Devitt, S.K., Pearce, T.R., Perez, T. & Bruza, P. (2016). Mitigating
against cognitive bias when eliciting expert intuitions. International
Conference on Thinking. Brown University, Providence RI, 4-6 Aug.
Abstract
Experts are increasingly being called upon to build decision support systems. Expert
intuitions and reflective judgments are subject to similar range of cognitive biases as
ordinary folks, with additional levels of overconfidence bias in their judgments. A
formal process of hypothesis elicitation is one way to mitigate against some of the
impact of systematic biases such as anchoring bias and overconfidence bias.
Normative frameworks for hypothesis or ‘novel option’ elicitation are available across
multiple disciplines. All frameworks acknowledge the importance and difficulty of
generating hypotheses that are a) sufficiently numerous b) lateral and c) relevant and
d) plausible. This paper explores whether systematic hypothesis generation can
generate the desired degree of creative, ‘out-of-the-box’ style options given that
abductive reasoning is one of the least tractable styles of thinking that appears to shirk
systematization. I argue that while there is no universal systematic hypothesis
generation procedure, experts can be exposed to deliberate and systematic
information ecosystems to reduce the prevalence of certain types of cognitive biases
and improve decision support systems.
Keywords: Abduction, cognitive bias, option generation
Mitigating against cognitive bias when eliciting expert intuitions
International Conference on Thinking , Brown University, 2-4 Aug 2016
1 Introduction
Decision-making research has thoroughly investigated how people choose from a set
of externally provided options (Tobler et al., 2013). However, ill-structured real-world
environments do not reveal all relevant options; instead options must be generated by
the agent, usually by abductive reasoning. Can we design a system for effective option
generation? We aim to build a back-end relational database with front-end social
interface that facilitates effective and novel option generation amongst teams of
experts (see visual inspiration Figure 1). The database provides a collective memory of
options and the contexts under which they were developed. The database can be
queried, e.g., to assist in the generation of options for the problem at hand. Our
database motivates higher order, strategic queries—see metadata structure Figure 2.
The social interface allows experts to build capacity in abductive reasoning. Our
research explores whether systematic social hypothesis generation can mitigate against
a subset of cognitive and motivational biases
Background
We draw on the sub-logics GEN, ENGAGE and DIS (Gabbay & Woods, 2005a,
2005b) to solve “The Cut Down” problem of abduction—see Table 1.
Sub-logic Description
GEN
Generation of hypotheses, resulting in a space of hypotheses Η.
Members of Η are each a “possible hypothesis for possible conjecture”
ENGAGE engages elements of Η relevant to the abduction problem into a proper
subset Ρ⊂Η, the set of relevant hypotheses for possible conjecture. In
turn, the plausibility filter contracts Ρ into a set of possibilities for actual
conjecture, represented by Δ.
DIS
transforms the plausible hypotheses into a premise (or premises) by
subjecting it (or them) to a test filter of which Woods and Gabbay
identify two varieties: ``The filter of independent confirmation" and
``The filter of theoretical fruitfulness".
Table 1. Description of Gabbay & Woods (2005a, 2005b) three sub-logics GEN, ENGAGE and DIS to
solve ‘The Cut Down’ problem of abduction.
Mitigating against cognitive bias when eliciting expert intuitions
International Conference on Thinking , Brown University, 2-4 Aug 2016
2 Figure 1: Visualization of Hypotheses in the user interface:
Combination of the elemental attraction of fire, the conceptual
spark of an idea and ethereal wild horse.
Hypotheses
are linked
to evidence
Figure 2: Diagram of underlying database logic. Evidence metadata includes both the
justification evidence relative to a given hypothesis (binary voting, subjective
description and drop-down argument type) and a quality assessment (subjective and
automatically scraped from documents).
Mitigating against cognitive bias when eliciting expert intuitions
International Conference on Thinking , Brown University, 2-4 Aug 2016
3 Case Study
We are running a case study by exposing experts working for an Online Travel
Agency (OTA) to a deliberate and systematic information ecosystem to reduce the
prevalence of certain types of cognitive and motivational biases to improve decisions.
The OTA’s philosophy endorses the scientific method, particularly hypothesis
generation and evaluation with data and experiments. Option generation is at risk at
this agency because:
1. hypotheses are not well grounded in evidence.
2. not enough hypotheses are generated
3. hypotheses iterate narrowly on previous hypotheses, leading to
canalisation (see Figure 3)
Figure 3: Canalisation in biology is a measure of the tendency of a population to produce the
same phenotype regardless of variability of its environment or genotype. In organisations, employee
beliefs can canalise, making them immune to evidence.
