Journal of the Academy of Marketing Science

, Volume 41, Issue 4, pp 436–455

Managerial decision making in customer management: adaptive, fast and frugal?

Authors

    • Institute of Retail ManagementUniversity of St.Gallen
  • Philipp Schmitt
    • School of Business and EconomicsGoethe University Frankfurt
  • Vicki G. Morwitz
    • Stern School of BusinessNew York University
  • Russell S. Winer
    • Stern School of BusinessNew York University
Original Empirical Research

DOI: 10.1007/s11747-012-0320-7

Cite this article as:
Bauer, J.C., Schmitt, P., Morwitz, V.G. et al. J. of the Acad. Mark. Sci. (2013) 41: 436. doi:10.1007/s11747-012-0320-7

Abstract

While customer management has become a top priority for practitioners and academics, little is known about how managers actually make customer management decisions. Our study addresses this gap and uses the adaptive decision maker as well as the fast and frugal heuristics frameworks to gain a better understanding of managerial decision making. Using the process-tracing tool MouselabWEB, we presented sales managers in retail banking with three typical customer management prediction tasks. The results show that a majority of managers in this study are adaptive in their decision making and that some managers use fast and frugal heuristics. Usage of adaptive decision making seems to be mainly driven by low objective task difficulty, the use of fast and frugal heuristics by experience. While adaptive decision making does not impact predictive accuracy, usage of fast and frugal heuristics is associated with proportionally greater use of high predictive quality cues and a significant increase in accuracy. Hence, the existing skepticism concerning heuristics should be questioned.

Keywords

Customer managementAdaptive decision makingFast and frugal heuristicsProcess-tracingMouselab

Introduction

In the past decades, marketing academia and practice have shifted their focus from a product- to a customer-centric view (Rust et al. 2004). Rather than maximizing brand equity, many firms have made customer management a top priority and aim to improve their performance by building long term, profitable customer relationships (e.g., Bolton 1998; Hanssens et al. 2008; Rust et al. 2010; Verhoef et al. 2010). The basic premise of this trend is that customers are one of the firm’s most important assets (Gupta et al. 2004). To assess the economic value of this asset, two key metrics are commonly used: customer lifetime value (CLV; the net present value of all earnings from a specific customer during the time of her/his relationship with the company minus the costs of attracting, retaining, and servicing this client) and customer equity (CE; the sum of all individual lifetime values generated by the company’s current and prospective customers) (e.g., Berger and Nasr 1998; Blattberg and Deighton 1996; Rust et al. 2004). To increase CLV and maximize CE, many firms strategically focus on customer acquisition, retention, and cross-selling (Blattberg et al. 2001). These activities are considered to be the main drivers of CLV (Gupta and Zeithaml 2006).

To that end, considerable research focuses on developing models to help firms make decisions about marketing resource allocation, customer segmentation, and customer selection (Kumar et al. 2006; for a review, see Jain and Singh 2002). However, despite extensive research linking firms’ marketing actions to CLV and CE, little is known about the decision making processes of individual managers when they are faced with important questions concerning how to acquire and retain customers and how to best cross-sell. This is surprising as managers today have access to extensive databases and are constantly challenged to process substantial amounts of customer information in order to make accurate decisions that increase CLV. Wierenga et al. (1999) rightfully point out that the marketing literature offers countless recommendations concerning how managerial decisions should be made but, regrettably, less insight into how managerial decisions actually are made. While there is a general shortage of studies investigating managers’ decision making processes, this gap is particularly profound in the area of customer management (Van Bruggen and Wierenga 2010).

Therefore, the goal of our research is to gain a better understanding of managerial decision making in customer management by investigating the decision making process, the factors impacting it, and the factors impacting the predictive accuracy of managers’ decisions. In order to do so, we employ two research frameworks which build on Simon’s (1955) concept of bounded rationality: (1) the adaptive decision maker framework (Payne et al. 1993) and (2) the fast and frugal heuristics framework (Gigerenzer and Goldstein 1996). The former states that decision makers are flexible and that their selection of decision strategies is contingent on a variety of factors; the latter asserts that people use simplifying rules to come up with smart decisions quickly (fast) and with a limited amount of information (frugal). We empirically test these two frameworks in the context of customer management decisions. Specifically, we try to answer the following questions with regard to managers’ decisions concerning customer acquisition, retention, and cross-selling: (1) Are managers adaptive in their decision making and/or do they use fast and frugal heuristics? (2) Which characteristics of the task and the decision maker drive the use of adaptive decision making and of fast and frugal heuristics? (3) Which characteristics of the task and the decision maker impact the predictive accuracy of managers’ decisions?

To answer these questions, we used a process-tracing approach and investigated the decision making behavior of 49 sales managers of a leading German bank by employing MouselabWEB (MouselabWEB 2012). Managers participated in three different tasks, one of each related to acquisition, retention, and cross-selling. In each task, they were provided with several pieces of customer information for a set of real bank clients and asked to select those customers they predicted would display a certain behavior—(1) obtain a new consumer loan in the cross-selling task, (2) cancel their checking account in the retention task, and (3) refer a new customer to the bank in the acquisition task—within a specified timeframe. Customer information was hidden on the computer screen and had to be requested by clicking on the respective box. This procedure allowed us to monitor (1) what information managers used, (2) in which order they accessed the information, and (3) how long they took to reach a decision. We then examined managers’ decision making processes and assessed their decision accuracy by comparing their predictions to the actual behaviors of the customers.

Our research makes several managerial and theoretic contributions. First, by explaining which factors drive good decisions in customer management, our results can improve managers’ decision making, which ultimately increases the lifetime value of the firm’s customers and thus customer equity. Second, combining a process-tracing method with a prediction task not only offers a rich, descriptive account of decision making in customer management but also allows us to simultaneously assess decision makers’ accuracy. Third, our study is the first to empirically test the adaptive decision maker framework as well as the fast and frugal heuristics framework in the context of customer management, an area of high strategic importance for many firms. In sum, by providing comprehensive insights into the way managers actually make decisions and the factors that determine decision quality, this research contributes not only to the academic literature on customer management but also to decision making research in the realm of managerial cognition. From a theoretical perspective, our study adds to the understanding of the relationship between people’s usage of fast and frugal heuristics and their decision accuracy. Due to inconsistent results in the judgment and decision making literature, the latter aspect is, to date, a subject of controversial debate.

The remainder of this article is organized as follows: We first provide an overview of the existing literature on managerial decision making in the fields of managerial cognition and customer management. Then, we provide the theoretical background by introducing the concepts of bounded rationality, adaptive decision making, and fast and frugal heuristics. A description of the method and the results of our study follow. We conclude by discussing managerial implications, limitations of our study, and directions for further research.

Literature review

Managerial cognition and strategic decision making

To arrive at the optimal solution for a decision problem, the managerial decision making literature proposes the following six steps: (1) definition of the problem, (2) identification of the criteria, (3) weighting of the criteria, (4) generation of alternatives, (5) rating of each alternative on each criterion, and (6) computation of the optimal decision (Bazerman 1998). In their daily decision making, however, managers are faced with a variety of complex problems. Scarcity of time (e.g., Mintzberg 1973) and cognitive limitations (e.g., Miller 1956) make it difficult for managers to process all available information and to make fully rational decisions concerning all possible alternatives (Simon 1955, 1997). Instead, decision makers typically depart from rational decision making and use multiple strategies which depend on the type of the problem (e.g., Anderson 1983; Hickson et al. 1986; Nutt 1984). Rather than optimizing decision outcomes, managers make trade-offs between decision accuracy and costs (i.e., time and cognitive effort) and choose strategies that lead to satisfactory (versus optimal) outcomes (for a review, see Eisenhardt and Zbaracki 1992; Walsh 1995).

