Decision Making by M2H14i0


									Decision Making

    chapter 8
                 Daniel Kahneman
• Dan Kahneman (e.g. Kahneman & Tversky, Kahneman
  & Treisman) won the 2002 Nobel Prize in Economics
  along with GMU Professor Vernon Smith.

“Kahneman has integrated insights from psychology into
  economics, especially concerning human judgment and
  decision-making under uncertainty, the Royal Swedish
  Academy of Sciences said in its citation.” –
               What makes a decision good?
1. Maximize the expected value of the return

2. Good decisions produce good outcomes; bad decisions
   produce bad outcomes.

3. Expertise -- Experts tend to produce better decisions that
           What makes a decision good?
• Maximizing expected value of return?
   – Over time, make sure the average payoff of a series of
     decisions is as high as possible?

   – Problems include:
      • the value/cost of a payoff might be hard to estimate
      • expected value is only realized over the long term;
        value of individual decisions may vary
      • it may be more important to avoid loss than to secure
        a benefit (“loss aversion”)
            What makes a decision good?
• Example
  – An investor might believe the higher return of a stock
    investment is less valuable than the security of a
    savings account.

  – The expected value of a series of roulette bets is
    negative, but a single bet might return a positive value
           What makes a decision good?
• Producing a good result for a given decision?
   – problem: since outcomes are probabilistic, a single
     good result might not reflect the value of average,
     long-term results.

   – Example: a single winning bet on red on a roulette
     wheel does not indicate that betting red consistently is
     a good strategy.
            What makes a decision good?
• Domain experts say that it‟s good?
   – Problem: experts’ opinions about what is a good
     decision might conflict with those of first two criteria

   – Example: circumstances might meet criteria for a
     police officer to fire a weapon, even if suspect is not a
         What Constitutes a Good Decision?
•   Maximizing Expected Value
•   Producing a „good‟ outcome
•   Meeting experts‟ criteria

•   Expected Value
    – To calculate expected value:
      1. List all potential outcomes, along with their value
      2. Calculate probability of each outcome
      3. Multiply probability of each outcome by value of
         each outcome.
      4. Sum the resulting values.
Assume that in a lottery, 6 numbers are randomly chosen from the
   range 1-52. To win, a player must match all 6 numbers. The prize
   is $1 million.
Q: What is the expected value of a $ 1 lottery ticket with 2 chances to

• Probability of winning = .0000001
• Value of winning = $999,999

• Probability of losing = .9999999
• Value of losing = -$1

• Expected Value of $1 bet = $999,999  .0000001  $1 .9999999
                            = -$0.90

“A tax on people who 
                     aren‟t very good at math” – Garrison Keillor
Assume that two people agree to bet on tosses of a coin. For
  every toss that lands heads, Person A wins $1. For every toss
  that lands tails, Person A loses $1

Q: What is the expected value of a coin toss to Person A?

P(heads) = .5        Value(heads) = $1
P(tails) = .5        Value(tails) = -$1

E(coin toss) = (.5 x $1)+(.5 x -$1)
               = $0.5 - $0.5
Note that at after any given coin toss, one person or the other
  may be momentarily ahead of the other
                  Perception of Cues
• To make a good decision, people need to be able to
  properly assess the situation.

• That is, they need to look for clues that will guide them
  in their decision making.

• However, humans being humans, there are some
  inherent biases and weaknesses in their ability to
  correctly perceive clues…
                  Perception of Cues
• Humans are good at
   – estimating the mean of multiple values
   – estimating proportions that aren’t too extreme

• Humans are poorer at:
   – estimating extreme proportions
      • If I have seen 99 normal parts, then detecting 1
        abnormal part will have more of an impact than the
        100th normal part
                      Perception of Cues
• Humans are poorer at:
  – estimating extreme proportions
  – extrapolating nonlinear trends








                  0    1   2   3   4   5   6   7
                  Perception of Cues
• Humans are poorer at:
  – estimating extreme proportions
  – extrapolating nonlinear trends

  – estimating variance
     • estimations are affected overall
     • tend to estimate ratio of
       variance to mean magnitude
                 Perception of Cues
• Humans are poorer at:
  – estimating extreme proportions
  – extrapolating nonlinear trends
  – estimating variance

