Davidhardisty.info



Neutralizing the Expense Prediction BiasChuck Howarda, David J. Hardistya, Abigail B. Sussmanb, and Melissa A. Z. KnollcaSauder School of Business, University of British Columbia, CanadabBooth School of Business, University of Chicago, United StatescConsumer Financial Protection Bureau, United StatesAuthor note: Address correspondence to Chuck Howard, chuck.howard@sauder.ubc.ca. This research was supported by the Social Sciences and Humanities Research Council of Canada and the True North Communications Inc. Faculty Research Fund at The University of Chicago Booth School of Business. The views expressed are those of the authors and do not necessarily represent the views of the Consumer Financial Protection Bureau or the United States.Abstract Expense prediction bias is the phenomenon that consumers tend to under-predict their future expenses. The present research theorizes that this bias occurs because people mentally represent the future using prototypes, and leverages this insight to develop a simple cognitive tool that improves expense prediction accuracy. Four studies (N = 2,022) provide support for this theory and validate the effectiveness of the tool. Study 1 demonstrates that consumers predict their future expenses will be both lower and more typical than their past expenses, and that higher perceived typicality is correlated with lower expense predictions. Study 2 replicates these findings in a nationally representative sample of Americans and shows that decreasing perceived typicality of future expenses increases expense predictions. Study 3 shows that the tool improves expense prediction accuracy (vs. control) in a week-long financial diary study that compares expense predictions against actual expenses incurred during the target week. Study 4, a five week longitudinal field study with clients of a Canadian bank, shows that without intervention mean prediction accuracy does not improve over time, but that the tool is capable of completely neutralizing the bias. Consumers tend to under-predict their future expenses, and this error can be costly. For example, approximately 25% of Americans with a 401(k) savings account withdraw funds early (i.e., before retirement), often to cover unexpected expenses (Fellowes and Willemin, 2013). These early withdrawals cost consumers approximately $7 billion a year in penalties. Similarly, each year almost 2 million Americans use a payday loan to cover an unexpected expense (Pew, 2012). The APR on these loans frequently exceeds 400% (Consumer Federation of America, 2018). Many consumers also hold the expectation that they will be able to pay off their credit card balance each month (Yang, Markoczy, and Qi, 2007). Yet American consumers collectively hold over one trillion dollars in credit card debt and pay associated interest costs (Federal Reserve Bank of New York, 2018). These examples suggest that increasing expense prediction accuracy can help consumers spend, save, and/or borrow money in a more efficient manner. An accurate assessment of future expenses can help consumers better allocate funds between their checking and 401(k) accounts to avoid penalties for early withdrawals. And if consumers had a clearer idea of how much they would spend in the future, they might choose to spend less in the present to avoid the costs associated with borrowing or using credit to cover expenses down the road. The prosocial value of helping consumers avoid these costs is evident. The rush by venture capital firms to fund Fintech start-ups offering app-based products that help consumers manage their expenses (CB Insights, 2018) indicates there is also firm value in improving expense prediction accuracy.Echoing the examples offered above, academic research also suggests that consumers tend to under-predict their future expenses (Ulkumen, Thomas, and Morowitz, 2008; Peetz and Buehler, 2009; Sussman and Alter, 2012), a phenomenon we label the expense prediction bias. The goal of the present research is to identify a key psychological driver of the bias, then leverage that theoretical insight to develop, test, and validate a simple cognitive tool that improves consumers’ expense prediction accuracy. To do so, we first theorize that the bias occurs in part because mental representations of the future are shaped by cognitive prototypes, then demonstrate that prediction accuracy can be improved by decreasing perceived typicality of future expenses.By developing and testing a prototype theory of expense prediction bias, this research makes the following contributions. First, we advance understanding of why this bias occurs by identifying perceived typicality of future expenses as an important cause of the bias. Second, we provide the first practical, effective, and field-tested intervention capable of neutralizing the bias. Third, we provide the first comprehensive knowledge of the bias itself. For example, this research is the first to identify the magnitude, prevalence, and persistence of the bias in non-student samples. It is also the first work to study the bias longitudinally and in the field, to measure monthly expense predictions against actual expenses incurred during the target month, and to explore correlates of the bias. Finally, the present research contributes to a nascent literature that demonstrates a temporal asymmetry in which people mentally represent the future in more prototypical terms than the past (Kane, Van Boven, and McGraw, 2012; Van Boven, Kane, and McGraw, 2009; Williams and Laboeuf, under review). By comparing perceived typicality of past vs. future expenses we extend this work to the domain of money. In the following section, we present our prototype theory of expense prediction bias and lay out our hypotheses. We then present four studies that test these hypotheses in the lab and the field. To conclude, we discuss the theoretical and practical implications of this work, and directions for future research.THEORETICAL DEVELOPMENTPredictions are based on mental representations of the future (Kahneman and Tversky, 1979; Buehler, Griffin, and Peetz, 2010; Dunning 2007). Therefore, understanding the nature of these representations is important for understanding prediction errors. The proposition at the core of our theory is that consumers under-predict their expenses in part because their mental representations of the future are shaped by cognitive prototypes. We define the term prototype broadly to mean a cognitive “blueprint” based on the central or typical instances of a category (Kane, Van Boven, and McGraw, 2012). We then conceptualize mental representations of the future as falling along a continuum that ranges from prototypical to exhaustive. So, for example, at one end of the continuum a consumer can base her prediction on a prototypical representation that includes only her typical, routine expenses like groceries, gas, and phone bill. Towards the other end of the spectrum, a consumer can base her prediction on a more exhaustive representation that includes the aforementioned expenses plus relatively atypical expenses which vary with respect to timing and/or amount, like new clothes, a spontaneous night out with friends, or a car repair. Importantly, mental representations can fall at either end of this spectrum or anywhere in between, with the implication being that, in general, more prototypical representations will lead to lower and less accurate expense predictions. In the subsections that follow we draw on the findings of previous research to develop our theory in detail and present our hypotheses. COMPARING REPRESENTATIONS OF THE PAST AND FUTURETo evaluate the veracity of the proposition that mental representations of the future are shaped by cognitive prototypes it is useful to begin by contrasting representations of the future (prospection) against those of the past (retrospection). People tend to retrospect in two ways: they remember by reconstructing the past in pursuit of factually accurate memories or they engage in counterfactual thinking about a past that could have been but was not. Although both of these processes leave room for imagination and abstraction, research suggests they are habitually constrained by reality (Van Boven, Kane, and McGraw, 2009; Johnson and Raye, 1981; Kahneman and Miller, 1986; Roese, 1997). Prospection should be similarly constrained by realistic details given the important role it plays in planning for the future, but research has demonstrated it is not (D’Argembeau and Van der Linden, 2004; Kane, Van Boven, and McGraw, 2012). This suggests a temporal asymmetry in which representations of the future are more prototypical than representations of the past, because prototypes tend to lack contextual detail (Kane, Van Boven, and McGraw, 2012). In a direct test of this temporal asymmetry hypothesis, Kane, Van Boven, and McGraw (2012; study 4) had one group of participants draw a picture of a tropical beach vacation in retrospect and another group draw the vacation in prospect. They then presented a third group of participants with randomly matched pairs of the drawings (one drawn in prospect, the other in retrospect) and had them evaluate which drawing was closer to a “prototypical vacation.” In support of the hypothesis that representations of the future are more prototypical than representations of the past, the pictures drawn in prospect were more frequently rated as more prototypical than those drawn in retrospect. Notably, this result and the hypothesis behind it is consistent with work showing that there is considerable neural differentiation during prospective (vs. retrospective) event construction (Addis, Wong, and Schacter 2007).In the context of expense prediction, the temporal asymmetry hypothesis leads to the expectation that consumers will predict their future expenses will be more typical than their past expenses. If true, it then follows that predicted expenses (e.g., for the next week) will be lower than recalled past expenses for the same time frame because, as argued above, higher perceived typicality of future expenses is expected to be associated with lower expense predictions. COMPARING PREDICTIONS WITH REALITYEvaluating how predictions systematically differ from what actually happens in reality offers a second way in which we can understand the nature of the representations upon which predictions are built. Below, we draw from research on affective forecasting, the planning fallacy, and consumer expense prediction to further advance our position that mental representations of the future are shaped by cognitive prototypes, and that prototypical representations lead consumers to under-predict their future expenses.Research on affective forecasting has found that people overestimate the impact of future emotional events (Gilbert, Pinel, Wilson, Blumberg, and Wheatly, 1998; Wilson, Wheatly, Meyers, Gilber, and Axsom, 2000). This supports the proposition that mental representations of the future are shaped by cognitive prototypes because the prototype of emotional events tends to be highly affective, and involves expecting the best of positive events and the worst of negative events (Gilbert, Morewedge, Risen, and Wilson, 2004; Schkade and Kahneman, 1998; Wilson et al., 2000). Research on the planning fallacy has documented that people systematically under-predict their task completion times (e.g., how long it will take to file their taxes, or finish their holiday shopping), even when equipped with the knowledge that similar tasks have taken longer than planned in the past (Kahneman and Tversky, 1979; Buehler, Griffin, and Ross, 1994; Buehler, Griffin, and MacDonald, 1997). This inability to incorporate relevant details from the past into predictions for the future has largely been attributed to a focus on plans for success (Buehler, Griffin, and Peetz, 2010). A complementary perspective is that people base their predictions on prototypes that neglect atypical outcomes. For example, a consumer’s prediction that it will take 3 hours to finish their holiday shopping may be based on typical instances of such a task that include going to the mall, finding presents, and even waiting in line to pay, but not less typical instances like discovering the Lego set that little Timmy really wanted is sold out and needs to be bought from another store, or getting rear-ended in the parking lot. In this case, prediction is not necessarily predicated on a plan to finish shopping quickly or by a certain time, it is simply a function of what is (and isn’t) included in the prototype on which the prediction is based. To parse the relative influence of plans vs. prototypes on expenses predictions, we measure short-term financial propensity to plan (Lynch et al., 2010) in addition to perceived typicality of future expenses in studies 3 and 4. Finally, although our work is theoretically distinct from past research on expense misprediction, our proposition that consumers’ expense predictions are based on prototypical representations of the future is consistent with previous findings in this literature. For example, Sussman and Alter (2012; study 1) demonstrated that people tend to under-predict their future spending on exceptional expenses (e.g., replacing a broken TV) but not their ordinary expenses (e.g., paying a monthly bill). Similarly, the results of Peetz et al. (2015) suggest that expense predictions often do not include “non-focal” expenses, and Peetz and Buehler (2013) found that prompting people to consider goals that require spending more money than they typically do (e.g., by buying a luxury, being more generous, or dressing more fashionably) increases spending estimates. Taken together, these findings suggest that consumers are attuned to prototypical expenses when making predictions, but they neglect to incorporate atypical expenses that fall outside of their prototype. Our theory is also consistent with the observation that higher prediction confidence can lead to lower (and less accurate) expense predictions (Ulkumen, Thomas, and Morowitz, 2008), because confidence indicates cognitive ease (Alter and Oppenheimer, 2009), and prototypes are “easy on the mind” (Winkielman, Halberstadt, Fazendeiro, and Catty, 2006). In sum, the research reviewed above supports the proposition that representations of the future are shaped by cognitive prototypes, and more specifically, that prototypical representations of the future lead consumers to under-predict their future expenses. Below, we outline how these insights can be leveraged to design an intervention that improves expense prediction accuracy. IMPROVING EXPENSE PREDICTION ACCURACYIf our theorizing is correct and expense under-prediction is due at least in part to prototypical representations of the future, then it follows that decreasing prototypicality should increase expense prediction accuracy. We reasoned that having people list three reasons why their expenses might be different from a typical week before making their expense prediction would serve as a simple cognitive tool that accomplishes this goal. The logic underlying this tool mirrors research demonstrating that “defocalizing” can improve affective forecasts. For example, Wilson et al. (2000) found that people induced to think about their post-event daily routines before they predicted how that event would make them feel were subsequently less likely to over-predict their affective reaction to that event. Because affective prototypes tend to be extreme, a reasonable interpretation of this finding is that having people consider non-prototypical information lessened their reliance on prototypes, and in doing so improved their prediction accuracy. In the context of expense prediction we therefore expected that because prototypical representations of future expenses are based on typical, routine expenses, having people consider atypical expenses would reduce prototypicality and increase expense prediction accuracy.PROTOTYPICALITY AND THE EXPENSE PREDICTION BIASThe findings reviewed above provide evidence from several distinct areas of research that converges on the proposition that consumers under-predict their expenses in part because their mental representations of the future are shaped by cognitive prototypes. This leads to several testable hypotheses. First, if the future is represented in more prototypical terms than the past, we would expect to find support for the hypothesis that:H1: Consumers, on average, predict their future expenses will be more typical than their past expenses.Our second hypothesis is that greater reliance on prototypical representations should be associated with lower and less accurate expense predictions. This follows from the fact that cognitive prototypes tend to lack detail, and our conceptualization of representations falling along a continuum ranging from prototypical to exhaustive. Because mental representations are clearly a latent construct, we use perceived typicality of future expenses as a proxy measure for them and hypothesize that: H2: Perceived typicality of future expenses is negatively correlated with the dollar amount of expense predictions.A clear implication of H1 and H2 is that consumers will under-predict their future expenses as compared to their past expenses. In other words, if people think their expenses will be more typical in the future, and higher perceived typicality is associated with lower expense predictions, then it follows that:H3a: Consumers predict lower expenses for the future as compared to the expenses they recall for the past. Note that if H3a holds, then it also follows that consumers will under-predict their expenses as compared to their actual expenses for the target week or month, because although retrospection is constrained by reality, people do tend to remember the past in slightly optimistic terms (Buehler, Griffin, and Ross, 1994). Therefore, we further hypothesize that:H3b: Consumers under-predict their expenses for the target week or month as compared to their actual expenses for that week or month.Finally, we hypothesize that our intervention will succeed because:H4: Experimentally decreasing perceived typicality of future expenses will increase expense prediction accuracy.OVERVIEW OF STUDIESTo test our hypotheses we conducted a series of 4 experiments. In study 1, we examine whether consumers predict more typical and/or lower expenses for the future as compared to the past, and the extent to which perceived typicality is associated with predictions. In study 2, we directly replicate study 1 and manipulate perceived typicality of future expenses to establish a causal relationship between perceived typicality and expense predictions. In study 3, we test the efficacy of our intervention (vs. control) in a week long financial diary study. In study 4, we examine the persistence of these effects in a repeated-measures, longitudinal field study. Given that we sample from economically diverse non-student populations, and that our expense measures are unbounded, free response questions, the distribution of expenses in our data displays strong positive skew, includes some highly influential outliers, and violates the homogeneity of variances (HOV) assumption in multi-condition studies. To address these concerns we perform a simple three-step data transformation that is in line with past research on expense prediction, and which we apply consistently to all data across all studies (Meyvis and Van Osselaer, 2017). First, to reduce the impact of the most extreme outliers we exclude the data of participants whose reported expenses exceed their predicted expenses by a factor of ten or more, or whose predicted expenses exceed their reported expenses by a factor of ten or more. Second, because the data still displays significant positive skew and still violates the HOV assumption after excluding these outliers, we natural-log transform all expense data for inferential statistical analyses (Peetz and Buehler, 2015). Finally, to make our findings easily interpretable, we exponentiate our results and present our descriptive findings in dollar terms. Notably, this procedure significantly reduces skew in the data and ameliorates concerns related to HOV violations and the influence of outliers, but it does not change the pattern of results. All exclusions are detailed in the study descriptions below, each step of the data transformation is described statistically in web appendix A, and all study materials, data, and syntax are available here: [insert hyperlink].Study 1The goal of this study was to test the hypotheses that consumers predict their future expenses will be more typical (H1) and lower (H3a) than their past expenses, and that perceived typicality of future expenses is negatively correlated with expense predictions (H2). In this study we defined EPB as the difference between recalled expenses for the past and predicted expenses for the next week. We chose weekly expenses (as opposed to monthly or yearly expenses) as our unit of analysis to test H3a for two reasons: 1) we collected survey data from our participant pool prior to running this study that suggested people commonly think about their expenses in weekly terms, and 2) we wanted participants to be able to recall their past expenses while their memories were fresh. MethodParticipants. We recruited 499 US residents via Amazon Mechanical Turk to participate in a short consumer expense survey (Mage = 33.51; 41.3% female). The predicted and/or recalled expenses of 14 participants (2.8% of the initial sample) fell outside of our inclusion criterion (as specified above), leaving us with an