Probability basics Probability theory offers a mathematical language for describing uncertainties: p(disease d) = 0.35 Probability theory and assessment 0 0.5 1 In a medical problem, a probability p(disease d) = 0.35 for a patient is interpreted as: • in a population of 1000 similar patients, 350 have the disease d; • the attending physician assesses the likelihood that the patient has the disease d at 0.35. 77 / 401 78 / 401 A frequentist’s probabilities • in theory, a probability Pr(d) is the relative frequency with which d occurs in an infinitely repeated experiment; • in practice, the probability is estimated from (sufficient) experimental data. Sources of information Sources of probabilistic information are: Example In a clinical study, 10000 men over 40 years of age have been examined for hypertension: • (experimental) data; • literature; • human experts. hypertension no hypertension 1983 8017 The first two sources are the most reliable. The probability of hypertension in a man aged 45 is estimated to be 1983 p(hypertension) = = 0.20 10000 79 / 401 80 / 401 Probability scales — an example For their new soda, an expert from Colaco has to assess the probability that the soda will turn out a national success: Subjective probabilities For assessing probabilities with experts, various tools are available: certain (almost) • probability scales; • the expert is asked, using mathematical notation, to assess: • probability wheels; • betting models; p(national success) = ? • lottery models. A subjective probability is based upon personal knowledge and experience. • the expert is asked to indicate the probability on a scale from 0 to 100% certainty: probable 100 85 75 expected fifty-fifty 50 uncertain 25 improbable (almost) impossible 81 / 401 15 0 82 / 401 Betting models — an example For their new soda, an expert from Colaco is asked to assess the probability of a national success: • the expert is offered two bets: Probability wheels A probability wheel is composed of two coloured faces and a hand: national success x euro d national failure national success d¯ national failure The expert is asked to adjust the area of the red face so that the probability of the hand stopping there, equals the probability of interest. 83 / 401 −y euro −x euro y euro ¯ then • if the expert is indifferent between d and d, x · Pr(n) − y · (1 − Pr(n)) = y · (1 − Pr(n)) − x · Pr(n) y from which we find Pr(n) = . x+y 84 / 401 Lottery models — an example For their new soda, an expert from Colaco is asked to assess the probability of a national success: • the expert is offered two lotteries: national success Hawaiian trip • subjective probabilities are coherent if they adhere to the d national failure p(outcome) d¯ p(not outcome) Coherence and calibration postulates of probability theory; • subjective probabilities are well calibrated if they reflect true frequencies. chocolate bar Hawaiian trip chocolate bar ¯ then • if the expert is indifferent between d and d, Pr(n) = p(outcome). • the second lottery is termed the reference lottery; 85 / 401 86 / 401 Overconfidence and underconfidence Heuristics Upon assessing probabilities, people tend to use simple cognitive heuristics: • representativeness: the probability of an outcome is based • a human expert is an overconfident assessor if, compared with the true frequencies, his subjective probabilities show a tendency towards the extremes; • a human expert is an underconfident assessor if, compared with the true frequencies, his subjective probabilities show a tendency away from the extremes. 87 / 401 upon the similarity with a stereotype outcome; • availability: the probability of an outcome is based upon the ease with which similar outcomes are recalled; • anchoring-and-adjusting: the probability of an outcome is assessed by adjusting an initially chosen anchor probability: 88 / 401 Pitfalls — cntd. Pitfalls Using the representativeness heuristic upon assessing probabilities, can introduce biases: • the prior probabilities, or base rates, are insufficiently taken Using the availability heuristic upon assessing probabilities, can introduce biases: • the ease of recall from memory is influenced by recency, rareness, and the past consequences for the assessor; • the ease of recall is further influenced by external stimuli: Example into account; • the assessments are based upon insufficient samples; • the weights of the characteristics of the stereotype outcome are insufficiently taken into consideration; • ... 