Several studies have also examined how experience affects managerial decision making. For example, Day and Lord (1992) show that experts (38 CEOs) categorize ill-structured problems faster than novices (30 MBA students). They explain that experienced managers expedite decisions by using well-developed heuristics (“rules of thumb”) in the early stages of the decision making process. Similarly, Isenberg (1986) finds that experienced managers request less information and act sooner than undergraduate students. Other researchers have found that experienced decision makers use both reasoning and intuition simultaneously (Dane and Pratt 2007; Fredrickson 1985) which, in turn, leads to faster decision making (Wally and Baum 1994). One reason for this might be that experts have richer knowledge structures than novices (e.g., Fiske et al. 1983; Lurigio and Carroll 1985; Wagner 1987). However, even though they are capable of making faster decisions, experienced managers also tend to be prone to systematic biases resulting from overconfidence (e.g., Einhorn and Hogarth 1978; Mahajan 1992). The literature on managerial cognition and strategic decision making provides considerable evidence that overconfidence not only is a common phenomenon among managers but also leads to lower accuracy (e.g., Neale and Bazerman 1985; Russo and Schoemaker 1992; Schwenk 1986). Overconfident managers overestimate their chances of success when making decisions about market entries (Camerer and Lovallo 1999), product introductions (Simon and Houghton 2003), and corporate investments (Malmendier and Tate 2005). Thus, more experience results in faster decisions but not necessarily in higher decision accuracy.

The literature on managerial cognition and strategic decision making has shown that managers use various decision strategies and accelerate decision making processes by relying on their experience and intuition. However, there is no empirical study which provides a comprehensive account of the task- and manager-related factors (and their interactions) that drive adaptive or fast and frugal decision making and their impact on decision accuracy.

Decision making in customer management

Researchers in customer management have developed numerous mathematical models to support managerial decision making (for an overview see Reinartz and Venkatesan 2008). Specifically, models have been developed to calculate CLV and CE (e.g., Berger and Nasr 1998; Dwyer 1997), to indentify the antecedents of CLV (e.g., Reinartz and Kumar 2003; Rust et al. 2004), to determine the optimal balance between customer acquisition and retention spending (e.g., Berger and Nasr Bechwati 2001; Blattberg and Deighton 1996; Reinartz et al. 2005), and to provide insights into a firm’s customer base (e.g., Reinartz and Kumar 2000; Schmittlein and Peterson 1994; Schmittlein et al. 1987). However, there is a shortage of descriptive studies that investigate how individual managers make customer management decisions. A notable exception is the work of Wübben and Wangenheim (2008), who show in a customer management context that if one were to apply certain heuristics (i.e., the hiatus and persistence heuristics), the decisions would be at least as accurate as results from the Pareto/NBD and BG/NBD statistical models. Yet they do not investigate if managers actually use these heuristics when making decisions. In the area of customer acquisition, Morwitz and Schmittlein (1998) show in a direct marketing context that an interplay of managerial judgments and statistical models lead to more accurate decisions than does managerial judgment alone and thus ultimately lead to increased profits. However, Morwitz and Schmittlein (1998) do not examine managers’ decision making processes. In a lab experiment, Hoch and Schkade (1996) come to a similar conclusion. Accordingly, they recommend providing managers with decision support systems.

Existing studies show the effectiveness of different decision strategies but essentially treat the process leading to a decision as a black box. To address this gap, we investigate individual decision making behavior and use a process-tracing method which allows us to examine the intervening steps between the informational input and the decision output.

Theoretical background

Bounded rationality

The concept of bounded rationality was first introduced in a seminal article by Herbert A. Simon (1955). It questions the unrealistic picture of human decision making that assumes full rationality and complete knowledge. Instead, Simon (1955) stressed that (1) a focus on how decisions are made is needed (instead of only looking at what decisions are made), (2) a more realistic view of the decision maker has to take into account their limited computational capacity (instead of assuming global rationality), and (3) understanding decision processes has to account for the interaction of cognitive limitations and the environment in which the decision is made (i.e., a variety of task- and context-related factors). This implies that any analysis of human decision making must be concerned not only with the content of decisions (substantive rationality) but also with the decision making process (procedural rationality) as Simon (1981) notes. Such an approach is needed to achieve the core goal of the concept of bounded rationality: a more realistic view of human decision making.

Adaptive decision making

A well-known framework that builds on Simon’s ideas of bounded rationality is adaptive decision making. It shows that decision makers use different decision strategies contingent on a variety of task, context, and individual difference factors (Payne et al. 1993), and that effort-accuracy considerations play a crucial role in deciding how to decide. In line with Simon (1955, 1990), adaptive decision making assumes that decision makers are highly flexible and adjust to their task environment. Not only do different decision makers use different strategies but the same individual will also use different decision strategies for different problems.

Payne (1976) argues that human decision strategies mainly differ along two dimensions: (1) the amount of information searched as either constant or variable across alternatives and (2) the sequence of information searched as either inter-dimensional (i.e., successively evaluating alternatives on the basis of different evaluation criteria) or intra-dimensional (i.e., comparing different alternatives on the basis of one evaluation criterion). He explains that each combination of these two information search dimensions characterizes a particular decision strategy that people use to evaluate a set of alternatives before making a final choice. Figure 1 illustrates Payne’s (1976) two-dimensional classification of decision strategies.
https://static-content.springer.com/image/art%3A10.1007%2Fs11747-012-0320-7/MediaObjects/11747_2012_320_Fig1_HTML.gif
Fig. 1

Two-dimensional classification of decision strategies

The four resulting decision strategies are characterized as follows:
  1. (1)

    The linear model is characterized by a constant and inter-dimensional information search: each alternative is evaluated separately by additively combining the values of all evaluation criteria into an overall value.

     
  2. (2)

    The conjunctive model is characterized by a variable and inter-dimensional information search: the first alternative that exceeds certain cutoff values on all evaluation criteria will be chosen.

     
  3. (3)

    The additive difference model is characterized by a constant and intra-dimensional information search: differences between alternatives for selected evaluation criteria are summed together in order to make a choice (Tversky 1969).

     
  4. (4)

    The elimination-by-aspects model is characterized by a variable and intra-dimensional information search: first the most important evaluation criterion is determined and then all alternatives that do not meet a certain cutoff value for this criterion are eliminated. The process continues with the second most important evaluation criterion, then the third, and so on (Tversky 1972).

     

Past research has shown that people are adaptive in their decision making for a variety of tasks, such as choosing an apartment (Sundström 1987) or loan candidates (Biggs et al. 1985). The following task and context factors have been found to determine people’s selection of decision strategy: the number of alternatives in the choice set (e.g., Johnson and Meyer 1984), the number of attributes (e.g., Sundström 1987), time pressure (e.g., Payne et al. 1996), and the similarity of alternatives (e.g., Bettman et al. 1993). Studies investigating the relationships between decision strategies, cognitive effort, and decision accuracy have shown that (1) under certain circumstances (e.g., time constraints), heuristic decision strategies that involve highly selective information processing (e.g., the elimination-by-aspects model) can be as accurate as more deliberate decision strategies (e.g., Johnson and Payne 1985; Payne et al. 1988, 1996; Rieskamp and Hoffrage 2008), and (2) people will shift decision strategies in an adaptive way to achieve reasonable performance in the two meta-goals of decision accuracy and cognitive effort (e.g., Bettman et al. 1998; Creyer et al. 1990; Payne and Bettman 2004).

Fast and frugal heuristics

Another framework building on the concept of bounded rationality is the research program of fast and frugal heuristics (Gigerenzer and Goldstein 1996). In general, a heuristic is a simple decision rule. However, not every rule is a heuristic. Contrary to heuristics, rules can also be complex, such as computational algorithms. According to Gigerenzer (2004), heuristics have three specific qualities: (1) Heuristics exploit evolved or learned capacities and are simple in relation to these capacities. (2) Heuristics exploit structures of the environment. A heuristic is not good or bad per se; it is only good or bad in a particular environment. (3) Heuristics are distinct from “as-if” optimization models. The latter do not give information about the actual process, while a heuristic is a problem-solving rule whose purpose is to describe both process and outcome. These qualities are incorporated in the concept of fast and frugal heuristics, which are defined as employing “a minimum of time, knowledge, and computation to make adaptive choices in real environments” (Gigerenzer et al. 1999, p. 14). Fast and frugal decision making deliberately ignores information through its embrace of stopping rules (Gigerenzer et al. 1999).