effective sample size of 485 (Mage = 33.67; 41.6% female). Procedure. Participants were first asked to report their expenses for the past week and then predict their expenses for the next week. Specifically, participants read the following instructions:Please take some time to estimate your expenses for the past [next] week (i.e., the past [next] 7 days).Please enter your total estimated expenses (in dollars) for the past [next] week. Your estimate should account for all the expenses you incurred [will incur] except monthly expenses like rent that happen[ed] to be due in the past [next] week. We asked participants to exclude monthly expenses like rent from their estimates to reduce the possibility that any observed bias could be due to a time of the month effect. We next measured perceived typicality of expenses by asking “How different or similar do you think your expenses were [will be] for the past [next] week, relative to a typical week?” (1=Very different; 7=Very similar). Finally, participants were asked to report basic demographic information.Results and DiscussionPerceived Typicality. As illustrated in the right hand side of figure 1, participants predicted that their future expenses would be more typical than their past expenses, (Mpastweek = 4.71, 95% CIpastweek = [4.56, 4.86], Mnextweek = 5.03, 95% CInextweek = [4.89, 5.16], t(484) = 4.22, p <.001, d = .20). Furthermore, correlational analysis showed that higher perceived typicality of future expenses was associated with lower expense predictions (r(483) = -.17, p < .01). Expense Prediction Bias (Recalled – Predicted Expenses). As illustrated in the left hand side of figure 1, predicted expenses for the next week were 10.80% ($19.55) lower than reported expenses for the past week, as confirmed with a paired t-test (Mpastweek = $180.98, 95% CIpastweek = [$165.41, $198.03], Mnextweek = $161.43, 95% CInextweek = [$147.57, $176.60], t(484) = 3.89, p < .001). Figure 1Mean Reported vs. Predicted Expenses and Mean Perceived Typicality of Past vs. Future Expenses in Study 1Error Bars Represent 95% Confidence IntervalsAncillary Analyses. We also investigated the distributional properties of reported vs. predicted expenses. Of the 485 participants in our sample for this study, 18.6% (n = 90) predicted that their expenses for the next week would be the same as their expenses for the past week. Of the remaining 395 participants, 59.5% predicted that their future expenses would be lower than their past expenses vs. 40.5% who predicted that their future expenses would higher. Thus, among participants who predicted that their expenses would fluctuate, significantly more participants predicted that their expenses would fall rather than rise in the future (z = 3.78, 95% CI = [54.48%, 64.38%], p < .001).Finally, we also examined the variability of expenses over time. Reported and predicted expenses were positively correlated (r(485) = .79, p < .001), demonstrating that participants with higher expenses in the past also predicted higher expenses for the future. This “correlational accuracy” echoes findings in work on the planning fallacy (Buehler, Griffin, and Peetz, 2010).To summarize, the results of study 1 provide support for our hypotheses that consumers predict their future expenses will be more typical (H1) and lower (H3a) than their past expenses. The data also support our hypothesis that higher perceived typicality of future expenses is associated with lower expense predictions (H2). However, study 1 is correlational in nature and therefore does not allow us to make causal inferences about the relationship between temporal asymmetry and expense prediction bias. Studies 2 – 4 address this limitation by experimentally manipulating perceived typicality of future expenses and measuring expense predictions against recalled expenses (study 2) and actual expense for the target week (studies 3 and 4).Study 2The purpose of study 2 was twofold. First, we sought to directly replicate the results of study 1 with a nationally representative sample of the adult population in the US. Second, we wanted to establish a causal relationship between perceived typicality and expense predictions by testing the hypothesis that experimentally decreasing perceived typicality will increase expense predictions (H4). As in study 1, we defined EPB in study 2 as the difference between recalled expenses for the past week and predicted expenses for the next week. MethodParticipants. A nationally representative sample of 1,108 US residents completed study 2 via Time-Sharing Experiments for the Social Sciences. The predicted and/or recalled expenses of 60 participants (5.4% of the initial sample) fell outside of our inclusion criterion, leaving us with an effective sample size of 1,048 (Mean age = 49.59; 53.0% female; 72.8% Caucasian, 9.4% Black, 10.7% Hispanic, 7.2% Other; Mode level of education = Bachelor’s degree; Median household annual income = $50-59,999).Procedure. Participants were randomly assigned to one of three conditions: control, typical, or atypical. In the control condition participants recalled and predicted their expenses for the past and next week as in study 1. Participants in the typical condition recalled and predicted their expenses but received the following instructions before making their prediction: “Now consider why your expenses for next week might be similar to that of any other week. In the spaces provided below, please type 3 reasons why your expenses for next week might be similar to that of any other week.” We hypothesized that this would not significantly impact predictions (as compared to control) because predictions in the control condition should already be swayed by the illusion of a more typical future (H1). The atypical condition paralleled the typical condition but instructed: “Now consider why your expenses for next week might be different from that of any other week. In the spaces provided below, please type 3 reasons why your expenses for next week might be different from that of any other week.” We hypothesized that this would decrease perceived typicality of future expenses and therefore increase expense predictions (H4). The order of prediction and recall was counterbalanced in all conditions. After they completed the expense recall and prediction task we asked participants “Is there anything that you spent money on during the past week that you believe you will NOT spend money on in the next week?” and “Is there anything you believe you will spend money on in the next week that you did NOT spend money on during the past week?” Participants were then given the opportunity to list a description and corresponding dollar amount for up to five such expenses. This measure was designed to let us gain insight into whether a typical future includes fewer expenses and/or lower expenses, and likewise, whether our atypical intervention operates on the number of expenses people consider and/or the size of expenses that people consider.Participants next completed measures that asked them to indicate how different or similar they thought their expenses were (and would be) for the past and upcoming week. Specifically, they were asked to “Please use the following scale to indicate how different or similar you think your expenses were [will be] for the past week [next week], relative to that of any other week” (1=Very different; 7=Very similar). Finally, participants completed five measures designed to let us to explore the relationship between EPB, financial slack (Zauberman and Lynch, 2005), various measures of spending (e.g., willingness to pay for an optional expense like a fancy dinner out with friends), and available resources. These measures yielded null results that are discussed in web appendix B.Results and DiscussionReplicating Study 1. We began by analyzing the study 2 control condition data in isolation to see if the results of study 1 were directly replicated. As illustrated in figure 2, they were. Participants in the control condition of study 2 predicted that their future expenses would be more typical than their past expenses (Mpastweek = 4.40, 95% CIpastweek = [4.23, 4.57], Mnextweek = 4.65, 95% CInextweek = [4.48, 4.81], t(415) = -3.42, p = .001, d = .17), and 8.6% ($20.44) lower than their past expenses (Mpastweek = $237.46, 95% CIpastweek = [$217.02, $262.43], Mnextweek = $217.02, 95% CInextweek = [$196.37, $239.85], t(415) = 2.76, p = .006). Furthermore, higher perceived typicality of future expenses was again associated with lower expense predictions (r(414) = -.21, p < .001). Taken together with the results of study 1, this provides converging support for our hypotheses that consumers believe their future expenses will be more typical (H1) and lower (H3a) than their past expenses, and that higher perceived typicality is associated with lower expense predictions (H2).The distributional results of study 1 were also replicated in the control condition of study 2. Of the 416 participants in the control condition of study 2, 29.3% (n = 122) predicted that their expenses for the next week would be the same as their expenses for the past week. Of the remaining 294 participants, 57.14% predicted that their future expenses would be lower than their past expenses vs. 42.86% who predicted that their future expenses would higher. Thus, as in study 1, a significantly higher proportion of participants who predicted that their expenses would fluctuate predicted that their expenses would fall rather than rise in the future (z = 2.45, 95% CI = [51.27%, 62.81%], p = .01). We next expand to our analyses to test for differences in perceived typicality and EPB across all three conditions.Figure 2Mean Reported vs. Predicted Expenses and Mean Perceived Typicality of Past vs. Future Expenses in the Control Condition of Study 2Error Bars Represent 95% Confidence IntervalsPerceived Typicality Manipulation Check. The manipulation in our atypical condition successfully decreased perceived typicality vs. the control and typical conditions, as confirmed by a one-way ANOVA with condition (control vs. typical vs. atypical) as the independent variable and perceived typicality of future expenses as the dependent variable (F(2, 1044) = 32.27, p < .001). Planned contrasts confirmed lower perceived typicality in the atypical condition (Matypical = 3.74, 95% CIatypical = [3.56, 3.93]) than in the control (Mcontrol = 4.64, 95% CIcontrol = [4.48, 4.81]; t(1044) = -7.24, p < .001, d = .55) and typical conditions (Mtypical = 4.65, 95% CItypical = [4.48, 4.83]; t(1044) = -6.91, p <.001, d = .56). Notably, perceived typicality did not differ between the control and typical conditions (t(1044) = .03, p = .97). This provides further evidence that consumers perceive the future as typical. If a consumers’ beliefs about future typicality were ambiguous, then