Hoffrage How causal knowledge simplifies decision making
Mind Mach (2006) 16:365 380 DOI 10.1007/s11023-006-9035-1 How causal knowledge simplifies decision-making Rocio Garcia-Retamero Ć Ulrich Hoffrage Received: 23 February 2005 / Accepted: 14 November 2005 / Published online: 11 August 2006 Ó Springer Science+Business Media B.V. 2006 Abstract Making decisions can be hard, but it can also be facilitated. Simple heuristics are fast and frugal but nevertheless fairly accurate decision rules that people can use to compensate for their limitations in computational capacity, time, and knowledge when they make decisions [Gigerenzer, G., Todd, P. M., & the ABC Research Group (1999). Simple Heuristics That Make Us Smart. New York: Oxford University Press.]. These heuristics are effective to the extent that they can exploit the structure of information in the environment in which they operate. Specifically, they require knowledge about the predictive value of probabilistic cues. However, it is often difficult to keep track of all the available cues in the environment and how they relate to any relevant criterion. This problem becomes even more critical if compound cues are considered. We submit that knowledge about the causal struc- ture of the environment helps decision makers focus on a manageable subset of cues, thus effectively reducing the potential computational complexity inherent in even relatively simple decision-making tasks. We review experimental evidence that tested this hypothesis and report the results of a simulation study. We conclude that causal knowledge can act as a meta-cue for identifying highly valid cues, either individual or compound, and helps in the estimation of their validities. Keywords Causal knowledge Ć Compound cue Ć Cue selection Ć Fast and frugal heuristics Ć Search processes Ć Take the Best Ć Take the Best Configural Ć Validity estimation R. Garcia-Retamero (&) Max Planck Institute for Human Development, Lentzeallee 94, D-14195 Berlin, Germany e-mail: rretamer@mpib-berlin.mpg.de U. Hoffrage Ecole des Hautes Etudes Commerciales, Université de Lausanne, CH-1015 Lausanne- Dorigny, Switzerland 123 366 R. Garcia-Retamero, U. Hoffrage Introduction When we are faced with a decision, for example, which restaurant to go to or which meal to order, it is often impossible to consider all the available alternatives and to gather and process all the information regarding these options. For instance, we generally do not consider every restaurant in the city, and when we do select one, we often do not have much detail about the entrees on the menu (e.g., the amount of cholesterol, fat, or preservatives in the dishes, the cooking methods used, or how they taste) to help us infer which one we would like most. In fact, in real-life situations such as this, we often make fast decisions based on little information. Recently, Gigerenzer, Todd, and the ABC Research Group (1999) have suggested that we use simple heuristics in these situations, that is, fast and frugal but never- theless fairly accurate strategies for making decisions with a minimum of informa- tion (see also Todd & Gigerenzer, 2000). These rules are fast because they do not involve much computation, and they are frugal because they search for only some of the available information in the environment. One of the fast and frugal heuristics proposed by the ABC research group is Take The Best (TTB; Gigerenzer & Goldstein, 1996, 1999). This heuristic is de- signed for so-called two-alternative forced-choice tasks and can be used to infer which of two alternatives has a higher value on a quantitative criterion, such as which of two university professors earns more money. The alternatives are de- scribed on several dichotomous cues such as gender or whether the professor is on the faculty of a state or a private university. These cues allow to make probabilistic inferences about the criterion. Like each of the fast and frugal heuristics that has been proposed in the context of this research program, TTB is constructed from building blocks, which are the precise steps of information gathering and pro- cessing involved in making a decision. Specifically, this heuristic has a search rule, which defines the order in which to search for information (TTB looks up cues in the order of their validity, i.e., the probability that a cue will point to the correct decision given that it discriminates between the alternatives); a stopping rule, which specifies when the search is to be stopped (TTB stops after the first discriminating cue); and a decision rule, which specifies how to use the information that has been looked up when it comes to making a decision (TTB chooses the alternative favored by the first discriminating cue). The TTB heuristic and an extension of TTB for comparisons among more than two alternatives have been subjected to empirical tests in a number of studies (e.g., Bröder, 2000, 2003; Bröder & Schiffer, 2003a; Newell, Rakow, Weston, & Shanks, 2004; Newell & Shanks, 2003; Rieskamp & Hoffrage, 1999). There is accumulating experimental evidence for the use of this heuristic, particularly when there are search costs for accessing cues (see Bröder, 2000, 2003; Bröder & Schiffer, 2003b), or when decisions have to be made under time pressure (e.g., Rieskamp & Hoffrage, 1999). In addition, Newell, Weston, and Shanks (2003) tested the building blocks of TTB separately and reported that 75% of participants followed TTB s search rule by validity. Furthermore its stopping and decision rules were obeyed in 80% and 89% of the trials, respectively (see also Newell & Shanks, 2003). However, these experimental results on the use of TTB need to be qualified. In most of these studies, participants were encouraged to use cues in the order of their validity by being informed about cue validities or the validity order (see Bröder, 123 Causal knowledge and simple heuristics 367 2000, 2003; Bröder & Schiffer, 2003b; Newell et al., 2003). In two studies that tested search by validity against alternative search orders, validity was not the