neuroeconomics { where science and economics meet
TRANSCRIPT
Neuroeconomics – where science and economicsmeet
Agnieszka Tymula
2017 European Economic Science Association Meeting
24,649 papers as of 21/08/2017
Outline
1. How does brain encode value?
2. Examples of neuroscience informing economics
neuroscience → economics
I divisive normalizationI choice set effectsI Prospect Theory-like behaviors
I gray matter volume
3. Neuroeconomics - useful resources
Outline
1. How does brain encode value?
2. Examples of neuroscience informing economics
neuroscience → economics
I divisive normalizationI choice set effectsI Prospect Theory-like behaviors
I gray matter volume
3. Neuroeconomics - useful resources
Outline
1. How does brain encode value?
2. Examples of neuroscience informing economics
neuroscience → economics
I divisive normalizationI choice set effectsI Prospect Theory-like behaviors
I gray matter volume
3. Neuroeconomics - useful resources
Outline
1. How does brain encode value?
2. Examples of neuroscience informing economics
neuroscience → economics
I divisive normalizationI choice set effectsI Prospect Theory-like behaviors
I gray matter volume
3. Neuroeconomics - useful resources
Outline
1. How does brain encode value?
2. Examples of neuroscience informing economics
neuroscience → economics
I divisive normalizationI choice set effectsI Prospect Theory-like behaviors
I gray matter volume
3. Neuroeconomics - useful resources
Outline
1. How does brain encode value?
2. Examples of neuroscience informing economics
neuroscience → economics
I divisive normalizationI choice set effectsI Prospect Theory-like behaviors
I gray matter volume
3. Neuroeconomics - useful resources
How does brain encode value?
I Cortex consists of neurons
I Neurons produce electrochemical impulses called actionpotentials
I Neurons convey information by the rate at which they produceaction potentials - firing rate (up to 200 Hz)
I Activity of neurons can be precisely recorded using implantedelectrodes and approximated using fMRI
I Cortex consists of neurons
I Neurons produce electrochemical impulses called actionpotentials
I Neurons convey information by the rate at which they produceaction potentials - firing rate (up to 200 Hz)
I Activity of neurons can be precisely recorded using implantedelectrodes and approximated using fMRI
I Cortex consists of neurons
I Neurons produce electrochemical impulses called actionpotentials
I Neurons convey information by the rate at which they produceaction potentials - firing rate (up to 200 Hz)
I Activity of neurons can be precisely recorded using implantedelectrodes and approximated using fMRI
I Cortex consists of neurons
I Neurons produce electrochemical impulses called actionpotentials
I Neurons convey information by the rate at which they produceaction potentials - firing rate (up to 200 Hz)
I Activity of neurons can be precisely recorded using implantedelectrodes and approximated using fMRI
Subjective value brain map
Bartra, McGuire & Kable, NeuroImage, 2013
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)xmin xmax
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)xmin xmax
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)xmin xmax
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)xmin xmax
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)xmin xmax
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)xmin xmax
x1 x2x x
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)xmin xmax
x1 x2x x
Four very important facts about neurons:
1. Firing rate is biophysically restricted by a maximum rate ofaction potentials
2. We have limited number of neurons in the brain
3. Brain activity is stochastic
4. Neurons are expensive - brain takes 3% of body mass and20% of total calorie consumption
sub
jec�
ve
va
lue
(H
z)
reward (x)xmin xmax
x1 x2
Neuroeconomic models ofdecision-making
Three dominant models of choice in neuroeconomics:
1. Divisive normalizationI Glimcher, Heeger
2. Stochastic network modelsI Rustichini, Padoa-Schioppa
3. Drift-diffusion modelI Rangel, Shadlen
I Divisive normalization originates from enormous amount ofresearch on sensory systems
I Perception of sensory stimulus depends not only on itsobjective features but also on:
I other currently presented stimuli (spatial normalization /choice-set effects)
I previously experienced stimuli (temporal normalization / riskattitudes)
Three dominant models of choice in neuroeconomics:
1. Divisive normalizationI Glimcher, Heeger
2. Stochastic network modelsI Rustichini, Padoa-Schioppa
3. Drift-diffusion modelI Rangel, Shadlen
I Divisive normalization originates from enormous amount ofresearch on sensory systems
I Perception of sensory stimulus depends not only on itsobjective features but also on:
I other currently presented stimuli (spatial normalization /choice-set effects)
I previously experienced stimuli (temporal normalization / riskattitudes)
Three dominant models of choice in neuroeconomics:
1. Divisive normalizationI Glimcher, Heeger
2. Stochastic network modelsI Rustichini, Padoa-Schioppa
3. Drift-diffusion modelI Rangel, Shadlen
I Divisive normalization originates from enormous amount ofresearch on sensory systems
I Perception of sensory stimulus depends not only on itsobjective features but also on:
I other currently presented stimuli (spatial normalization /choice-set effects)
I previously experienced stimuli (temporal normalization / riskattitudes)
Spatial normalization
Ri (x1, x2, ..., xn) = rmaxxαi
σα+∑
j 6=i Xαj
+ εCarandini & Heeger, Nature Reviews, 2012
This general form occurs from retina to cortex and mediatescontrast adaptation, surround suppression, visual attention,multisensory integration, ...