we would expect the manipulation in our typical condition to increase perceived typicality vs. control. That perceived typicality was virtually identically in these two conditions suggests that consumers’ natural belief is that the future will be typical.Expense Prediction Bias (Recalled – Predicted Expenses). Predicted expenses were 8.6% ($20.44) lower than recalled expenses in the control condition (t(415) = 2.76, p = .006), and 6.4% ($13.68) lower than recalled expenses in the typical condition (t(331) = 2.07, p = .039), but predicted and recalled expenses were statistically equivalent in the atypical condition (t(299) = -1.49, p = .14). In other words, EPB was neutralized (and even slightly reversed) by our intervention in the atypical condition, as illustrated in Figure 3. A 3(condition: control vs. vs. typical vs. atypical) x 2(time period: past week vs. next week) between-within ANOVA with expenses as the dependent variable confirmed a significant main effect of condition (F(2, 1045) = 4.64, p = .010), no main effect of time period (F(1, 1046) = 1.69, p =.19), and a significant condition by time period interaction (F(2, 1045) = 5.22, p = .006). Planned contrasts further confirmed that predicted expenses in the atypical condition (Matypical = $273.14, 95% CIatypical = [$242.26, $307.97]) were 25.9% ($56.12) higher than in the control condition (Mcontrol = $217.02, 95% CIcontrol = [$196.37, 239.85]; t(1045) = 2.91, p = .004), and 35.0% ($70.79) higher than in the typical condition (Mtypical = $202.35, 95% CItypical = [$181.27, $223.63]; t(1045) = 3.67, p < .001). Predictions did not differ between the control and typical conditions (t(1045) = .98, p = .33). Planned contrasts also revealed that recalled expenses did not differ between the atypical (Matypical = $254.68, 95% CIatypical = [$223.63, $287.15]) and control conditions (Mcontrol = $237.46, 95% CIcontrol = [$217.02, $262.43]; t(1045) = .79, p = .43), but they were somewhat lower in the typical condition (Mtypical = $214.86, 95% CItypical = [$192.48, $239.85]) than in the atypical condition (t(1045) = 2.01, p = .045). While we cannot explain this difference in recalled expenses, it is worth noting that this makes our test of EPB between these two conditions quite conservative. This is because lower (higher) recalled expenses decreases (increases) the size of that bias. However, despite lower recalled expenses in the typical condition and higher recalled expenses in the atypical condition, we observe a significant bias in the former but not in the latter.Figure 3Mean Recalled vs. Predicted Expenses in the Control vs. Typical vs. Atypical Conditions in Study 2Error Bars Represent 95% Confidence IntervalsMediation Analysis. The results above confirm that our atypical manipulation succeeded in decreasing perceived typicality of future expenses, and that expense predictions were significantly higher in the atypical condition as well. To further investigate the relationship between perceived typicality and expense predictions we tested a mediation model with condition (atypical = 1 vs. control = 0) as the independent variable, expense prediction as the dependent variable, and perceived typicality as the mediating variable. The indirect effect of condition on expense prediction via perceived typicality was significant (indirect effect =.12, SE = .03, 95% CI = [.07, .18]). Specifically, the model confirms that our atypical condition intervention succeeded in decreasing perceived typicality of future expenses (b = -.90, 95% CI = [-1.15, -.66]; t(713) = -7.18, p < .001), and demonstrates that lower perceived typicality is associated with higher expense predictions even while controlling for condition (b = -.14, 95% CI = [-.18, -.09]; t(1103) = -5.79, p < .001). A categorical mediation model that includes all three conditions also produces a significant indirect effect (omnibus test of indirect effect =.01, SE = .002, 95% CI = [.003, .012]; indirect effect of atypical dummy =.11, SE = .02, 95% = [.067, .161]; indirect effect of typical dummy = -.005, SE = .015, 95% CI = [-.028, .030]).Expense Listing Task. To analyze the results of the expense listing task we began by examining the correlation between perceived typicality of future expenses and the number of expenses participants listed for the next week, as well as the mean dollar amount of the expenses they listed for the next week. Higher perceived typicality was negatively correlated with both the number of expenses listed (r(658) = -.20, p < .001) and the mean dollar amount (r(591) = -.34, p < .001), suggesting that a more typical mental representation of the future includes fewer and lower expenses. Contrast analyses then revealed that participants in the atypical condition listed more expenses (contrast = .49, t(658) = 2.08, p = .04, d = .18) and higher expenses (contrast = .86, t(591) = 2.96, p = .003, d = .26) for the coming week than participants in the control and typical conditions. This provides some insight into how our intervention works: it seems that it not only prompts participants to consider more expenses, it also prompts them to consider larger expenses.The results of study 2 are notable in several ways. First, study 2 provides additional support for H1-H3a by directly replicating study 1 in the control condition. Study 2 also expands on these findings by manipulating perceived typicality of future expenses and identifying a causal link between perceived typicality and expense predictions. Specifically, study 2 provides evidence that our atypical manipulation increases expense predictions by decreasing perceived typicality (H4). The results of study 2 also provide insight into how our intervention works by showing that it increases both the number and mean dollar amount of expenses that consumers consider when contemplating the future.The operationalization of EPB in studies 1 and 2 (recalled – predicted expenses) is useful for both theory building and hypothesis testing, but its natural limitation is that it doesn’t reveal true prediction accuracy. To address this, studies 3 and 4 operationalize EPB as the difference between predicted expenses and the expenses each participant actually incurs during the target week (studies 3 and 4) or month (study 4). This will allow us to more accurately measure prediction accuracy and better understand the economic significance (i.e., the true magnitude) of the bias. Study 3The results of studies 1 and 2 show that consumers believe their future expenses will be more typical (H1) and lower (H3a) than their past expenses, and that higher perceived typicality is associated with lower expense predictions (H2). Furthermore, study 2 experimentally manipulates participants’ perception of the future and demonstrates that decreasing perceived typicality of future expenses increases expense predictions to the level of reported expenses for the past week (H4). Study 3 builds off of study 2 by measuring predictions against actual expenses for the target week (H3b), as recorded in a daily expense diary. Thus, EPB in study 3 is defined as the difference between actual and predicted expenses so that we can measure the impact that experimentally decreasing perceived typicality has on prediction accuracy. We also used this study to investigate the relationship between EPB and financial decisions and outcomes related to savings and debt (H5). Participants completed a financial decision making task that asked them to imagine that they had just unexpectedly inherited $1,000, then asked them how they would allocate the $1,000 between spending, savings, and debt repayment. The financial outcome variables were self-reports of how much money participants’ tried to save each month, how much money they actually saved each month, how frequently they incurred credit card interest charges, their credit score, their current level of credit card, car loan, and payday loan debt, how much money they currently owed to friends and family, their subjective financial well-being (Consumer Financial Protection Bureau, 2017), perceived financial security, financial slack (Zauberman and Lynch, 2005), and ability to meet their savings goals. Finally, study 3 also investigated the relationship between EPB and several individual difference measures of potential theoretical and practical importance. These measures were short-term financial propensity to plan (Lynch et al., 2010), trait optimism (Scheier, Carver, and Bridges, 1994), spendthrift-tightwad tendencies (Rick, Cryder, and Lowenstein, 2008), numeracy (Schwartz et al., 1997), linear vs. cyclical time orientation (adapted from Tam and Dholakia, 2014), openness to experience (John, Donahue, and Kentle, 1991), temporal discounting for both losses and gains (Kirby & Marakovi?, 1996), preferred form of payment (credit card, debit card, cash, other), budgeting behavior, perceived expense predictability and variability, socio-economic status, gender, and education. These measures yielded almost exclusively null results, and are therefore summarized in web appendix C.MethodParticipants. For study 3 we recruited 402 people from 21 different countries via the online platform prolific.ac to participate in a week long consumer finance diary study that required completing eight surveys (one survey per day from Sunday to Sunday). Payment for completing the first seven surveys was ?0.50 and payment for completing the eighth survey was ?7.00. Each survey took approximately 5 minutes to complete. 