89 / 401 • ... 90 / 401 Pitfalls — cntd. Continuous chance variables Using the anchoring-and-adjusting heuristic upon assessing probabilities, can introduce biases: • the assessor does not choose an appropriate anchor; • the assessor does not adjust the anchor to a sufficient extent: Example In decision trees, chance variables are discrete: • a discrete variable C takes a single value from a non-empty finite set of values {c1 , . . . , cn }, n ≥ 2; • the distribution associated with C is a probability mass function, which assigns a probability to each value ci of C. In reality, chance variables can also be continuous: • a continuous variable C takes a single value within a non-empty range of values [a, b], a < b; • the distribution associated with C is a probability distribution function, which defines a probability for any interval [x, y] ⊆ [a, b]. • ... 91 / 401 92 / 401 Pivoting on values For a real-estate agency, the demand for luxury apartments is a continuous chance variable A: The following procedure provides for modelling a continuous variable C in a decision tree: 1 construct a cumulative distribution function for C: • pivoting on the values of C, or • pivoting on the cumulative probabilities for C; 2 approximate the probability distribution function for C by a probability mass function for a discrete chance variable C ′ : • using the Pearson-Tukey method, or • using bracket medians. 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 50 60 70 80 90 100 no. of appartments 110 120 130 A cumulative distribution function is constructed by • assessing the cumulative probabilities for a number of values of A: Pr(A ≤ 65) = 0.08 Pr(A ≤ 90) = 0.90 Pr(A ≤ 80) = 0.65 Pr(A ≤ 110) = 0.97 • and drawing a curve through them: 1 cum. distribution Using continuous chance variables 0.8 0.6 0.4 0.2 0 50 60 70 93 / 401 80 90 100 no. of appartments 110 120 130 94 / 401 Pivoting on cumulative probabilities Pivoting on values — cntd. 0.35 For assessing cumulative probabilities, by pivoting on the values of a variable under study, lottery models can be used. Example Reconsider the demand for luxury apartments, modelled by a continuous chance variable A: 0.3 0.25 0.2 0.15 0.1 0.05 0 50 d A > 90 apartments p(outcome) d¯ p(not outcome) 70 80 90 100 no. of appartments 110 120 130 A cumulative distribution function is constructed by • finding the values for A that give a number of cumulative probabilities: Hawaiian trip chocolate bar Pr(A ≤ a1 ) = 0.05 Pr(A ≤ a3 ) = 0.50 Pr(A ≤ a2 ) = 0.35 Pr(A ≤ a4 ) = 0.95 Hawaiian trip • and drawing a curve chocolate bar through them: Other elicitation tools can be used as well. 1 cum. distribution A ≤ 90 apartments 60 0.8 0.6 0.4 0.2 0 50 95 / 401 60 70 80 90 100 no. of appartments 110 120 130 96 / 401 The Pearson-Tukey method Pivoting on cumulative probabilities — cntd. The Pearson-Tukey method approximates the distribution function for a continuous variable C by a probability mass function over a discrete variable C ′ : For assessing cumulative probabilities, by pivoting on these cumulative probabilities, lottery models can be used. Example • find the values c1 , c2 , c3 for which A ≤ x apartments Pr(C ≤ c1 ) = 0.05 Pr(C ≤ c3 ) = 0.95 Pr(C ≤ c2 ) = 0.50 Hawaiian trip d A > x apartments p(outcome) = 0.35 d¯ chocolate bar • construct the discrete chance variable C ′ with values Hawaiian trip c1 , c2 , c3 and probabilities p(not outcome) = 0.65 chocolate bar Pr(C ′ = c1 ) = 0.185 Pr(C ′ = c3 ) = 0.185 Pr(C ′ = c2 ) = 0.63 Other elicitation tools can be used as well. 97 / 401 98 / 401 Bracket medians (Clemen) The Pearson-Tukey method — an example For the real-estate agency, the demand for luxury apartments is a continuous chance variable A, for which probabilities need to be assessed: • for the cumulative probabilities Pr(A ≤ a1 ) = 0.05 Pr(A ≤ a3 ) = 0.95 Pr(A ≤ a2 ) = 0.50 we have found a1 = 62 a2 = 77 The method of bracket medians approximates the distribution function for a continuous variable C by a probability mass function over a discrete variable C ′ : 1 for a number of equally likely intervals, for example five, find the values of C for which Pr(C ≤ c1 ) = 0 Pr(C ≤ c4 ) = 0.60 Pr(C ≤ c2 ) = 0.20 Pr(C ≤ c5 ) = 0.80 Pr(C ≤ c3 ) = 0.40 Pr(C ≤ c6 ) = 1.00 2 a3 = 99 for each interval [ci , ci+1 ], i = 1, . . . , 5, establish the bracket median mi such that Pr(ci ≤ C ≤ mi ) = Pr(mi ≤ C ≤ ci+1 ) • the