Decision Analysis Flashcards
Problem: Clarify Objectives
Which method?
Value trees
AHP hierarchies
Problem: Compare alternatives over several objectives
Which method?
SMART
SMARTER
Multi-attribute utility theory
AHP
Problem: Structure options and outcomes
Which method?
Decision trees
Influence diagrams
Problem: Assess and manage uncertainty
Which method?
Risk analysis Probability wheels Event trees Fault trees Log-odds scales Uncertainty management
Problem: Compare option under uncertainty
Which methods?
Expected values Maximin criterion Utility Decision trees Influence diagrams Stochastic dominance Mean-standard deviation approach AHP
Problem: Group decision making
Which method?
Delphi
Decision Conferencing
Mathematical aggregation
Problem: Negotiation
Which method?
Negotiation models
What are the different heuristics?
Recognition heuristic Minimalist strategy Take-the-last strategy Lexicographic strategy Semi-lexicographic strategy Elimination-by-aspects strategy Sequential strategy – Satisficing Reason-based choice
Which of the heuristics are used for simultaneous problems?
Recognition heuristic Minimalist strategy Take-the-last strategy Lexicographic strategy Semi-lexicographic strategy Elimination-by-aspects
Which of the heuristics are used for sequential problems?
Satisficing
What influences which strategy is chosen?
Time Effort Knowledge Importance of accuracy Minimize conflict
What is bounded rationality
The human mind limited and not capable of always finding the optimal solution. Therefore, we often use approximations and heuristics (decision strategies and rules of thumb) to make satisfactory, not optimal, solutions.
What is compensatory strategies?
The decision maker does allow an option’s poor performance on one attribute to be compensated by a good performance on other attributes.
What is non-compensatory strategies?
The decision maker does not allow an option’s poor performance on one attribute to be compensated by a good performance on other attributes
Give an example of non-compensatory decision.
This could also be if decision maker is vegan and thereby only looking for vegan options – the price, taste, brand (other attributes) of non-vegan options is irrelevant as it cannot compensate for the fact that the product is not vegan.
What is the main difference between compensatory and non-compensatory?
Compensatory allow the DM to evaluate trade-offs (e.g. am I willing to accept lower quality if price is lower), where non-compensatory looks at attributes in isolation (e.g. I am not willing to accept low quality no matter the price).
Non-compensatory (most heuristics) are thereby more simple and less complex.
Explain recognition heuristics
: A heuristic applied when there are two options to choose among. If A is known and B in unknown, then A is chosen simply because it is recognized and requires less effort than to research option B and evaluate.
Minimalist strategy
A strategy that builds upon recognition heuristic in cases where either none or both options are recognized. If none are, then the decision maker simply guesses on the best option. If both are, a single attribute is chosen at random to be the decisive attribute. If A is better than B on this single attribute, then A is chosen.
Take-the-last strategy:
A version of minimalist (and thereby recognition heuristic) where instead of choosing the decisive attribute at random the decision maker identifies that attribute that helped make the decision last time the same choice was made.
Lexicographic strategy
This strategy requires the decision maker to rank the attributes in order of most to least important. The decision maker then evaluates the options based on the most important attribute and chooses the best-performing option.
Semi-lexicographic strategy
The same as lexicographic but with a tolerance limit. This could be that two options are said to be tied on the attribute price if they are within 50 cents.
Elimination by aspects (EBA)
This heuristic is based upon narrowing down the field of options by eliminating options that do not meet certain criteria. The decision is made as an iterative process where the decision maker identifies the most important attribute first and then defines an acceptable range (e.g. price is most important – eliminate all options that cost more than 100$ and less than 50$). This then narrows down the number of options. The process is then repeated for the second-most important attribute and so on.