We aim to reduce cognitive and motivational biases1 by using the Scientific Social
System to:
•
•
•
•
•
•
•
•
•
•
Provide multiple and counter anchors
Prompt employees to consider reasons in conflict with anchors
Build explicit probability competence
Provide counterexamples and statistics
Capitalise on multiple experts with different points of view about hypotheses
Challenge probability assessments with counterfactuals
Probe evidence for alternative hypotheses
Encourage decision makers to think about more objectives, new alternatives
and other possible states of the future
Prompt for alternatives including extreme or unusual scenarios
Functions of the user interface shape how experts generate, evaluate and justify
decisions.
Anchoring bias, Availability bias, Confirmation bias, Myopic problem
representation, Omission of important variables & Overconfidence bias
1
Mitigating against cognitive bias when eliciting expert intuitions
International Conference on Thinking , Brown University, 2-4 Aug 2016
4 Discussion
Though the generation and evaluation of hypotheses is the basis of Bayesian
rationality (Devitt, 2013; Gwin, 2011; Hajek & Hartmann, 2009; Henderson, 2013;
Oaksford & Chater, 2009) we recognize the limits of Bayesian rationality as a
satisfactory epistemology of reasoning and choice. Thus, we are designing an interface
and database aligned with normative Bayesian principles and subjective explanatory
justifications stemming from social interaction (see Douven & Wenmackers, 2015).
In addition to better option generation, there are significant benefits to decision
makers of having a robust and dynamic set of evaluated hypotheses across teams and
work hierarchies to amplify collective intelligence (Nielsen, 2012).
The work is especially relevant in terms of how decision makers can utilize employee
subjective probabilities (potentially elicited from our interface and stored in our
database) to evaluate the likelihood of strategic hypotheses being true with
uncertainties explicit—see Figure 3
Figure 3 Ascertaining the truth or falsity of a set of hypotheses given a set of outcomes characterised by
binary data. In this example (N=20; R=15 positive answers). The top plot shows the outcomes. The
bottom plot shows the posterior distribution for θ (Perez, 2016).
In addition to better option generation, there are significant benefits to decision
makers of having a robust and dynamic set of evaluated hypotheses across teams and
work hierarchies to amplify collective intelligence (Nielsen, 2012)
Contact
Dr Kate Devitt
Website: http://skdevitt.com
Email: [email protected]
Mitigating against cognitive bias when eliciting expert intuitions
International Conference on Thinking , Brown University, 2-4 Aug 2016
5 References
Devitt, S. K. (2013). Homeostatic epistemology: Reliability, coherence and coordination in a
Bayesian virtue epistemology. (Ph.D.), Rutgers The State University of New Jersey New Brunswick. Retrieved from http://eprints.qut.edu.au/62553/ QUT
ePrints database.
Douven, I., & Wenmackers, S. (2015). Inference to the Best Explanation versus
Bayes’s Rule in a Social Setting. The British Journal for the Philosophy of Science.
doi:10.1093/bjps/axv025
Gabbay, D., & Woods, J. (2005a). Filtration structures and the cut down problem in
abduction. In J. Woods, K. A. Peacock, & D. Irvine (Eds.), Mistakes of Reason:
Essays in Honour of John Woods (pp. 398-417): University of Toronto Press.
Gabbay, D., & Woods, J. (2005b). The reach of abduction: insight and trial (A practical logic of
cognitive systems, vol 2). London: Elsevier.
Gwin, M. F. (2011). The virtues of Bayesian epistemology. (PhD), University of Oklahoma,
ProQuest, UMI Dissertations Publishing. (3488133)
Hajek, A., & Hartmann, S. (2009). Bayesian epistemology. In J. Dancy, E. Sosa, & M.
Steup (Eds.), A Companion to Epistemology (pp. 93-105). Chicester: John Wiley &
Sons, Ltd.
Henderson, L. (2013). Bayesianism and Inference to the Best Explanation. The British
Journal for the Philosophy of Science. doi:10.1093/bjps/axt020
Nielsen, M. (2012). Reinventing discovery: the new era of networked science: Princeton
University Press.
Oaksford, M., & Chater, N. (2009). Précis of bayesian rationality: The probabilistic
approach to human reasoning. The Behavioral and brain sciences, 32(1), 69-84.
doi:10.1017/s0140525x09000284
Perez, T. (2016). Testing Hypotheses with Binary Data - A Bayesian Approach. Technical
Note. Electrical Engineering and Computer Science. QUT. Brisbane,
Australia.
Tobler, P. N., Simon, J., Schweizer, T. S., Kaiser, S., Kalis, A., & Mojzisch, A. (2013).
The cognitive and neural basis of option generation and subsequent choice.
Cognitive, Affective and Behavioral Neuroscience, 13(4), 814-829. doi:10.3758/s13415013-0175-5
Mitigating against cognitive bias when eliciting expert intuitions
International Conference on Thinking , Brown University, 2-4 Aug 2016
6