35
1 Restricted Branch Independence Michael H. Birnbaum California State University, Fullerton

Restricted Branch Independence

Embed Size (px)

DESCRIPTION

Restricted Branch Independence. Michael H. Birnbaum California State University, Fullerton. RBI is Violated by CPT. EU satisfies RBI as does SWU and PT, extended to 3-branch gambles. Cancellation CPT violates RBI (it MUST to explain the Allais Paradoxes) - PowerPoint PPT Presentation

Citation preview

Page 1: Restricted Branch Independence

1

Restricted Branch Independence

Michael H. BirnbaumCalifornia State University,

Fullerton

Page 2: Restricted Branch Independence

2

RBI is Violated by CPT

• EU satisfies RBI as does SWU and PT, extended to 3-branch gambles. Cancellation

• CPT violates RBI (it MUST to explain the Allais Paradoxes)

• RAM and TAX violate RBI in the opposite direction as CPT.€

⇒ RBI

Page 3: Restricted Branch Independence

3

′ z > ′ x > x > y > ′ y > z > 0

S → (x, p;y, p;z,1− 2p)

R → ( ′ x , p; ′ y , p;z,1− 2p)

In this test, we move the common branch from lowest, z, to highest,

z’ consequence.

Page 4: Restricted Branch Independence

4

Restricted Branch Independence (3-RBI)

S = (x, p,y,q;z,1− p − q) f

R = ( ′ x , p; ′ y ,q;z,1− p − q)

′ S = ( ′ z ,1− p − q;x, p, y,q) f

′ R = ( ′ z ,1− p − q; ′ x , p; ′ y ,q)

Page 5: Restricted Branch Independence

5

Types of Branch Independence

• The term “restricted” is used to indicate that the number of branches and probability distribution is the same in all four gambles.

• When we further constrain z and z’ to keep the same ranks in all four gambles, it is termed “comonotonic” (restricted) branch independence.

Page 6: Restricted Branch Independence

6

A Special Case

• We can make a still more restricted case of restricted branch independence, in order to test the predictions of any weakly inverse-S weighting function. Let p = q.

• This distribution has been used in most, but not all of the studies.

Page 7: Restricted Branch Independence

7

Example Test

S: .80 to win $2

.10 to win $40

.10 to win $44

R: .80 to win $2

.10 to win $4

.10 to win $96

S’: .10 to win $40

.10 to win $44

.80 to win $100

R’: .10 to win $4

.10 to win $96

.80 to win $100

Page 8: Restricted Branch Independence

8

Generic Configural Model

w1u(x) + w2u(y) + w3u(z) > w1u( ′ x ) + w2u( ′ y ) + w3u(z)

The generic model includes RDU, CPT, EU, RAM, TAX, GDU, as special cases.

S f R ⇔

⇔w2

w1

>u( ′ x ) − u(x)

u(y) − u( ′ y )

Page 9: Restricted Branch Independence

9

Violation of 3-RBI

A violation will occur if S f R and

′ S p ′ R ⇔

⇔′ w 3′ w 2

<u( ′ x ) − u(x)

u(y) − u( ′ y )€

′ w 1u( ′ z ) + ′ w 2u(x) + ′ w 3u(y) < ′ w 1u( ′ z ) + ′ w 2u( ′ x ) + ′ w 3u( ′ y )

Page 10: Restricted Branch Independence

10

2 Types of Violations:

S f R∧ ′ S p ′ R ⇔w2

w1

>u( ′ x ) − u(x)

u(y) − u( ′ y )>

′ w 3′ w 2

S p R∧ ′ S f ′ R ⇔w2

w1

<u( ′ x ) − u(x)

u(y) − u( ′ y )<

′ w 3′ w 2

SR’:

RS’:

Page 11: Restricted Branch Independence

11

EU allows no violations

• In EU, the weights equal the probabilities; therefore

w2

w1

=p

p=

p

p=

′ w 3′ w 2

Page 12: Restricted Branch Independence

12

RAM Weights

w1 = a(1,3)t(p) /T

w2 = a(2,3)t(p) /T

w3 = a(3,3)t(1− 2p) /T

T = a(1,3)t(p) + a(2,3)t(p) + a(3,3)t(1− 2 p)

