Adapting de Finetti's Proper Scoring Rules for Measuring Bayesian Subjective Probabilities when Those Probabilities Are not Bayesian Peter P. Wakker (& Gijs van de Kuilen, Theo Offerman, Joep Sonnemans) April 6, 2006 Rotterdam Topic: How to measure your subjective probability p of event E ("rain tomorrow")? Such that also acceptable to frequentists?
36
Embed
Adapting de Finetti's Proper Scoring Rules for Measuring Bayesian Subjective Probabilities when Those Probabilities Are not Bayesian Peter P. Wakker (&
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Adapting de Finetti's Proper Scoring Rules for Measuring
Bayesian Subjective Probabilities when Those Probabilities Are not
BayesianPeter P. Wakker (& Gijs van de Kuilen,
Theo Offerman, Joep Sonnemans)April 6, 2006Rotterdam
Topic: How to measure your subjective probability p of event E ("rain tomorrow")?Such that also acceptable to frequentists?
Method 1. Ask directly (introspection).Problem. No clear empirical meaning; no incentive(-compatibility).
Economists don't like such things.
2
Method 2. Reveal (binary) preferences.€ 0.30 (€ 1 under E) € 0.20 0.30 > p > 0.20 (if EV = expected value)Problem. Much work. Get only inequalities. (And: EV may be violated.)
Method 3. Reveal indifferences.€ 0.25 ~ (€ 1 under E) p = 0.25 (if EV …)Problem. Indifferences are hard to observe (BDM …). (And: EV may be violated.)
3
Method 4. de Finetti's book making. Not explained here.Problem. Not implementable in practice. (And: EV may be violated.)
4
Method 5. Proper scoring rules.Choose 0 r 1 as you like (called reported probability).Next, you receive
E Ec
1 – (1– r)2 1 – r2
EV: Take probability p of E;subjective if must be. Then maximize EV.Optimization of r:
Avoids all problems mentioned above, except Problem. EV may be violated …Proper scoring rules tractable & widely used:
See Hanson (Nature, 2002), Prelec (Science 2005).
In accounting (Wright 1988), Bayesian statistics (Savage 1971), business (Stael von Holstein 1972), education (Echternacht 1972), medicine (Spiegelhalter 1986), psychology (Liberman & Tversky 1993; McClelland & Bolger 1994).
6
Before analyzing nonEV descriptively, a:
Theoretical Example [EV].Urn K ("known") with 100 balls:25 G(reen), 25 R(ed), 25 S(ilver), 25 Y(ellow). One ball is drawn randomly.E: ball is not red; E = {G,S,Y}. p = 0.75.Under expected value,optimal rE is 0.75.rG + rS + rY = 0.25 + 0.25 + 0.25 = 0.75 = rE:r satisfies additivity; as probabilities should!
7
Example reanalyzed later.
0.25 0.50 0.75 10p
8Reported probability R(p) = rE as function of true probability p, under:
nonEU
0.69
R(p)
EU
0.61
rEV
EV
rnonEU
rnonEUA
rnonEUA refers to nonexpected utility for unknown probabilities ("Ambiguity").
(c) nonexpected utility for known probabilities, with U(x) = x0.5 and with w(p) as common;
(b) expected utility with U(x) = x0.5 (EU);
(a) expected value (EV);
reward: if E true if not E true EV rEV=0.75 0.94 0.44 0.8125 rEU=0.69 0.91 0.52 0.8094
Deviations from EV and Bayesianism so far were at level of behavior; not of beliefs. Now for something different; more fundamental.
3rd violation of EV: Ambiguity (unknown probabilities)
Preparation for theoretical example.Random draw from additional urn A.100 balls, ? Ga, ? Ra, ? Sa, ? Ya. Unknown composition ("Ambiguous").Ea: {Ga,Sa,Ya}; p = ?.How deal with unknown probabilities?
Have to give up Bayesian beliefs.
16
1st Proposal (trying to maintain Bayesian beliefs) Assign "subjective" probabilities to events. Then behave as if known probs (may be nonEU)("probabilistic sophistication," Machina & Schmeidler 1992; they did it for normative, not as we do now descriptively).By symmetry then P(G) = P(R) = P(S) = P(Y) = ¼.Treated same as known urn K.Empirically violated (Ellsberg)!
17
>< Probabilistic soph. Bayesian beliefs!
18
Instead of additive beliefs p = P(E), nonadditive beliefs B(E), with B(E) B(G) + B(S) + B(Y).(Dempster&Shafer, Tversky&Koehler, etc.)