  – estimating degree of correlation in scatter plots
     • tend to overestimate small correlations and
       underestimate large correlations
              Cues and Cue Integration
• Observer must attend to & integrate cues to diagnose a
  situation & establish an hypothesis about the state of the
• Cues can be characterized by 3 properties:
   – Diagnosticity
      • How much evidence does cue provide for a given
   – Reliability
      • How much can the cue be trusted?
      • Information Value = Diagnosticity x Reliability
   – Salience
      • How conspicuous is the cue?
               Cues and Cue Integration
Example: A financial website predicts that some company‟s
  stock will rise. Should you consider investing in this stock?

   – Diagnosticity
      • Does prediction say that rise is almost certain, or is
        only a bit more likely than not?
   – Reliability
      • Have this website‟s predictions been accurate in the
   – Salience
      • Is prediction written in big letters at the top of the
        page or in smaller letters near bottom of page?
         Quality of situation assessment?
• How well a person assesses the situation is dependent
   – Comprehensiveness of Info
      • Is important information missing?
   – Quantity of information
      • too little? too much?
   – Relative Salience of cues
      • Are some cues more salient than others
   – Relative weighting (importance) given to cues
      • Are some cues considered more important?
          Quality of situation assessment?
• Comprehensiveness of info
   – Important information might be lacking
   – a good decision maker should know what info is

   – Example
      • A computer user calling for tech support might
        neglect to report a valuable symptom of his
        computer's problem. An experienced technician
        should recognize that symptom has not been discussed
        and ask for relevant info.
          Quality of situation assessment?
• Quantity of information
   – Cues rarely have an Information Value of 1.0, so we
     need additional cues to help us make a decision.
   – However, there might be too much info for decision
     maker to attend to and remember.
   – After about two cues, ability to integrate addition info
   – Decision makers might seek more info than they can
     effectively utilize

   – Example
      • a financial web site might provide minutiae which
        cannot be read & comprehended.
          Quality of situation assessment?
• Relative Salience
   – Some cues might be bigger/brighter/louder than other
     info, regardless of relative value.
   – Salient cues tend to be weighted higher in decision

   – Example:
      • website might place info at top of page, above other
        info of equal value.
      • brochure might place some info in small type at
        bottom of page.
          Quality of situation assessment?
• Relative weighting given to cues
   – DM might fail to properly discount low value cues.
   – Often do not give more reliable cues enough
   – People tend to weight cues of equal salience equally.

   – Example:
      • Nurses might make diagnoses based on number of
        symptoms present, rather than on Diagnosticity of
                        Cue Conclusions
• Humans good at estimating: mean, non-extreme proportions
• Humans bad at estimating:
   –   extreme proportions
   –   extrapolating non-linear trends
   –   estimating variance
   –   estimating degree of correlation in scatter plots
• Cues have 3 characteristics
   – Diagnosticity
   – Reliability
   – Salience
• Cue Assessment (person side of things)
   –   Comprehensiveness of info
   –   Quantity of Info
   –   Relative Salience
   –   Relative Weighting
            Expertise and Automaticity
• Recognition-Primed DM
  – Rather than making a calculated decision, experts
    sometimes employ pattern matching to make rapid
  – That is, “I’ve encountered this in the past, and here is
    the solution I used”
  – Because detailed analysis is not performed, matching
    can sometimes occur in error and lead to incorrect or
    suboptimal choices.
          Heuristics in Decision Making
• Decision makers seem to behave as if guided by a
  number of Heuristics or “rules of thumb”
   – simplify decision making
   – don’t always produce a correct decision, but might be
     good enough most of the time
•   If a coin is flipped 6 times, which of the following
    outcomes is most likely?
    1. H T T H T H
    2. H H H T T T
    3. H H H H H T

     All are equally likely.
     Gamblers fallacy: assumes independent events
     are correlated.
                     Example 2
• A group of people contains 70 engineers and 30
  lawyers. A person is drawn at random. This person is a
  bit shy, and enjoys math and science. What is the
  likelihood that this person is a lawyer?

