1 / 50

Decision making Outside the Laboratory

Decision making Outside the Laboratory. Marcel Boumans University of Amsterdam. What’s the problem?. Problem = Model Target probability Information structure Sample space Rationality = Optimal solution within a model. Real-life problem. Model 1. ?. Model 2. ?. Model 3.

santos
Télécharger la présentation

Decision making Outside the Laboratory

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Decision makingOutside the Laboratory Marcel Boumans University of Amsterdam

  2. What’s the problem? Problem = Model • Target probability • Information structure • Sample space Rationality = Optimal solution within a model

  3. Real-life problem Model 1 ? Model 2 ? Model 3

  4. Experiment ? = Model of the experiment Model?

  5. Example 1 Decision making in medicine (Boumans 2008, Battle in the Planning Office: Field Experts versus Normative Statisticians, Social Epistemology 22.4)

  6. Harvard Medical School Test We asked 20 house officers, 20 fourth-year medical students and 20 attending physicians, selected in 67 consecutive hallway encounters at four Harvard Medical School teaching hospitals, the following question: “If a test to detect a disease whose prevalence is 1/1000 has a false positive rate of 5 per cent, what is the chance that a person found to have a positive result actually has the disease, assuming that you know nothing about the person’s symptoms or signs?”

  7. The ‘Wrong’ Answer The most common answer, given by 27: 95%. The average of all answers: 55.9%

  8. The ‘Correct’ Answer Only 11 of 60 (18%) participants gave this answer. P: disease is present, A: disease is absent, +: positive test result

  9. David Eddy’s test Probability that the breast mass is malignantPr(ca)= .01 Mammography Accuracy: sensitivity: Pr( + | ca) = 79.2% specificity: Pr ( - | benign) = 90.4% Pr(ca | + ) ?

  10. The ‘Wrong’ Answer Most physicians (approximately 95 out of 100) estimated the posterior probability to be about 75%.

  11. The ‘Correct’ Answer

  12. Base-Rate Fallacy Pr(ca | + ) and Pr( + | ca) “The latter probability is the one measured in clinical research programs and is very familiar, but it is the former probability that is needed for clinical decision making. It seems that many if not most physicians confuse the two.”

  13. Diagnostic Testing Question: Can I apply this valid, important test to a specific patient? Could it move us across a test-treatment threshold?

  14. Most Informative Test Pr(P | + ) and Pr(A |  ) are optimal: 40% < Pr(P) < 60%

  15. Reasonable parity 40% < Pr(P) < 60% Sensitivity  Specificity Pr(P | + )  Pr( + | P)

  16. Reevaluation of Tests 40% < Pr(P) < 60% Harvard Medical School TestPr(  | A) = 95% 93% < Pr(P | + ) < 95% Eddy’s testPr( + | P) = 79.2% and Pr(  | A) = 90.4% 84.6% < Pr(P | + ) < 92.5%

  17. Reasonable assumption “The medical students and staff that were given this question would know full well that patients having a disease are almost more likely to have a test for their disease administered to them than the general public […]. The majority response of 95% is consistent with the assumption that persons having the test applied to them have a 50% chance of actually having the disease […]. Certainly these assumptions are more reasonable than those needed to support the 2% answer. Perhaps this illustration shows not that medically trained people don’t understand probability but that some statisticians don’t understand medicine.”

  18. Example 2 Existence of Hot Hand Statistical Reality or Cognitive Illusion?

  19. Tversky & Gilovich: NO “Sequences of hits and misses in a basketball game offer an interesting context for investigating the perception of randomness outside the psychological laboratory.” Player’s performance records should differ from sequences of heads and tails produced by coin tossing in two essential respects. • P(H|H) > P(H|M) • #streaks hits/misses > #streaks by chance process

  20. Larkey, Smith and Kadane: YES Not only looking at isolated player shooting sequences, but taking account of context. where Pi = probability of a hit given a shot by player i. L = Length of run m = Number of possible shots = 20 i = probability of player i taking a shot G = Number of games. where Ag = All field goal attempts in game g.

  21. Example 3 Monty Hall problem

  22. Daniel Friedman (1998)The American Economic Review “Sometimes people make decisions that seem inconsistent with rational choice theory. We have a “choice anomaly” when such decisions are systematic and well documented.” “The problem is not just theoretical. Economists’ traditional analytical tools are based squarely on rational choice theory, so anomalies provoke debates about the practical value of these tools. The present paper offers some new laboratory evidence bearing on the anomalies debates It focuses on a particular choice task, chosen (after a search) for its robustness and persistence.”