Spatial normalization
Ri (x1, x2, ..., xn) = rmaxxαi
σα+∑
j 6=i Xαj
+ εCarandini & Heeger, Nature Reviews, 2012
This general form occurs from retina to cortex and mediatescontrast adaptation, surround suppression, visual attention,multisensory integration, ...
Spatial normalization
Ri (x1, x2, ..., xn) = rmaxxαi
σα+∑
j 6=i Xαj
+ εCarandini & Heeger, Nature Reviews, 2012
This general form occurs from retina to cortex and mediatescontrast adaptation, surround suppression, visual attention,multisensory integration, ...
Spatial normalization
Ri (x1, x2, ..., xn) = rmaxxαi
σα+∑
j 6=i Xαj
+ εCarandini & Heeger, Nature Reviews, 2012
This general form occurs from retina to cortex and mediatescontrast adaptation, surround suppression, visual attention,multisensory integration, ...
Spatial normalization
Ri (x1, x2, ..., xn) = rmaxxαi
σα+∑
j 6=i Xαj
+ εCarandini & Heeger, Nature Reviews, 2012
This general form occurs from retina to cortex and mediatescontrast adaptation, surround suppression, visual attention,multisensory integration, ...
Spatial normalization
Ri (x1, x2, ..., xn) = rmaxxαi
σα+∑
j 6=i Xαj
+ εCarandini & Heeger, Nature Reviews, 2012
This general form occurs from retina to cortex and mediatescontrast adaptation, surround suppression, visual attention,multisensory integration, ...
Spatial normalization in valuation
Normalization outperforms other models in predicting neuronalfiring rates (Louie et al., 2011)
It leads to specific choice and brain activity patterns - now testedin monkeys and humans (Louie et al., 2013, 2015, Webb et al.,2012, 2014)
u1u2u3
u1u2
u3 u4u5
It is optimal maximization tool under costly value representation,equivalent to reducing Shannon entropy (Steverson et al. 2016)
References: Kenway Louie at NYU
Temporal normalization
Stimulus history has strong effect in sensory processing
Temporal normalization
Stimulus history has strong effect in sensory processing
Temporal normalization
contrast level
!ri
ng
ra
te
Ohzawa, Scalar & Freeman, Journal of neurophysiology, 1985
R(x) = rmaxxα
xα+t−1∑i=1
γ(1−γ)t−ixi
LoFaro et al., Letters in Biomathematics, 2014
Louie et al., Journal of Neuroscience, 2014
Temporal normalization
contrast level
!ri
ng
ra
te
Ohzawa, Scalar & Freeman, Journal of neurophysiology, 1985
R(x) = rmaxxα
xα+t−1∑i=1
γ(1−γ)t−ixi
LoFaro et al., Letters in Biomathematics, 2014
Louie et al., Journal of Neuroscience, 2014
Temporal normalization of value
Hypothesis: After high (low) adapt block, bids in test block will belower (higher). This effect will gradually wear off.
Khaw, Glimcher, Louie, under revision
Temporal normalization of value
Hypothesis: After high (low) adapt block, bids in test block will belower (higher). This effect will gradually wear off.