Page 13: Restricted Branch Independence

13

RAM Violations

• RAM model violates 3-RBI.

w2

w1

=a(2,3)t(p)

a(1,3)t(p)≠

a(3,3)t(p)

a(2,3)t(p)=

′ w 3′ w 2

a(i,n) = i ⇒w2

w1

=2

1>

3

2=

w3

w2

⇒ S ′ R

Page 14: Restricted Branch Independence

14

CPT/ RDU

w1 = W ( p) −W (0)

w2 = W (2p) −W ( p)

w3 =1−W (2p)

′ w 1 = W (1− 2p)

′ w 2 = W (1− p) −W (1− 2p)

′ w 3 =1−W (1− p)

Page 15: Restricted Branch Independence

15

Inverse-S Weighting Function

0

1

0 1

Decumulative Probability

Decumulative Weight

p 2p

W(2p)

W(p)

1-p1-2p

W(1-p)

W(1-2p)

w1 > w2

⇒w2

w1

<1

′ w 3 > ′ w 2

⇒′ w 3′ w 2

>1

Page 16: Restricted Branch Independence

16

CPT implies RS’ violation

• If W(P) = P, CPT reduces to EU.

• However, if W(P) is any weakly inverse-S

function, CPT implies the RS’ pattern.

• (A strongly inverse- S function is weakly inverse-S plus it crosses the identity line. If we reject weak, then we reject the strong as well.)

Page 17: Restricted Branch Independence

17

CPT Analysis of Table 1, #9 and 15: RBI

0

0.5

1

1.5

2

0.5 1.0 1.5

Weighting Function Parameter, γ

, Utility Function Exponent

β

RS'

RR'

SR'

SS'

Page 18: Restricted Branch Independence

18

Transfer of Attention Exchange (TAX)

• Each branch (p, x) gets weight that is a function of branch probability

• Utility is a weighted average of the utilities of the consequences on branches.

• Attention (weight) is drawn from one branch to others. In a risk-averse person, weight is transferred to branches with lower consequences.

Page 19: Restricted Branch Independence

19

“Special” TAX Model

Assumptions:

U(G) =Au(x) + Bu(y) + Cu(z)

A + B + C

A = t( p) −δt(p) /4 −δt(p) /4

B = t(q) −δt(q) /4 + δt(p) /4

C = t(1− p − q) + δt(p) /4 + δt(q) /4

G = (x, p;y,q;z,1− p − q)

Page 20: Restricted Branch Independence

20

“Prior” TAX Model

u(x) = x; $0 < x < $150

t( p) = pγ ; γ = 0.7

δ =1Parameters were chosen to give a rough approximation to Tversky & Kahneman (1992) data. They are used to make new predictions.

Page 21: Restricted Branch Independence

21

TAX Model Weights

Each term has the same denominator; middle branch gives up what it receives when p = q.€

A = t( p) − 2δt( p) /4

B = t( p) + δt(p) /4 −δt(p) /4

C = t(1− 2p) + δt( p) /4 + δt( p) /4

Page 22: Restricted Branch Independence

22

Special TAX: SR’ Violations

• Special TAX model violates 3-RBI when delta is not zero.

w2

w1

=t(p)

t( p) − 2δt( p) /4>

t(p) + δt(p) + δt(1− 2p)

t(p) −δt(p) + δt(1− 2p)=

′ w 3′ w 2

Page 23: Restricted Branch Independence

23

Summary of Predictions

• EU, SWU, OPT satisfy RBI• CPT violates RBI: RS’

• TAX & RAM violate RBI: SR’

• Here CPT is the most flexible model, RAM and TAX make opposite prediction from that of CPT.