325 participants completed the study in full. 23 of these participants (7.1% of the initial sample) had expenses that fell outside of our data inclusion criterion (as specified above). Thus, our effective sample size was 302 (Mage = 35.6, 44.4% female).Procedure. Participants were randomly assigned to predict their expenses for the coming week in one of three conditions: control, typical, or atypical. Participants in the control condition were first asked to “Please take some time to consider your expenses for the next week (i.e., the next 7 days).” Then, on the following page, they were asked to “Please enter your total estimated expenses for the next week.” Note that in this study we did not instruct participants to exclude monthly expenses from their predictions (as we did in studies 1 and 2) because this study measures predicted expenses against actual expenses, not past expenses. Therefore, natural variation in the timing of monthly expense payments posed no threat to the validity of our EPB measure in this study. Participants in the typical and atypical conditions were asked to “Please take some time to consider why your expenses for the next week (i.e., the next 7 days) might be similar to [different from] a typical week.” Then, on the same page of the study, they received the following instructions: “In the spaces provided below, please type 3 reasons why your expenses for next week might be different from a typical week.” Finally, on the following page, they were asked to “Please enter your total estimated expenses for the next week.” Participants in all conditions then proceeded to a new page on which they were asked “How different or similar do you think your expenses will be for the next week, relative to a typical week?”, “How sure or confident are you that your estimate of your total expenses for next week is accurate?”, and “What currency did you report your expense estimate in?” The different/similar question was used as a manipulation check. The confidence question was included so that we could test the possibility that our manipulation was simply decreasing participants’ confidence in their predictions, because past research has shown that lower prediction confidence can lead to higher expense predictions (Ulkumen et al., 2008). The currency question was included because we recruited an international sample for this study and we wanted to be able to convert all expenses into a common currency (USD) for analysis.On a series of separate pages participants were then asked to indicate how much spare money they expected to have in the next week (1 = Very little spare money, 11 = A lot of spare money), how they would allocated the $1,000 in the decision task described above, what form of payment they most often use (credit card, debit card, cash, other), if they set a weekly budget (yes/no), how much time they spend thinking about and planning their expenses each week, how much money they could come up with on short notice to pay an unexpected bill immediately (i.e., a measure of available resources), and where they would place themselves on a 10 step ladder representing socio-economic status. Participants then completed standard demographic questions. Over the following seven days (Monday to Sunday) participants reported their expenses online at the end of every day, and completed a subset of the individual difference and financial outcome measures described above. The expense reporting instructions were as follows: In the column below titled "Expense Description" please provide a brief description of each expense that you incurred on Monday, June 5th (e.g., Coffee, Lunch, Groceries, Hydro Bill, Car Repair). In addition to expenses paid for with cash, credit cards, debit cards, cheques, etc..., please also be sure to include any expenses that automatically came out of your bank account today (e.g., bank fees, interest payments, monthly recurring bills).In the column titled "Amount" please enter the total amount of each expense including all taxes, fees, and "extras" like tips that you might have left for good service at a restaurant. Please only type numbers into this column. Once you have listed all your expenses for Monday, June 5th, please leave any remaining spaces blank and proceed to the next page of the survey. NOTE: For expenses such as groceries you do not need to list each individual item (e.g., bread, milk, eggs...), you can just record the total amount you spent at the market and label the whole expense as "groceries".Results and DiscussionPerceived Typicality Manipulation Check. The manipulation in our atypical condition successfully decreased perceived typicality vs. the control and typical conditions, as confirmed by a one-way ANOVA with condition (control vs. typical vs. atypical) as the independent variable and perceived typicality of future expenses as the dependent variable (F(2, 299) = 61.38, p < .001). Planned contrasts confirmed lower perceived typicality in the atypical condition (Matypical = 3.53, 95% CIatypical = [3.20, 3.86]) than in the control (Mcontrol = 5.09, 95% CIcontrol = [4.79, 5.39]; t(299) = 7.59, p < .001, d = .98) and typical conditions (Mtypical = 5.73, 95% CItypical = [5.50, 5.95]; t(299) = 10.77, p < .001, d = 1.53). Expense Prediction Bias (Actual – Predicted Expenses). Predicted expenses were significantly lower than actual expenses in all three conditions (p’s < .001), but the mean prediction bias was 31.3% lower in the atypical condition than in the control condition, as can be seen in Figure 4. A (condition: control vs. typical vs. atypical) x 2(expenses: predicted vs. actual) between-within ANOVA confirmed a main effect of condition (F(2, 299) = 3.92, p = .021) and a main effect of expenses (F(1, 299) = 186.43, p < .001), which were qualified by a condition by expenses interaction (F(2, 299) = 3.26, p = .040). Planned contrasts further confirmed that expense predictions in the atypical condition (Matypical = $210.61, CIatypical = [$179.47, $247.15]) were 27.1% ($44.94) higher than in the control condition (Mcontrol = $165.67, 95% CIcontrol = [$138.38, $196.37]; t(299) = 2.01, p = .046), and 47.7% ($68.02) higher than in the typical condition (Mtypical = $142.59, 95% CItypical = [$120.30, $169.02]; t(299) = 3.21, p = .001). Actual expenses did not differ between the control and atypical conditions (Mcontrol = $330.30, 95% CIcontrol = [$273.14, $403.43], Matypical = $323.76, 95% CIatypical = [$267.74, $395.44], t(299) = .14, p = .89), but they were somewhat lower in the typical condition (Mtypical = $252.14, CItypical = [$212.72, $298.87]) than in the control (t(299) = 2.03, p = .043) and atypical conditions (t(299) = 1.91, p = .057). A second measure of prediction accuracy that should be considered in evaluating the effectiveness of our intervention is correlational accuracy, which can be indexed by the correlation between predicted and actual expenses (Epley and Dunning, 2006; Kruger and Evans, 2004; Peetz et al., 2015). An intervention that improves mean prediction accuracy but worsens correlational accuracy is undesirable because it could lead to higher variance in predictions, thus increasing error. Put differently: it is possible for an intervention to improve mean EPB but simultaneously increase error and so make the predictions of many consumers (those in the tails) tangibly worse. Thus, it is notable that our atypical intervention did not sacrifice correlational accuracy (ratypical(100) = .68, rcontrol(96) = .65, z = .36, p = .71). Figure 4Mean Predicted vs. Actual Expenses for Each Condition in Study 3Error Bars Represent 95% Confidence IntervalsMediation Analysis. The results above confirm that our atypical manipulation succeeded in decreasing perceived typicality of future expenses, and that expense predictions were significantly higher in the atypical condition as well. To further investigate the relationship between perceived typicality and expense predictions we tested the same mediation model as in study 2, with condition (atypical = 1 vs. control = 0) as the independent variable, expense prediction as the dependent variable, and perceived typicality as the mediating variable. The indirect effect of condition on expense prediction via perceived typicality was significant (indirect effect =.18, SE = .07, 95% CI = [.06, .34]). Specifically, the model confirms that our atypical condition intervention succeeded in decreasing perceived typicality of future expenses (b = -1.56, 95% CI = [-2.01, -1.12]; t(198) = -6.95, p < .001), and demonstrates that lower perceived typicality is associated with higher expense predictions even while controlling for treatment condition (b = -.12, 95% CI = [-.19, -.04]; t(198) = -3.17, p = .002). A categorical mediation model (Hayes and Preacher, 2014) that includes all three conditions also produces a significant indirect effect (omnibus test of indirect effect =.02, SE = .01, 95% CI = [.00, .004]; indirect effect of atypical dummy = .11, SE = .06, 95% = [.00, .23]; indirect effect of typical dummy = -.05, SE = .03, 95% CI = [-.12, .00]).Robustness Checks. As robustness checks we also examined whether our intervention influenced participants’ confidence in their predictions, and/or the amount of time they spent making their predictions. A one-way ANOVA with condition (control vs. typical vs. atypical) as the independent variable and prediction confidence as the dependent variable revealed a significant effect of condition (F(2, 299) = 17.10, p < .001). Follow-up tests showed that prediction confidence was lower in the atypical condition (Matypical = 4.29, 95% CIatypical = [4.00, 4.59]) than in the control condition (Mcontrol = 4.91, 95% CIcontrol = [4.67, 5.15], p = .003), which was in turn lower than in the typical condition (Mtypical = 5.36, 95% CItypical = [5.13, 5.60], p = .039). Because higher confidence has been linked to lower expense predictions by previous research (Ulkumen et al., 2008), we opted to run a 3(condition: control vs. typical vs. atypical) x 2(expenses: predicted vs. actual) between-within ANOVA with prediction confidence as a covariate to test the veracity of our study 3 EPB results. Notably, the condition by expenses interaction remained significant (F(2, 298) = 4.30, p = .014), as did the planned contrasts demonstrating that expense predictions were significantly higher in the atypical condition vs. the control (p = .044) and typical conditions (p = .002). We also re-ran the categorical mediation analysis with prediction confidence as a covariate and found that the relative indirect effect of condition on expense predictions via perceived typicality remained significant (omnibus indirect effect = -.02, SE = .01, 95% CI = [-.04, .00]; indirect effect of atypical dummy = -.10, SE = .06, 95% = [-.23, .00]; indirect effect of typical dummy = .04, SE = .03, 95% CI = [.00, .11]). Finally, we also ran a categorical mediation analysis with prediction confidence replacing perceived typicality as the mediator, and found that the indirect effect was not significant (indirect effect = .001. SE = .004. 