Khaw, Glimcher, Louie, under revision
Temporal normalization of value
Khaw, Glimcher, Louie, under revision
Expected Subjective Value Theory: A Representation ofDecision in Time under Risk and Certainty
by Paul Glimcher and Agnieszka Tymula
Theoretically explore the implications of divisive normalizationalgorithm for:
I risk attitudes,
I loss aversion, and
I probability weighting
Model
The subjective value (V ) of a prize x ∈ IR+ at time t is given by:
V (x) =xα
xα + Mαt
α - predisposition (free parameter)Mt - reference point / reward expectation
Mt =t−1∑i=1
γ(1− γ)t−ixi
γ ∈ (0, 1) sets the forgetting rate
The expected subjective value of a lottery (x , p) with outcomesx1, x2, ..., xn occurring with corresponding probabilitiesp1, p2, ..., pn:
EV (x , p) =n∑
i=1
piV (xi )
The subjective value (V ) of a prize x ∈ IR+ at time t is given by:
V (x) =xα
xα + Mαt
α - predisposition (free parameter)Mt - reference point / reward expectation
Mt =t−1∑i=1
γ(1− γ)t−ixi
γ ∈ (0, 1) sets the forgetting rate
The expected subjective value of a lottery (x , p) with outcomesx1, x2, ..., xn occurring with corresponding probabilitiesp1, p2, ..., pn:
EV (x , p) =n∑
i=1
piV (xi )
Examples of subjective value functions: predisposition
0.5
1V(x)
5x
α=1, M=5
α=2, M=5
α=5, M=5
α=10, M=5
A
V ∈ [0, 1]
Examples of subjective value functions: reference point
0.5
1V(x)
3 5 7x
α=5, M=5
α=5, M=3
α=5, M=7
B
For all Mt ,V (Mt) = 12
Risk taking
Arrow-Pratt index of relative risk aversion (RRA):
RRAt = −xV ′′
V ′=
(1 + α)xα − (α− 1)Mαt
xα + Mαt
Proposition 1 Individual is risk averse if
x > α
√α− 1
α + 1Mt
I predisposition (α), in addition to the reference (Mt),determines risk attitude
Arrow-Pratt index of relative risk aversion (RRA):
RRAt = −xV ′′
V ′=
(1 + α)xα − (α− 1)Mαt
xα + Mαt
Proposition 1 Individual is risk averse if
x > α
√α− 1
α + 1Mt
I predisposition (α), in addition to the reference (Mt),determines risk attitude
Predisposition and risk attitude
x > α
√α− 1
α + 1Mt
α ≤ 1
α ∈ (1,∞) α =∞
risk averse for x > 0
x > α
√α−1α+1Mt risk averse for x > Mt
0.2
.4.6
.81
v(x)
0 2 4 6 8 10x
Concave u�lity for
all x
(Bernoulli)
0.2
.4.6
.81
v(x
)
0 2 4 6 8 10x
Prospect Theory
Predisposition and risk attitude
x > α
√α− 1
α + 1Mt
α ≤ 1
α ∈ (1,∞)
α =∞risk averse for x > 0
x > α
√α−1α+1Mt
risk averse for x > Mt
0.2
.4.6
.81
v(x)
0 2 4 6 8 10x
Concave u�lity for
all x
(Bernoulli)
0.2
.4.6
.81
v(x
)
0 2 4 6 8 10x
Prospect Theory
Predisposition and risk attitude
x > α
√α− 1
α + 1Mt
α ≤ 1 α ∈ (1,∞) α =∞risk averse for x > 0 x > α
√α−1α+1Mt risk averse for x > Mt
0.2
.4.6
.81
v(x)
0 2 4 6 8 10x
Concave u�lity for
all x
(Bernoulli)
0.2
.4.6
.81
v(x
)
0 2 4 6 8 10x
Prospect Theory
Reference point and risk attitudeProposition 4 For any given gamble, individual is more risktolerant the larger his reference point is.
Proof: ∂RRAt
∂Mt< 0
0.5
1V(x)
3 5 7x
α=5, M=5
α=5, M=3
α=5, M=7
B
A kind of “wealth effect”, but not exactly
Reference point and risk attitude
Risk attitudes depend not only on the total rewards earned butalso their timing and value
total t=1 t=2 t= 3
increasing income 30 5 10 15 more risk tolerantdecreasing income 30 15 10 5 more risk averse
Proposition 5 Keeping the total past income and predispositionconstant, increasing income streams lead to more risk tolerancethan decreasing income streams.
Reference point and risk attitude
Risk attitudes depend not only on the total rewards earned butalso their timing and value
total t=1 t=2 t= 3
increasing income 30 5 10 15 more risk tolerantdecreasing income 30 15 10 5 more risk averse
Proposition 5 Keeping the total past income and predispositionconstant, increasing income streams lead to more risk tolerancethan decreasing income streams.
Reflection effects
People on aggregate tend to be risk averse in gains and riskseeking in losses (Kahneman and Tversky, 1979)
But not on individual level (Baucells and Villasis, 2010; Cohen etal., 1987; Schoemaker, 1990; Tymula et al., 2013)!
