26
1 Upper Cumulative Independence Michael H. Birnbaum California State University, Fullerton

Upper Cumulative Independence

  • Upload
    saddam

  • View
    32

  • Download
    0

Embed Size (px)

DESCRIPTION

Upper Cumulative Independence. Michael H. Birnbaum California State University, Fullerton. UCI is implied by CPT. CPT, RSDU, RDU, EU satisfy UCI. RAM and TAX violate UCI. Violations are direct internal contradiction in RDU, RSDU, CPT, EU. - PowerPoint PPT Presentation

Citation preview

Page 1: Upper Cumulative Independence

1

Upper Cumulative Independence

Michael H. BirnbaumCalifornia State University,

Fullerton

Page 2: Upper Cumulative Independence

2

UCI is implied by CPT

• CPT, RSDU, RDU, EU satisfy UCI.• RAM and TAX violate UCI. • Violations are direct internal

contradiction in RDU, RSDU, CPT, EU.

Page 3: Upper Cumulative Independence

3

′ z > ′ x > x > y > ′ y > z > 0

′ S → ( ′ z ,1− p − q;x, p;y,q)

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

In this test, we reduce z’ in both gambles and coalesce it with x’ (in R’), and we decrease x and coalesce it with y (in S’ only).

Page 4: Upper Cumulative Independence

4

Lower Cumulative Independence (3-LCI)

′ S = ( ′ z ,1− p − q;x, p;y,q) p

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

′ ′ ′ S = ( ′ x ,1− p − q;y, p + q) p

′ ′ ′ R = ( ′ x ,1− q; ′ y ,q)

Page 5: Upper Cumulative Independence

5

UCI implied by any model that satisfies:

• Comonotonic restricted branch independence

• Consequence monotonicity• Transitivity• Coalescing• (Proof on next page.)

Page 6: Upper Cumulative Independence

6

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

( ′ x ,1− p − q;x, p;y,q) p ( ′ x ,1− p − q; ′ x , p; ′ y ,q)

( ′ x ,1− p − q;y, p;y,q) p ( ′ x ,1− p − q;x, p;y,q)

( ′ x ,1− p − q;y, p;y,q) p ( ′ x ,1− p − q; ′ x , p; ′ y ,q)

( ′ x ,1− p − q;y, p + q) p ( ′ x ,1− q; ′ y ,q)

′ ′ ′ S p ′ ′ ′ R

Comonotonic RBI

Consequence monotonicity

Transitivity

Coalescing

Page 7: Upper Cumulative Independence

7

Example Test

′ S : 10 to win $40 10 to win $44

80 to win $110

′ R : 10 to win $10 10 to win $98

80 to win $110

′ ′ ′ S : 20 to win $40 80 to win $98

′ ′ ′ R : 10 to win $10 90 to win $98

Page 8: Upper Cumulative Independence

8

Generic Configural Model

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

′ S p ′ R ⇔

⇔w3

w2

<u( ′ x ) − u(x)

u(y) − u( ′ y )

Page 9: Upper Cumulative Independence

9

3-2-LCI in CPT

Suppose

CPT satisfies coalescing;

′ ′ ′ S f ′ ′ ′ R ⇔

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

⇔ w2u(y) + w3u(y) > w2u( ′ x ) + w3u( ′ y )

⇔w3

w2

>u( ′ x ) − u(y)

u(y) − u( ′ y )>

u( ′ x ) − u(x)

u(y) − u( ′ y )⇒⇐ contradiction

′ S p ′ R ⇔w3

w2

<u( ′ x ) − u(x)

u(y) − u( ′ y )

Page 10: Upper Cumulative Independence

10

2 Types of Reversals:

R’S’’’: This is a violation of UCI. It refutes CPT.

S’R’’’: This reversal is consistent with LCI. (S’ made worse relative to R’.)

Page 11: Upper Cumulative Independence

11

RAM Weights

w1 = a(1,3)t(1− p − q) /T

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

w3 = a(3,3)t(q) /T

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

′ ′ ′ w 1 = a(1,2)t(1− p − q) / ′ ′ ′ T

′ ′ ′ w 2 = a(2,2)t(p + q) / ′ ′ ′ T

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

Page 12: Upper Cumulative Independence

12

RAM Violations• RAM violates 3-2-UCI. If t(p) is

negatively accelerated, RAM violates coalescing: coalescing branches with better consequences makes the gamble worse and coalescing the branches leading to lower consequences makes the gamble better. Even though we made S relatively worse, the coalescings made it relatively better.

