Title: The influence of hierarchy on probability judgment
1The influence of hierarchy on probability
judgment
- David A. Lagnado
- David R. Shanks
- University College London
2Level of hierarchy can modulate judgment
- Consider two statements about the next World Cup
- It is most likely that Brazil will win
- It is most likely that a European team will win
- These appear to support opposing predictions, but
both may be true - Shows the importance of the level at which
probabilistic information is represented
3Hierarchical structure
- Pervasive feature of how we represent the world
- Reflects pre-existing physical and social
hierarchies - Readily generated through conceptual combination
- Category hierarchies serve both to organize our
knowledge, and to structure our inferences
4Inference using a hierarchy
- One powerful feature of a category hierarchy is
that given information about categories at one
level, you can make inferences about categories
at another level. - This allows you to exclude alternatives, or
reduce the number you need to consider
5Probabilistic Inference using a hierarchy
- In many real-world situations we must base our
initial category judgments on imperfect cues,
degraded stimuli, or statistical data. - What effect do such probabilistic contexts have
on the hierarchical inferences that we are
licensed to make?
6Commitment heuristic
- Commitment heuristic - When people select the
most probable category at the superordinate
level, they assume that it contains the most
probable subordinate category. - This leads to the neglect of subordinates from
the less probable superordinate.
7How adaptive is this heuristic?
- The efficacy of such a heuristic depends on the
precise structure of the environment. - In certain environments it confers considerable
advantages - increases inferential power by focus on
appropriate subcategories - reduces computational demands by avoiding complex
Bayesian calculations. - But in some environments it can lead to anomalous
judgments and inferences.
8Non-aligned hierarchy
Tabloid 60
Broadsheet 40
Times 5
Guardian 35
Mirror 30
Sun 30
- In the above sample the most frequently read type
of paper is a Tabloid, but the most frequently
read paper is a Broadsheet (the Guardian). - Non-aligned hierarchy the most probable
superordinate category does not contain the most
probable subordinate category.
9Real world examples
- Word frequencies the superordinate BE- is more
frequent than BU-, but the subordinate BUT is
more frequent than any of the other subordinates
(BET, BEDetc.) - NHS statistics on survival rate for operations
for different areas sub-areas - You are more likely to survive a hip operation in
Surrey rather than Essex, but the best sub-area
for survival is Colchester (in Essex).
10Experiments 1 and 2
- Learning phase - participants exposed to a
non-aligned hierarchical environment in which
they learn to predict voting behavior from
newspaper readership. - 100 trials reading/voting profiles
11Screen during learning phase
Broadsheet
Tabloid
Chronicle
Herald
Reporter
Globe
? Liberal
? Progressive
12Screen during learning phase
Reading profile for J. K.
Broadsheet
Tabloid
Chronicle
Herald
Reporter
Globe
? Liberal
? Progressive
13Screen during learning phase
Reading profile for J. K.
Broadsheet
Tabloid
Chronicle
Herald
Reporter
Globe
? Liberal
Outcome feedback
? Progressive
14Structure of environment
Tabloid 60
Broadsheet 40
Times 5
Guardian 35
Mirror 30
Sun 30
Party A Party B
50
50
15Judgment phase
What is the probability that X votes for one
party rather than the other?
Baseline
X is selected at random
Which type of paper is X most likely to read?
Type
Which paper is X most likely to read?
Paper
16Results of Experiment 1
- Probability ratings for Party B rather than Party
A with judgments divided into those based on
aligned and non-aligned choices
17Experiment 2
- Replication of Experiment 1, with frequency as
well as probability response formats - Frequentist hypothesis that probability biases
reduced with frequency format
18Results of Experiment 2
- Mean ratings for Party B rather than Party A
collapsed across probability and frequency ratings
19Summary of Results
- Participants allow their initial probability
judgment about category membership (newspaper
readership) to shift their rating of the
probability of a related outcome (voting
preference), even though all judgments are made
on the basis of the same statistical data. - When their prior choices were non-aligned this
led to a switch in predictions about the outcome
category
20Conclusions
- These biases are explicable by the Commitment
heuristic - The priming question commits people to just one
inferential path, leading them to compute an
erroneous estimate for the final probability. - This is understandable given the complexity of
the normative Bayesian computation.
21Comparison of Bayesian and commitment heuristic
computations (just type level inference)
Type of paper? Type of paper?
0.4
0.6
0.6
Tabloid Broadsheet
Tabloid
0.23
0.1
0.9
0.77
0.77
Party A Party B
Party A
P(A) (0.6 . 0.77) (0.4 . 0.1) 0.46
0.04 0.5
P(A) 0.77
Bayesian computation
Simplified heuristic computation
22Conclusions
- Simplifying heuristic that assumes that
environment is aligned - Empowers inference when hierarchical structure is
aligned, otherwise can lead to error - Suggests tendency to reason as if a probable
conclusion is true
23Process level accounts
- Associative model
- People learn predictive relations between
category options (at both levels of hierarchy)
and outcome. At test responses to category
questions prime the appropriate associations and
lead to a biased rating of the outcome. - Frequency-based model
- People encode event frequencies in the learning
phase. At test responses to the category question
serves as the reference class for subsequent
conditional probability judgments about voting
preferences.
24Implications
- Importance of the level at which probabilistic
data is represented to (or by) a decision maker - E.g., using NHS statistics to decide on hospital
- How do people search through hierarchical
statistical data? - Peoples judgments can be manipulated by the
level at which statistical information is
represented - More generally, in multi-step inferences people
are susceptible to biased probability judgments