0
Ris
k a
ttitu
de
- g
ain
s
0
Risk attitude − losses
F. data from Tymula et al. (PNAS, 2013)
seeking averse
see
kin
ga
ve
rse
Tymula et al., PNAS, 2013
correlation = 0.254, p = 0.003; 79/128 in PT region
Reflection effects
People on aggregate tend to be risk averse in gains and riskseeking in losses (Kahneman and Tversky, 1979)
But not on individual level (Baucells and Villasis, 2010; Cohen etal., 1987; Schoemaker, 1990; Tymula et al., 2013)!
0
0
F. data from Tymula et al. (PNAS, 2013)
Ris
k a
ttitu
de
- g
ain
s
Risk attitude − losses
seeking averse
see
kin
ga
ve
rse
Tymula et al., PNAS, 2013
correlation = 0.254, p = 0.003; 79/128 in PT region
Reflection effects: model simulation
−1
0−
50
51
0
RR
A −
ga
ins
−10 −5 0 5 10
RRA − losses
averseseeking
averse
seeking
α = 5M = 10
xl = 5 lossxg = 15 gain
Reflection effects: model simulation
−1
0−
50
51
0
RR
A −
ga
ins
−10 −5 0 5 10
RRA − losses
−1
0−
50
51
0
RR
A −
ga
ins
−10 −5 0 5 10
RRA − losses
A. alpha~U(0,10), M~U(0,20) B. alpha~N(5,2), M~N(10,2)
Reflection effects: model simulation
−1
0−
50
51
0
RR
A −
ga
ins
−10 −5 0 5 10
RRA − losses
A. alpha~U(0,10), M~U(0,20)
high M and
su!ciently high α
low α and/or
low M
correlation = 0.287, p < 0.001 - opposite of reflection effect
Loss aversion
Common approach to loss aversion
The most commonly used utility specification:
U(x) =
{u(x) if x ≥ 0
λu(x) if x < 0
where λ > 1 - loss aversion
But the kink in the utility function is not necessary to produce lossaverse behavior (e.g. Blavatsky and Pogrebna, 2009; Ert and Erev,2013)
U
xxk kM
Common approach to loss aversion
The most commonly used utility specification:
U(x) =
{u(x) if x ≥ 0
λu(x) if x < 0
where λ > 1 - loss aversion
But the kink in the utility function is not necessary to produce lossaverse behavior (e.g. Blavatsky and Pogrebna, 2009; Ert and Erev,2013)
U
xxk kM
ESVT and loss aversion
We use traditional definition of loss aversion:
λt =V (Mt)− V (Mt − k)
V (Mt + k)− V (Mt)
k ∈ IR+ - distance from the reference point Mt
λt =(Mα
t − (Mt − k)α)(Mαt + (Mt + k)α)
((Mt − k)α + Mαt )((Mt + k)α −Mα
t )
Individual is loss averse if λt > 1
Proposition 7 An individual with preferences represented by valuefunction V (x) is always loss averse, though not necessarily to asignificant degree.
ESVT and loss aversion
We use traditional definition of loss aversion:
λt =V (Mt)− V (Mt − k)
V (Mt + k)− V (Mt)
k ∈ IR+ - distance from the reference point Mt
λt =(Mα
t − (Mt − k)α)(Mαt + (Mt + k)α)
((Mt − k)α + Mαt )((Mt + k)α −Mα
t )
Individual is loss averse if λt > 1
Proposition 7 An individual with preferences represented by valuefunction V (x) is always loss averse, though not necessarily to asignificant degree.
ESVT and loss aversion
We use traditional definition of loss aversion:
λt =V (Mt)− V (Mt − k)
V (Mt + k)− V (Mt)
k ∈ IR+ - distance from the reference point Mt
λt =(Mα
t − (Mt − k)α)(Mαt + (Mt + k)α)
((Mt − k)α + Mαt )((Mt + k)α −Mα
t )
Individual is loss averse if λt > 1
Proposition 7 An individual with preferences represented by valuefunction V (x) is always loss averse, though not necessarily to asignificant degree.
ESVT and loss aversion
11
.52
2.5
3lo
ss a
ve
rsio
n (
lam
bd
a)
0 2 4 6 8 10k
α=1
α=2
α=5
α=10
Proposition 8 For any stake, an individual with lowerpredisposition is more loss averse. Loss aversion first increases andthen decreases with stake size. Loss aversion peaks at lower stakesfor individuals with higher predisposition.