All currently existing decision models:For r 0.5, nonEU(r) = w(B(E))U(1 – (1–r)2) + (1–w(B(E)))U(1–r2).(Commonly written with W(E) for w(B(E)), soB(E) = w–1(W(E)); matter of notation.)
19
U´(1–r2)U´(1 – (1–r)2)
(1–w(B(E)))w(B(E)) +
w(B(E))rE =
U´(1–r2)U´(1 – (1–r)2)
(1–r)r +
rB(E) =
Explicit expression:
w –1(
)
Example continued [Ambiguity, nonEUA]. (urn A, 100 balls, ? G, ? R, ? S, ? Y).E: {G,S,Y}; p = ?rEV = 0.75.rEU = 0.69.rnonEU = 0.61 (under plausible assumptions).Typically, w(B(E)) << w(B(E)) (E from known urn). rnonEUA is, say, 0.52.rG + rS + rY = 0.48 + 0.48 + 0.48 = 1.44 > 0.52 = rE:additivity is very extremely violated!r's are close to always saying fifty-fifty.Belief component B(E) = 0.62.
20
go to p. 8, with figure of R(p)
B(E): ambiguity attitude /=/ beliefs??Before entering that debate, first:How measure B(E)?Our contribution: through proper scoring rules with "risk correction."
Earlier proposals for measuring B:
21
Proposal 1 (common in decision theory).Measure U,W, and w from behavior.Derive B(E) = w–1(W(E)) from it.Problem: Much and difficult work!!!
22
Proposal 2 (common in decision analysis of the 1960s, and in modern experimental economics). Measure "canonical probabilities": For ambiguous event Ea, find objective prob. p such that (€100 if Ea) ~ (€100 with prob. p). Then (algebra …) B(E) = p.Problem: measuring indifferences is difficult.
23
Proposal 3 (common in proper scoring rules): Calibration …Problem: Need many repeated observations.
24
25
Our proposal: Take the best of all worlds!
Get B(E) = w–1(W(E)) without measuring U,W, and w from decision making.
Get canonical probability without measuring indifferences, or BDM.
Calibrate without needing long repeated observations.
Do all that with no more than simple proper-scoring-rule questions.
26
We reconsider explicit expressions:
U´(1–r2)U´(1 – (1–r)2)
(1–r)r +
rp = w
–1(
)
U´(1–r2)U´(1 – (1–r)2)
(1–r)r +
rB(E) = w
–1(
)
Corollary. p = B(E) if related to the same r!!
27
Because p = B(E) if related to same: We simply measure the R(p) curves, and use their inverses:B(E) = R–1(rE).
Appling R–1 is called risk correction.
Directly implementable empirically.
We did so in an experiment, and found plausible results.
Experimental Test of Our Correction Method
28
Method
Participants. N = 93 students. Procedure. Computarized in lab. Groups of 15/16 each. 4 practice questions.
29
30Stimuli 1. First we did proper scoring rule for unknown probabilities. 72 in total.
For each stock two small intervals, and, third, their union. Thus, we test for additivity.
31Stimuli 2. Known probabilities: Two 10-sided dies thrown. Yield random nr. between 01 and 100.Event E: nr. 75 (etc.).Done for all probabilities j/20.
Motivating subjects. Real incentives. Two treatments. 1. All-pay. Points paid for all questions.6 points = €1. Average earning €15.05.2. One-pay (random-lottery system).One question, randomly selected afterwards, played for real. 1 point = €20. Average earning: €15.30.
Figure 9.1. Empirical density of additivity bias for the two treatments
Fig. b. Treatment t=ALL
0.60
20
40
60
80
100
120
140
160
0.4 0.2 0 0.2 0.4 0.6
Fig. a. Treatment t=ONE
0
20
40
60
80
100
120
140
160
0.6 0.4 0.2 0 0.2 0.4 0.6
For each interval [, ] of length 0.05 around , we counted the number of additivity biases in the interval, aggregated over 32 stocks and 89 individuals, for both treatments. With risk-correction, there were @ > 60 additivity biases between 0.375 and 0.425 in the treatment t=ONE, and without risk-correction there were @<100 such; etc.
corrected
corrected
uncorrected
uncorrected
Summary and ConclusionModern decision theories show that proper scoring rules are heavily biased.We present ways to correct for these biases (get right beliefs even if EV violated).Experiment shows that correction improves quality and reduces deviations from ("rational"?) Bayesian beliefs.Correction does not remove all deviations from Bayesian beliefs. Beliefs seem to be genuinely nonadditive/nonBayesian/sensitive-to-ambiguity.