Page 24: Restricted Branch Independence

24

Results: n = 1075

SR’ (CPT predicted RS’)

No. S R % R

9 .80 to win $2

.10 to win $40

.10 to win $44

.80 to win $2

.10 to win $4

.10 to win $96

42.4

15.10 to win $40

.10 to win $44

.80 to win $100

.10 to win $4

.10 to win $96

.80 to win $100

56.0

Page 25: Restricted Branch Independence

25

Lab Studies of RBI

• Birnbaum & McIntosh (1996): 2 studies, n = 106; n = 48, p = 1/3

• Birnbaum & Chavez (1997): n = 100; 3-RBI and 4-RBI, p = .25

• Birnbaum & Navarrete (1998): 27 tests; n = 100; p = .25, p = .1.

• Birnbaum, Patton, & Lott (1999): n = 110; p = .2.

• Birnbaum (1999): n = 124; p = .1, p = .05.

Page 26: Restricted Branch Independence

26

Web Studies of RBI

• Birnbaum (1999): n = 1224; p   = .1, p = .05

• Birnbaum (2004b): 12 studies with total of n = 3440 participants; different formats for presenting gambles probabilities; p = .1, .05.

• Birnbaum (2004a): 3 conditions with n = 350; p = .1. Tests combined with Allais paradox.

Page 27: Restricted Branch Independence

27

Additional Replications

• SR’ pattern is significantly more frequent than RS’ pattern in judgment studies as well. (Birnbaum & Beeghley, 1997; Birnbaum & Veira, 1998; Birnbaum & Zimmermann, 1999).

• A number of as yet unpublished studies have also replicated the basic findings with a variety of different procedures in choice.

Page 28: Restricted Branch Independence

28

S = ($ 44 , . 1 ; $ 40 , . 1 ; $ 2 , . 8 ) .vs R = ($ 98 , . 1 ; $ 10 , . 1 ; $ 2 , . 8 )

′ S = ($ 110 , . 8 ; $ 44 , . 1 ; $ 40 , . 1 )) . vs ′ R = ($ 110 ; . 8 ; $ 98 , . 1 ; $ 10 , . 1 ) ,Choice Pattern

Condition n S ′ S S ′ R R ′ S R ′ R

New Tickets 141 34 54 14 37

Aligned Matrix 141 28 51 13 46

Unaligned Matrix 151 28 53 14 52

Losses (reflected) 200X 2

74 104 45 174

Page 29: Restricted Branch Independence

29

Error Analysis

• We can fit “true and error” model to data with replications to separate “real” violations from those attributable to “error”.

• Model estimates that SR’ violations are

“real” and probability of RS’ is equal to zero.

Page 30: Restricted Branch Independence

30

Violations predicted by RAM & TAX, not CPT

• EU, SWU, OPT are refuted in this case by systematic violations.

• Editing “cancellation” refuted.• TAX & RAM, as fit to previous data

correctly predicted the modal choices.• Violations opposite those implied by CPT

with its inverse-S W(P) function.• Fitted CPT correct when it agrees with

TAX, wrong otherwise.

Page 31: Restricted Branch Independence

31

To Rescue CPT:

• CPT can handle the result of any single test, by choosing suitable parameters.

• For CPT to handle these data, let γ

> 1; i.e., an S-shaped W(P) function, contrary to previous inverse- S.

Page 32: Restricted Branch Independence

32

CPT Analysis of Table 1, #9 and 15: RBI

0

0.5

1

1.5

2

0.5 1.0 1.5

Weighting Function Parameter, γ

, Utility Function Exponent

β

RS'

RR'

SR'

SS'

Page 33: Restricted Branch Independence

33

Adds to the case against CPT/RDU/RSDU

• Violations of RBI as predicted by TAX and RAM but are opposite predictions of CPT.

• Maybe CPT is right but its parameters are just wrong. As we see in the next program, we can generate internal contradiction in CPT.

Page 34: Restricted Branch Independence

34

Next Program: LCI

• The next programs reviews tests of Lower Cumulative Independence (LCI).

• Violations of 3-LCI contradict any form of RDU, CPT.

• They also refute EU but are consistent with RAM and TAX.