ESVT and loss aversion
11
.52
2.5
3lo
ss a
ve
rsio
n (
lam
bd
a)
0 2 4 6 8 10k
α=1
α=2
α=5
α=10
Proposition 8 For any stake, an individual with lowerpredisposition is more loss averse. Loss aversion first increases andthen decreases with stake size. Loss aversion peaks at lower stakesfor individuals with higher predisposition.
Endowment effect: literature
People endowed with a mug demand more money to sell it thanthose who do not have a mug are willing to pay for it
Traditional explanation: loss aversion (Thaler, 1980). Impliesendowment effect should always occur
This explanation has been the subject of a heated debate(Engelmann and Hollard, 2010; Isoni et al., 2011; List, 2002 &2003; Plott and Zeiler, 2005, 2007 & 2011). Reverse endowmenteffect has been documented in Lerner et al. (2004), Lin et al.(2006), Martinez et al. (2011)
Sense of ownership emerges as the key determinant of theendowment effect (review by Ericson & Fuster (2014))
Endowment effect: literature
People endowed with a mug demand more money to sell it thanthose who do not have a mug are willing to pay for it
Traditional explanation: loss aversion (Thaler, 1980). Impliesendowment effect should always occur
This explanation has been the subject of a heated debate(Engelmann and Hollard, 2010; Isoni et al., 2011; List, 2002 &2003; Plott and Zeiler, 2005, 2007 & 2011). Reverse endowmenteffect has been documented in Lerner et al. (2004), Lin et al.(2006), Martinez et al. (2011)
Sense of ownership emerges as the key determinant of theendowment effect (review by Ericson & Fuster (2014))
Endowment effect: literature
People endowed with a mug demand more money to sell it thanthose who do not have a mug are willing to pay for it
Traditional explanation: loss aversion (Thaler, 1980). Impliesendowment effect should always occur
This explanation has been the subject of a heated debate(Engelmann and Hollard, 2010; Isoni et al., 2011; List, 2002 &2003; Plott and Zeiler, 2005, 2007 & 2011). Reverse endowmenteffect has been documented in Lerner et al. (2004), Lin et al.(2006), Martinez et al. (2011)
Sense of ownership emerges as the key determinant of theendowment effect (review by Ericson & Fuster (2014))
Endowment effect: literature
People endowed with a mug demand more money to sell it thanthose who do not have a mug are willing to pay for it
Traditional explanation: loss aversion (Thaler, 1980). Impliesendowment effect should always occur
This explanation has been the subject of a heated debate(Engelmann and Hollard, 2010; Isoni et al., 2011; List, 2002 &2003; Plott and Zeiler, 2005, 2007 & 2011). Reverse endowmenteffect has been documented in Lerner et al. (2004), Lin et al.(2006), Martinez et al. (2011)
Sense of ownership emerges as the key determinant of theendowment effect (review by Ericson & Fuster (2014))
Endowment effect in ESVTNotation:
before endowment after endowment
ownership x0 x1reference M0 M1
Proposition 9 The endowment effect occurs if(Mα
0 −Mα1 )(Mα
0 Mα1 − xα0 x
α1 ) > 0.
Endowment effect does not occur if:
(Mα0 −Mα
1 )(Mα0 M
α1 − xα0 x
α1 ) = 0
I there is no change in reference point after endowment:M0 = M1
I when the reference point is equal to the current ownershiplevel: M0 = x0 and M1 = x1
Endowment effect in ESVTNotation:
before endowment after endowment
ownership x0 x1reference M0 M1
Proposition 9 The endowment effect occurs if(Mα
0 −Mα1 )(Mα
0 Mα1 − xα0 x
α1 ) > 0.
Endowment effect does not occur if:
(Mα0 −Mα
1 )(Mα0 M
α1 − xα0 x
α1 ) = 0
I there is no change in reference point after endowment:M0 = M1
I when the reference point is equal to the current ownershiplevel: M0 = x0 and M1 = x1
Endowment effect in ESVTNotation:
before endowment after endowment
ownership x0 x1reference M0 M1
Proposition 9 The endowment effect occurs if(Mα
0 −Mα1 )(Mα
0 Mα1 − xα0 x
α1 ) > 0.