variable A′ is constructed with 3 Pr(A′ = 62) = 0.185 Pr(A′ = 99) = 0.185 Pr(A′ = 77) = 0.63 99 / 401 construct the discrete chance variable C ′ with Pr(C ′ = m1 ) = Pr(C ′ = m2 ) = Pr(C ′ = m3 ) = Pr(C ′ = m4 ) = Pr(C ′ = m5 ) = 0.20 100 / 401 Brackets medians — an example Bracket medians – more general and practical With the bracket medians method, using n equally likely intervals, steps 1 and 2 basically provide for finding values mi , i = 1, . . . , n, of the chance variable C, such that Pr(C ≤ mi ) = 1 + k · n, 2n k = 0, . . . , n − 1 1 , n Pr(A ≤ m1 ) = 0.10 Pr(A ≤ m4 ) = 0.70 Pr(A ≤ m2 ) = 0.30 Pr(A ≤ m5 ) = 0.90 Pr(A ≤ m3 ) = 0.50 we have found m1 = 65 m3 = 77 m5 = 89 m2 = 73 m4 = 81 to construct, in step 3, the discrete chance variable C ′ with Pr(C ′ = mi ) = For the real-estate agency, the demand for luxury apartments is a continuous chance variable A, for which probabilities need to be assessed: • for the five cumulative probabilities • the variable A′ is constructed with i = 1, . . . , n Pr(A′ = 65) = 0.20 Pr(A′ = 81) = 0.20 Pr(A′ = 73) = 0.20 Pr(A′ = 89) = 0.20 Pr(A′ = 77) = 0.20 101 / 401 M.C. Airport: probability assessments Ia The real-estate agency revisited For each of the identified chance variables, the outcome will depend on the chosen strategy, that is, on activity and site in 1975, 1985 and 1995. Reconsider the decision problem for the real-estate agency: • with the Pearson-Tukey method, the agency’s decision tree includes: the expected demand is 78 apartments; Pr(A′ = 62) = 0.185 It is assumed that the chance variables are all probabilistically independent of each other, therefore Pr(A′ = 77) = 0.63 Pr(A′ = 99) = 0.185 Pr(C) = Pr(C1 ∧ C2 ∧ C3 ∧ C4 ∧ C5 ∧ C6 ) = Pr(A′ = 65) = 0.20 • using bracket medians, the agency’s tree includes: the expected demand is 77 apartments; 102 / 401 Pr(A′ = 73) = 0.20 6 Y Pr(Ci ) i=1 Probabilities are to be assessed for • the outcome of each Ci , i = 1, . . . , 6 for each of the ∼ 100 decision alternatives of D (where D captures the sequence D75 , D85 , D95 ) Pr(A′ = 77) = 0.20 Pr(A′ = 81) = 0.20 Pr(A′ = 89) = 0.20 103 / 401 104 / 401 M.C. Airport: probability assessments Ib M.C. Airport: probability assessments IIa For each of the identified chance variables, the outcome will depend on the chosen strategy, that is, on activity and site in 1975, 1985 and 1995. Alternatively, probabilities can be assessed for • the outcome of each Ci , i = 1, . . . , 6, for the 16 decision alternatives of each D j , j = 75, 85, 95. This second option • requires less assessments; • requires easier assessments (Cij vs Ci ); • assumes that Cik is independent of Cij given Dk , j = 75, 85, k = 85, 95; • requires that for some functions fi , i = 1, . . . , 6, Pr(Ci ) = fi (Pr(Ci75 ), Pr(Ci85 ), Pr(Ci95 )). The required probabilities were established from • information from previous studies; • government administrators (group consensus). For each of the Cij (i = 1, . . . , 6, j = 75, 85, 95) cumulative distributions were assessed using • the fractile method, and • consistency checks Distributions for Ci were derived from Cij by defining Ci ≡ Ci75 + Ci85 + Ci95 3 105 / 401 106 / 401 M.C. Airport: probability assessments IIb Example Consider the 1975 noise impact of the ’all activity at Texcoco’ alternative. To establish Pr(C675 | D 75 = T-IDMG), the following numbers are assessed: min #people = ? max #people = ? Pr(#people ≤ a1 ) = 0.5 ⇒ a1 = ? Pr(#people ≤ a2 ) = 0.25 ⇒ a2 = ? Pr(#people ≤ a3 ) = 0.75 ⇒ a3 = ? etc. 400.000 800.000 640.000 540.000 700.000 107 / 401 Introduction to utility theory and assessment 108 / 401 The appraisal of consequences For various decision problems, the fundamental objectives have a natural numerical scale: • money; • percentages; • length of life; • ... For other decision problems, not all objectives have such a scale: • reputation; • attractiveness; • quality of life; • ... For such objectives, a proxy scale may be used, or a new numerical scale need be designed. The gangrene problem A 68-year old man is suffering from diabetic gangrene at an injured foot. The attending physician has to choose between two decision alternatives: • to amputate the leg below the knee, which involves a small risk of death; • to wait: • if untreated, the gangrene may cure; • if the gangrene expands, an amputation above the knee becomes necessary, which involves a larger risk of death. 