Decoy effects
A decoy effect is the result of asymmetric dominance between options. If you compare a sport-car A, and a station-car B, at the same price level with free insurance, then it might be difficult to choose. If you then see the station-car B at another car dealer at the same price but without the free insurance, your tendency to choose the station-car B over sports-car A is greater
satisficing
When making choices where all options are not available at the same time, we must make sequential choices (e.g. looking for a job or a house, where options come and go). Satisficing is then choosing the first option that meets your requirement (e.g. if you look for a 3-bedroom house in Aarhus for 2mil, you will take the first house fitting this description). These requirements are referred to as aspiration level.
Reason-based choice
A different way of explaining how people make decisions. When decision markers are choosing, they are aware that they need to construct reasons for their choices in order to resolve conflict and justify their decisions. Therefore, they deviate from rational decision making as they seek the choice that is easiest to reason / justify.
Implications of reason based choice
Decision makers become sensitive to framing
Decision makers change decision when alternatives are introduced
Explain the distinction between event and outcome
Outcome are everything that can occur given an event.
An event consists of different outcomes
E.g. at least one store closes (event). Outcomes are then:
- Store 1 can close
- Store 2 can close
- Both stores can close
What is mutually exclusive or disjoint events?
The occurrence of one of the events precludes the simultaneous occurrence of the other (e.g. If sales exceeds 10 units it cannot also be less than 10 units or if a product is found to be in working well it cannot also be defective).
When can the addtion rule be applied?
used for cases where A OR B occurs
When can the mutliplication rule be applied?
used for cases where A AND B occurs
When are two event independent?
Two events, A and B, are independent if the probability of event A occurring is unaffected by the occurrence or non-occurrence of event B.
What are the axioms of probability theory=
•Positiveness: The probability of events occuring must be non-negative
Certainty: If an event is certain, then the probability must be 1
Unions: If events A and B are mutually exclussive then: p(A or B)=P(A)+P(B)
What is the maximin criterion?
The maximin criterion is choosing the maximum of the minimum outcome, hence maximizing the worst-case scenario.
What are the downsides for maximin criterion?
The method is 100% risk adverse! It does not look at the probability of the worst-case occurring, it assumes that the worst-case will occur and choose the best outcome then. If we have 99,99% chance of winning 1000$ and 0,01% chance of losing 1$ compared to 100% of 0$ if doing nothing, then maximin would suggest doing nothing as 0$ is better than -1$.
What is EMV?
Execpted Monetary Value
Prob. * Value
Cons of EMV?
This method ignores the risk for the DM. The EMV method simply chooses the course of action that yields the highest expected monetary value, but it fails to ignore a potential worst-case outcome might influence the decision maker:
Critique of utility measures?
Hypothetical situations
Chaining
Allais paradox
Two methods for deriving single-attribute utility?
Probability equivalence
Certainty equivalence
Explain the rollback method
This method states that the decision tree should be analyzed from right to left - hence considering the later decisions first.
Pros of rollback
The rollback method can be used for EMV, single-attribute utility, multi-attribute utility, and NPV if cashflows are to be considered.
How can we use continuous probabilities in decision tree
This could be by translating continuous into discrete (e.g. market share groups high, medium, or low) instead.
Another method is to use extended Pearson-Tukey (EP-T) approximation
Explain the EP-T method
Another method is to use extended Pearson-Tukey (EP-T) approximation stating that the value which has 95% chance of being exceeded should be assigned a probability of 0,185, 50% chance is prob. 0,63 and 5% chance is prob. 0,185 (pp. 155).
What is the process of Monte Carlo simulations?
1 Find factors 2 Model factor relations 3 Sensitivity Analysis to screen factors 4 Prob distr. factors 5 Simulation 6 Sensitivity Analysis (not necessary) 7 Compare
What methods can be used to compare findings in simulations?
Simple plot
Stochastic dominance
Mean-std. dev. method
What are cons of simple plot when looking at courses of action?
It only focuses on money and not risk
When is something first-order dominant?
If the cumulative probability distribution curve is constantly to the right
How should first-order dominance be interpreted?