Endowment effect does not occur if:
(Mα0 −Mα
1 )(Mα0 M
α1 − xα0 x
α1 ) = 0
I there is no change in reference point after endowment:M0 = M1
I when the reference point is equal to the current ownershiplevel: M0 = x0 and M1 = x1
Endowment effect in ESVTNotation:
before endowment after endowment
ownership x0 x1reference M0 M1
Proposition 9 The endowment effect occurs if(Mα
0 −Mα1 )(Mα
0 Mα1 − xα0 x
α1 ) > 0.
Endowment effect does not occur if:
(Mα0 −Mα
1 )(Mα0 M
α1 − xα0 x
α1 ) = 0
I there is no change in reference point after endowment:M0 = M1
I when the reference point is equal to the current ownershiplevel: M0 = x0 and M1 = x1
Endowment effect in ESVTNotation:
before endowment after endowment
ownership x0 x1reference M0 M1
Proposition 9 The endowment effect occurs if(Mα
0 −Mα1 )(Mα
0 Mα1 − xα0 x
α1 ) > 0.
Endowment effect does not occur if:
(Mα0 −Mα
1 )(Mα0 M
α1 − xα0 x
α1 ) = 0
I there is no change in reference point after endowment:M0 = M1
I when the reference point is equal to the current ownershiplevel: M0 = x0 and M1 = x1
No endowment effect in ESVT
There is no endowment effect if M0 = x0 and M1 = x1
A
0.5
1V(mugs)
V0
V1
WTP
WTA
3 4
mugs
Endowment effect in ESVT
Assume M0 = x0. Endowment effect occurs if M1 < x1
B
0.5
1V(mugs)
3 4
V1
V0
WTP
WTA
mugs
Reverse endowment effect in ESVT
Let M0 < M1. RRE if M0 > x0 and M1 > x1 (e.g. after a loss)
0.5
1V(mugs)
3 4
WTP
WTA
V0
V1
mugs
Diminishing WTP and WTA: divisive versus difference
divisive model - ESVT
0.5
1V(mugs)
WTP
WTP
WTP
WTP
WTP
WTP
WTP
2 3 4 5mugs
difference-based model
0.5
1
mugs
WTP
WTP
2 3 4
U(mugs)
Diminishing WTP and WTA: divisive versus difference
divisive model - ESVT
0.5
1V(mugs)
WTP
WTP
WTP
WTP
WTP
WTP
WTP
2 3 4 5mugs
difference-based model
0.5
1mugs
WTP
WTP
2 3 4U(mugs)
Probability weighting
Probability weighting - theory versus practice
Theory
Practice
Tobler et al., Journal of Neuroscience (2008)
There is a lot of unexplained heterogeneity in the estimatedprobability weighting functions on the individual level
Probability weighting - theory versus practice
Theory Practice
Tobler et al., Journal of Neuroscience (2008)
There is a lot of unexplained heterogeneity in the estimatedprobability weighting functions on the individual level
Probability weighting - theory versus practice
Theory Practice
Tobler et al., Journal of Neuroscience (2008)
There is a lot of unexplained heterogeneity in the estimatedprobability weighting functions on the individual level
Probability weighting - neuro evidence
Essentially all neuroeconomic studies agree that neurons arecapable of encoding probability without any distortions
I Abler et al. 2006; Berns et al. 2008; Preuschoff, Bossaerts,and Quartz 2006; Tobler et al. 2008
There seem to be brain regions that distort probabilities, but either
I distortions have no influence on choice (Tobler et al. 2008)
I cannot reject the null of linear coding in the brain overall(Berns et al. 2008)
I estimates may be biased by the functional assumptions (Hsuet al. 2008)
Probability weighting - neuro evidence
Essentially all neuroeconomic studies agree that neurons arecapable of encoding probability without any distortions
I Abler et al. 2006; Berns et al. 2008; Preuschoff, Bossaerts,and Quartz 2006; Tobler et al. 2008
There seem to be brain regions that distort probabilities, but either
I distortions have no influence on choice (Tobler et al. 2008)
I cannot reject the null of linear coding in the brain overall(Berns et al. 2008)
I estimates may be biased by the functional assumptions (Hsuet al. 2008)
Probability weighting an artefact of normalization?
1. We simulated choices for 9 individuals (3 reference points x 3predispositions) with ESVT-preferences in the Gonzalez andWu (1999) lottery task
EV (x , p) =n∑
i=1
piV (xi )
w(p) = p
2. We fitted these choices with
U(x) = x r
w(p) =δpγ
δpγ + (1− p)γ
Probability weighting an artefact of normalization?