109 / 401 110 / 401 The gangrene problem — continued The gangrene problem — continued The elements of the gangrene problem are organised in the following decision tree: survive amputation below knee amputated below knee p = 0.99 0.89 amputation die death p = 0.01 below knee survive 0.9 p = 0.99 die 0 p = 0.01 recover wait Reconsider the gangrene problem and compare the following appraisals of the consequences cured p = 0.70 survive worsen amputation p = 0.30 above knee p = 0.90 die p = 0.10 0.92 amputated above knee wait death recover 1.0 p = 0.70 survive worsen amputation p = 0.30 above knee 0.72 Before the decision tree can be evaluated, the various consequences need be assigned numerical appraisals. 111 / 401 p = 0.90 die p = 0.10 0.8 0 112 / 401 Lotteries D EFINITION A lottery is simply a probability distribution over a known, finite set of outcomes, that is, it is a pair L = (R, Pr) with • R = {r1 , . . . , rn }, n ≥ 2, is a set of rewards; • Pr is a probability distribution over R, with P i=1,...,n Pr(ri ) = 1. A lottery is graphically represented as: A lottery L is commonly denoted L = [ p 1 , r1 ; . . . ; p i , ri ; . . . ; p n , rn ] p1 Types of lottery There exist different types of lottery: • a lottery is termed a certain lottery if it has a single reward r with Pr(r) = 1; • a lottery is called a simple lottery if all its rewards are certain lotteries; • a lottery is coined a compound lottery if at least one of its rewards is not a certain lottery. r1 where pi = Pr(ri ), i = 1, . . . , n; pi ··· occasionally, we write [R] for short. pn ··· ri rn 113 / 401 114 / 401 The gangrene problem revisited A preference ordering Reconsider the gangrene problem. In the problem, several types of lottery occur: • the certain lottery [1.0, amputated above knee]; • the simple lottery [0.99, amputated below knee; 0.01, death]: p = 0.99 amputated below knee p = 0.01 • for all Li , Lj , Lk ∈ L, if Li Lj and Lj Lk , then Li Lk ; 0.7, cured; 0.3, [0.9, amputated above knee; 0.1, death] : p = 0.70 If Li Lj , we say that Li is preferred over Lj . cured p = 0.90 If Li Lj and Lj Li , we say that Li and Lj are equivalent; we then write Li ∼ Lj . amputated above knee p = 0.30 p = 0.10 Let L be a set of lotteries. A binary relation is termed a preference ordering on L if adheres to the following properties: • for all Li , Lj ∈ L, we have that Li Lj or Lj Li ; death • the compound lottery D EFINITION death 115 / 401 116 / 401 The gangrene problem revisited The gangrene problem — continued Reconsider the gangrene problem: • a preference ordering on the set of all lotteries of the problem specifies a total ordering on the four certain lotteries: L1 = amputated below knee L2 = death L3 = cured L4 = amputated above knee it seems evident that L3 L1 L4 L2 ; • the ordering also specifies a total ordering on the more complex lotteries: p = 0.99 p = 0.90 amputated below knee L5 Reconsider the gangrene problem, with L3 L1 L4 L2 : L1 L2 L3 L4 death p = 0.99 amputated above knee p = 0.90 amputated above knee p = 0.30 L7 p = 0.10 death death Does L7 L8 hold, or is L8 L7 ? 117 / 401 The continuity axiom (aka Archimedean axiom) for all Li , Lj , Lk ∈ L with Li Lj Lk , there is probability p such that [p, Li ; (1 − p), Lk ] ∼ Lj Consider three (certain) lotteries Li Lj Lk with rewards ri , rj , rk . The axiom states that there exists a probability p such that p 118 / 401 The independence axiom (or: substitutability) Let L be a set of lotteries and let be a preference ordering on L. Then, the continuity axiom asserts: rj L8 amputated below knee cured death it seems evident that L5 L6 . 1.0 p = 0.70 p = 0.01 p = 0.10 amputatied below knee death cured amputated above knee For L6 p = 0.01 = = = = ri Let L be a set of lotteries and let be a preference ordering on L. Then, the independence axiom asserts: for all Li , Lj , Lk ∈ L with Li ∼ Lj and for each probability p, we have that [p, Li ; (1 − p), Lk ] ∼ [p, Lj ; (1 − p), Lk ] Consider three lotteries Li , Lj , Lk . The independence axiom states that if Li and Lj are equivalent, then so are p ∼ 1−p Li p Lj rk 1−p • p is termed the calibration probability for Lj and [p, Li ; (1 − p), Lk ]; • Lj is termed the certainty equivalent for [p, Li ; (1 − p), Lk ]. 