95% CI = [-.006, .009]). Thus, we are comfortable concluding that experimental differences in prediction confidence were not the driving force behind expense predictions in this study.Another alternative explanation for our results is that our atypical manipulation simply makes people think harder about their expenses, and that this, rather than decreased perceived typicality, is what increases prediction accuracy. This is one reason why we included our typical condition – by mirroring the atypical condition in terms of structure but differing in terms of content, we reasoned that the time participants spent predicting their expenses would not differ between these two conditions. This was confirmed by a t-test comparing ln transformed prediction time between these two conditions (t(299) = .86, p = .39). In tandem with the EPB analysis above demonstrating that expense predictions were significantly higher in the atypical vs. typical conditions, this suggests that thinking harder is not sufficient to increase expense predictions; the thought content matters too. Consequential Correlates of EPB. Table 1 summarizes the correlations between EPB and the consequence measures collected in study 3. Higher EPB is associated with: 1) a tendency to allocate less money to savings and more money to debt repayment in the $1,000 allocation decision task, 2) (marginally) lower credit scores and significantly higher frequency of credit card interest charges, and 3) lower monthly saving goals and lower actual savings per month. These results support our hypothesis that EPB is associated with financial decisions and outcomes related to savings and debt (H5). EPB was not correlated with car loan, credit card, money owed to friend and family, or payday loan debt, but the small number of people who reported on these variables (for example, only 12 participants reported having ever used a payday loan) doesn’t provide much power to detect a relationship with EPB. EPB was also not correlated with the subjective measures of financial well-being that we collected, which suggests that if there is a relationship between EPB and these measures, it involves an interaction between EPB and other variables. As we did not a priori hypothesize what those variable might be, we leave the exploration of this possibility for future research. Table 1Correlation between EPB and Study 3 Consequence MeasuresIn sum, the results of study 3 provide support for our hypothesis that consumers under-predict their future (vs. actual) expenses (H3a), that experimentally decreasing perceived typicality of future expenses increases mean expense prediction accuracy (H4), and that EPB (actual – predicted expenses) is correlated with important financial decisions and outcomes related to savings and debt. Study 3 also rules out decreased prediction confidence and increased prediction effort (time) as alternative explanations for the success of our intervention. In study 4 we examine the bias and test the impact of our intervention in a longitudinal field study.Study 4: A Longitudinal Study of EPB in the FieldThe purpose of study 4 was to test our hypotheses in the field. To do so, we partnered with a Canadian bank to run a five week longitudinal field study with 187 of their clients. This study also let us observe the magnitude and persistence of the bias in a context with high external validity across multiple points in time. Finally, study 4 also allowed us to test and validate our intervention in the field. Method Participants. Participants for this study were clients of a midsized Canadian bank (~60,000 clients) recruited through an online panel of clients (~5,000) that the bank maintains in order to conduct market research. We targeted a sample size of 200 based on the effect sizes observed in our previous studies. Each participant completed 6 surveys over the course of 5 weeks as illustrated by each time period marked in Figure 5. Figure 5Data Collection Schedule in Study 4Because we had no prior experience sampling from this population, data collection took place in two waves. In wave 1, we sent a survey at time zero (T0) to 400 randomly selected clients at noon on Sunday, September 10th, 2017. Ninety-three people completed the survey before it was deactivated at 11:59pm on Monday, September 11th, 2017. We then monitored attrition for two weeks before calculating that the T0 survey should be sent to another 800 randomly selected clients (the max number allowed by our field partner) in the second wave of data collection so that we could recruit as close to 200 total participants as possible. 219 clients completed the T0 survey in wave 2. At the end of both waves of data collection we had collected complete data from 187 participants (61 from wave 1 and 126 from wave 2, Mage = 51.12, 57.8% female). Compensation for each participant included a personalized spending report (provided at the end of the study) that served as an incentive to predict and report expenses as accurately as possible. Participants also received a $10 Amazon gift certificate for each completed survey. Procedure. Each of the six surveys in this five week field study were emailed to participants at noon on a Sunday, and required completion before 11:59 pm the next day. The first survey asked participants to predict their expenses for the coming week and month, indicate how similar or different they expected their expenses to be relative to a typical week and month, and indicate how confident they were in their predictions. Participants also completed measures of financial slack (Zauberman and Lynch, 2005), trait openness to experience (John, Donahue, and Kentle, 1991), questions about their spending and borrowing behavior, and demographics. The remaining five surveys began by asking participants to log into their online bank account and report their expenses for the previous week, then predict their expenses for the upcoming week. Both expense reports and predictions were followed immediately by the same measures of typicality and confidence used in survey 1. These follow-up surveys also included measures of short-term financial propensity to plan (Lynch et al., 2010), trait optimism (Scheier, Carver, and Bridges,1994), tightwad-spendthrift tendencies (Rick, Cryder, and Lowenstein, 2008), consideration of future consequences (Strathman et al, 1994), linear vs. cyclical time orientation (adapted from Tam and Dholakia, 2014), temporal discounting (Kirby & Marakovi?, 1996), and risk aversion (adapted from Lauriola, Levin, and Hart, 2007). In the second to last survey half of our sample was randomly assigned to receive our atypical intervention, making the final week of our study a 2(condition: control vs. atypical) x 2(expenses: predicted vs. actual) between-within design.Results and DiscussionPerceived Typicality. To test our hypothesis that people predict their expenses will be more typical in the future than in the past (H1), we compared reported vs. predicted expense typicality at T1, T2, T3, and T4. In other words, we tested whether or not participants predicted their expenses would be more typical in week 2 than week 1, week 3 than week 2, and so on. As illustrated in Figure 6 and Table 2, mean predicted typicality of future expenses was significantly higher than mean reported typicality of past expenses at each stage of the study. Figure 6 also illustrates that our atypical intervention at T4 succeeded in neutralizing (and even slightly reversing) this tendency (t(88) = -1.78, p = .08). In sum, these results provide strong support for H1, and demonstrate that our intervention successfully decreased perceived typicality of future expenses. To test our hypothesis that perceived typicality of future expenses is negatively correlated with expense predictions (H2), we analyzed the correlation between perceived typicality of future expenses and weekly expense predictions for each week of the study, as well as for the month. Perceived typicality of future expenses was negatively correlated with weekly expense predictions at time zero (r(187) = -.30, p < .001), time two (r(187) = -.28, p < .001), and time four (r(187) = -.25, p = .001). The correlation at time one was marginally significant (r(187) = -.12, p = .09) as was the correlation between perceived typicality of monthly expenses and expense predictions for the month (r(187) = -.12, p = .09). The correlation at time three was directionally consistent though not significant (r(187) = -.03, p = .74). In aggregate, these results replicate the descriptive findings of studies 1-3 and provide further support for H2. Figure 6Mean Reported Expense Typicality for the Past Week vs. Mean Predicted Expense Typicality for the Next Week for each Week of Study 4 Error Bars Represent 95% Confidence IntervalsTable 2 T-tests Comparing Reported Past- Week vs. Predicted Next Week Typicality in Study 4Expense Prediction Bias (Recalled – Predicted Expenses). To test our hypothesis that consumers predict their future expenses will be lower than their past expenses (H3a), we compared reported expenses against predicted expenses at T1, T2, T3, and T4. That is, we tested whether participants predicted their expenses would be lower in week 2 than week 1, week 3 than week 2, and so on. As illustrated in Figure 7 and Table 3, mean predicted expenses were significantly lower than mean recalled expenses at each stage of the study until our intervention was deployed. This longitudinally replicates the findings of studies 1 and 2 and provides strong support for H3a.Figure 7Mean Expenses Incurred in the Past Week vs. Mean Predicted Expenses for the Next Week for each Week of Study 4 Error Bars Represent 95% Confidence IntervalsTable 3T-tests Comparing Reported Past-Week vs. Predicted Next-Week Expenses in Study 4Expense Prediction Bias (Actual – Predicted Expenses). As illustrated in Figure 8 and Table 4, mean predicted expenses were significantly lower than mean incurred expenses in each week of the study, except during week 5 in the atypical condition, in which our intervention completely neutralized mean expense prediction bias. A 2(condition: control vs. atypical) x 2(expenses: predicted vs. actual) between-within ANOVA confirmed a significant condition by expenses interaction (F(1, 181) = 5.08, p = .025), and planned contrasts further confirmed that actual expenses did not differ by condition (F(1, 181) = .44, p = .51), but that predicted expenses were 36.7% higher in the atypical condition as compared to the control condition (F(, 181) = 4.48, p = .036). EPB in the control condition was $79.99 (different from zero, t(91) = 3.19, p = .002) vs. –$6.65 (not different from zero, t(90) = -.20, p = .85) in the atypical condition. It is also once again notable that our intervention did not require sacrificing correlational accuracy (rcontrol(90) = .81, rtreatment(89) = .80, z = .19, p = .85). These findings provide support for H3b and H4.Mediation Analysis. To more deeply understand the relationship between perceived typicality of future expenses and expense predictions in the experimental phase of this study (week 5), we tested a mediation model with condition (atypical = 1, control = 0) as the independent variable, expense prediction as the dependent variable, and perceived typicality as the mediator. The indirect effect of condition on expense prediction via perceived typicality was significant (indirect effect = .13, SE = .07, 95% CI = [.03, .29]). Specifically, the model replicated the results of studies 2 and 3 by demonstrating that our intervention succeeded in decreasing perceived typicality of future expenses (b = -1.00, 95% CI = [-1.45, -.56]; t(181) = -4.41, p < .001), and that lower perceived typicality was associated with higher expense predictions, even while controlling for condition (b = -.13, 95% CI = [-.22, -.04]; t(181) = -2.75, p = .007). Notably, these results also held after controlling