The idea that people make judgments and decisions in a fast and frugal way has become increasingly popular in recent years in part because the underlying heuristics are both elegant and straightforward (Oppenheimer 2003). Fast and frugal heuristics can be used in a vast variety of situations, ranging from classifying incoming heart attack patients (Breiman et al. 1993) to assessing the authenticity of antique statues (Hoving 1996). When making a decision, people are assumed to select a decision strategy from a repertoire of fast and frugal heuristics, the “adaptive toolbox” (Gigerenzer and Selten 2001). Past research has identified several of these relatively simple heuristics and investigated their use and accuracy (for a review, see Gigerenzer and Gaissmaier 2011). However, the evidence is mixed. On the one hand, there is evidence that these simple heuristics work even better than more deliberate decision strategies despite requiring less information and computation (e.g., Bröder 2000, 2003; Bröder and Schiffer 2003; Dhami and Ayton 2001; Gigerenzer and Goldstein 1996; Gigerenzer et al. 1999). On the other hand, other research questions the fast and frugal approach by showing that most people’s behavior is inconsistent with its theoretical assumptions (e.g., Newell et al. 2003; Newell and Shanks 2003) and that the decision accuracy of these simplifying heuristics is worse than predictions made by chance (e.g., Oppenheimer 2003). By investigating the relationship between people’s usage of fast and frugal heuristics and their decision accuracy, our research will add to our understanding of heuristic decision making.

Method

To empirically test the adaptive decision making and the fast and frugal heuristics frameworks in the context of customer management, we needed a sample of practicing managers and a method that allows us to capture their decision making processes when being faced with customer management decisions. We chose to run a MouselabWEB study with sales managers working in the retail banking industry. This section provides details on the participants of our study, the data collection using MouselabWEB, and the design of the experiment.

Participants

All participants were sales managers of the retail banking division of a German bank. Hence, all sampled managers were from one functional area, namely the retail bank’s sales department. In their daily work, sales managers in retail banking are primarily concerned with customer management activities, such as client care and advisory on a range of financial products. Thus, the participants of this study routinely make customer management decisions similar to those used in the experiment. The bank itself has been a leading player in the German banking market for more than 130 years. It is one of the larger European banks and offers all major banking services. In recent years, customer management has become a critical success factor for the bank and the focus of top management. Given this, the senior management of the bank was interested and supportive of our research and allowed us to select participants from the bank’s entire pool of sales managers. In order to limit external factors (different focus areas, regional differences, etc.), we decided to focus on the region that was most representative of the bank’s overall population of sales managers and clients. From this region, 70 managers were randomly selected from the firm’s internal database and contacted by the office of their regional sales director. They were told that participation in the study was voluntary but would help to improve the sales management process of the bank. Of 70 managers initially contacted, 49 mangers agreed to take the study (i.e., 70% participation rate).

Participants are nearly equally distributed among three different age groups: 21–30 years (34.7%), 31–40 years (36.7%), and 41–50 years (28.6%). Twenty-eight participants were female (57.1%); twenty-one were male (42.9%). Participants varied in their years of work experience: less than 1 year (4.1%), 1 to 3 years (4.1%), 4 to 6 years (18.8%), 7 to 9 years (14.6%), 10 or more years (58.3%). While a majority of participants (58.3%) have 10 or more years of work experience in the sales area, there is enough variation for us to examine the role of expertise in the subsequent analyses. While confidentiality agreements preclude us from reporting actual numbers, senior management confirmed that our sample is representative of the firm’s overall population of sales managers. Specifically, there were no significant differences between participants and non-participants regarding age, gender, and sales experience.

Data collection using process-tracing

Managers were directed via email to a self-programmed website using the MouselabWEB technology. After a short explanation about the background of the study and their role, they were presented with three tasks, which appeared in the same order for all participants. Each task was first briefly described and then participants were shown a matrix such as the one in Fig. 2.
https://static-content.springer.com/image/art%3A10.1007%2Fs11747-012-0320-7/MediaObjects/11747_2012_320_Fig2_HTML.gif
Fig. 2

MouselabWEB layout of the cross-selling task (low complexity condition)

The matrix contained information about a set of customers. All information was initially hidden behind the boxes. To reveal a piece of information, participants had to click on the respective box. Once a box was opened, it remained open for the duration of that task. During the time managers worked on each of the three tasks, MouselabWEB recorded (1) the number of boxes participants opened to solve each task (i.e., the amount of information accessed per task), (2) the order in which the boxes were opened (i.e., the sequence of accessed information), (3) the time participants spent on each box they opened, (4) the time participants took to categorize each customer, (5) the total time to complete the task, and (6) the chosen options. Thus, MouselabWEB traced managers’ information search and problem solving strategies without interfering with the decision making processes.

After completing each task, managers responded to several questions on five-point semantic differential scales about how realistic (task authenticity: 1 = not realistic at all/5 = very realistic) and difficult they perceived the task (self-reported task difficulty: 1 = very easy/5 = very difficult). Managers next indicated how often they make decisions similar to those in this study (task familiarity: 1 = daily; 2 = weekly; 3 = monthly; 4 = yearly; 5 = less than once per year). Then managers estimated the percentage of customers that they classified correctly (estimated hit rate in %). After having completed all three tasks, managers reported their age (coded 1 = 21–30 years; 2 = 31–40 years; 3 = 41–50 years; 4 = 51 years and above), gender (coded 0 = female; 1 = male), and sales experience (coded 1 = less than 1 year; 2 = one to three years; 3 = four to six years; 4 = seven to nine years; 5 = 10 or more years).

Manipulation

In order to investigate context-related effects on managers’ decision making processes, we manipulated task complexity. Managers were randomly assigned to either a low (n = 25) or high complexity (n = 24) condition. In the low (high) complexity condition, managers were shown seven pieces of information for 10 customers (10 pieces of information for 16 customers) per task. The number of metrics and customers were chosen based on discussions with the bank’s management to reflect realistic tasks at different levels of complexity. Furthermore, our experimental manipulation is in line with Miller’s (1956) findings which show that seven pieces of information are a natural limit for the working memory capacity in human information processing. Managers in both the low and high complexity conditions saw the same seven customer metrics. These metrics were selected by the bank as ones with high predictive quality. The three additional customer metrics shown only to those managers in the high complexity group were ones with low predictive quality. Therefore, the low and high complexity group differed only in the number of customer metrics provided, not in the overall predictive quality of the available information. Based on this information, participants in the low (high) complexity condition had to select those five (eight) customers they expected to display a certain behavior within a specified time frame. In our analyses, we account for the task complexity manipulation with a dummy variable (coded 0 = low complexity; 1 = high complexity).

Customer sample

The bank’s retail customers are all in the consumer market. In the experiment, managers had to make decisions concerning the bank’s real customers, and their corresponding customer information was retrieved from the bank’s internal database. For each task, a pool of customers was selected using the following procedure: In the first step, a small subset of customers was randomly selected from the retail bank’s overall population of customers. The bank double-checked the selection and confirmed that these customers were representative of the overall consumer client base. In the second step, 10 (or 16) customers were chosen from the random draw in the first step; however, selected customers had to fulfill the following two requirements: (1) they had to show variation on the two customer metrics with the highest predictive quality for the respective task, and (2) for each task, half of the customers (five in the low complexity condition and eight in the high complexity condition) had to display the behavior in question.

Note that since managers were told beforehand to select either five (low complexity condition) or eight customers (high complexity condition), the pattern in the data is somewhat stronger than in reality. While this might increase managers’ overall predictive accuracy, it should neither affect the use of adaptive decision making or fast and frugal heuristics nor should it affect task- and manager-related influences. A purely random draw of customers would not have been appropriate for our experimental design as it most likely would not show any meaningful pattern (e.g., customers might not differ on the most important metrics) and thus would have made it impossible for managers to reach a reasonable decision. We discuss the limitations associated with this procedure at the close of the article.

Tasks

The tasks used in the experiment were selected based on several criteria. They should (1) increase in difficulty, so that the first task is clearly the easiest and the third task is clearly the most difficult, (2) deal with customer management decisions that sales managers in retail banking routinely face and are familiar with, (3) be solvable with information well-known to managers and available in the bank’s internal database, and (4) be as realistic as possible.

To ensure that these four requirements were met, a small pretest was conducted. Participants were seven managers from sales, marketing, and senior management. Based on their feedback, several changes were made to the prediction tasks. After this refinement, all seven managers agreed that the tasks met the four criteria. Thus, our pretest confirmed that the three tasks were appropriate for our research goal of capturing real managerial decision making.