119 / 401 Lk 1−p Lk for all probabilities p. 120 / 401 The compound lottery axiom The unequal probability axiom (or: monotonicity) Let L be a set of lotteries and let be a preference ordering on L. Then, the unequal probability axiom asserts: for all Li , Lj ∈ L with Li Lj and for all probabilities p, p′ with p ≥ p′ , we have [p, Li ; (1 − p), Lj ] [p′ , Li ; (1 − p), Lj ] Let L be a set of lotteries and let be a preference ordering on L. Then, the compound lottery axiom asserts: for all Li , Lj ∈ L, Lj = [q, Lm ; (1 − q),Ln ], Lm , Ln ∈ L, 0 ≤ q ≤ 1, and for each probability p, we have [p, Li ; (1−p), Lj ] ∼ [p, Li ; (1−p)·q, Lm ; (1−p)·(1−q), Ln ]. Consider two lotteries Li , Lk with Lk = [q, Li ; (1 − q), Lj ], Lj ∈ L, 0 ≤ q ≤ 1. The compound lottery axiom states that Consider two lotteries Li , Lj with Li Lj . The unequal probability axiom states that p p p′ Li 1−p Lj Li q 1−p 1 − p′ for all probabilities p ≥ p′ . 1−q Lj p + (1 − p) · q Li ∼ Li (1 − p) · (1 − q) Li Lj Lj for all probabilities p. The axiom is also termed “no fun in gambling”. 121 / 401 122 / 401 The assumption of finiteness A rational preference ordering Consider the following decision problem: D EFINITION lose all your money Let L be a set of lotteries. A preference ordering on L is a rational preference ordering if it adheres to the Von Neumann – Morgenstern utility axioms: • the continuity axiom; • the independence axiom; • the unequal probability axiom; • the compound lottery axiom. 123 / 401 win billions gamble p = 0.99999 everything bad p = 0.00001 bankruptcy riches infinite misery Bayes criterion for choosing between decision alternatives does not help much if the problem involves consequences of infinite appraisal. 124 / 401 The main theorem — a sketchy proof Consider two lotteries L and L′ with the rewards r1 · · · rn , n ≥ 1: The main theorem of utility theory L′ = [p′1 , r1 ; · · · ; p′n , rn ] L = [p1 , r1 ; · · · ; pn , rn ] T HEOREM Let L be a set of lotteries and let be a rational preference ordering on L. Then, there exists a real function u on L such that • for all Li , Lj ∈ L, we have that Li Lj iff u(Li ) ≥ u(Lj ); X i=1,...,n [ui , r1 ; (1 − ui ), rn ] ∼ ri For the two lotteries, we then have that # " X X pi · ui ), rn pi · ui , r1 ; (1 − L ∼ • for each [p1 , L1 ; . . . ; pn , Ln ] ∈ L, we have that u([p1 , L1 ; . . . ; pn , Ln ]) = For each reward ri we can find a value 0 ≤ ui ≤ 1, such that pi · u(Li ). i=1,...,n i=1,...,n L The function u is termed a utility function on L. ′ ∼ " X i=1,...,n p′i · ui , r1 ; (1 − So, L L′ if and only if 125 / 401 X i=1,...,n X · ui ), rn X p′i · ui . i=1,...,n p i · ui ≥ # p′i i=1,...,n 126 / 401 Utility versus expected reward Example Consider the set of euro rewards R = {0, 10, 20, 50}, and the following two lotteries over R: Some notes The main theorem of utility theory implies that: L1 = [0.0, 0; 1.0, 20], • a lottery with highest utility is a most preferred lottery; and L2 = [0.7, 10; 0.3, 50] Although IEr(L2 ) > IEr(L1 ), the decisionmaker may express the preference L1 ≻ L2 without being irrational! Consider two possible utility functions over the rewards, expressing that more money is preferred to less: • any rational preference ordering on a set of lotteries is encoded uniquely by the utilities of its certain lotteries. u1 (0) = 0, u1 (10) = 0.2, u1 (20) = 0.4, u1 (50) = 1 The main theorem does not imply that: • a lottery with highest expected reward is a most preferred lottery 127 / 401 u2 (0) = 0, u2 (10) = 0.1, u2 (20) = 0.6, u1 (50) = 1 Note that u1 (L1 ) < u1 (L2 ), but u2 (L1 ) > u2 (L2 ). Note: even if objectives have a natural numerical scale, preferences (over lotteries) may be such that a utility function is required! 128 / 401 Strategic equivalence D EFINITION Let L be a set of lotteries. Two utility functions ui , uj on L are strategically equivalent, written ui ∼ uj , if they imply the same preference ordering on L. Example: Consider the following two utility functions for the gangrene problem: u1 (amputated below knee) = 0.9 u1 (death) = 0 u1 (cured) = 1.0 u1 (amputated above knee) = 0.8 and u2 (amputated below knee) = 2.98 u2 (death) = 1 u2 (cured) = 3.2 u2 (amputated above knee) = 2.76 A linear transformation of utilities T HEOREM Let L be a set of lotteries and let ui , uj be two utility functions on L, then ui ∼ uj ⇐⇒ uj = a · ui + b, for some constants a, b with a > 0 A utility function is unique up to a positive linear transformation. The functions u1 and u2 are strategically equivalent. 