1. We simulated choices for 9 individuals (3 reference points x 3predispositions) with ESVT-preferences in the Gonzalez andWu (1999) lottery task
EV (x , p) =n∑
i=1
piV (xi )
w(p) = p
2. We fitted these choices with
U(x) = x r
w(p) =δpγ
δpγ + (1− p)γ
Probability weighting an artefact of normalization?
01
w(p)
0 1p
low reference point
01
w(p)
0 1p
01
w(p)
0 1p
medium reference point high reference point
α=1α=2α=5
Summary of ESVT
I Our model, based on established neuronal computationdivisive normalization, captures many of the observedbehavioral phenomena with only two parameters
I It can capture:I reflection effects in risk attitudesI loss aversionI probability weightingI choice set effects
I Additionally, it captures heterogeneity in all of the above dueto individual differences in reference points and predispositions
I Lack of discontinuity in the value function allows for easyestimation of the reference point
Summary of ESVT
I Our model, based on established neuronal computationdivisive normalization, captures many of the observedbehavioral phenomena with only two parameters
I It can capture:I reflection effects in risk attitudesI loss aversionI probability weightingI choice set effects
I Additionally, it captures heterogeneity in all of the above dueto individual differences in reference points and predispositions
I Lack of discontinuity in the value function allows for easyestimation of the reference point
Summary of ESVT
I Our model, based on established neuronal computationdivisive normalization, captures many of the observedbehavioral phenomena with only two parameters
I It can capture:I reflection effects in risk attitudesI loss aversionI probability weightingI choice set effects
I Additionally, it captures heterogeneity in all of the above dueto individual differences in reference points and predispositions
I Lack of discontinuity in the value function allows for easyestimation of the reference point
Summary of ESVT
I Our model, based on established neuronal computationdivisive normalization, captures many of the observedbehavioral phenomena with only two parameters
I It can capture:I reflection effects in risk attitudesI loss aversionI probability weightingI choice set effects
I Additionally, it captures heterogeneity in all of the above dueto individual differences in reference points and predispositions
I Lack of discontinuity in the value function allows for easyestimation of the reference point
Neuroeconomic variables - graymatter volume
Study set up
I n=61 young adults (half exploration / half replication)
Tymula∗, Gilaie-Dotan∗,Cooper, Kable, Glimcher & Levy. J of Neuroscience, 2014
Gray matter volume estimation
I 10 minute structural scans
I Siemens Allegra 3T head-only scanner
Tymula∗, Gilaie-Dotan∗,Cooper, Kable, Glimcher & Levy. J of Neuroscience, 2014
Study 1 Result
Tymula∗, Gilaie-Dotan∗,Cooper, Kable, Glimcher & Levy. J of Neuroscience, 2014
Gray matter volume and risk attitudes in ageing
I Gray matter volume decreases with age
Salat et al., 2012
I Risk aversion increases with age (von Gaudecker et al. 2011,Dohmen et al., 2011, Tymula et al., 2013, ...)
Study 2
I n=52 (30 female)I 18-88 years old, mean: 54.7, s.d.:22.1
Grubb, Tymula, Gilaie-Dotan, Glimcher & Levy, Nature Communications, 2016
Study 2 Result
Grubb, Tymula, Gilaie-Dotan, Glimcher & Levy, Nature Communications, 2016
Rationality and gray matter volume
I 39 subjects (14 male)I 65-92 years old (mean: 72.46)I no signs of dementia (tested on mini-mental)
Chung, Tymula, Glimcher, r&r
Rationality and gray matter volume
Rationality measurement
I Harbaugh, Krause, Berry, AER 2002
good A
go
od
B
good A
go
od
B
u(a)>u(d) choice
u(d)>u(b) monotonicity
u(b)>u(c) choice
u(c)>u(a) monotonicity
a
b
c
d
good A
go
od
B
Rationality and gray matter volume
Rationality measurement
I Harbaugh, Krause, Berry, AER 2002
good A
go
od
B
u(a)>u(d) choice
u(d)>u(b) monotonicity
u(b)>u(c) choice
u(c)>u(a) monotonicity
a
b
c
d
good A
go
od
B
Rationality and gray matter volume
Rationality measurement
I Harbaugh, Krause, Berry, AER 2002
good A
go
od
B
u(a)>u(d) choice
u(d)>u(b) monotonicity
u(b)>u(c) choice
u(c)>u(a) monotonicity
a
b
c
d
good A
go
od
B
I Rationality is positively correlated with grey matter volumes inleft anterior prefrontal cortex (a47r)
Chung, Tymula, Glimcher, r&r
Why is this interesting?
