Causal inference in neuroimaging Sebastian Weichwald sweichwald.de - - PowerPoint PPT Presentation

c o p e n h a g e n c a u s a l i t y l a b

university of copenhagen

Causal inference in neuroimaging

Sebastian Weichwald

sweichwald.de

@sweichwald

Global Excellence Seminar, DRCMR, Hvidovre 2020-01-24

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Lab Causality Copenhagen

Sebastian Weichwald — CI in NI — Slide 2 Drawing by Anna-Julia Plichta. Keyword cloud created on scimeter.org.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Sebastian Weichwald — CI in NI — Slide 3 Messerli (2012). Chocolate Consumption, Cognitive Function, and Nobel Laureates. New England Journal of Medicine.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Sebastian Weichwald — CI in NI — Slide 3 Messerli (2012). Chocolate Consumption, Cognitive Function, and Nobel Laureates. New England Journal of Medicine.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Bobby goes on a cruise to another country..

Sebastian Weichwald — CI in NI — Slide 3 Messerli (2012). Chocolate Consumption, Cognitive Function, and Nobel Laureates. New England Journal of Medicine.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Bobby goes on a cruise to another country.. seeing: ..and reports back that year’s chocolate consumption.

Sebastian Weichwald — CI in NI — Slide 3 Messerli (2012). Chocolate Consumption, Cognitive Function, and Nobel Laureates. New England Journal of Medicine.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Bobby goes on a cruise to another country.. seeing: ..and reports back that year’s chocolate consumption. doing: ..and brings enormous amounts of chocolate for a year.

Sebastian Weichwald — CI in NI — Slide 3 Messerli (2012). Chocolate Consumption, Cognitive Function, and Nobel Laureates. New England Journal of Medicine.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Bobby goes on a cruise to another country.. seeing: ..and reports back that year’s chocolate consumption. doing: ..and brings enormous amounts of chocolate for a year. Can we predict #country’s Nobel Laureates?

Sebastian Weichwald — CI in NI — Slide 3 Messerli (2012). Chocolate Consumption, Cognitive Function, and Nobel Laureates. New England Journal of Medicine.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

“Correlation does not imply causation.”

Sebastian Weichwald — CI in NI — Slide 4

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

“Correlation does not imply causation.”

seeing vs doing

Sebastian Weichwald — CI in NI — Slide 4

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hippocampal activity in this study was correlated with amygdala activity, supporting the view that the amygdala enhances explicit memory by modulating activity in the hippocampus. amygdala hippocampus explicit memory

Sebastian Weichwald — CI in NI — Slide 5

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hippocampal activity in this study was correlated with amygdala activity, supporting the view that the amygdala enhances explicit memory by modulating activity in the hippocampus. amygdala hippocampus explicit memory Can we enhance explicit memory by amygdala stimulation?

Sebastian Weichwald — CI in NI — Slide 5

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hippocampal activity in this study was correlated with amygdala activity, supporting the view that the amygdala enhances explicit memory by modulating activity in the hippocampus. amygdala hippocampus explicit memory Can we enhance explicit memory by amygdala stimulation?

Sebastian Weichwald — CI in NI — Slide 5

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hippocampal activity in this study was correlated with amygdala activity, supporting the view that the amygdala enhances explicit memory by modulating activity in the hippocampus. amygdala hippocampus explicit memory h Can we enhance explicit memory by amygdala stimulation?

Sebastian Weichwald — CI in NI — Slide 5

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

“Correlation does not imply causation.”

seeing vs doing

Causal questions require causal answers.

Sebastian Weichwald — CI in NI — Slide 6

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What’s the cause and what’s the effect?

Sebastian Weichwald — CI in NI — Slide 7 Mooij, Janzing, Zscheischler, and Schölkopf (2014). CauseEffectPairs repository at webdav.tuebingen.mpg.de/cause-effect/.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What’s the cause and what’s the effect? X (Altitude) → Y (Temperature)

Sebastian Weichwald — CI in NI — Slide 7 Mooij, Janzing, Zscheischler, and Schölkopf (2014). CauseEffectPairs repository at webdav.tuebingen.mpg.de/cause-effect/.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What’s the cause and what’s the effect?

Sebastian Weichwald — CI in NI — Slide 8 Mooij, Janzing, Zscheischler, and Schölkopf (2014). CauseEffectPairs repository at webdav.tuebingen.mpg.de/cause-effect/.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What’s the cause and what’s the effect? Y (Solar Radiation) → X (Temperature)

Sebastian Weichwald — CI in NI — Slide 8 Mooij, Janzing, Zscheischler, and Schölkopf (2014). CauseEffectPairs repository at webdav.tuebingen.mpg.de/cause-effect/.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What’s the cause and what’s the effect?