• A group of people contains 30 engineers and 70
  lawyers. A person is drawn at random. This person is a
  bit shy, and enjoys math and science. What is the
  likelihood that this person is a lawyer?
             Heuristics & Biases in DM
• Representativeness Heuristic
   – “If it walks like a duck, and quacks like a duck, it’s
     probably a duck.”
   – Assessments of situation based on similarity to mental
     representation of hypothesized situation.
   – Ignores base rate of events (e.g. what if ducks are

   – Example:
      • If a patient has 4 symptoms that match disease A and
        2 symptoms that match disease B, doctor might
        diagnose A even though B is much more common.
             Heuristics & Biases in DM
• Availability Heuristic
   – judged likelihood of event might be based on the ease
     with which event comes to mind.
   – incorporates base-rate info, because common events
     come to mind more easily
   – might be biased by irrelevant characteristics (salience,
     recency, or simplicity of diagnosis).

   – Example:
      • people tend to overestimate the frequency of deaths by
        an exciting cause (e.g. airplane, sniper) and
        underestimate frequency of death by mundane cause
        (e.g. heart disease).
              Heuristics & Biases in DM
• Anchoring Heuristic
   – after forming a belief, people are biased not to
     abandon it.
   – in other words, primacy of info increases its weighting
     in judgments

   – Example:
      • When surprised by reported earnings, analysts
        naturally anchor to their old earnings until they are
        convinced the earnings change is due to permanent
        rather than temporary factors.
               Heuristics & Biases in DM
• Anchoring Heuristic
   – When a long series of simple cues must be integrated,
     primacy effects are shown.
   – When cues are complex (detailed, unfamiliar), recency
     effects are more likely.

• Punchline:
      • Beliefs might depend on order in which info is
      • Where info should be equally weighted, should be
        presented simultaneously.
              Heuristics & Biases in DM
• Confirmation Bias
   – After forming a belief, people tend to seek evidence
     consistent with that belief, and discount inconsistent

   – Example:
      • A person that believes in the abilities of a psychic will
        tend to pay attention to the successes and ignore the
             Heuristics & Biases in DM
• Overconfidence bias
   – People tend to assume that their judgments are much
     more accurate than is true

   – Example:
      • of all instances when a given stock picker estimates a
        a stock is 90% likely to climb, it might actually climb
        only 60% of the time.
           Heuristics & Biases Conclusions

• Experts might rely on automaticity (time-stressed situations)

• Heuristics & Biases
   –   Representativeness
   –   Availability
   –   Anchoring
   –   Confirmation Bias
   –   Overconfidence Bias
                     Choice of Action
•   Certain Choice
    – Results of action are known with certainty
    – A DM can optimize choice:
      1. List attributes of potential choices, decide how
         important each attribute is
      2. For each potential choice, multiply value of
         attributes by importance of attribute.
      3. 3 Sum products, and choose option which
         maximizes this sum.
                   optimize choice
                                BMW 325i     Yamaha
                     w eight     Wagon       YZF-R1      Miata
Room                    0.25          100          10            50
Gas mileage               0.2          30         100            50
insurance                 0.1          50          10            60
Sexiness                  0.2          60          80            90
crashw orthiness        0.25          100          10            70

                                BMW 325i     Yamaha
                                 Wagon       YZF-R1      Miata
Room                                   25          2.5        12.5
Gas mileage                              6          20          10
insurance                                5           1           6
Sexiness                               12           16          18
crashw orthiness                       25          2.5        17.5
                    total              73           42          64
                 Choice of Action
Or more likely…
   • Satisfice, or pick a choice which might not be optimal,
     but is good enough.
       – BMW dealer is 100 miles away, Mazda dealer is
         down the street…
• or
   • choose by heuristic elimination by aspects: choose a
     single attribute, discard any choices which do not
     meet criterion for that attribute.
   • Example:
       • I only have $20k to spend on a vehicle. Therefore,
         the BMW is out of my price range and the Miata is
                  Choice of Action
• Uncertain Choice
  – Consequences of choice are not certain
  – That is, the consequences are probabilistic
  – Maximize expected utility of outcome
     Utility = subjective value
  – Utility is not always the same as objective value
           Distortions of Value and Cost
1. People underestimate gains in