The specific information shown for each task was based on the predictive quality of the respective customer metrics. For each task, the predictive quality of the incorporated customer metrics was assessed by the bank using regression analyses. The seven customer metrics shown to both the low and high complexity group were high in predictive quality and explained a significant proportion of variance in the respective regression analysis. The three customer metrics shown only to the high complexity group were low in predictive quality and did not further contribute to the variance explained. Therefore, rather than arbitrarily selecting customer metrics, our procedure ensured (1) that we provided the appropriate customer information for each task and (2) that the low and high complexity group only differed in the amount of information and not in its predictive quality.

The first task—the cross-selling task—concerned obtaining a new consumer loan. The available metrics for each customer for both complexity groups were the customers’ number of fully repaid loans, number of current loans, gross margins for the last four quarters, current usage of a consumer loan, average account balance, number of products, and average overdraft in the last month. The high complexity group additionally saw information on the customers’ age, the number of children they have, and the number of credit cards they used in the last three months. Based on these metrics managers had to predict which five (low complexity condition) or eight (high complexity condition) customers actually obtained a new consumer loan in the next three months. Figure 2 illustrates the actual layout of the task in MouselabWEB.

The second task—the retention task—concerned predicting which customers cancelled a checking account. All participants saw the customers’ number of checking account transactions per month, whether they had an overdraft facility, the number of saving accounts they had with a cancellation period, the number of negative credit reports they received, the sum of their positive account balance, the sum of their negative account balance, and their total number of products. Participants in the high complexity condition were also provided with information about the customers’ age, whether they had a provision for depreciation, and whether they had bought a featured product. Managers had to predict which five (low complexity condition) or eight (high complexity condition) customers actually cancelled their checking account within the next six months.

The third task—the acquisition task—concerned whether or not the customer referred a new customer to the bank. Participants in both conditions saw the customers’ age, their household size, their account volume, their types of checking accounts, their total number of products, whether they had bought a featured product, and the number of months since their last transaction. Those in the high complexity condition also could view the number of children the customers have, their number of checking account transactions per month, and their number of remittances. Based on these metrics managers had to predict which five (low complexity condition) or eight (high complexity condition) customers actually referred a new customer through the referral program of the bank within the next 18 months.

In order to investigate how task difficulty affects the use of adaptive decision making, fast and frugal heuristics, and managers’ predictive accuracy, we coded the tasks based on their objective task difficulty with an ordinal variable as follows: 1 = easy (cross-selling task), 2 = difficult (retention task), 3 = very difficult (acquisition task).

Measures

Usage of adaptive decision making is measured using a dummy variable to indicate if there was any change in decision strategy by the decision maker across the three tasks (coded 1 = two or more different decision strategies were used for the three tasks; 0 = the same decision strategy was used for all three tasks). A change in decision strategy occurred if participating managers switched between the linear, conjunctive, additive difference, or elimination-by-aspects model (Payne 1976). Note that a change in decision strategy can happen for a number of reasons (e.g., learning, different information needs) but that the focus of our study is to investigate if a change in decision strategy happened. Usage of adaptive decision making is measured as a dummy variable which is constant for each manager over all three tasks and serves as an independent variable in a generalized estimating equation (GEE) model investigating managers’ predictive accuracy.

Strategy changed is a dummy variable which captures any change in decision strategy between the first and the second, and between the second and the third task (coded 1 = manager changed her/his decision strategy from the previous task; 0 = there was no change in strategy from the previous task). Hence, strategy changed produces only values for the second (retention) and the third (acquisition) task because no change in decision strategy can occur for the first task, which serves as reference point for the transition to the second task. Strategy changed is the dependent variable in a GEE (binary logistic) model investigating the task- and manager-related factors which drive adaptive decision making.

Usage of fast and frugal heuristics is measured using a dummy variable (coded 1 = fast and frugal heuristic was used; 0 = otherwise). A fast and frugal heuristic was assumed to be used if managers’ predictions in a specific task were (1) based on 20% or less of the available information (i.e., 14 cells in the low and 32 cells in the high complexity group) and (2) made in 150 seconds or less for the low or 300 seconds or less for the high complexity group.1 Thus, for each task there is one observation indicating whether or not a specific manager solved the decision making problem using a fast and frugal heuristic. We cannot rule out the possibility that participants who were coded as not using a fast and frugal heuristic actually did so since we do not know if they based their decision on all the information they looked at (Rieskamp and Hoffrage 1999). Specifically, managers who opened more than 14 (32) boxes in the low (high) complexity condition might not have used all of these accessed pieces of information for their predictions. Thus, they may still have used a fast and frugal heuristic without being detected by MouselabWEB and classified as fast and frugal according to our measure. However, if participants reached a decision within a short amount of time and only looked at a limited amount of information, they must have used a fast and frugal heuristic. Therefore, we believe that our measure is conservative because it reflects the minimum number of incidents where fast and frugal heuristics were used. Usage of fast and frugal heuristics serves as the dependent variable in a GEE (binary logistic) model examining the task- and manager-related factors which drive fast and frugal decision making. Usage of fast and frugal heuristics is also an independent variable in a GEE (linear) model investigating the impact of fast and frugal decision making on managers’ predictive accuracy.

Predictive accuracy is measured as the actual hit rate of each manager for the respective task (one observation per task). The hit rate is defined as the proportion of correctly classified customers. Predictive accuracy is the dependent variable in a GEE (linear) model which identifies the main determinants for managers’ decision quality.

The difference between each manager’s actual and estimated hit rate serves as a measure of each manager’s over- and underconfidence, respectively. Whereas positive differences indicate that managers display underconfidence in their estimates, negative differences imply overconfidence. By using a metric variable, we detect not only whether or not managers over- or underestimate their predictive accuracy but also the degree of their over-/underconfidence bias.

Results

Manipulation check

To test if our manipulation of task complexity was successful, we examined whether participants in the low (n = 25) and high complexity (n = 24) condition differed in their ease of processing the provided customer information (for a review, see Alter and Oppenheimer 2009). We used managers’ reaction times for classifying the first customer (in seconds) in each task as a measure of processing ease (Winkielman et al. 2006). Participants in the high complexity condition (10 pieces of information per customer in each task) required more time to make their first prediction than those in the low complexity condition (seven pieces of information per customer in each task). We used a Mann–Whitney test in order to examine whether or not managers’ reaction times in all three tasks (three decisions per participant; n = 147) were equally distributed across the two conditions of our experimental manipulation. The analysis revealed a significant effect (U = 3326.00, z = 2.43, p < .05) with participants in the high complexity condition (MedHigh complexity = 126.55, MHigh complexity = 209.37) requiring more time to classify the first customer than participants in the low complexity condition (MedLow complexity = 87.99, MLow complexity = 130.66). In other words, managers in the high complexity group experienced less ease, suggesting they found the tasks to be more complex than did managers in the low complexity group. Thus, our manipulation of task complexity worked as intended.

Managers’ task perceptions and task familiarity

In general, managers considered all three tasks realistic and stated that they were very familiar with such decisions. Over half of the tasks (53.8%) were rated as very realistic or realistic, and nearly all (91.7%) participants reported that they deal with similar customer management decisions on at least a monthly basis, 62.5% at least once per week.

Managers’ decision strategies and decision accuracy

In order to provide a detailed description of managers’ decision strategies, we first examined whether managers accessed a constant or a variable amount of information for evaluating each customer in a specific task (Payne 1976). We concluded that managers followed a constant information search pattern if they distributed the total number of boxes that they opened during the task equally across all customers of the entire set (10 in the low and 16 in the high complexity condition). If the amount of accessed information was unequally distributed across customers, we concluded that managers followed a variable information search pattern. Across all three tasks, we found that managers preferred to gather a constant amount of information for each customer when making their predictions (66.0% of managers). Only 34.0% of managers used a variable information search pattern. Furthermore, there were no significant differences in information search patterns between participants in the low (constant = 68.0%, variable = 32.0%) and high (constant = 63.9%, variable = 36.1%; χ2 = .277, df = 1, p > .10) complexity conditions. Thus, most managers first gathered an equal amount of information for all alternatives in the choice set before making their decisions.