Sebastian Weichwald — CI in NI — Slide 9 Mooij, Janzing, Zscheischler, and Schölkopf (2014). CauseEffectPairs repository at webdav.tuebingen.mpg.de/cause-effect/.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What’s the cause and what’s the effect? X (Age) → Y (Income)

Sebastian Weichwald — CI in NI — Slide 9 Mooij, Janzing, Zscheischler, and Schölkopf (2014). CauseEffectPairs repository at webdav.tuebingen.mpg.de/cause-effect/.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

“Correlation does not imply causation.”

seeing vs doing

Causal questions require causal answers.

Sebastian Weichwald — CI in NI — Slide 10

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

“Correlation does not imply causation.”

seeing vs doing

Causal questions require causal answers. Correlation(s) may tell us something about causation.

Sebastian Weichwald — CI in NI — Slide 10

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

“Correlation does not imply causation.”

seeing vs doing

Causal questions require causal answers. Correlation(s) may tell us something about causation.

Causal inference: assumptions & data causal hypotheses

Sebastian Weichwald — CI in NI — Slide 10

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Reichenbach’s principle of common cause (1956) If two variables X and Y are statistically dependent then either

Sebastian Weichwald — CI in NI — Slide 11

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Reichenbach’s principle of common cause (1956) If two variables X and Y are statistically dependent then either

X Y I

Sebastian Weichwald — CI in NI — Slide 11

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Reichenbach’s principle of common cause (1956) If two variables X and Y are statistically dependent then either

X Y I X Z Y II

Sebastian Weichwald — CI in NI — Slide 11

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Reichenbach’s principle of common cause (1956) If two variables X and Y are statistically dependent then either

X Y I X Z Y II X Y III

Sebastian Weichwald — CI in NI — Slide 11

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Reichenbach’s principle of common cause (1956) If two variables X and Y are statistically dependent then either

X Y I X Z Y II X Y III

• every statistical dependence is due to a causal relation

Sebastian Weichwald — CI in NI — Slide 11

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Reichenbach’s principle of common cause (1956) If two variables X and Y are statistically dependent then either

X Y I X Z Y II X Y III

• every statistical dependence is due to a causal relation
• cases I, II, and III can also occur simultaneously

Sebastian Weichwald — CI in NI — Slide 11

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Reichenbach’s principle of common cause (1956) If two variables X and Y are statistically dependent then either

X Y I X Z Y II X Y III

• every statistical dependence is due to a causal relation
• cases I, II, and III can also occur simultaneously
• distinction between the 3 cases is a key problem in

scientific reasoning

Sebastian Weichwald — CI in NI — Slide 11

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Metaphor for the local Markov condition

Person X Father Mother Brother Grand- mother

If someone knows the genes of X’s parents, neither the genes

• f the grandmother nor the genes of the brother contain

additional information about X

Sebastian Weichwald — CI in NI — Slide 12

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X =⇒ path between S and X w/o collider

(Markov)

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X =⇒ path between S and X w/o collider

(Markov)

• S ⊥

⊥ Y

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X =⇒ path between S and X w/o collider

(Markov)

• S ⊥

⊥ Y =⇒ path between S and Y w/o collider

(Markov)

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X =⇒ path between S and X w/o collider

(Markov)

• S ⊥

⊥ Y =⇒ path between S and Y w/o collider

(Markov)

• S ⊥

⊥ Y |X

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X =⇒ path between S and X w/o collider

(Markov)

• S ⊥

⊥ Y =⇒ path between S and Y w/o collider

(Markov)

• S ⊥

⊥ Y |X =⇒ all paths between S and Y blocked by X

(faithfulness)

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X =⇒ path between S and X w/o collider

(Markov)

• S ⊥

⊥ Y =⇒ path between S and Y w/o collider

(Markov)

• S ⊥

⊥ Y |X =⇒ all paths between S and Y blocked by X

(faithfulness)

• Can rule out cases such as S → X ← h → Y

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X =⇒ path between S and X w/o collider

(Markov)

• S ⊥

⊥ Y =⇒ path between S and Y w/o collider

(Markov)

• S ⊥

⊥ Y |X =⇒ all paths between S and Y blocked by X

(faithfulness)

• Can rule out cases such as S → X ← h → Y
• Can formally prove that X indeed is a cause of Y

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Hidden confounding and constraint-based CI in NI

• Randomised stimulus S
• Observe neural activity X and Y
• Assume we find
• S ⊥

⊥ X =⇒ path between S and X w/o collider

(Markov)

• S ⊥

⊥ Y =⇒ path between S and Y w/o collider

(Markov)

• S ⊥

⊥ Y |X =⇒ all paths between S and Y blocked by X

(faithfulness)

• Can rule out cases such as S → X ← h → Y
• Can formally prove that X indeed is a cause of Y