                                    Loss               Gain
2. Potential losses are perceived
   as having greater subjective
                   Uncertain Choice
• Biases in setting Utility
   – people are loss aversive, prefer to avoid risk of loss
     rather than gamble on a gain.
   – Example:
      • Imagine that you're a contestant on a TV game show.
        You have just won $10,000. The host offers you a
        choice: You can quit now and keep the $10,000, or
        you can play again. If you play again, there is a 0.5
        probability that you will win again, and wind up with
        $20,000. If you play again and lose, you lose your
        $10,000 and take home nothing. You quickly calculate
        that the expected value of playing again is $10,000,
        the same as sticking with the $10,000 you have won
        so far. Which do you chose?
            Distortions of Value and Cost
3. There are progressively
   smaller gains in utility as
   value increases.                        Utility
   A gift of $100 seems like a lot
      if you are a poor graduate
      student, but is a pittance of
      you are a movie star.         Loss               Gain

              Perceptions of Probability
1. Probability of rare events
   is overestimated
   e.g. probability of needing

                                   Subjective Probability
      to file insurance claim is

                                                            0                          1.0
                                                                  Stated Probability
             Perceptions of Probability
• Example: Insurance
  – The estimated value of purchasing insurance is
    negative: on average, you will lose money (other wise
    the insurer would not make money).
• Example:
  – the stock market might appear more risking than a
    savings account. However, with interest rate below
    the rate of inflation, a savings account guarantees
             Perceptions of Probability
2. Reduced Sensitivity to
   probability changes at low

                                Subjective Probability
   e.g. sluggish beta,
      heuristic, & ignorance
      of base rates

                                                         0                          1.0
                                                               Stated Probability
                Perceptions of Probability
3. Perceived probability is
   less than real probability
   e.g. the probability of a risky

                                     Subjective Probability
      but high-payoff outcome
      might be
      underestimated, and a
      surefire (but low-paying)
      outcome might be
      chosen instead.

                                                              0                          1.0
                                                                    Stated Probability
                   Uncertain Choice
• Biases in setting Utility
   – Perception of outcome as a gain or loss is determined
     by perceived neutral point. This is the framing effect.
   – Example:
      • people see a treatment with 90% survival rate as being
        preferable to a treatment with 10% mortality rate.
             Biases in Uncertain Choice
• Direct Retrieval/automaticity
   – Experts might automatically retrieve a solution when
     the cues fit a given pattern, rather than analyzing the
   – Error example:
       • Policeman fires his gun when a person reaches into his
                     Choice Conclusions
• Certain Choice
   –   outcomes are known
   –   Optimize choice
   –   Satisfice
   –   Elimination by aspect
• Uncertain Choice
   –   outcomes are probabilistic
   –   maximize expected utility
   –   gain is underestimated
   –   loss is overestimated
   –   rare events are overestimated
   –   if not rare, overall probability is underestimated
• Direct Retrieval/automaticity
           Improving Decision Making
• Practice in DM doesn‟t necessarily make perfect.
• Expertise in a DM task doesn‟t make one immune from
  the effects of biases and heuristics.
            Improving Decision Making
In many domains, feedback is limited in several ways:
   1. Feedback is ambiguous:
      • poor choices sometimes reinforced, good choices
         sometimes punished.
      • Example: lottery sometimes pays off
            Improving Decision Making
In many domains, feedback is limited in several ways:
   1. Feedback is ambiguous

   2. Feedback is often delayed:
      Example: effects of smoking might not be felt for
            Improving Decision Making
In many domains, feedback is limited in several ways:
   1. Feedback is ambiguous:
   2. Feedback is often delayed:

   3. Feedback is provided selectively:
      • some choices receive feedback and others don‟t.
      •   Example: no feedback emerges from a decision to
          not hire somebody.
• Expert Systems
   – Statistical systems, that when fed clues (e.g.
     symptoms), will recommend a choice of action

• Problem: Quality of decision making is only as good as
  the quality of the model
   – Long Term Capital Management (LCTM)
      • Hedge fund company set up by 2 winners of Nobel
        Prize in Economics.
      • Decision model did not understand how to deal with
        Russian default on bonds (1998).
      • Result: $100 billion in debt.

To top