In the second step, we determined whether managers had a tendency to search for information by customer (i.e., inter-dimensional or line-by-line with respect to Fig. 2) or by metric (i.e., intra-dimensional or column-by-column with respect to Fig. 2). Based on these sequences recorded by MouselabWEB and the corresponding Payne-index (a measure of participants’ tendency to either search within or across alternatives; for computational details, see Payne 1976), we found that 77.6% (22.4%) of the managers gathered information in an intra-dimensional (inter-dimensional) way. There were significant differences in managers’ information acquisition sequences between the low (intra-dimensional = 66.7%, inter-dimensional = 33.3%) and high (intra-dimensional = 88.9%, inter-dimensional = 11.1%; χ2 = 10.420, df = 1, p < 0.01) complexity conditions. Even though inter-dimensional search patterns were used more frequently in the low (33.3%) than in the high (11.1%) complexity condition, managers overall tended to process information based on metrics (i.e., column-by-column) rather than customers (i.e., line-by-line).

Based on the two-dimensional classification of managers’ information search behaviors (amount and sequence of accessed information), we categorized the specific decision strategies that managers used for each task. Table 1 shows the results for both the low and high complexity conditions.
Table 1

Decision strategies, accessed information, and decision accuracy (by task complexity)

  

Decision strategy

Task complexity

Linear model

Conjunctive model

Additive difference model

Elimination-by-aspects model

Total

1. Low complexity

Usage in % (n)

20.0% (15)

13.3% (10)

48.0% (36)

18.7% (14)

100.0% (75)

Accessed information in % (mean)

100.0%

59.1%

45.6%

47.1%

58.6%

Hit rate (mean)

92.0%

96.0%

85.0%

82.9%

87.5%

2. High complexity

Usage in % (n)

.0% (0)

11.1% (8)

63.9% (46)

25.0% (18)

100.0% (72)

Accessed information in % (mean)

n/a

52.1%

46.7%

42.8%

46.4%

Hit rate (mean)

n/a

85.9%

83.7%

86.8%

84.7%

Total

Usage in % (n)

10.2% (15)

12.2% (18)

55.8% (82)

21.8% (32)

100.0% (147)

Accessed information in % (mean)

100.0%

56.0%

46.3%

44.7%

52.6%

Hit rate (mean)

92.0%

91.5%

84.3%

85.1%

86.1%

Consistent with previous results on intra- and inter-dimensional search patterns, managers’ choices of decision strategies significantly varied with respect to the task’s complexity (χ2 = 16.89, df = 3, p < 0.01). Whereas in the high complexity condition managers did not follow an information search pattern consistent with the linear model, the latter was used in 20.0% of all tasks in the low complexity condition. Interestingly, five managers consistently used the linear model across all three tasks (n = 15). They not only evaluated customer by customer but also accessed all available information in each task. Hence, providing less information in a decision making problem seems to make managers more likely to engage in full information processing. In contrast, providing more information in a decision making problem seems to foster the use of intra-dimensional decision strategies. Specifically, the additive difference (high complexity = 63.9% vs. low complexity = 48.0%) and elimination-by-aspects (high complexity = 25.0% vs. low complexity = 18.7%) models have higher choice shares in the high complexity condition. Managers’ overall preference for the additive difference (55.8%) and elimination-by-aspects (21.8%) models may be because both decision strategies require relatively little information to make accurate decisions. As confirmed by a one-way ANOVA, the overall hit rates of the four decision strategies did not significantly differ (linear = 92.0% vs. conjunctive = 91.5% vs. additive difference = 84.3% vs. elimination-by-aspects = 85.1%; F(3, 146) = 1.38, p > .10), even though the corresponding amount of accessed information ranged from 44.7% to 100.0% (F(3, 146) = 27.81, p < 0.001). Thus, more deliberate decisions with respect to the amount of accessed information do not necessarily translate into significantly better customer management decisions. Surprisingly, the complexity of the task did not significantly affect managers’ hit rates (low complexity = 87.5% vs. high complexity = 84.7%; F(1, 146) = 0.83, p > .10).

Table 2 provides a similar descriptive overview of managers’ decision strategies, accessed information, and decision accuracy by task.2 A chi-square test provides evidence that managers use different decision strategies which depend on the task to be solved (χ2 = 19.05, df = 6, p < 0.01). As shown in Table 2, the conjunctive (24.5%), the additive difference (34.7%), and the elimination-by-aspects (30.6%) models have significant choice shares in the cross-selling task. However, as the tasks became more demanding, managers increasingly switched from the conjunctive and the elimination-by-aspects model to the additive difference model. Thus, whereas gathering a variable amount of information might be a viable decision making strategy for relatively easy tasks, it seems to be less useful for more difficult decisions. The transition to the additive difference model implies that, when facing relatively difficult decisions, managers prefer to compare all options along a certain number of key metrics. Furthermore, the fact that managers switch between different decision strategies provides initial evidence for adaptive decision making. Table 2 also shows that managers varied in the percent of available information they used for each task (cross-selling = 61.5% vs. retention = 48.1% vs. acquisition = 48.3%; F (2,146) = 4.13, p < 0.05). Recall that the three tasks the mangers completed were designed to increase sequentially in difficulty. Thus, it is not surprising that participants’ hit rates decreased across the tasks (cross-selling = 94.9% vs. retention = 85.7% vs. acquisition task = 77.7%; F(2, 146) = 12.74, p < 0.001).
Table 2

Decision strategies, accessed information, and decision accuracy (by task)

  

Decision strategy

Task

Linear model

Conjunctive model

Additive difference model

Elimination-by-aspects model

Total

1. Cross-selling

Usage in % (n)

10.2% (5)

24.5% (12)

34.7% (17)

30.6% (15)

100.0% (49)

Accessed information in % (mean)

100.0%

55.1%

66.7%

47.8%

61.5%

Hit rate (mean)

100.0%

96.9%

90.0%

97.3%

94.9%

2. Retention

Usage in % (n)

10.2% (5)

10.2% (5)

63.3% (31)

16.3% (8)

100.0% (49)

Accessed information in % (mean)

100.0%

59.7%

41.2%

35.1%

48.1%

Hit rate (mean)

96.0%

84.5%

85.7%

80.0%

85.7%

3. Acquisition

Usage in % (n)

10.2% (5)

2.0% (1)

69.4% (34)

18.4% (9)

100.0% (49)

Accessed information in % (mean)

100.0%

48.8%

40.7%

48.1%

48.3%

Hit rate (mean)

80.0%

62.5%

80.1%

69.2%

77.7%

Total

Usage in % (n)

10.2% (15)

12.2% (18)

55.8% (82)

21.8% (32)

100.0% (147)

Accessed information in % (mean)

100.0%

56.0%

46.3%

44.7%

52.6%

Hit rate (mean)

92.0%

91.5%

84.3%

85.1%

86.1%

Overall, managers performed very well. As Table 2 also indicates, the participants’ actual hit rate across all three tasks was 86.1%. While many studies have shown people to be overconfident (for an overview, see Hoffrage 2004), our study found managers to be underconfident (Moore and Cain 2007). Specifically, they significantly underestimated their actual hit rates by an average of 13.2 percentage points (actual hit rate = 86.1% vs. estimated hit rate = 72.9%; t = 7.08, p < 0.001).

Task- and manager-related drivers of adaptive decision making

In this study, managers clearly were adaptive in their decision making. Only twenty managers (40.8%) used the same decision strategy in all three tasks. Twenty-six managers (53.1%) used two, and three participants (6.1%) even used three different decision strategies to solve the three tasks. Thus, across all tasks, a majority of managers (59.2%) changed their decision strategy at least once and were classified as adaptive decision makers.

In order to identify the task- and manager-related factors which drive adaptive decision making, we estimated a generalized estimating equation (GEE) model for binary responses which accounts for the repeated measures of the dependent variable strategy changed across two tasks for each manager (Liang and Zeger 1986; Zeger and Liang 1986). Since multiple observations from the same individual are correlated, our GEE (binary logistic) model accounts for the within-subject dependence by incorporating a “working” correlation matrix (i.e., compound symmetry) of the responses from the same manager (e.g., Ballinger 2004; Zorn 2001). In other words, the model adjusts for the within-subject correlations of the repeated binary outcomes when estimating the relationship between a set of independent variables (i.e., task- and manager-related factors) and the dependent variable strategy changed. The following task- and manager-related characteristics were the independent variables in the model:
  • objective task difficulty, self-reported task difficulty, and the difference of actual and estimated hit rate as within-subject variables which are task-dependent,

  • age, gender, and sales experience as between-subject variables which control for individual differences between managers,

  • task complexity as manipulated between-subject factor.