Robust against hidden confounding

Sebastian Weichwald — CI in NI — Slide 13 Weichwald et al. (2015). NeuroImage; Grosse-Wentrup et al. (2016). NeuroImage; Weichwald et al. (2016). IEEE Selected Topics in Signal Processing.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Neural Dynamics of Probabilistic Reward Prediction

Sebastian Weichwald — CI in NI — Slide 14 Bach, Symmonds, Barnes, and Dolan (2017). Whole-brain neural dynamics of probabilistic reward prediction. Journal of Neuroscience.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Neural Dynamics of Probabilistic Reward Prediction

Sebastian Weichwald — CI in NI — Slide 14 Bach, Symmonds, Barnes, and Dolan (2017). Whole-brain neural dynamics of probabilistic reward prediction. Journal of Neuroscience.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Neural Dynamics of Probabilistic Reward Prediction

S

Sebastian Weichwald — CI in NI — Slide 15 Bach, Symmonds, Barnes, and Dolan (2017). Whole-brain neural dynamics of probabilistic reward prediction. Journal of Neuroscience.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Causal interpretation of encoding and decoding models Xi ⊥ ⊥ C “Significant variation explained by experimen- tal condition?” Xi ⊥ ⊥ C|X¬i “Does removal impair decoding performance?”

Sebastian Weichwald — CI in NI — Slide 16 Weichwald et al. (2015). Causal interpretation rules for encoding and decoding models in neuroimaging. NeuroImage.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Causal interpretation of encoding and decoding models Xi ⊥ ⊥ C marginal corr “Significant variation explained by experimen- tal condition?” Xi ⊥ ⊥ C|X¬i partial corr “Does removal impair decoding performance?”

Sebastian Weichwald — CI in NI — Slide 16 Weichwald et al. (2015). Causal interpretation rules for encoding and decoding models in neuroimaging. NeuroImage.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Causal interpretation of encoding and decoding models Xi ⊥ ⊥ C marginal corr “Significant variation explained by experimen- tal condition?” Xi ⊥ ⊥ C|X¬i partial corr “Does removal impair decoding performance?” relevant feature ?

cognitive process

Sebastian Weichwald — CI in NI — Slide 16 Weichwald et al. (2015). Causal interpretation rules for encoding and decoding models in neuroimaging. NeuroImage.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What else can go wrong? Cholesterol and Heart Disease diet LDL HDL Heart Disease − +

Sebastian Weichwald — CI in NI — Slide 17 Rubenstein, Weichwald, et al (2017). Causal Consistency of Structural Equation Models. Uncertainty in Artificial Intelligence.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What else can go wrong? Cholesterol and Heart Disease diet Total Chol. Heart Disease − + diet LDL HDL Heart Disease − +

Sebastian Weichwald — CI in NI — Slide 17 Rubenstein, Weichwald, et al (2017). Causal Consistency of Structural Equation Models. Uncertainty in Artificial Intelligence.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

What else can go wrong? Cholesterol and Heart Disease diet Total Chol. Heart Disease − +

⇑

diet LDL HDL Heart Disease − + Macro-variables can be problematic.

Sebastian Weichwald — CI in NI — Slide 17 Rubenstein, Weichwald, et al (2017). Causal Consistency of Structural Equation Models. Uncertainty in Artificial Intelligence.

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Causal inference: assumptions & data causal hypotheses

Sebastian Weichwald — CI in NI — Slide 18

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Causal inference: assumptions & data causal hypotheses

• 4 lectures on causality by J Peters (8 h)

MIT Statistics and Data Science Center, 2017

stat.mit.edu/news/four-lectures-causality

• causality tutorial by D Janzing and S Weichwald (4 h)

Conference on Cognitive Computational Neuroscience 2019

sweichwald.de/ccn2019

• course on causality by D Janzing and B Schölkopf (3 h)

Machine Learning Summer School 2013

mlss.tuebingen.mpg.de/2013/speakers.html

Sebastian Weichwald — CI in NI — Slide 18

u n i v e r s i t y o f c o p e n h a g e n c o p e n h a g e n c a u s a l i t y l a b

Causal inference: assumptions & data causal hypotheses

• 4 lectures on causality by J Peters (8 h)

MIT Statistics and Data Science Center, 2017

stat.mit.edu/news/four-lectures-causality

• causality tutorial by D Janzing and S Weichwald (4 h)

Conference on Cognitive Computational Neuroscience 2019

sweichwald.de/ccn2019

• course on causality by D Janzing and B Schölkopf (3 h)

Machine Learning Summer School 2013

mlss.tuebingen.mpg.de/2013/speakers.html

• Come and talk to us at the

Lab Causality Copenhagen

Sebastian Weichwald — CI in NI — Slide 18