  23. Run 1: 40 cents versus 10 cents Directions: • In front of you are three cards labeled A, B, C. • Two of the three cards are the Low Cards and are worth 10 cents. • One of the three cards is the Prize Card and is worth 40 cents. • Pick the card that you believe is the prize Card. • The conductor will flip over a Low Card that you did not choose initially • Decide whether or not you want to switch or remain with your card you chose initially. • The conductor will record your responses and your earnings. • Repeat 10 periods.

  24. Result

  25. Discussion “One conclusion seems clear. The three-door task now deserves a place among the leading choice anomalies. … Indeed, I am not aware of any anomaly that has produced stronger departures from rationality in a controlled laboratory environment. The ultimate reasons for its strength remain unclear, but a proximate reason is that the task ties together several classic anomalies, including the gambler’s fallacy, commitment escalation and the endowment effect, Bayesian updating failures, and probability matching. It would appear difficult to accommodate the three-door anomaly into a useful generalization of expected utility theory.”

  26. Crucial Question “The anomalous laboratory evidence confronts economists with at least two hard questions. Why are people so irrational? When (if ever) can people become more rational in tasks of this sort?” But are people so irrational? Does laboratory evidence show that this is the case?

  27. The first version A Problem in Probability S. Selvin (1975) The American Statistician It is “Let’s Make a Deal” – a famous TV show starring Monte Hall. Monte Hall: One of the three boxes labeled A, B, and C contains the keys to that new 1975 Lincoln Continental. The other two are empty. If you choose the box containing the keys, you win the car. Contestant: Gasp! Monte Hall: Select one of these boxes. Contestant: I’ll take box B. Monte Hall: Now box A and box C are on the table and here is box B (contestant grips box B tightly). It is possible the car keys are in that box! I’ll give you $100 for the box. Contestant: No, thank you.

  28. Monte Hall: how about $200? Contestant: No! Audience: No! Monte Hall: remember that the probability of your box containing the keys to the car is 1/3 and the probability of your box being empty is 2/3. I’ll give you $500. Audience: No! Contestant: No, I think I’ll keep this box. Monte Hall: I’ll do you a favor and open one of the remaining boxes on the table (he opens box A). It’s empty! (Audience: applause). Now either box C or your box B contains the car keys. Since there are two boxes left, the probability of your box containing the keys is now ½. I’ll give you $1000 cash for your box. WAIT!!! Is Monte right? The contestant knows at least one of the boxes on the table is empty. He now knows it was box A. Does this knowledge change his probability of having the box containing the keys from 1/3 to ½? One of the boxes on the table has to be empty. Has Monte done the contestant a favor by showing him which of the two boxes was empty? Is the probability of winning the car ½ or 1/3? Contestant: I’ll trade you my box B for the box C on the table. Monte Hall: That’s weird!!

  29. Marilyn vos Savant “Ask Marilyn” in Parade Magazine 1990 Suppose you're on a game show, and you're given the choice of three doors. Behind one door is a car, behind the others, goats. You pick a door, say #1, and the host, who knows what's behind the doors, opens another door, say #3, which has a goat. He says to you, “Do you want to pick door #2?” Is it to your advantage to switch your choice of doors?

  30. Marilyn’s Answer “Yes; you should switch. The first door has a 1/3 chance of winning, but the second door has a 2/3 chance. Here's a good way to visualize what happened. Suppose there are a million doors, and you pick door #1. Then the host, who knows what's behind the doors and will always avoid the one with the prize, opens them all except door #777,777. You'd switch to that door pretty fast, wouldn't you?” She also gave 3 other explanations why to switch.

  31. Rao & Rao (1992)Mathematical Scientist Scenario 1: The participant contacts a statistician as to the best course of action to be taken to maximize the probability of winning the automobile before getting on to the stage. Answer: ⅔ Scenario 2: The situation in which the participant is actually on the stage. Answer: Pr(C1 | G3) = Pr(C1 | C2 or C1) = Pr(C1)/[Pr(C2) + Pr(C1)] = (⅓)/(⅔) = ½

  32. What is the problem? What is the conditional probability of winning a car by switching doors when door 3 shows a goat? Rao & Rao: not the correct interpretation, but the one of those who criticized vos Savant’s solution.