for prediction confidence (indirect effect = .12, SE = .06, 95% CI = [.02, .28]).Figure 8Mean Predicted Expenses vs. Mean Incurred Expenses for each Week of Study 4 Error Bars Represent 95% Confidence IntervalsTable 4T-tests Comparing Predicted Expenses vs. Incurred Expenses in Study 4.Monthly Expense Prediction Accuracy. We were also able to test the accuracy of participants’ expense predictions for the next month by comparing them to the expenses they actually incurred in weeks 1 through 4 of the study. Results showed that on average predicted expenses for the target month (Mpred = $2276.74, 95% CI = [2031.64, 2551.15]) were $416.77 lower than expenses actually incurred (Mactual = $2693.51, 95% CI = [2376.06, 3053.37]; t(184) = 3.85, p < .001). To put the size of this monthly EPB in perspective, consider that 46% of Americans report they do not have enough money to cover a $400 emergency expense (US Federal Reserve, 2016). To the best of our knowledge this provides the first evidence that consumers under-predict their monthly as well as weekly expenses. Persistence of EPB. To analyze the extent to which weekly EPB persists over time we coded (predictions < actual expenses) as 1 and (predictions >= actual predictions) as 0. We then created an index variable to represent the tendency to under vs. over-predict weekly expenses by summing these binary variables for each week and dividing by 5, the number of weeks for which we have prediction data. On average, participants were prone to under-estimate their expenses more often than not (M = .55, t(187) = 2.88, p = .004). Distributional analyses revealed a median and mode of .60, further suggesting that under-prediction is the norm for a majority of participants. Notably, 28.9% of participants under-predicted their expenses in at least 4 out of 5 weeks. We next analyzed the extent to which expense predictions, actual expenses, and EPB scores (actual – predicted) were correlated week to week. Expense predictions for each week of the study were strongly correlated (r’s > .40, p’s < .001), as were actual expenses (r’s > .45, p’s < .001), but EPB scores were not (p’s > .17). Interestingly, each single shot EPB score (i.e., past – predicted expenses) was strongly correlated with the subsequent “accuracy” EPB score (i.e., actual – prediction). So, predicting lower expenses for the future vs. the past was associated with less accurate predictions (r’s > .45, p’s < .001). Consequential Correlates of EPB. Finally, we analysed the extent to which EPB was associated with financial decisions and outcomes related to savings and debt. To begin, we analysed the week 1 data in isolation to see if the results of the diary study could be directly replicated. In the $1,000 allocation task the results of the field study were consistent with the results of the diary study: EPB (reported expenses at T1 – predicted expenses at T0) was marginally associated with a tendency to allocate less money to savings (r(178) = -.12, p = .096) and more money to debt repayment (r(178) = .13, p = .076). We next analyzed the relationship between monthly EPB (actual – predicted expenses for the month) and the financial outcome variables. Most notably, monthly EPB was correlated with use of the “fair and fast” loans that our field partner offers as a payday loan substitute (r(185) = .15, p = .043). In sum, the results of this field study provide longitudinal evidence with high external validity supporting each of our central hypotheses. The results also suggests that the magnitude of the bias is economically significant. From a practical perspective, it is also highly notable that our intervention is capable of completely neutralizing mean expense prediction bias in a real world setting.Summary of Descriptive ResultsA notable contribution of the present research is that it provides that first comprehensive understanding of the expense prediction bias itself. That is, it is the first research to identify the magnitude, prevalence, and persistence of EPB in non-student samples, in a field study, and over time. Table 6 summarizes the descriptive results from the control conditions in studies 1-4, and several observations stand out. First, the magnitude of the bias varies substantially across studies. Although it is statistically significant in all studies it is notably smaller when measured as the difference between recalled and predicted expenses than when it is measured as the difference between actual and predicted expenses. Of course, each study samples from a different population so we cannot say for sure if these differences are due to measurement method or sample characteristics, but systematically investigating measurement of the bias is a fruitful avenue for future research. The prevalence of the bias is also notable. Across studies the percentage of participants who under-predicted (vs over-predicted) their expenses ranges from just over 50% to 83.7%. This suggests that EPB is a widespread phenomenon, and not just driven by a handful of consumers who are particularly bad at expense prediction. Finally, the week-by-week results from study 4 suggest that the bias persists over time. In sum, these descriptive statistics provide evidence that the magnitude of EPB is meaningful, that it is prevalent, and that it is persistent. In other words, it is a phenomenon of substantive importance.Table 5General DiscussionThe present research develops, tests, and validates a simple cognitive tool that helps consumers make more accurate expense predictions. The tool is based on the theory that expense prediction bias———defined as the underestimation of future expenses———is driven by mental representations of the future that rely on cognitive prototypes. In support of this theory, we demonstrate that consumers predict their weekly expenses will be lower and more typical in the next week as compared to the past week, and that higher perceived typicality of future expenses is associated with lower expense predictions (study 1). We then replicate these findings in a nationally representative sample of Americans, and show that experimentally decreasing perceived typicality of future expenses can increase expense predictions for the next week to the level of recalled expenses for the past week (study 2). Next, we compare consumers’ expense predictions to the expenses they actually incur during a week-long online financial diary study (study 3), and a five week longitudinal field study with clients of a mid-sized North American bank (study 4). We find that consumers significantly under-predict both their weekly and monthly expenses. Furthermore, we find that our cognitive tool – listing three reasons why expenses for the next week will be different from a typical week – successfully decreases perceived typicality of future expenses, and increases mean expense prediction accuracy without sacrificing correlational accuracy. Underlining the substantive importance of this cognitive tool, studies 3 and 4 also provide evidence that the expense prediction bias is correlated with lower savings and more frequent use of high interest debt. Theoretical Contributions This work makes theoretical contributions to three areas of research. First and foremost, we contribute to the literature on consumer (mis)prediction by developing and testing a prototype theory of expense prediction bias. This is important in an immediate sense because it helps explain why this bias occurs and how it can be neutralized. In a broader sense, the prototype theory has the potential to unify currently disparate findings in the expense misprediction literature. For example, it is reasonable to hypothesize that consumers’ overconfidence in their monthly budget estimates (Ulkumen et al., 2008) can be explained, at least in part, by prototypical mental representations of the future, because prototypical expenses seem to be those that are most certain. Likewise, under-prediction of exceptional expenses (Sussman and Alter, 2012) may be explained in part by consumers over-relying on cognitive prototypes that do not include atypical outcomes. Taking an even broader view, our theory is compatible with findings in the planning fallacy and affective forecasting literatures, and therefore may present a single parsimonious explanation for consumer mis-prediction across multiple domains. Testing these possibilities presents many exciting avenues for future research. The present research also contributes to the literature on expense misprediction by providing the first comprehensive understanding of the expense prediction bias itself. That is, these are the first studies to identify the magnitude, prevalence, and persistence of the bias in non-student samples. This work is also the first to study the bias longitudinally and in the field, to measure monthly expense predictions against actually expenses for the target month, and to investigate consequential correlates of the bias. This approach allows us to claim with some certainty that EPB is meaningful in a very real-world sense, and is therefore an important context in which scholars should conduct future research. The present research also contributes to a nascent literature on temporal asymmetry which hypothesizes that people mentally represent the future in more prototypical terms than the past (Kane, Van Boven, and McGraw, 2012; Van Boven, Kane, and McGraw, 2009; Williams and Laboeuf, under review). By comparing perceived typicality of past vs. future expenses we extend this work to the domain of money, which provides a notably conservative test of the proposition that people mentally represent the future in more prototypical terms than the past. This is because money is relatively concrete and predictable (Zauberman and Lynch, 2005), whereas the hypothetical people, places, and events that participants have been asked to mentally represent in past studies are arguably much more ambiguous. Therefore, because prototypes are generalizations, it is reasonable to believe that people will rely on prototypes less when they are mentally representing the uses of a resource like money, which has very specific uses.Finally, by demonstrating that consumers rely on prototypes when predicting their future expenses, and that this leads to the systematic under-prediction of future expenses, we contribute to the prototype heuristic literature. Past work in this area has demonstrated that people use prototypes when making intuitive judgments (for a review, see Kahneman, 2003). However, to the best of our knowledge, we are the first to show that people rely on prototypes when predicting the future, and that this can lead to systematically biased predictions. Our intervention also provides valuable insight into the content of prototypes that consumers use when prospecting. Consider for example