Anchoring and overconfidence in probability elicitation.
Engineering Honours Degree 2008
University of Adelaide
The anchoring and adjustment heuristic is described by Tversky and Kahneman (1974) as the tendency of people to base estimates on any number they have just seen, resulting in an insufficient adjustment away from such an ‘anchor’. It has been suggested that this particular heuristic may induce bias when constructing estimates of uncertain quantities, and may also lead to overconfidence. Research has investigated whether such anchors, both external and internal (or implicit), may be hindering the construction of confidence intervals by technical experts operating in industry. The currently accepted advice is to directly provide the boundaries to a confidence interval, so as to avoid implicit anchor effects imposed by the specification of a ‘best-guess’ value for the quantity.
The present study investigates the effects of both internal and external anchors on estimation through an elicitation exercise presented to 56 petroleum industry experts. This exercise took the form of an online survey, in which subjects were required to provide confidence interval estimates for actual field quantities such as porosity and recovery factor. Real-world field data taken from open-access sources was provided in scenario-based instances, by which subjects must draw upon their knowledge of partial dependencies between relevant parameters to construct their confidence intervals.
Results indicate that requiring subjects to provide a ‘best-guess’ value prior to specifying a range did increase overconfidence, as Tversky and Kahneman (1974) predicted. However, the predicted detrimental effects of external anchors on estimation success did not eventuate in the expected fashion. The provision of relatively high external anchor values did reduce estimation success, but low anchors statistically increased a subject’s capacity to estimate the true value accurately. Furthermore, industry experience did not significantly improve one’s success rate, and it is shown that experienced personnel seem just as susceptible to cognitive bias in this vein as the layperson.