We used a top-down strategy for model selection, which reduces a full model incorporating the maximum number of independent variables (or model effects) to only those predictors which significantly contribute to the explanation of the dependent variable (Verbeke and Molenberghs 2000). Accordingly, we first estimated the full model and determined the significance of model effects using a Wald Chi2 test. In the next step, we reduced the model by removing all insignificant predictors (e.g., West et al. 2007). The final model was then selected based on the quasi-likelihood (QIC) and the corrected quasi-likelihood under the independence model criterion (QICC); both represent the standard criteria for model selection in the GEE framework (Pan 2001). As modifications of Akaike’s Information Criterion, QIC and QICC also penalize the number of parameters in the model in order to account for the opposing needs of parsimony and model fit. Table 3 summarizes the results of the model effects for the full and the reduced GEE (binary logistic) models.
Table 3

Task- and manager-related drivers of adaptive decision making (test of model effects)

Dependent variable: strategy changed

Full model

Reduced model

Test of model effects

Wald Chi2

df

p-value

Wald Chi2

df

p-value

Intercept

2.035

1

.154

3.309*

1

.069

Objective task difficulty

5.495**

1

.019

6.170**

1

.013

Self-reported task difficulty

.100

1

.752

   

Complexity

.001

1

.978

   

Age

2.603

1

.107

   

Gender

.943

1

.332

   

Sales experience

2.821*

1

.093

1.079

1

.299

Difference between actual and estimated hit rate

.707

1

.400

   

Model summary (n = 95):

Model summary (n = 96):

QIC = 127.770

QIC = 122.801

QICC = 127.828

QICC = 123.292

Standard errors (s.e.) in parentheses; significance levels: *** p ≤ .01, ** p ≤ .05, * p ≤ .10

By directly comparing the results of the two models in Table 3, we can see that objective task difficulty is the main predictor of adaptive decision making (full model: Wald Chi2 = 5.495, p < .05; reduced model: Wald Chi2 = 6.170, p < .05). Whereas sales experience was marginally significant in the full model (Wald Chi2 = 2.821, p < .10), the effect became insignificant in the reduced model (Wald Chi2 = 1.079, p > .10). All other model effects (including the demographic control variables age and gender) were insignificant in the full model and were not included in the reduced model. Based on the lower QIC and QICC values, we indentified the reduced model reported in Table 3 as the final model. Therefore, we interpret significant parameters and their odds ratios on the basis of the results derived from the reduced model (cf. Table 4).
Table 4

Task- and manager-related drivers of adaptive decision making (parameter estimates)

Dependent variable: strategy changed

Full model

Reduced model

Parameter

b (s.e.)

Wald Chi2

Exp b

b (s.e.)

Wald Chi2

Exp b

Intercept

−.804 (1.101)

.533

.447

−.860 (.801)

1.152

.423

Acquisition task

−1.264** (.539)

5.495

.283

−1.140** (.459)

6.170

.320

Retention task (reference category)

0

 

1

0

 

1

Self-reported task difficulty

.137 (.433)

.100

1.147

   

High complexity

−.013 (.480)

.001

.987

   

Low complexity (reference category)

0

 

1

   

Age

−.681 (.422)

2.603

.506

   

Male

−.448 (.461)

.943

.639

   

Female (reference category)

0

 

1

   

Sales experience

.494* (.294)

2.821

1.639

.185 (.178)

1.079

1.203

Difference between actual and estimated hit rate

−.011 (.013)

.707

.989

   

Standard errors (s.e.) in parentheses; significance levels: *** p ≤ .01, ** p ≤ .05, * p ≤ .10

Based on the results of the reduced model, we can conclude that the more difficult the task, the less likely managers were to change their decision strategies. Specifically, as the odds ratio (Exp b) of objective task difficulty suggests, managers’ likelihood to switch decision strategy decreases on average by 68% (=.320–1.000) as tasks become more difficult. This result is consistent with previous descriptive findings (cf. Table 2), which showed that percentage changes in decision strategies between the first (cross-selling) and the second (retention) task were considerably higher than between the second (retention) and third (acquisition) task.

Task- and manager-related drivers of fast and frugal heuristics

We observed 18 instances of fast and frugal heuristics being used (only 12% of all observations). Interestingly, of the 18 times fast and frugal heuristics were used, 17 times were by managers who had at least 7 years of sales experience. In total, 11 managers used fast and frugal heuristics; seven of them twice and four of them once.

To see which task- and manager-related characteristics impact the use of fast and frugal heuristics, we estimated another GEE (binary logistic) model with usage of fast and frugal heuristics as the dependent variable. Again, the GEE approach was preferred over a binary logistic regression because it can account for the repeated measurements which managers made across the three tasks. We included the previously used task- and manager-related factors as independent variables and followed the same top-down strategy for model selection. Table 5 summarizes the results of the model effects for the full and the reduced GEE (binary logistic) models.
Table 5

Task- and manager-related drivers of fast and frugal heuristics (test of model effects)

Dependent variable: usage of fast and frugal heuristics

Full model

Reduced model

Test of model effects

Wald Chi2

df

p-value

Wald Chi2

df

p-value

Intercept

1.530

1

.216

3.851**

1

.050

Objective task difficulty

3.948

2

.139

   

Self-reported task difficulty

18.336***

1

.000

6.256**

1

.012

Complexity

2.076

1

.150

   

Age

2.294

1

.130

   

Gender

1.457

1

.227

   

Sales experience

4.287**

1

.038

4.359**

1

.037

Difference between actual and estimated hit rate

.388

1

.533

   

Model summary (n = 140):

Model summary (n = 142):

QIC = 97.268

QIC = 98.167

QICC = 96.515

QICC = 97.839

Standard errors (s.e.) in parentheses; significance levels: *** p ≤ .01, ** p ≤ .05, * p ≤ .10

As Table 5 indicates, both models identified self-reported task difficulty (full model: Wald Chi2 = 18.336, p < .01; reduced model: Wald Chi2 = 6.256, p < .05) and sales experience (full model: Wald Chi2 = 4.287, p < .05; reduced model: Wald Chi2 = 4.359, p < .05) as the main drivers of fast and frugal decision making. Since the full model performs better than the reduced model on both information criteria QIC and QICC (see bottom of Table 5), we interpret significant parameter estimates and their odds ratios on the basis of the results derived from it (cf. Table 6).
Table 6

Task- and manager-related drivers of fast and frugal heuristics (parameter estimates)

Dependent variable: usage of fast and frugal heuristics

Full model

Reduced model

Parameter

b (s.e.)

Wald Chi2

Exp b

b (s.e.)

Wald Chi2

Exp b

Intercept

−4.048** (1.998)

4.103

.017

−3.014** (1.536)

3.851

.049

Acquisition task

1.567* (.816)

3.686

4.793

   

Retention task

1.024 (.816)

1.576

2.784

   

Cross-selling task (reference category)

0

 

1

   

Self-reported task difficulty

−1.330*** (.311)

18.336

.265

−.904** (.362)

6.256

.405

High complexity

1.223 (.848)

2.076

3.396

   

Low complexity (reference category)

0

 

1

   

Age

−.595 (.393)

2.294

.552

   

Male

.854 (.708)

1.457

2.350

   

Female (reference category)

0

 

1

   

Sales experience

.898** (.434)

4.287

2.454

.681** (.326)

4.359

1.977

Difference between actual and estimated hit rate

.011 (.018)

.388

1.011

   

Standard errors (s.e.) in parentheses; significance levels: *** p ≤ .01, ** p ≤ .05, * p ≤ .10

The GEE results confirm our descriptive finding that the more experience managers have, the more likely they are to use a fast and frugal heuristic (b = .898, p < .05). In particular, every three years of sales experience (corresponding to the intervals of our sales experience groups) increases the likelihood of making fast and frugal decisions by a factor 2.45 or 145% (on average). Furthermore, the easier the task was perceived, the more likely a heuristic was used (b = −1.330, p < .01). Thus, managers rely on their own perceptions of task difficulty when deciding whether or not to use a fast and frugal heuristic. Surprisingly, the parameter estimate of acquisition task was marginally significant (b = 1.567, p < .10) in the full model, which suggests that managers were more likely to use a fast and frugal heuristic in the acquisition than in the cross-selling task. However, as the left panel of Table 5 indicates, the overall effect of objective task difficulty was too small to infer that managers are more likely to use fast and frugal heuristics as tasks become more difficult.

