The Influence of Reasons on Interpretations of Probability Forecasts

When providing a probability estimate for an event, experts often supply reasons that they expect will clarify and support that estimate. We investigated the possible unintended influence that these reasons might have on a listener’s intuitive interpretation of the event’s likelihood. Experiments 1 and 2 demonstrated that people who read positive reasons for a doctor’s probability estimate regarding a hypothetical surgery were more optimistic than those who read negative reasons for the identical estimate. Experiment 3 tested whether a doctor’s failure forecast for a surgery would result in differing levels of pessimism when the potential risk was attributed to one complication that had a probability of 0.30 versus three complications that had a disjunctive probability of 0.30. Overall, the findings are consistent with the argument that a probability estimate, albeit numerically precise, can be flexibly interpreted at an intuitive level depending on the reasons that the forecaster provides as the basis for the estimate. Copyright # 2003 John Wiley & Sons, Ltd. key words probability; optimism; risk perception; reasons; forecasts; evaluability When facing an important decision, people must often rely on an expert to provide the most informed estimate of the likelihood of a relevant outcome. For example, most people facing surgery have little idea how to determine their chances of a successful surgery, and therefore, need to rely on their doctor to provide the ‘bottom-line’ information about the likelihood that a surgery will be a success. Hence, it is critical that doctors and other experts provide likelihood information that is readily understood and not easily misinterpreted. For experts and laypeople, the best form of bottom-line information would appear to be a numeric probability or relative-frequency estimate. A numeric estimate provides people with precise information on a standard metric that has a shared and externally-referenced interpretation. The weakness of the main alternative to numeric probability forecasts—verbal uncertainty expressions—is that their interpretations can be quite vague (see e.g. Beyth-Marom, 1982; Bryant & Norman, 1980; Budescu & Wallsten, 1985; Sutherland et al., 1991; see Budescu & Wallsten, 1995, for a review). An expression such as ‘quite unlikely’ might be interpreted as suggesting a 0.05 probability by one person and a 0.20 probability by another. Relatedly, the Copyright # 2003 John Wiley & Sons, Ltd. * Correspondence to: Paul D. Windschitl, Department of Psychology, University of Iowa, Iowa City, IA 52242, USA. E-mail: [email protected] Contract/grant sponsor: National Science Foundation; contract/grant number: SES 99-11245. meaning of verbal likelihood expressions can vary depending on context (Brun & Teigen, 1988; Budescu & Wallsten, 1995; Fox & Irwin, 1998; Mullet & Rivet, 1991; Wallsten et al., 1986; Weber & Hilton, 1990). The term ‘likely’ might be interpreted differently when it is used to describe the chances of a snowstorm in Minnesota than when it is used to describe the chances of a snowstorm in Missouri. For these and other reasons, many researchers have explicitly argued that physicians should, whenever possible, communicate uncertainty information, such as the likelihood that a surgery will be successful, with numeric values rather than verbal expressions (Marteau et al., 2000; Nakao & Axelrod, 1983; Welkenhuysen et al., 2001). While noting the overwhelming support for the idea that verbal likelihood expressions are somewhat vague and susceptible to contrast effects, it is also important to note that numeric probability forecasts are not immune to interpretation problems. A forecaster’s probability estimate, while being very precise in terms of a formal numerical system, is somewhat ambiguous in terms of how it should be intuitively interpreted (see Moxey & Sanford, 2000; Teigen & Brun, 1999, 2000; Windschitl et al., 2002; Windschitl & Weber, 1999). For example, learning that a speaker said that there is an 18% chance of X does not tell a listener whether the speaker is trying to convey a possibility, a doubt, or neither. In contrast, verbal expressions of likelihood tend to be less ambiguous in terms of direction; a speaker who uses the phrase ‘small chance’ is interpreted as expressing a possibility, albeit small, that something might happen, whereas a speaker who uses the phrase ‘highly unlikely’ is interpreted as expressing a doubt (see Moxey & Sanford, 2000; Teigen & Brun, 2000). More broadly, Windschitl et al. (2002) argued that any numeric probability—whether it is a communicated forecast, an internal belief regarding the objective likelihood of an event, or external information on which a belief is based—can be ambiguous from an intuitive perspective even though it is numerically precise. This contention can be connected to work by Hsee and colleagues on the broader notion of evaluability (Hsee, 1996; Hsee et al., 1999). We suggest that an isolated numeric probability forecast is often difficult to evaluate and therefore does not have strong affective or intuitive implications (see also Kunreuther et al., 2001; Slovic et al., 2002). This makes probability forecasts susceptible to the influence of context. Learning that you have a 15% chance of experiencing a specific negative outcome may lead to alarm or complacency depending on contextual factors. Any context information that reduces the directional ambiguity of a forecaster’s numeric estimate or enhances an audience’s ability to meaningfully evaluate that numeric estimate relative to some standard will shape intuitive reactions. Windschitl and Weber (1999) demonstrated how one type of context information can have an assimilation effect on people’s intuitive responses to an expert’s precise numeric forecast. In one of their studies, participants read about a woman who was told by her doctor that, because of a diagnosed blood condition, she would have a 30% chance of developing a disease related to malaria on an upcoming trip. Participants then indicated their certainty on a nonnumeric scale that she would develop the disease on her trip. Participants who read that the doctor’s 30% estimate referred to a trip to India expressed greater certainty that she would get the disease than did participants who read that the 30% estimate referred to a trip to Hawaii. Results from studies using a variety of scenarios indicated that when respondents held an a priori association between the event in question and the described context (e.g. snow in the Colorado mountains), the context information would influence respondents’ certainty about the event in the direction of assimilation (Windschitl & Weber, 1999). Other recent experiments have demonstrated how context information of a different form can produce contrast rather than assimilation effects in the interpretations