The curve is always to the right, hence the prob. of earning more at any level of payoff is greatest for the dominant
What are assumptions of mean-std. dev approach?
Payoff to be approx. normal distributed
There should be risk averision in utility function
Explain Bayes’ Theorem
Explaination
What determines the effect of new information?
The vagueness of priors.
If all priors are equal, then the new information dictates the posterior
What is availability heuristic?
Events that are easier to recall are (more available) are overestimated. People judge the probability of an event occurring by how easily events alike are recalled
This can be beneficial when there is a correspondence between recollection and probability (e.g. cross the road and not being hit by a car is easy to remember and is unlikely).
Which biases arise from availability heuristic?
No link between ease of recall and associated probability
Ease of imagination is not related to probability
llusory correlation
Which biases arise from Representativeness heuristic?
Ignoring base-rate frequencies
Expecting sequences of events to be random
Expecting chance to be self-correcting
Ignoring regression to the mean
Which biases arise from Anchoring and adjustment heuristic
Insufficient adjustments
Overestimating the probability of conjunctive events
Underestimating probabilities for disjunctive events
Overconfidence
How can individual probabilities be derived?
Direct assessment
Probability wheel
How can probability distributions be derived?
The probability method
method of relative heights
How can validity of probability assessments be checked?
calibration
Brier Score (BS)
Which method is best for testing validity of probability assessments?
The issue with calibration is that a forecaster can simply stick to always predicting the mean, which means that in the very long run, the calibration would be perfect (as everything tends to the mean). This corresponds to the weather forecaster predicting a 28% chance of rain tomorrow every day. In the long run, there will be an average of 28%, hence perfect calibration even though he is wrong most days.
Using the BS method, a forecast is rewarded for making prediction that are honest. If a weather forecaster predicts 28% chance of rain and there is rain, then the BS score is 1-(1-0,28)^2=0,48. However, if the weather forecast really believed it to rain and predicts 70% chance, the score would be 0,91.
Therefore, the BS rewards predictions that are certain if the event predicts occurs.
How can one assess probability of rare events?
Event Trees
Fault Trees
Log-odds scales
How can one identify sources of uncertainty?
Exploratory trees and break-downs
How can one assess value of risk management
Calculate the effect of perferct control in risky elements and use Tornado Diagrams
What are the two main ways of aggregating individual jdugments to group decisions
Mathematical: Getting individual judgments and then aggregating
Behavioral: Getting inviduals to reach decision together
What are problems with weighted averages in aggregation?
How to assign weight?
Can we really trust that more knowledge = better estimate?
What is Condorcet’s paradox?
We then compare A to B and see that A > B two times vs. B > A one time = group prefers A over B
We then compare A to C and see that A > C one time vs. C > A two times = group prefers C over A
We then compare B to C and see that B > C two times vs. C > B one time = group prefers B over C
Combining this, we get A > B > C > A, which cannot be true. The group prefers A over B over C over A, hence A is both better than and worse than C. This is the paradox of sequential comparison called Condorcet’s paradox.
What is the issue with common strength preference scales?
We use them to try an convert individual preferences into a common scale to overcome the issue that my A > B might not be as strong as your B > A.
However, it is very difficult to determine the strength of perferrences
What are the methods to facilitate good behavioral aggregation
Delphi
Prediction Markets
Decision conferencing
What can happen is behavioral aggregation is not managed?
- Powerful individual in the group might inhibit less powerful individual to contribute
- Introvert and less talkative people might fell dominated and thereby not contribute or be ignored
- Seating arrangements have found to be influential on individuals’ contribution
- There is a risk of groupthink – ideas that are critical of the current direction in which the group is moving is suppressed. People agree with the chosen path and cannot criticize it.
a. Leads to incomplete investigation of other courses of action
What is Delphi method?
People give estimates
Feedback on group opinion is presented
Repeat until stable estimates