129 / 401 130 / 401 The gangrene problem revisited Normalisation Let X be an attribute with values x1 . . . xn , n > 1. Often, utility functions are normalised such that, for example, Reconsider the gangrene problem with u(amputated below knee) = 0.9 u(death) = 0 u(cured) = 1.0 u(amputated above knee) = 0.8 For the two lotteries u(x1 ) = 0 and u(xn ) = 1 p = 0.99 L5 These values only set the origin of u(X) and the unit of measurement. amputated above knee L6 p = 0.01 If we decide to consider, for example, xi ≺ x1 or xj ≻ xn , then u(xi ) < 0, and u(xj ) > 1, respectively. p = 0.90 amputated below knee death p = 0.10 death we have that u(L5 ) = 0.99 · u(amputated below knee)+ +0.01 · u(death) = 0.99 · 0.9 = 0.89 u(L6 ) = 0.90 · u(amputated above knee)+ +0.10 · u(death) = 0.90 · 0.8 = 0.72 131 / 401 So, L5 L6 . 132 / 401 The gangrene problem — continued Reconsider the gangrene problem with u(amputated below knee) = 0.9 u(death) = 0 u(cured) = 1.0 u(amputated above knie) = 0.8 For the two lotteries p = 0.99 • subjective assessment • direct methods • magnitude estimation/production p = 0.70 amputated below knee L7 p = 0.01 Utility assessment L8 cured p = 0.90 • ratio estimation/production amputated above knee p = 0.30 death p = 0.10 • indirect (behavioural) methods • based on reference gambles • ”objective” assessment death we have that u(L7 ) = 0.99 · 0.9 = 0.89 • choose a mathematical function u(L8 ) = 0.70 · 1.0 + 0.3 · 0.9 · 0.8 = 0.92 So, L8 L7 . 133 / 401 134 / 401 Direct methods Direct methods Ratio estimation or production: Magnitude estimation or production can be done using a utility scale: 1.0 0.5 Example Reconsider the diabetic gangrene treatment example. To assess the utilities for the different treatment consequences, a patient is asked one of the following types of question: • ”How do you value life after an above-knee amputation?” (estimation); • ”Which consequence do you associate with a utility of 0.2?” (production) 0 135 / 401 Example My sister is interested in buying a new car. To assess the utilities for the different car options, she is asked one of the following types of question: • ”How much more do you value a Volvo than a Fiat?” (estimation); • ”Which car seems to you twice as valuable as a Fiat?” (production) This method was used to assess the empirical utility of money (Galanter, 1962). √ empirical utility function for monetary gain: ∼ x empirical utility function for monetary loss: ∼ −x2 136 / 401 Assessment using certainty equivalents Reference gamble Consider a set of consequences for which utilities are to be assessed. Let ci , cj , ck be consequences from that set such that ci cj ck . A reference gamble is a choice between two lotteries: 1 the certain lottery L = [1.0, cj ] ′ 2 the simple lottery L = [p, ci ; (1 − p), ck ] The utilities of a decision maker for the possible consequences in a decision problem can be assessed using several certainty equivalents: 1 2 elicit a preference order on consequences from most preferred (c+ ) to least preferred (c− ); assign the first two points of the utility function: u(c+ ) = 1 and u(c− ) = 0 • L′ is the reference lottery of the ”gamble” — if u(ci ) = 1.0 3 • for fixed ci , cj , and ck , the probability p for which L ∼ L′ is 5 and u(ck ) = 0, then L is called a standard reference lottery; ′ the indifference probability for L and L′ ; • for fixed p and ci , ck , the consequence cj for which L ∼ L′ is the certainty equivalent, or indifference point, for L′ . 4 6 7 create a standard reference gamble with p = 0.5; elicit the certainty equivalent cCE for the reference lottery; compute the utility of this indifference point; create two reference gambles with p = 0.5 and consequences c+ and cCE , and cCE and c− , respectively; repeat steps 4 – 6 until enough points are found to draw the utility curve. 137 / 401 138 / 401 Computing a utility An example Consider the lotteries L and L′ of a reference gamble: ′ L = [1.0, cj ] and L = [p, ci ; (1 − p), ck ] where ci cj ck . Assume that the utilities u(ci ) and u(ck ) for consequences ci and ck are known. Question: How do we compute u(cj ) for consequence cj ? Suppose you have an old computer for which components are bound to need replacement in the near future. You have a number of replacement alternatives that will cost you between e 50 and e 500. What is your utility function? Two points can be fixed: u(e 500) = 0 and u(e 50) = 1 Answer: If either You are then presented with the following gamble: p is the indifference probability for L and L′ , or cCE 0.5 0.5 cj is the certainty equivalent of L′ 500 Suppose that cCE = 200 is the indifference point for this gamble. We then have that then L ∼ L′ and consequently 1.0 · u(cj ) = p · u(ci ) + (1 − p) · u(ck ). From this equation we can solve u(cj ). 