Determinants of managers’ predictive accuracy

We estimated a GEE (linear) model with predictive accuracy, measured by managers’ hit rate in each task, as the dependent variable. We again accounted for repeated observations with a “working” correlation matrix and included task- and manager-related characteristics as independent variables. Additionally, usage of adaptive decision making and usage of fast and frugal heuristics served as independent variables in order to investigate their influence on managers’ decision quality. Table 7 summarizes the results of the model effects for the full and the reduced GEE (linear) models.
Table 7

Determinants of managers’ predictive accuracy (test of model effects)

Dependent variable: predictive accuracy

Full model

Reduced model

Test of model effects

Wald Chi2

df

p-value

Wald Chi2

df

p-value

Intercept

166.129***

1

.000

446.291***

1

.000

Objective task difficulty

10.691***

2

.005

11.706***

2

.003

Self-reported task difficulty

5.453**

1

.020

5.536**

1

.019

Complexity

2.137

1

.144

   

Age

.139

1

.709

   

Gender

.731

1

.392

   

Sales experience

.247

1

.619

   

Difference between actual and estimated hit rate

53.385***

1

.000

54.863***

1

.000

Usage of adaptive decision making

.518

1

.472

   

Usage of fast and frugal heuristics

12.965***

1

.000

10.225***

1

.001

Model summary (n = 140):

Model summary (n = 143):

QIC = 21015.404

QIC = 22021.039

QICC = 21008.606

QICC = 22018.806

Standard errors (s.e.) in parentheses; significance levels: *** p ≤ .01, ** p ≤ .05, * p ≤ .10

Both models consistently revealed significant effects for objective task difficulty (full model: Wald Chi2 = 10.691, p < .01; reduced model: Wald Chi2 = 11.706, p < .01), self-reported task difficulty (full model: Wald Chi2 = 5.453, p < .05; reduced model: Wald Chi2 = 5.536, p < .05), the difference between actual and estimated hit rate (full model: Wald Chi2 = 53.385, p < .01; reduced model: Wald Chi2 = 54.863, p < .01), and the usage of fast and frugal heuristics (full model: Wald Chi2 = 12.965, p < .01; reduced model: Wald Chi2 = 10.225, p < .01). For the interpretation of significant parameter estimates, we refer to the results of the full model as suggested by the lower values for QIC and QICC.
Table 8

Determinants of managers’ predictive accuracy (parameter estimates)

Dependent variable: predictive accuracy

Full model

Reduced model

Parameter

b (s.e.)

Wald Chi2

b (s.e.)

Wald Chi2

Intercept

90.453*** (6.984)

167.737

92.656*** (4.395)

444.419

Acquisition task

−10.997*** (3.635)

9.154

−11.129*** (3.468)

10.298

Retention task

−1.724 (2.255)

.584

−1.903 (2.222)

.734

Cross-selling task (reference category)

0

 

0

 

Self-reported task difficulty

−4.814** (2.062)

5.453

−4.578** (1.946)

5.536

High complexity

−3.930 (2.689)

2.137

  

Low complexity (reference category)

0

   

Age

.690 (1.850)

.139

  

Male

−2.291 (2.679)

.731

  

Female (reference category)

0

   

Sales experience

.647 (1.300)

.247

  

Difference between actual and estimated hit rate

.576*** (.079)

53.385

.577*** (.078)

54.863

Two or more decision strategies

2.090 (2.903)

.518

  

One decision strategy (reference category)

0

   

Fast and frugal heuristics

9.143*** (2.539)

12.965

8.185*** (2.560)

10.225

No heuristics used (reference category)

0

 

0

 

Standard errors (s.e.) in parentheses; significance levels: *** p ≤ .01, ** p ≤ .05, * p ≤ .10

As the left panel of Table 8 shows, the parameter estimates for the retention and the acquisition task suggest that moving forward to the next task reduced managers’ predictive accuracy. Whereas there is no significant difference in managers’ decision quality between the cross-selling task and the retention task (b = −1.724, p > .10), predictive accuracy drops by an average of nearly 11.0% in the acquisition task relative to the cross-selling task (b = −10.997, p < .01). Also, self-reported task difficulty has the expected negative association with predictive accuracy. Thus, the more difficult managers perceived the tasks to be, the less accurate were their predictions. Another interesting finding is that the difference between the actual and estimated hit rate not only has a positive impact on predictive accuracy but also its effect size is the largest of all variables. Hence, managers who underestimated the quality of their decisions were the most accurate. Also, the more prudent managers were in their estimated hit rate (i.e., the larger the underconfidence bias was), the more accurate they were in their predictions. Whereas usage of adaptive decision making was insignificant in this analysis, usage of fast and frugal heuristics was significant and associated with an increase in predictive accuracy. All other things being equal, making fast decisions and perhaps ignoring seemingly irrelevant pieces of information increased managers’ predictive accuracy by an average of 9.14 percentage points. Based on the GEE results, on average, managers have a predictive accuracy of 84.5% if they did not use a fast and frugal heuristic and have a predictive accuracy of 93.7% if they did.

The role of low and high predictive quality metrics in fast and frugal decision making

In order to examine why usage of fast and frugal heuristics was associated with increased decision accuracy, we further explored which pieces of customer information managers accessed in the high complexity condition. It is possible that managers who use a fast and frugal heuristic are more accurate because they focus more on the high predictive quality metrics and less on the low predictive quality metrics. To test this, we examined what percent of the total metrics each manager looked at for each task was low versus high predictive quality. A one-way ANOVA analysis with the percentage of low predictive quality metrics as the dependent variable and usage of fast and frugal heuristics as the independent variable revealed that managers who used a fast and frugal heuristic accessed a smaller percentage of low predictive quality metrics (MUsage of fast and frugal heuristics = 4.2%) than managers who used other decision strategies (MOtherwise = 22.2%; F(1, 70) = 15.27, p < .001).

Discussion

Theoretical contributions

Since the 1960s researchers have tried to explain managerial decision making in marketing (e.g., Howard 1963). Even though there is considerable work investigating marketing decision making at the organizational level (e.g., Frankwick et al. 1994; Hutt et al. 1988; Ronchetto et al. 1989; Ruekert and Walker 1987), there is little published research about how individual managers process information in order to make decisions. This gap is particularly profound in the area of customer management and is presumably due to two major challenges associated with examining managerial decision making (Van Bruggen and Wierenga 2010): (1) observing the decision making processes of real managers in an unobtrusive manner is difficult and (2) empirical studies need to account for an evaluation criterion on which to judge the quality of managers’ decisions. In this article, we presented a methodology which fulfills both requirements. By using a process-tracing tool and presenting real managers with realistic customer management tasks that involve a prediction, we were able not only to provide a rich descriptive account of managers’ decision strategies but also to assess decision accuracy and the factors influencing it. Furthermore, our study is the first to empirically test the fast and frugal heuristics and the adaptive decision maker frameworks in the context of customer management decisions.

In our study, a majority of managers were adaptive in their decision making. Rather than consistently following one decision strategy throughout the experiment, they switched among the linear, conjunctive, additive difference, and elimination-by-aspects models and used more than one of these four decision strategies for solving the three different customer management tasks. Our findings further indicate that managers adapted their decision strategies to the objective difficulty of the task. However, the more difficult the tasks became, the less likely the participating managers were to switch to another decision strategy.

Across all three tasks, we also identified 18 instances where managers used a fast and frugal heuristic. That is, managers made decisions quickly and on the basis of only a very limited amount of information. The results show that the use of fast and frugal heuristics was largely driven by managers’ experience. Thus, our findings support Gigerenzer’s assumption that people learn to choose the appropriate heuristic from their adaptive toolbox in order to make fast, frugal, and accurate decisions (e.g., Gigerenzer and Gaissmaier 2011).

