Introduction
Formal models of judgement and decision-making hold that judgements of probability and utility should be assessed using all of the information available to the decision-maker, with the application of appropriate statistical rules.122 However, humans are not perfect information processors. The amount of information processed can be affected by time pressure, limitations in cognitive capacity, lack of motivation and personal desire for a particular outcome. When it comes to probabilistic reasoning, specifically the failure to recognise when a statistical rule should be applied and unfamiliarity with the processes for making statistical inferences, probability judgements do not always conform to normative rules.123 Experts, being human, are not immune to this. Indeed, even among highly educated populations, awareness of how to make simple statistical inferences can be limited.124 In the context of HCDM, those practitioners with the greatest relevant knowledge and expertise (e.g. nurses, physiotherapists) may not necessarily have a high level of training in statistics or experience with elicitation.
Humans often make judgements using simple rules of thumb (or ‘heuristics’).123,125 These strategies are usually effective in appropriately guiding judgement,126 especially among experts who have a large base of experience and knowledge to draw on.127 However, in some contexts they can lead to systematic errors known as ‘biases’. SEE should seek to elicit probability judgements in a way that minimises the effect of these systematic errors. This is increasingly recognised in the literature on HCDM, in which SEE can be used to inform health policy and treatment recommendations.12,44,60,85,128 However, although heuristics, biases and strategies for bias reduction have been widely studied in the broader risk, judgement and decision-making literature, there is a dearth of evidence for HCDM and what does exist has not been summarised in this context.
This chapter reviewed evidence relating to the psychological biases of greatest relevance to SEE for HCDM, specifically evidence on how these can be minimised. First, key cognitive and motivational biases that have the potential to negatively impact on the quality of expert elicitation for HCDM are outlined (see Cognitive and motivational biases), then potential strategies for addressing them (see Addressing psychological biases in structured expert elicitation) through technical measures (see Technical bias reduction strategies) and behavioural bias reduction techniques (see Behavioural bias reduction strategies with consistent support). Reflecting the fact that some behavioural bias reduction techniques have a large amount of evidence to support them whereas others are more tentative, techniques are categorised into those for which a high degree of consensus exists and those for which evidence is lacking or conflicted. Finally, the key recommendations are summarised in Conclusions.
Cognitive and motivational biases
A distinction may be drawn between cognitive biases that result from how information is processed, and motivational biases that come about as a result of preferences for particular outcomes.77,129 Both have been implicated in systematic overconfidence, which poses a threat to calibration in SEE.
Cognitive biases
Cognitive biases arise when decision-makers do not process the full range of information available to them. This may result from limitations in cognitive capacity, time pressure or a lack of motivation to expend cognitive effort on a task. They may also arise as a result of decision-makers lacking the normative skill to make appropriate probabilistic inferences. In the context of SEE, cognitive biases of particular importance include availability and anchoring, and insufficient adjustment, first, because they are both implicated in overconfidence, which leads to the systematic underestimation of uncertainty in probability judgements, and, second, because unlike biases that may result from deficits in substantive knowledge of a subject area, or from a lack of knowledge about how to reason with statistical information, both have the potential to affect expert judgement.77,130
In making probabilistic judgements, people may rely on how easily examples of an outcome come to mind as a guide to how likely it is (the availability heuristic).131 Although this is often a good guide to frequency, it means that probability judgements can easily be distorted by very recent or very prominent events.132 For instance, a clinician may focus on particularly memorable examples of treatment success or treatment failure when making probability judgements, neglecting instances that come less readily to mind. Availability bias has been linked to the systematic underestimation of uncertainty.133 Anchoring and insufficient adjustment occurs when people fix (‘anchor’) on an initial value and fail to sufficiently adjust their estimates away from it to provide an accurate judgement. For example, in judging the success of an intervention, a clinician may ‘anchor’ on a value provided by a source that they know to be flawed (e.g. a poor-quality empirical study) and fail to sufficiently adjust their own experienced-based estimate from this point, despite being aware of the flaws and adjusting in the right direction.125 Anchoring has proved challenging to de-bias, with even arbitrary and irrelevant values being found to affect judgement (see Kahneman and Egan123 for an overview). This can decrease accuracy in judgements of location and central tendency (e.g. mean, median).