  33. Role of Monty Hall? Granberg and Brown (1995) Personality and Social Psychology Bulletin: “Most people seem to ignore, or at least do not adequately take into account, the knowledge of the host as a cue.” Vos Savant: “pure probability is the paradigm, and we published no significant reason to view the host as anything more than an agent of chance who always opens a losing door and offers the contestant the opportunity to switch”

  34. How to model Monty Hall? subset of C  H or subset of C ? specification of ?

  35. Morgan, Chaganty, Dahiya and Doviak (1991)The American Statistician [Vos Savant gave] a dubious analogy to explain the choice. [She] went on to defend her original claim with a false proof and also suggested a false simulation as a method of empirical verification. Vos Savant does not back down, and for good reason, as, given a certain assumption, her answer is correct. Her methods of proofs, however, are not. [T]his apparently innocuous little problem can be erroneously “solved” in a variety of ways and that the nature of these errors can be quite subtle, at least to those not accustomed to thinking in terms of conditional probability (and sometimes even those of us who are!).

  36. General solution for the conditional problem vos Savant’s strategy: p23 = 1

  37. Vos Savant’s reply “strong attempts at misinterpretation” “Nearly all of my critics understood the intended scenario, and few raised questions of ambiguity.”

  38. Gillman (1992)American Mathematical Monthly Game I: Pr(Ws | G3) = 1/(1 + q) where q = Pr(H3 | C1) = p13 “depends on [the host’s] selection strategy when he has this choice – on the probabilityq that he will then open door #3” “Marilyn did not address this question” Game II: Pr(Ws) = 2/3 (Morgan et al. misrepresented this probability)

  39. Chun (1999)The American Statistician • Information Economics Approach • Information matrix • Decision matrix • Payoff matrix • Prior probabilities matrix

  40. Modeling problems When the player has selected door No. 1, the game ends immediately if (a) the host promptly opens door No. 1, or (b) the host opens one of the two remaining doors and shows the car behind it. If the host opens a different door that does not have the prize behind it, then the player is offered a chance to switch to the remaining, unselected door.

  41. Real-world Problems versus Textbook Exercises “Textbook problems are often set in artificial situations that neither require nor inspire real-world thinking on the student’s part”(Morgan et al.)

  42. Textbook Exercises vs.Real-life Problems Outside the never-never land of textbooks, however, conditioning events are not handed out on silver platters. They have to be inferred, determined, extracted. In other words, real-life problems (or textbook problems purporting to describe real life) need to be modeled before they can be solved formally. And for the selection of an appropriate model (i.e., probability space), the way in which information is obtained (i.e. the statistical experiment) is crucial. – Bar-Hillel and Falk (1982)

  43. Bohl, Liberatore and Nydick (1995) Journal of Recreational Mathematics We believe that the issue of clearly stating, then questioning, and finally agreeing on, one or more sets of appropriate assumptions is critical to the success of any modeling effort. “Buying in” to the assumptions is tantamount to buying in to the solution. In the long run, you cannot have the second without the first.

  44. Reasonable Set of Assumptions • The game show host will always open a door that was not selected by the contestant. • The opened door will always reveal a goat. • If the contestant selects a car, the game show host will select one of the other two doors with equal probability. • The car is more valuable than the goats. • The position of the car will not change once the game begins. • The contestant cannot offer to sell the door to the host or to anyone else. • The game show does not have the option to offer the contestant money to purchase his or her door. • The game show host will always offer the opportunity to switch. • The game is not repeatable, that is, the contestant only has the opportunity to play the game once. • The initial location of the car was randomly determined prior to the start of the game.

  45. Modeling Target Probability: Pr(Ws) Pr(Ws | G3), Pr(Ws | G2orG3) Strategy of the Host: pij = Pr(Hj | Ci) Sample Space: subset of C  H

  46. General

  47. vos Savant: p12 = p13 = ½, p23 = p32 = 1

  48. Random agent: pi1 = pi2 = pi3

  49. What should be learned? Friedman (1998): Assertion. – Every choice “anomaly” can be greatly diminished or entirely eliminated in appropriately structured learning environments. People are “not hardwired to behave irrationally” “there is no such thing as an anomaly in the traditional sense of stable behavior that is inconsistent with rationality. There are only pseudo-anomalies describing transient behavior before the learning process has been completed.”

  50. Thank you for your attention!

More Related