50 1.0 · u(200) = 0.5 · u(50) + 0.5 · u(500) = 0.5 139 / 401 140 / 401 An example Assessment using probability equivalents The utilities of a decision maker for the possible consequences in a decision problem can be assessed using several probability equivalents: 1 2 elicit a preference order on consequences from most preferred (c+ ) to least preferred (c− ); assign the first two points of the utility function: 4 5 amputated below knee The patient is given the following gamble: u(c+ ) = 1 and u(c− ) = 0 3 Reconsider the decision problem for treatment of diabetic gangrene. The possible consequences of treatment are: cured 1.0 amputated below knee ? amputated above knee ? death 0.0 create a standard reference gamble for enough intermediate outcomes; elicit the indifference probability for the lotteries in each of the gambles; compute the utility of an intermediate outcome using this indifference probability. p cured 1−p death Suppose that p = 0.9 is the indifference probability for the lotteries. We then have that u(amputated below knee) = 0.9·u(cured)+0.1·u(death) = 0.9 141 / 401 142 / 401 Example: extreme utilities Direct vs indirect methods direct: • roots in psychophysics • inferior in both validity & reliability • easily applied to risky tasks with complex consequences indirect: • roots in utility axioms • time consuming • irrelevant ”gaming” effect • distasteful / unethical • unsuitable for measuring very small or very large utilities 143 / 401 For the diabetic gangrene decision problem, we could use the following gambles: p 1−p amputated above knee amputated below knee death amputated below knee q cured 1−q amputated above knee 144 / 401 Subjective utility assessments • can change overtime Risk attitudes • are required for unknown / unexperienced outcomes • are influenced by framing and certainty effects • are influenced by third parties • are not comparable from person to person • ... 145 / 401 146 / 401 Risk-neutral preferences Let X be an attribute with values x1 . . . xn , n ≥ 2, measured in some unit. Let L = [p, xi ; (1 − p), xj ] be a lottery over X. An Example Suppose you are given the choice between the following two “games”: 0.5 0.5 0.5 0.5 A decision maker is risk-neutral if each additional unit of X is valued with the same increase in utility: Example e5 D EFINITION A decision maker is risk-neutral, if e −1 u(p · xi + (1 − p) · xj ) = e6 p · u(xi ) + (1 − p) · u(xj ) e −2 that is, the utility function is linear. Do you have a clear preference for playing one or the other? Note that u′′ (X) = 0. 147 / 401 148 / 401 Risk-averse preferences An Example Let X be an attribute with values x1 . . . xn , n ≥ 2, measured in some unit. Let L = [p, xi ; (1 − p), xj ] be a lottery over X. Suppose you are given the choice between the following two “games”: 0.5 0.5 0.5 0.5 A decision maker is risk-averse if each additional unit of X is valued with a smaller increase in utility: Example e 52 D EFINITION A decision maker is risk-averse, if u(p · xi + (1 − p) · xj ) > e −2 e 5000 p · u(xi ) + (1 − p) · u(xj ) e −4950 that is, the utility function is concave. Which game would you prefer to play? Note that u′′ (X) < 0. 149 / 401 150 / 401 Risk-prone preferences Let X be an attribute with values x1 . . . xn , n ≥ 2, measured in some unit. Let L = [p, xi ; (1 − p), xj ] be a lottery over X. An Example Suppose you are given the choice between the following two “games”: A decision maker is risk-prone, or risk-seeking, if each additional unit of X is valued with a larger increase in utility: Example 0.2 0.8 0.8 0.2 D EFINITION A decision maker is risk-prone, if e 10 e 0, 10 e 2, 50 u(p · xi + (1 − p) · xj ) < e1 p · u(xi ) + (1 − p) · u(xj ) that is, the utility function is convex. Which game would you prefer to play? Note that u′′ (X) > 0. 