Motivational biases
Motivational biases, sometimes referred to as ‘self-serving’ biases, result from being invested in a specific outcome (e.g. a particular treatment being successful) (see Bazerman and Moore129 for discussion). In situations where individuals are aware of potential conflicts of interest and strive to make objective and honest judgements, motivational biases can still distort judgements through rendering some information and experiences more salient (cognitively ‘available’) and easier to recall than others. Confirmation bias, for instance, leads individuals to focus on information that is consistent with their existing beliefs and preferences and, therefore, subject it to a less critical appraisal than inconsistent information. Desirability bias (also referred to as ‘optimistic bias’ or ‘wishful thinking’) leads people to overestimate the likelihood of positive outcomes. Undesirability bias, meanwhile, leads to an overestimation of the likelihood of negative outcomes and worst case scenarios (e.g. owing to a focus on taking a precautionary approach). These biases result from motivated reasoning rather than a lack of knowledge or experts.77,129 Hence, they have the potential to adversely affect the outcomes of SEE. In HCDM, those with greatest knowledge of a particular treatment or procedure may be those most invested.
Overconfidence bias
As a consequence of limiting the amount of information considered by decision-makers, both availability133 and confirmation bias134 may lead to the uncertainty surrounding future outcomes being underestimated. This is known as ‘overconfidence bias’. It leads to interval judgements and probability distributions that are too narrow (e.g. estimates of 80% confidence intervals containing < 50% of subsequent realisations). Overconfidence is prevalent among experts as well as novices,35,135 making it an important consideration for any form of SEE.
Addressing psychological biases in structured expert elicitation
Strategies for reducing psychological biases could be said to fall into three categories: (1) technical (e.g. using formal statistical procedures to correct for systematic errors in judgement); (2) directly changing individual behaviour and perceptions (e.g. through training, incentives, feedback); and (3) changing the structure of the judgement or decision task (e.g. how questions are asked).136,137 In practice, however, they represent two fundamental approaches: (1) post-hoc statistical techniques to make corrections after the fact, most notably through calibration (discussed in Chapter 5) (technical); and (2) interventions to change judgement and behaviour (behavioural).
In reviewing approaches for reducing psychological bias (or ‘debiasing’), we restricted our search to studies that provide empirical evidence for the efficacy of bias reduction in the context of SEE. For this reason, we have excluded papers that suggest approaches but do not present empirical evidence to support them. We also exclude studies that focus on biases in decision from description (i.e. when choices can be made through analysis of a complete information set), rather than elicited judgements. Relevant papers that did not appear in the searches but that were cited in the papers identified, were examined and included when appropriate. A potential weakness of this approach is that bias reduction techniques that are relevant to SEE, but that do not mention expert elicitation directly, may have been missed if they were not cited in other papers identified through the search. However, a full review of the heuristics and biases literature, which often focuses on novice rather than expert judgement, is beyond the scope of this targeted search.
Technical bias reduction strategies
Technical bias reduction strategies are commonly discussed with respect to overconfidence. These can involve statistical bias correction and the weighting of experts based on their performance on seed questions, as is the case in Cooke’s classic model.138–140 These approaches do not require interventions at the individual or task level, as the procedures are applied post hoc. However, they do rely on the availability of appropriate seed questions from which the level of experts propensity to overconfidence can measured.141 This may be relatively easy in contexts in which past realisations of the same or similar target variables are available (e.g. probabilistic weather forecasting). In HCDM, however, it could prove challenging to implement, as contextually similar seed variables with appropriate realisations are not always readily available. Likewise, HTA brings together diverse sets of experts who have specialist knowledge of specific treatments, interventions or procedures. They are not, therefore, guaranteed to have similar expertise on the subject of seed questions.53
Behavioural bias-reduction strategies with consistent support
Given the challenges in applying technical approaches to bias reduction, which are outlined above, it is important for those implementing SEE in the context of HCDM to consider behavioural approaches. In this section, we outline bias reduction strategies for which there is consistent empirical support. In Behavioural bias reduction techniques with conflicting evidence we briefly discuss debiasing approaches for which there is conflicting evidence.
