interventions with causal backdoors
![Page 16: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/16.jpg)
D
C
B
AA
C
B
D D
C
B
AA
C
B
D
Observations and interventions with causal backdoors
![Page 17: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/17.jpg)
D
C
B
C
Observing („seeing“)
D
C
B
C
Intervening („doing“)
A AA
B
D DA
B
?
?
?
?
?
?
?
?
Observations and interventions with causal backdoors
P(D|C) = P(D|do C)
Observational probability
Interventionalprobability
P(A|C) > P(A|do C)
P(B|C) > P(B|do C)
![Page 18: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/18.jpg)
D
C
B
C
Observing („seeing“)
D
C
B
C
Intervening („doing“)
A AA
B
D A
B
D?
?
?
?
?
?
?
?
Observations and interventions with causal backdoors
P(D|¬C) < P(D|do ¬C)
Observational probability
Interventionalprobability
P(A|¬C) < P(A|do ¬C)
P(B|¬C) < P(B|do ¬C)
![Page 19: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/19.jpg)
Are people sensitive to the normative difference between observations and interventions when reasoning with causal backdoors?
Diagnostic and predictive inferences
Observational and interventional inferences with backdoor paths
D
C
B
AA
C
B
D D
C
B
AA
C
B
D
![Page 20: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/20.jpg)
?
?
?
?
Lekanoid
Ceranat
Renoxin Desulfan
Observational Learning
![Page 21: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/21.jpg)
?
?
?
?
Lekanoid
Ceranat
Renoxin Desulfan
Observational Learning
![Page 22: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/22.jpg)
0
0.2
0.4
0.6
0.8
1
C ¬C C ¬CObservation Intervention
* n.s.
Prob
abili
ty o
f A
A D
C
B
People understand the difference between observations and interventions in diagnostic reasoning from effect to cause
Reasoning from C to A
?
Results: Diagnostic Inferences(Meder, Hagmayer & Waldmann, 2008)
![Page 23: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/23.jpg)
Results: Predictive inferences(Meder, Hagmayer & Waldmann, 2008)
A C
C
B
0
0.2
0.4
0.6
0.8
1
C ¬C C ¬CObservation Intervention
*
Prob
abili
ty o
f D
The influence of the confounding causal pathway is takeninto account when reasoning from C to D
Reasoning from C to D
?
![Page 24: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/24.jpg)
People are sensitive to the normative difference between conditional inferences based on observed states of of the world, and the very same states generated by means of intervention
Diagnostic and predictive conditional inferences take into account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009)
People reason differently about causal and logical conditionals („A causes B“ vs. „If A then B“) (Sloman & Lagnado, 2005)
Causal reasoning with the “do”-operator
![Page 25: Causal reasoning with the „do“-operatorpfn23853/rffcw/mhw.pdf · account causal structure and data (Waldmann & Hagmayer, 2005; Meder, Hagmayer, & Waldmann, 2008, 2009) People](https://reader034.vdocument.in/reader034/viewer/2022050207/5f97da6288d92037a530595b/html5/thumbnails/25.jpg)
The rationality of causal conditionals
People reason differently about causal and logical conditionals (Sloman & Lagnado, 2005)
If cause then effect.No effect.Therefore, no cause.
Logical theories—like associative or purely probabilistic models—lack the expressive power to distinguish between observation and intervention