151 / 401 152 / 401 Discounting Risk attitudes • the process of translating future rewards to their present value; • necessary to perform ”cost-benefit” analyses now. The zero-illusion curve: u(X) u(X) RA RA Examples RS • the time value of money value of an euro depends on when it is available euros can be invested to yield more euros RA RS 0 X 0 • the time value of life X life years in the future less valuable than today (?) life years are valued relative to money 153 / 401 154 / 401 An example Consider the following ‘standard reference gamble’: A discountingfactor 0.50 A discountingfactor indicates how much less a patient values each successive year of life, compared to the previous year: The utility function for length of life with a constant discountingfactor δ is approximated by Z x u(x) = e−δ·t dt t=0 for life-expectancy x. 155 / 401 1.0 x years ∼ 0.50 25 years 0 years For a utility function for length of life with a constant discountingfactor δ = 0.02, we find: u(25 years) = u(0 years) = and Z 25 Z 0 −0.02·t e−0.02·t dt = 19.67 e dt = 0 t=0 t=0 The patient should be indifferent about the choice between the two lotteries for a life-expectancy x for which u(x) = 0.5 · 19.67 + 0.5 · 0 = 9.835 Z x We find from e−0.02·t dt = 9.835 that x ≈ 11 years. t=0 156 / 401 An example The risk-premium Let X be an attribute with values x1 . . . xn , n ≥ 2, measured in some unit. Let u(X) be a utility function over X. Consider a lottery L = [p, xi ; (1 − p), xj ] over X. Let xC be the certainty equivalent of lottery L and xE the expected value of L. Let u(x) = −e for all x ∈ [0 . . . 50] Consider the lottery L = [0.5, 0; 0.5, 10]. Compute the risk premium for this lottery. −0.2x The expected value xE of the lottery is 0.5 · 0 + 0.5 · 10 = 5 The certainty equivalent xC of the lottery is determined from The risk premium RP of L is defined as: xE − xC if u(X) is increasing RP = xC − xE if u(X) is decreasing u(xC ) = 0.5 · u(0) + 0.5 · u(10) = −0.5 − 0.5 · e−2 ≈ −0.57 and equals 2.83. As u(x) is an increasing function, the risk premium for L is RP = 5 − 2.83 = 2.17 Exercise Let u(x) = − log(x + 30) for all x > −30. Consider the lottery L = [0.5, −20; 0.5, −10]. What is the lottery’s risk premium? Is the decision maker risk averse or risk prone? 157 / 401 Risk averseness & risk premium 158 / 401 Risk proneness & risk premium A decision maker is • risk averse iff his/her risk premium is positive for all nondegenerate lotteries; • decreasingly risk averse iff he/she is risk averse and • his/her risk premium for any lottery [0.5, x − h; 0.5, x + h] decreases (↓ 0) as x increases; • increasingly risk averse iff he/she is risk averse and • his/her risk premium for any lottery [0.5, x − h; 0.5, x + h] increases (↑ ∞) as x increases; • constantly risk averse iff he/she is risk averse and • his/her risk premium for any lottery [0.5, x − h; 0.5, x + h] remains constant for all x. 159 / 401 A decision maker is • risk prone iff his/her risk premium is negative for all nondegenerate lotteries; • decreasingly risk prone iff he/she is risk prone and • his/her risk premium for any lottery [0.5, x − h; 0.5, x + h] increases (↑ 0) as x increases; • increasingly risk prone iff he/she is risk prone and • his/her risk premium for any lottery [0.5, x − h; 0.5, x + h] decreases (↓ −∞) as x increases; • constantly risk prone iff he/she is risk prone and • his/her risk premium for any lottery [0.5, x − h; 0.5, x + h] remains constant for all x. 160 / 401 The risk-aversion function The degree of risk aversion / proneness – Introduction Information regarding a decision maker’s risk attitude is given by RP: sign indicates aversion vs proneness; magnitude captures the degree of this behaviour, for one specific lottery! u′′ (X): sign indicates aversion vs proneness; magnitude conveys no relevant information, since strategically equivalent functions capture same risk attitude: u(X) u(x) = −3e−x u(X) u(x) = 1 − e−x D EFINITION Consider a utility function u(X), and let σ ∈ {+, −} denote the sign of u′ (x). The risk-aversion function R(X) for u(X) is then defined by R(X) = −σ u′′ (X) u′ (X) • if R(X) > 0 then the decision maker is risk averse; • if R(X) < 0 then the decision maker is risk prone; • if R(X) = 0 then the decision maker is risk neutral; T HEOREM RP x (xE ) x−h R(X) is increasing (decreasing, constant) iff the decision maker’s risk premium for any lottery [0.5, x − h; 0.5, x + h] is increasing (decreasing, constant) for increasing x. RP x+h x−h X certainty equivalent x (xE ) x+h X certainty equivalent 161 / 401 An example Let u(x) = 1 − e , x > 0, be a utility function. Find the risk-aversion function for X. −x/900 We have that u′ (x) = 1 · e−x/900 , 900 x>0 and −1 · e−x/900 , x > 0 9002 Since u(x) is an increasing function in x (u′ (x) > 0 for all x > 0), the risk-aversion function is defined as u′′ (X) R(X) = − ′ u (X) u′′ (x) = 1 and that u(X) models constant We conclude that R(X) = 900 risk aversion. 163 / 401 162 / 401