Consider more information
It has been found that individuals with a greater predisposition towards open-minded thinking demonstrate better calibration on judgement tasks.142 Increasing the amount of information considered by participants may therefore be effective in countering these biases. Behavioural bias reduction techniques that prompt experts to consider more information (increasing the range of possibilities considered) have perhaps been the most frequently tested in the context of expert judgement.
Early research with student samples failed to find added value from instructing groups of participants to consider why their estimates may be wrong, or appointing one member to be a ‘devil’s advocate’.143 However, more structured approaches have had far greater success.134,144,145 Soll and Klayman134 found that asking student participants to separately give lowest plausible estimates, highest plausible estimates and median estimates for an almanac question with which students were likely to have some familiarity led to lower levels of overconfidence than simply asking for a single 80% confidence interval. It was suggested that making people consider lowest, highest and median estimates sequentially focuses attention on a wider range of possibilities than asking for a single range [e.g. forcing participants to think of reasons why a value might be below (or above) a specific value]. Building on this, Haran et al.144 found that further increasing the number of considerations by asking participants to make judgements about the likelihood of different local seasonal temperature intervals reduced overconfidence. Adding a fourth step to the procedure suggested by Soll and Klayman134 and Speirs-Bridge et al.145 found that ranges were widened further when participants (epidemiologists and ecologists) were asked how likely it was that the ‘true’ value would fall within their specified range and were allowed to revise their estimates accordingly. This is consistent with research suggesting that people may be better at evaluating confidence intervals than providing them.146,147 More recently, Ferreti et al.148 noted reductions in overconfidence when environmental science students were instructed to (1) actively think of reasons why their initial highest and lowest estimates of sea level rise may be incorrect; and (2) consider their willingness to place hypothetical bets on elicited confidence intervals.
Together, these studies provide strong evidence that structuring tasks in a way that increases consideration of a wider range of possibilities can reduce bias and improve calibration. They demonstrate that confidence intervals should not be elicited as a single-stage process. Lower and upper bounds should be elicited individually,134,145 or multiple smaller intervals should be considered individually.144 Likewise, they show that participants should be given the opportunity to evaluate and adjust their confidence intervals.
Feedback
There is extensive evidence that receiving repeated feedback on one’s judgements both improves accuracy and reduces overconfidence.35,123,141 Experts, such as weather forecasters, who receive direct and timely feedback on the accuracy of their judgements tend to be well calibrated in their domain of expertise,149 although this does not result in a domain general improvement.137 One suggestion for reducing the overconfidence bias in expert elicitation is to provide feedback on a set of practice questions.150 A challenge in doing this is the fact that domain-specific seed variables may be more readily available in some contexts than in others (e.g. past realisations in forecasting tasks). Hence, although this approach may be broadly effective in improving the calibration of expert judgement, it could be difficult to implement in some HTA contexts in which identifying appropriate seed questions that a diverse set of experts will be familiar with could be challenging. Nonetheless, in cases in which these are available, the existing evidence suggests that providing feedback seed questions can reduce overconfidence.
Avoid unnecessary anchors
Ensuring that elicitation materials do not contain unnecessary anchor values is a ‘common sense’ approach to reducing biases caused by anchoring and insufficient adjustment.77 For instance, elicitation tools should not feature pre-set values that participants are then asked to adjust to match their views. However, it may not always be possible to eliminate anchors entirely. In the case of ‘carryover’ effects, for example, experts may use their own judgement on a previous question as an anchor.151 Although there is some evidence to suggest that self-generated median anchors do not threaten accuracy and calibration to the same extent as those that are externally imposed,134,152 Morgan35 advises that measures of central tendency (i.e. the median) should only be elicited after lower and upper bounds have been estimated. Hence, although it may not be possible to eliminate all potential anchor values in an elicitation task, a clear recommendation to avoid unnecessary anchors can be made. Likewise, when eliciting confidence intervals, eliciting lower and upper bounds before the median may reduce the tendency to anchor on the median value.