In the classic view, the use of heuristics is associated with biases in judgment and decision making (e.g., Tversky and Kahneman 1974). Therefore, many scholars consider heuristics as inaccurate decision rules which are best avoided (e.g., Hutchinson et al. 2010). Contrary to this, our results show that (1) adaptive decision making does not have a negative impact on decision quality, and (2) fast and frugal heuristics can lead to a significant increase in decision accuracy. In particular, we found that the four decision strategies under investigation performed equally well so that switching strategies did not impact predictive accuracy. Even more surprisingly, the use of fast and frugal heuristics was associated with an increase in predictive accuracy of over nine percentage points. This may be because managers focus more on the important and less on the unimportant pieces of information. This result is in line with the theoretical assumption that fast and frugal heuristics exploit evolved capacities (e.g., Gigerenzer 2004). In other words, experienced managers have learned over time to deliberately ignore irrelevant pieces of information and to focus on the relevant metrics when making customer management decisions. Overall, our study showed that simple decision strategies performed well in reducing cognitive effort without jeopardizing decision quality. Thus, the existing skepticism concerning heuristics should be questioned.

Managerial implications

Having investigated the factors that drive good decisions, the results of this study provide some practical implications about how individual managers can improve their decision making in order to make valuable customer acquisition, retention, and cross-selling decisions.

In particular, our findings suggest that being too confident can negatively affect decision quality and thus harm customer relationships. Identifying and successfully managing overconfidence seems promising for increasing the decision quality of a firm’s managers. As suggested by Russo and Schoemaker (1992), timely and precise feedback can be an effective tool to reduce overconfidence. Counterargumentation can be another effective debiasing technique (Russo and Schoemaker 1992). Overconfidence is often the result of managers being motivated to support and successfully defend their initial opinion. In this process, contradicting evidence is largely neglected. Asking managers to think of reasons why their opinions or estimates might be wrong can reduce overconfidence (Koriat et al. 1980). Thus, reminding managers to account for both supporting and contradicting reasons in their decision making processes will increase decision quality.

We showed not only that managers use fast and frugal heuristics but that they do so successfully. The beneficial impact of fast and frugal heuristics may be because they reduce information overload. The latter is a common phenomenon in today’s managerial decision making (Reuters 1996). By focusing only on important decisions inputs, usage of fast and frugal heuristics avoids (1) the overweighting of irrelevant pieces of information and (2) data-rich but poor decisions made on the basis of an unmanageable amount of information. Therefore, rather than just accumulating data, managers should be encouraged to trust their intuition about which information is important when making decisions.

While we focused on customer management decisions, our results have implications more generally for managerial decision making. Consider a senior manager of a corporate group who has to make an investment decision about whether to acquire a company in order diversify the group’s business portfolio. Such complex investment decisions usually demand the processing of a substantial amount of information from financial statements (e.g., information about financial ratios, earnings, expenses, cash flows), annual reports (e.g., information about the diversification strategy, product portfolio), and a range of other sources (e.g., rating reports) in order to judge the risk and return of the investment. However, there is also some evidence which suggests that fast and frugal heuristics can even improve decision quality for such complex investment decisions. In an interview that we conducted, a very successful senior manager of a leading financial institution explained that he only uses four criteria to judge the health and prospects of a company before making the investment decision: (1) market share gain to see if the company is growing faster than the competition, (2) revenue growth to assess the company’s top-line health, (3) cost growth to see if costs are growing faster than revenues, and (4) flexibility of the cost base to judge whether the business is able to quickly adapt to changing circumstances (e.g., sudden drop in demand). Thus, rather than processing all available financial information, the manager implicitly accounts for risk and return considerations by assessing more “global” decision criteria. This anecdote is in line with research on venture capital decision making. For example, Hall and Hofer (1993) find that venture capitalists make go/no-go investment decisions in an average of less than 6 minutes on initial proposal screening and less than 21 minutes on business proposal assessment. They also conclude that finances are implicitly assessed by venture capitalists’ judgments about the nature of the proposed business, its strategy, and its economic environment (e.g., the industry’s risk and rate of return). Despite relying on relatively few pieces of information, the success rate of venture capital-backed businesses is significantly higher than the success rate of other new ventures (Hall and Hofer 1993).

Limitations and directions for further research

Even though our research contributes to the understanding of managerial decision making, we want to point out some potential limitations of our study. First, all three tasks in our experiment were predictive in nature. That is, managers in the low (high) complexity condition knew beforehand that they had to predict a certain behavior for five (eight) customers in each task. Although this study design allowed us to assess managers’ decision accuracy, it is associated with less uncertainty in decision making compared to (1) a task where the number of objects to be classified is unknown or (2) a task that does not have a direct predictive dimension (e.g., investment decisions). Whereas this reduction in uncertainty might have positively affected managers’ hit rates, we do not think it should have influenced the way in which managers searched for information (e.g., the number and sequence of accessed information, decision time). Thus, we feel confident that the conclusions we drew for the use of adaptive decision making and fast and frugal heuristics should also hold for other types of decisions. Also, we think that the determinants of managers’ decision accuracy should be unaffected by a possible overestimation of hit rates which would only be reflected in a higher value for the intercept of the GEE (linear) model. Therefore, the tasks we used accounted not only for a normative evaluation criterion of managers’ decision quality but also for a predictive component which is involved in most managerial decisions. Consider the following example: if a retailer wants to decide whether or not to enter into a new market, this investment decision will be driven to a large extent by managers’ predictions about the anticipated sales volume, expansion costs, operating costs, etc. As the example illustrates, most managerial decisions are either directly or indirectly characterized by making predictions. Nevertheless, since it remains possible that the nature of the tasks we used influenced the outcome, future research should examine whether the nature of the task affects managerial decision making by utilizing a more diverse set of tasks, including ones that do not involve an explicit prediction.

Because we held the order of the tasks constant, we did not directly test or control for learning. Since we invited real managers to participate in our research, we did not have the opportunity to administer several experimental sessions where we could alter the order of the tasks. Therefore, future research should test whether the usage of adaptive decision making and fast and frugal heuristics increases with learning.

Although we were fortunate to have access to a sample of real managers, we would like to note that all participants were from the same functional area (i.e., sales department) of one bank. Thus, rather than drawing general conclusions about the entire population of managers, our findings should be considered as initial evidence for managers’ use of adaptive decision making and fast and frugal heuristics. Future research should therefore investigate the two frameworks in different contexts (e.g., different industries, different functional areas) in order to further enhance our understanding of managerial decision making.

Finally, as we noted, it is possible that managers who took more time and opened more information boxes also used fast and frugal heuristics. Future research could manipulate the use of specific fast and frugal heuristics, rather than infer their use, to help validate whether use of fast and frugal heuristics can increase decision accuracy.

In addition to the extensions implied by these limitations, there are several other directions for further research that we would like to highlight. Using MouselabWEB, future studies could investigate how demographic (e.g., age, gender, educational background) and personality-related differences between managers (e.g., sensing-thinking vs. intuition-thinking vs. sensing-feeling vs. intuition feeling; Myers and McCaulley 1985) affect their choice of decision strategy and decision accuracy. Another interesting avenue for future research would be to manipulate participants’ level of confidence and to employ MouselabWEB for identifying differences in the decision making processes and accuracy of under- and overconfident managers.

Footnotes
1

Using a cutoff value of 20% was based on the fact that (1) a lower cutoff was not possible since that would have not yielded enough participants to draw meaningful conclusions (only four participants used 10% of the information), (2) a higher cutoff does not seem consistent with the frugal heuristics framework, and (3) the number of participants using 20% or less of the information formed exactly the 10th percentile of all participants. Using the cutoff value of 150 or 300 seconds was based on discussion with the bank’s management. In order to provide stronger support for our findings, we conducted a sensitivity analysis which examined the robustness of our results with respect to different definitions of usage of fast and frugal heuristics. Across different definitions of usage of fast and frugal heuristics based on a range from (1) 14% to 26% (in 2% intervals) for information use, (2) 120 to 180 seconds (in 10 second intervals) for decision time in the low complexity condition, and (3) 270 to 330 seconds for decision time in the high complexity condition, all analyses reported in this article yielded similar results. Thus, our findings are robust to changes in how we operationalize usage of fast and frugal heuristics.

 
2

A 2x3 contingency table analyzing the usage of decision strategies by task complexity and by task would not produce reliable results because there were too many cells with an expected value of less than five.

 

Acknowledgments

The authors thank Walter Herzog, Jan R. Landwehr, Bernd Skiera, and the four anonymous reviewers for their helpful comments.

Copyright information

© Academy of Marketing Science 2012