search cri- terion that predicted participants searches best (Läge, Hausmann, Christen, & Daub, 2005; Newell et al., 2004) because participants were instead making use of simple rules for ordering cues based on trial-by-trial learning (Dieckmann & Todd, 2004; see also Garcia-Retamero, Takezawa, & Gigerenzer, 2006). The cue orderings established through such rules do not necessarily converge toward the ordering established by validity. Therefore, participants might have had difficulties computing cue validities and then ordering cues accordingly even though they were dealing in those experiments with relatively few cues (i.e., four to six). The problem of finding a good cue ordering appears even more severe if one considers that in most situations, there are myriad potential cues that could be used to make a decision, and it is practically impossible to keep track of them all and to compute their validities for any potentially relevant criterion (Juslin & Persson, 2002). Cue selection is further complicated if potential combinations of cues (i.e., compound cues) are taken into account. Yet sometimes an accurate decision re- quires us to do so. For example, some medications might have side effects, such as nausea, if ingested together with alcohol, whereas neither the drug nor the alcohol would cause any problems if ingested alone (of course, this would also depend on the amount of alcohol that is consumed). Thus, the relationship between one cue (the ingestion of a medication) and the criterion (nausea) depends on the presence of another cue (the ingestion of alcohol). The problem is that in real-world environ- ments, there exist a multitude of potential combinations of cues to form compounds, rendering it nearly impossible to keep track of them all. As a consequence, a strategy that processes all possible compound cues as configurations would be too compu- tationally demanding. Nor is it plausible to assume that the brain comes prewired with a representation for each of the combinations of elementary stimulus inputs (Kehoe & Graham, 1988). Bearing these comments in mind, the interesting question is whether there is a way in which the relevant cues, individual or compounded, can be selected from the abundance of possibilities in the environment. We hypothesize that people do not process all possible cues in their natural environments but rather use their causal knowledge, that is, their knowledge about causal relationships between events in the environment, to focus on a small and manageable subset of relevant cues. We further assume that causal knowledge might also aid learning of cue validities. In sum, causal knowledge might allow decision makers to deal adaptively with the huge number of individual and compound cues that appear in the environment by directing them to those that are potentially relevant. In the remainder of this paper, we offer more precise predictions about how causal knowledge can influence decision-making processes and review several experiments and a simulation study, all conducted within the fast-and-frugal heuristic framework, in which these predictions were tested. The adaptive value of knowledge about the causal texture of the environment The adaptive importance of causal processing has been stressed by many authors in a wide range of disciplines (see Gopnik & Schulz, in press, for a review), including computer science (e.g., Pearl, 2000), philosophy (Glymour, 1998; Harre & Madden, 1975; Hume, 1987; Kant, 1965; Mackie, 1974; White, 1995), developmental 123 368 R. Garcia-Retamero, U. Hoffrage psychology (Gopnik et al., 2004; Koslowski & Masnick, 2002; Schlottmann, 1999), and cognitive psychology (Ahn & Kalish, 2000; Cheng, 1997; Garcia-Retamero, in press; Waldmann, Holyoak, & Fratianne, 1995). We have a general tendency to consider connections between events in terms of causal relationships, and we understand, predict, and control our environment by positing underlying causal mechanisms that generate our sensory experience (Lagnado & Sloman, 2004). When it is said that a cause brings about an effect, the implication is that there is a stable causal link between the cause and the effect and an underlying causal mechanism that is an essential property of this link (Ahn & Kalish, 2000; Glymour & Cheng, 1999). Such a link between a cause and an effect goes beyond the mere covariation between them as the cause produces the effect (Cheng, 1997; Novick & Cheng, 2004). Having causal knowledge about the cues in the environment is adaptive for individuals because it allows them to make predictions about future events and to intervene in the ecology to bring about new events (Gopnik et al., 2004; Lagnado & Sloman, 2004; Steyvers, Tenenbaum, Wagenmaker, & Blum, 2003; Waldmann & Hagmayer, 2005). Therefore, it is quite conceivable that natural selective pressures have, over the course of evolution, established some genetic basis for causal thinking (Shultz, 1982). Causal beliefs are not isolated but are tightly connected with other causal beliefs in a broad base of knowledge that represents the causal structure of the environment, henceforth referred to as a causal mental model. There are two main approaches in the psychological literature to explain how causal links between events can be inferred. The bottom-up approach assumes that observing or experiencing correlations among events could help in the generation of these causal links or in the adjustment of existing ones (Cheng, 1997; Glymour & Cheng, 1999; Gopnik et al., 2004; Koslowski & Masnick, 2002; Novick & Cheng, 2004; Shanks & Dickinson, 1987; Spellman, 1996). The top down approach was advanced by Waldmann (Waldmann & Holyoak, 1992; Waldmann et al., 1995; see also Ahn & Kalish, 2000; Harre & Madden, 1975; White, 1995), who argued that people s abstract knowledge about causality (such as causal directionality) shapes how data are interpreted. When it comes to decision-making, we posit that causal knowledge is advanta- geous for three reasons. First, causal knowledge might act as a meta-cue for iden- tifying valid cues