I TheoryI Constrains possible neural mechanisms underlying risk
attitudes and rationality
I Provides biological evidence for incorporating neural capacityin models of choice
I Simon QJE 1955, Rayo & Becker JPE 2007, Netzer AER2009, Woodford AER 2012, Robson & Whitehead, 2016
I Opportunity to unite numerous previously documentedassociations between host of brain-changing variables and riskpreferences and rationality into one coherent theory
I EmpiricalI New datasets - simple measurement of risk attitudes in
existing brain scans
I PolicyI New interventions
Why is this interesting?
I TheoryI Constrains possible neural mechanisms underlying risk
attitudes and rationalityI Provides biological evidence for incorporating neural capacity
in models of choiceI Simon QJE 1955, Rayo & Becker JPE 2007, Netzer AER
2009, Woodford AER 2012, Robson & Whitehead, 2016
I Opportunity to unite numerous previously documentedassociations between host of brain-changing variables and riskpreferences and rationality into one coherent theory
I EmpiricalI New datasets - simple measurement of risk attitudes in
existing brain scans
I PolicyI New interventions
Why is this interesting?
I TheoryI Constrains possible neural mechanisms underlying risk
attitudes and rationalityI Provides biological evidence for incorporating neural capacity
in models of choiceI Simon QJE 1955, Rayo & Becker JPE 2007, Netzer AER
2009, Woodford AER 2012, Robson & Whitehead, 2016
I Opportunity to unite numerous previously documentedassociations between host of brain-changing variables and riskpreferences and rationality into one coherent theory
I EmpiricalI New datasets - simple measurement of risk attitudes in
existing brain scans
I PolicyI New interventions
Why is this interesting?
I TheoryI Constrains possible neural mechanisms underlying risk
attitudes and rationalityI Provides biological evidence for incorporating neural capacity
in models of choiceI Simon QJE 1955, Rayo & Becker JPE 2007, Netzer AER
2009, Woodford AER 2012, Robson & Whitehead, 2016
I Opportunity to unite numerous previously documentedassociations between host of brain-changing variables and riskpreferences and rationality into one coherent theory
I EmpiricalI New datasets - simple measurement of risk attitudes in
existing brain scans
I PolicyI New interventions
Why is this interesting?
I TheoryI Constrains possible neural mechanisms underlying risk
attitudes and rationalityI Provides biological evidence for incorporating neural capacity
in models of choiceI Simon QJE 1955, Rayo & Becker JPE 2007, Netzer AER
2009, Woodford AER 2012, Robson & Whitehead, 2016
I Opportunity to unite numerous previously documentedassociations between host of brain-changing variables and riskpreferences and rationality into one coherent theory
I EmpiricalI New datasets - simple measurement of risk attitudes in
existing brain scans
I PolicyI New interventions
Perspective
I Specific population of participants - we need to study differentsamples
I Dementia sample at U of Sydney
I Findings do not imply a fixed deterministic relation betweengenetics and risk attitude
I Environment, own behavior may shape brain structure
I We cannot infer causality from this data
“The task is to replace the global rationality of economic man witha kind of rational behavior that is compatible with the access toinformation and the computational capacities that are actually
possessed by organisms.”
Herbert Simon, 1955, p. 241
Neuroeconomics - usefulresources
Neuroeconomics - useful resources
Books
Neuroeconomics - useful resources
Join the Society for Neuroeconomics - https://neuroeconomics.org
and Society for Neuroeconomics Annual Meetings (since 2005)
Neuroeconomics - useful resources
Do a postdoc - https://neuroeconomics.org/career/
Join us for the summer school - next edition in 2019
Collaborate (to learn and avoid embarrassment)
Be respectful and ask questions
Thank you!
Coauthors:
Paul Glimcher, Ifat Levy, Hui-Kuan Chung, Joe Kable, Sharon Gilaie-Dotan,
Michael Grubb, Natalie Cooper
Funding bodies:
How is reference point, Mt , computed in the brain?
I Dopaminergic reward prediction error system computes arecency-weighted average of past rewards
I Typically, it is modeled using discrete time, where at each stept, the brain computes future reward expectation, Pt , as:
Pt = Pt−1 + γ(xt − Pt−1)
= (1− γ)tP0 +t∑
i=1
γ(1− γ)t−ixi
I Assuming that initial expectation P0 = 0 (Sutton and Barto,1998):
Mt = Pt−1 =t−1∑i=1
γ(1− γ)t−ixi
Back to model