of numeric probability/frequency information (Windschitl et al., 2002). Participants in one experiment read about various diseases’ prevalence rates for a target group (e.g. women) and a context group (e.g. men). The prevalence rates for a given disease within the target group were held constant (e.g. 12%), but the rates for the disease within the context group were manipulated (either 4% or 20%). Participants used the prevalence rate for the context group as an immediate comparison standard for interpreting the rate for the target group. Consequently, participants intuitively perceived the target group as more vulnerable to the disease when the prevalence rate for the context group was low (4%) rather than high (20%). In another study, Windschitl et al. demonstrated that these contrast-induced shifts in intuitive perceptions could create situations in which participants’ beliefs about which of two events 108 Journal of Behavioral Decision Making Copyright # 2003 John Wiley & Sons, Ltd. Journal of Behavioral Decision Making, 16: 107–126 (2003) had a higher objective probability conflicted with their own intuitive expectations about which of the events was more likely. To summarize, although interpretations of numeric probability expressions may be less vulnerable to context effects than are verbal probability expressions, they are not immune. The above-mentioned research demonstrated how context information that matches a priori associations in memory can yield assimilation effects (Windschitl & Weber, 1999) and context information that constitutes an immediate comparison standard can yield contrast effects (Windschitl et al., 2002). In the present work, we investigated the potential influence of a different type of context information: the reasons that a forecaster cites for a probability estimate. Rarely do people receive a numeric probability forecast without some form of an accompanying explanation. To clarify and explain their forecasts, meteorologists show maps of pressure systems, stock market forecasters discuss various economic trends and signs, and physicians discuss test results, symptomology, and other factors with their patients. We suspected that in such cases, an audience would use these reasons or forms of supporting information as strong contextual cues for intuitively interpreting the expert’s probability forecast. For example, we suspected that the optimism of a 37-year-old person will differ if he or she is told ‘There is a 40% chance your joint will recover its full motion given you are under 40 years old’ or ‘There is a 40% chance your joint will recover its full motion given you are over 35 years old’. The former statement contains a positive cue given the situation under consideration, whereas the latter statement contains a negative cue. Similar to research showing that positively valenced primes lead to more favorable evaluations of ambiguous stimuli than do negative primes, we expected that a positive rather than a negative reason for a 40% estimate would lead to more favorable intuitive reactions to that intuitively ambiguous estimate (e.g. Bargh & Pietromonaco, 1982; Krosnick et al., 1992; Murphy & Zajonc, 1993). It is also the case that participants might use positive and negative reasons as a reflection of how the expert perceives the 40% estimate (e.g. unusually high relative to the norm), which would shape participants’ reactions to the estimate. As we will discuss more in the general discussion, processes of these two types have also been proposed as explanations for attribute framing effects (see Levin et al., 1998; McKenzie & Nelson, 2002). Our proposal is broadly related to, but distinct from, research on the influence of reasons for choices under uncertainty (see Shafir, 1993; Shafir et al., 1993; Tversky & Shafir, 1992). Shafir (1993), for example, investigated people’s choices between two options—one richly described with both positive and negative attributes and another sparsely described. Whether people were told to select or reject an option, they tended to indicate the richly described option, which illustrates how the compatibility of a reason with the operation (selecting or rejecting) can influence choice. It is important to emphasize that our focus is on the possibility that the reasons cited by an expert can influence people’s intuitive optimism even when the forecaster indicated that the information underlying those reasons was already incorporated into, or formed the basis of, his or her numeric estimate. From the forecaster’s perspective, the reasons he or she cites for a forecast should not be used by the audience to adjust the estimate. In other words, we focus on the impact of reasons cited by the forecaster for his or her bottom-line estimate. Although we expected that reasons offered for a bottom-line estimate would influence participants’ optimism, we did not assume that participants would necessarily think that the experts’ bottom-line estimate was wrong or in need of adjustment. Instead, we assumed that participants would generally believe that the experts’ numeric estimate was the best estimate regarding objective numeric probability, but that their intuitive interpretation of the estimate would differ depending on the additional information that accompanied it. OVERVIEW OF EXPERIMENTS We conducted three experiments in which participants read a hypothetical scenario that suggested they were about to have surgery on a duct leading to their gall bladder. The scenario mentioned various pieces of A. R. Flugstad and P. D. Windschitl Reasons and Probability Interpretations 109 Copyright # 2003 John Wiley & Sons, Ltd. Journal of Behavioral Decision Making, 16: 107–126 (2003) information provided by their doctor, including a precise numeric estimate reflecting the doctor’s ‘bottomline’ assessment of the probability that the surgery would fail. The primary dependent measure asked participants to report what their intuitive optimism/pessimism would be regarding the surgery. In Experiment 1, we manipulated the primary valence of the reasons supporting the numeric forecast as well as the actual numeric forecast regarding failure. In Experiment 2, we held the failure forecast constant and manipulated participant’s relative social-comparison status as well as the primary valence of the mentioned reasons. In Experiment 3 we again held the failure forecast constant, but we manipulated the reasons in a slightly different manner. Specifically, we manipulated the number, not the valence, of complications that could lead to an unsuccessful surgery. It is important to note that in these experiments, we did not directly manipulate the stated overall optimism/pessimism of the doctor (except in the forecast manipulation of Experiment 1). We simply inserted and manipulated information that a doctor might commonly mention in support of his or her assessment of a patient’s risks related to surgery.