that there are many reasons to believe that the future will contain atypically low expenses, as when consumers encounter sales, discounts, and promotions. However, despite this very real possibility, participants in our studies consistently increased their expense predictions when prompted to consider atypical expenses, suggesting that the content of their prototype was inaccurately optimistic rather than pessimistic. Practical Implications for Consumers and ManagersIn addition to the theoretical contributions outlined above, our work has clear practical implications for consumers and the firms that serve them. First and foremost, any consumer can easily use the tool we have developed to improve their expense predictions, the importance of which is underlined by our results showing that under-prediction is associated with undesirable financial outcomes. Second, firms with a mandate to help consumers improve their financial outcomes, such as those developing budgeting apps in the FinTech sector, can also use the results we have presented here to improve their products’ prediction functions. In fact, testing visual displays and textual cues that decrease perceived typicality and therefore improve prediction accuracy is a promising avenue for future research. Finally, because our theory is compatible with past research on misprediction in multiple contexts, we believe that our intervention can also be used to inform the design of apps that aim to improve consumers’ predictions with respect to calories, exercise, time management, and a host of other variables that can positively impact consumers’ well-being. ConclusionThe results of our investigation into the causes and consequences of the expense prediction bias led to the development of a simple cognitive tool that can be used to help consumer make more accurate expense predictions, the importance of which is underlined by the fact that expense prediction bias is associated with negative financial outcomes. We believe that the theoretical and practical implications of this work extend beyond the domain of expense prediction, opening many interesting avenues for future research. It is our belief that the prototype theory of misprediction is not only capable of advancing our understanding of misprediction, it is also capable of concretely improving consumer well-being. References Addis, D. R., Wong, A. T., & Schacter, D. L. (2007). Remembering the past and imagining the future: Common and distinct neural substrates during event construction and elaboration. Neuropsychologia, 45, 1363–1377.Alter, A. L., & Oppenheimer, D. M. (2009). Uniting the tribes of fluency to form a metacognitive nation.?Personality and social psychology review,?13(3), 219-235.Bland, J. M., & Altman, D. G. (1996). Transformations, means, and confidence intervals.?BMJ: British Medical Journal,?312(7038), 1079.Buehler, R., Griffin, D., & MacDonald, H. (1997). The role of motivated reasoning in optimistic time predictions. Personality and Social Psychology Bulletin, 23, 238–247.Buehler, R., Griffin, D., & Peetz, J. (2010). The planning fallacy: Cognitive, motivational, and social origins. In?Advances in experimental social psychology?(Vol. 43, pp. 1-62). Academic Press.Buehler, R., Griffin, D., & Ross, M. (1994). Exploring the “planning fallacy”: Why people underestimate their task completion times. Journal of Personality and Social Psychology, 67, 366–381.CB Insights. (2018). Global Fintech Report Q1 2018. , S. J., & Wyer, R. S. (2002). The role of prototypes in the mental representation of temporally related events. Cognitive Psychology, 44, 67–103.Consumer Financial Protection Bureau (2017). CFPB Financial Well-Being Scale: Scale Development Technical Report. Accessed April 14, 2018. ’Argembeau, A., & Van der Linden, M. (2004). Phenomenal characteristics associated with projecting oneself back into the past and forward into the future: Influence of valence and temporal distance. Consciousness and Cognition, 13, 844–858.Dunning, D. (2007). Prediction: The inside view. In A. W. Kruglanski & E. T. Higgins (Eds.),Social psychology (pp. 69–90). New York, NY: Guilford Press.Epley, N., & Dunning, D. (2006). The mixed blessings of self-knowledge in behavioral prediction: Enhanced discrimination but exacerbated bias. Personality and Social Psychology Bulletin, 32(5),641–655. Federal Reserve Bank of New York. (2018). Household Debt Continues Its Increase in the First Quarter of 2018. , M., & Willemin, K. (2013). The Retirement Breach in Defined Contribution Plans.?HelloWallet Research Reports.Fiske, S. T., & Taylor, S. E. (1991). Social cognition. New York, NY: Random House.Gilbert, D. T., Morewedge, C. K., Risen, J. L., & Wilson, T. D. (2004). Looking forward to looking backward: The misprediction of regret.?Psychological Science,?15(5), 346-350.Gillespie, P. (2017). Intuit: Gig economy is 34% of US workforce.?CNN Money. John, O. P., Donahue, E. M., & Kentle, R. L. (1991). Hayes, A.F., & Preacher, K.J. (2014). Statistical mediation analysis with a multicategorical independent variable. British Journal of Mathematical and Statistical Psychology, 67, 451-470.John, O.P., Donahue, E.M., & Kentle, R.L. (1991) The Big Five Inventory - Versions 4a and 54. Berkeley, CA: University of California, Berkeley, Institute of Personality and Social Research. Johnson, M. K., & Raye, C. L. (1981). Reality monitoring. Psychological Review, 88(1), 67–85.Kahneman, D. (2003). A perspective on judgment and choice: mapping bounded rationality.?American psychologist,?58(9), 697.Kahneman, D., & Miller, D. T. (1986). Norm theory: Comparing reality to its alternatives.?Psychological review,?93(2), 136.Kahneman, D., & Tversky, A. (1979). Intuitive prediction: Biases and corrective procedures. TIMS Studies in Management Science, 12, 313-327.Kane, J., Van Boven, L., & McGraw, A. P. (2012). Prototypical prospection: future events are more prototypically represented and simulated than past events.?European Journal of Social Psychology,?42(3), 354-362.Kirby, K. N., & Marakovic, N. N. (1996). Delay-discounting probabilistic rewards: Rates decrease as amounts increase. Psychonomic Bulletin and Review, 3(1), 100-104.Kruger, J., & Evans, M. (2004). If you don’t want to be late, enumerate: Unpacking reduces the planning fallacy. Journal of Experimental Social Psychology, 40, 586–598.Kunda, Z. (1999). Social cognition: Making sense of people. Cambridge, MA: MIT Press.Lauriola, M., Levin, I. P., & Hart, S. S. (2007). Common and distinct factors in decision making under ambiguity and risk: A psychometric study of individual differences. Organizational Behavior and Human Decision Processes, 104(2), 130-149.external image clear.gifdoi:10.1016/j.obhdp.2007.04.001Lynch Jr, J. G., Netemeyer, R. G., Spiller, S. A., & Zammit, A. (2010). A generalizable scale of propensity to plan: the long and the short of planning for time and for money.?Journal of Consumer Research,?37(1), 108-128.Peetz, J., & Buehler, R. (2009). Is there a budget fallacy? The role of savings goals in the prediction of personal spending.?Personality and Social Psychology Bulletin,?35(12), 1579-1591.Peetz, J., & Buehler, R. (2013). Different goals, different predictions: Accuracy and bias in financial planning for events and time periods.?Journal of Applied Social Psychology,?43(5), 1079-1088.Peetz, J., Buehler, R., Koehler, D. J., & Moher, E. (2015). Bigger not better: Unpacking future expenses inflates spending predictions.?Basic and Applied Social Psychology,?37(1), 19-30.Pew Charitable Trusts. (2012). Payday Lending in America: Who Borrows, Where They Borrow, and Why.?Safe Small-Dollar Loans Research Project.Rick, S. I., Cryder, C. E., & Loewenstein, G. (2007). Tightwads and spendthrifts.?Journal of Consumer Research,?34(6), 767-782.Roese, N. J. (1997). Counterfactual thinking. Psychological Bulletin, 121, 133–148.Scheier, M. F., Carver, C. S., & Bridges, M. W. (1994). Distinguishing optimism from neuroticism (and trait anxiety, self-mastery, and self-esteem): a reevaluation of the Life Orientation Test.?Journal of personality and social psychology,?67(6), 1063.Schkade, D. A., & Kahneman, D. (1998). Does living in California make people happy? A focusing illusion in judgments of life satisfaction. Psychological Science, 9, 340–346.Schwartz, L. M., Woloshin, S., Black, W. C., & Welch, H. G. (1997). The role of numeracy in understanding the benefit of screening mammography. Annals of Internal Medicine, 127(11), 966-972.Smith, E. R. (1998). Mental representation and memory. In D. T. Gilbert, S. T. Fiske, & G. Lindzey (Eds.), Handbook of social psychology (4th ed., Vol. 1, pp. 391–445). New York, NY: McGraw-Hill.Sussman, A. B., & Alter, A. L. (2012). The exception is the rule: Underestimating and overspending on exceptional expenses.?Journal of Consumer Research,?39(4), 800-814.Tam, L., & Dholakia, U. (2014). Saving in cycles: How to get people to save more money.?Psychological science,?25(2), 531-537.Thomas, M., Desai, K. K., & Seenivasan, S. (2010). How credit card payments increase unhealthy food purchases: Visceral regulation of vices. Journal of consumer research, 38(1), 126-139.?lkümen, G., Thomas, M., & Morwitz, V. G. (2008). Will I spend more in 12 months or a year? The effect of ease of estimation and confidence on budget estimates.?Journal of Consumer Research,?35(2), 245-256.US Federal Reserve. (2016). Report on the Economic Well-Being of US Households in 2015. Boven, L., & Ashworth, L. (2007). Looking forward, looking back: Anticipation is more evocative than retrospection. Journal of Experimental Psychology. General, 136, 289–300.Van Boven, L., Kane, J., & McGraw, A. P. (2008). Temporally asymmetric constraints on mental simulation: Retrospection is more constrained than prospection. In K. Markman, W. Klein, & J. Suhr (Eds.), The handbook of imagination and mental simulation (pp. 131–149). New York, NY: Psychology Press.Williams, E. F., & LeBoeuf, R. A. (under review). Starting Your Diet Tomorrow: Consumers Believe They Will Have More Control Over the Future than They Did Over the Past.Wilson, T. D., Wheatley, T., Meyers, J. M., Gilbert, D. T., & Axsom, D. (2000). Focalism: A source of the durability bias in affective forecasting. Journal of Personality and Social Psychology, 78, 821–836.Winkielman, P., Halberstadt, J., Fazendeiro, T., & Catty, S. (2006). Prototypes are attractive because they are easy on the mind.?Psychological science,?17(9), 799-806.Yang, S., Markoczy, L., & Qi, M. (2007). Unrealistic optimism in consumer credit card adoption.?Journal of Economic Psychology,?28(2), 170-185.Zauberman, G., & Lynch Jr, J. G. (2005). Resource slack and propensity to discount delayed investments of time versus money.?Journal of Experimental Psychology: General,?134(1), 23. ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download