Page 13: Upper Cumulative Independence

13

TAX Model

w1 =t(1− p − q)[1−δ /4 −δ /4]

t(1− p − q) + t( p) + t(q)

w2 =t( p) −δt(p) /4 + δt(1− p − q) /4

t(1− p − q) + t(p) + t(q)

w3 =t(q) + δt(1− p − q) /4 + δt(p) /4

t(1− p − q) + t(p) + t(q)

′ ′ ′ w 1 =t(1− p − q) −δt(1− p − q) /3

t(1− p − q) + t( p + q)

′ ′ ′ w 2 =t( p + q) + δt(1− p − q) /3

t(1− p − q) + t( p + q)

Page 14: Upper Cumulative Independence

14

TAX: Violates UCI

• Special TAX model violates 3-2-UCI. Like RAM, the model violates coalescing.

• Predictions were calculated in advance of the studies, which were designed to investigate those specific predictions.

Page 15: Upper Cumulative Independence

15

Summary of Predictions

• EU, CPT, RSDU, RDU satisfy UCI• TAX & RAM violate UCI• CPT defends the null hypothesis

against specific predictions made by both RAM and TAX.

Page 16: Upper Cumulative Independence

16

Birnbaum (‘99): n = 124

No.No Choice %R

10

′ S: 10 to win $40

10 to win $44

80 to win $110

′ R : 10 to win $10 10 to win $98

80 to win $110

72*

9

′ ′ ′ S : 20 to win $40 80 to win $98

′ ′ ′ R : 10 to win $10 90 to win $98

34*

Page 17: Upper Cumulative Independence

17

Lab Studies of UCI

• Birnbaum & Navarrete (1998): 27 tests; n = 100; (p, q) = (.25, .25), (.1, .1), (.3, .1), (.1, .3).

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

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

Page 18: Upper Cumulative Independence

18

Web Studies of UCI

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

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

Page 19: Upper Cumulative Independence

19

Additional Replications A number of as unpublished

studies (as of Jan, 2005) have replicated the basic findings with a variety of different procedures in choice.

Page 20: Upper Cumulative Independence

20

′ S = ($ 110 , . 8 ; $ 44 , . 1 ; $ 40 , . 1 ) versus ′ R = ($ 110 ; . 8 ; $ 98 , . 1 ; $ 10 , . 1 ) , ′ ′ ′ S = ($ 98 , . 8 ; $ 40 , . 2 ) versus ′ ′ ′ R = ($ 98 , . 9 ; $ 10 , . 1 )

Choice Pattern

Choices 10 and 9

Condition

n ′ S ′ ′ ′ S ′ S ′ ′ ′ R ′ R ′ ′ ′ S ′ R ′ ′ ′ R

new tickets 141 38 11 71 21

aligned 141 36 5 74 23

unaligned 151 34 9 81 25

Negative

(reflected)200 X

277 42 157 123

Page 21: Upper Cumulative Independence

21

Error Analysis

• “True and Error” Model implies violations are “real” and cannot be attributed to error.

Page 22: Upper Cumulative Independence

22

Violations predicted by RAM & TAX, not CPT

• EU, CPT, RSDU, RDU are refuted by systematic violations of UCI.

• TAX & RAM, as fit to previous data correctly predicted the violations. Predictions published in advance of the studies.

• Violations are to CPT as the Allais paradoxes are to EU.

Page 23: Upper Cumulative Independence

23

To Rescue CPT:

For CPT to handle these data, make

it configural. Let < 1 for two-

branch gambles and > 1 for three-branch gambles.

Page 24: Upper Cumulative Independence

24

Add to the case against CPT/RDU/RSDU

• Violations of Upper Cumulative Independence are a strong refutation of CPT model as proposed.

Page 25: Upper Cumulative Independence

25

Next Program: UTI

• The next programs reviews tests of Upper Tail Independence (UTI).

• Violations of 3-UTI contradict any form of CPT, RSDU, RDU, including EU.

• Violations contradict Lower GDU.• They are consistent with RAM and

TAX.