Reduce bias through expert selection
Addressing biases through expert selection means that experts are included or excluded based on their potential susceptibility to bias (see Chapter 5). As noted above, motivational biases, such as desirability bias and confirmation bias, are difficult to eliminate. Restricting participation to those without any conflicts of interest is therefore one recommended approach to reducing motivational77 biases. In HCDM this may be challenging, as those with the greatest knowledge about a particular treatment or technology may also be those with the greatest vested interest in the elicitation’s outcome.44 Rejecting those with any conflict of interest or strong opinions may eliminate those with the greatest relevant knowledge. In such cases an alternative strategy is to ensure that a range of viewpoints are represented in the sample, with the intention of ‘balancing out’ or at least diluting the effect of motivational biases.77
Behavioural bias reduction techniques with conflicting evidence
Bias warnings and training
Within the HCDM literature that considers heuristics and biases, training is the most commonly referenced approach to behavioural debiasing.85 Simply warning experts not to be biased (e.g. by stating that many people make their confidence intervals too narrow) is largely ineffective.143,152,153 However, in-depth training on the nature of biases and strategies for avoiding them has been found to be more effective. When biases occur as a result of experts not being familiar with rules for using and expressing probabilities, training on how to do so can reduce errors.154 Likewise, educating participants about biases and explicitly outlining strategies for combating them (i.e. through systematically considering more information) reduced overconfidence in a study of petroleum engineering students.155 However, this education programme was not effective for reducing anchoring, possibly because the student sample lacked the substantive knowledge of the field to give a more accurate value. Nonetheless, a study with a general population sample156 found evidence that interactive training interventions, explaining what anchoring and confirmation bias were, reduced instances of these biases on post-intervention tests relative to pre-intervention tests. These tests comprised tasks from the wider literature that were found to elicit the psychological biases.157–159 Hence, although the available evidence on the effectiveness of warnings and training for reducing psychological biases is not always consistent, it does provide an indication of the conditions under which bias avoidance training may be effective. First, it must go beyond simple warnings and admonitions not to be biased and explain the causes and consequences of biases. Second, it should provide instruction as to how to avoid bias (e.g. consider why upper and lower bounds may be incorrect). Third, it is useful only if participants have the substantive expertise to produce accurate responses.
Fixed value compared with fixed probability methods
A small number of studies have examined whether the fixed-value method (in which one must allocate probabilities to potential values of a target variable) or the fixed-probability method (in which one allocates values of the target variable to probabilities) affects overconfidence. In eliciting cumulative probability judgements from students regarding forecast variables with which they were expected to have some familiarity (i.e. local temperature and the Dow Jones), Abbas et al.160 found less evidence of overconfidence using the fixed-value method. However, Ferretti et al.148 found that this resulted in relatively little improvement in performance. Hence, although there is some evidence that fixed-value approaches may reduce overconfidence, this is limited.
Face-to-face compared with online elicitation
In one recent study161 it was found that face-to-face elicitation of energy demand with sectoral experts led to lower overconfidence than online elicitation. However, this finding was not replicated in a recent comparison of face-to-face and online SEE.60
Conclusions
The objective of this review has been to synthesise existing knowledge on the clinical effectiveness of different behavioural bias reduction techniques for expert elicitation, focusing specifically on their potential usefulness in the context of HCDM. Although the efficacy of some of these approaches remains undertested, the following five recommendations are supported based on the available evidence:
Confidence intervals should not be elicited as a single-stage process, as doing so leads participants to focus on a narrow set of salient possibilities. Instead, lower bounds, upper bounds and median values should be elicited separately. Eliciting lower and upper bounds before median values may also prevent participants from anchoring on median values.
Participants should be allowed to evaluate and revise their confidence intervals or probability distributions.
In selecting experts, those with pronounced conflicts of interest should be excluded. However, excluding all participants who may have strong feelings or vested interests in the outcome may result in the exclusion of those individuals with the greatest expertise in the subject. Hence, it is important to ensure that different viewpoints will be represented.
When suitable seed questions are available, these may be useful in providing practice feedback to participants on their performance and thus reduce overconfidence. However, care should be taken to ensure that all participants are familiar with the topic of these seed questions.
Bias training may reduce biases, but only if this goes beyond simple warnings, and explains what bias is and provides strategies for avoiding it.