in the environment. Second, causal knowledge might help us focus on certain cues, which, in turn, facilitates learning of cue validities. Third, causal knowledge might guide people in the selection of the relevant compound cues in the environment. We now elaborate on each of these three advantages in more detail. Considering the first advantage, we hypothesize that cues that are causally linked to a criterion tend to be more valid than other cues lacking such a connection to the criterion (Garcia-Retamero, Wallin, & Dieckmann, 2006; see also Ahn & Kalish, 2000; Wallin & Gärdenfors, 2000). For instance, lung cancer (here, an effect) is more likely to be predicted from a well-established smoking habit (i.e., a cause) than from yellowed fingers (i.e., a second effect of the common cause; see Boyle, 1997). Fur- thermore, correlations between events that are causally linked are likely to be more robust across environments, that is, less sensitive to contextual changes, than those without such a connection (Pearl, 2000; Reichenbach, 1956). Following our example, the correlation between smoking and lung cancer would be more robust across different series of patients than the correlation between lung cancer and yellowed fingers would be. We could expect this to be the case even if we control for other alternative causes that could bring about yellowed fingers (e.g., being a painter) that 123 Causal knowledge and simple heuristics 369 might reduce their predictability for lung cancer. We hypothesize that this asym- metry between causal and non-causal cues that holds in the physical world would be reflected in human cognitive processes. We therefore expect decision makers to use their causal knowledge as a meta-cue for selecting highly valid and robust cues in the environment. Besides facilitating the selection of valid cues in the environment, causal knowledge might reduce the number of cue criterion correlations to keep track of when computing cue validities (Garcia-Retamero et al., 2006). This hypothesis is supported by research using multiple cue probability learning (MCPL) tasks. In this paradigm, participants have to predict the criterion of a given object from multiple cues that are probabilistically related to this criterion. Previous empirical studies using this paradigm (see Kruschke & Johansen, 1999, for a review) suggest that there exists interference effects when multiple cues are available and, consequently, cue validities have to be learned concurrently. For instance, if irrelevant cues are present in such a task, the utilization of valid cues is reduced and, consequently, the accuracy of people s judgments is lower as compared to a condition in which these irrelevant cues are not included (Castellan, 1973; Edgell & Hennessey, 1980). One can explain this finding, which can be observed even after a large number of learning trials, by assuming that in the condition with the irrelevant cues it is harder for participants to identify and focus on the valid cues. In contrast, when participants can learn cue criterion relationships sequentially, that is, for one cue after another, their judgments more closely correspond to the ecological correlations (Brehmer, 1973). Based on this finding we suggest that in multiple-cue settings people equipped with causal knowledge might be able to use this knowledge to focus on certain cues, which, in turn, might facilitate learning of cue validities. Note, however, that causal knowledge about the cues in the environment also has to be learned. Therefore, our argument holds only if the acquisition of causal knowledge is simpler than the learning of cue validities. We think that this is, in fact, the case. Consider, for instance, learning of causal Bayes nets. Such learning is certainly not necessarily simple, but it could be simplified if prior specific or abstract domain knowledge about the structure of the environment (e.g., causal direction- ality) constrains the number of potential causal relations that need to be considered (see Tenenbaum, Griffiths, & Niyogi, in press; Waldmann, 1996; Waldmann & Martignon, 1998).1 The third advantage of causal knowledge is that it can guide people in the selection of the relevant compound cues in the environment that should be repre- sented as configurations (see Garcia-Retamero, Hoffrage, Dieckmann, & Ramos, in press b). Specifically, we hypothesize that when people perceive several cues to act through a common causal mechanism in bringing about an effect, they will consider the possibility that these cues might also interact with each other in bringing about that effect. This possibility may lead them to check whether the accuracy of their predictions would be increased through representing these cues as a configuration. 1 Along these lines, research in the field of artificial intelligence has recently proposed a number of algorithms capable of easily inferring causal relations from covariation patterns (e.g., the TETRAD II program; Spirtes, Glymour, & Scheines, 1993, 2000). These algorithms use causal models to generate a certain pattern of statistical dependencies and then search for certain clues that reveal fragments of the underlying structure. These fragments are pieced together to form a coherent causal model. Obviously, these systems do not provide information about how humans learn causal links, but they do tell us how such a task might be solved. 123 370 R. Garcia-Retamero, U. Hoffrage This might be the case in the medicine/alcohol example mentioned above. Specifi- cally, the effect of a medicine would be expected to be modulated by the ingestion of alcohol with both agents being taken up in the blood, and only the presence of both might cause nausea. Therefore, processing a compound cue as a configuration is an emergent property of a causal mental model of the environment when the compound cue s components are perceived to act jointly through the same underlying causal mechanism. As configural processing requires extra cognitive effort, we expect that when cues are perceived to act through different causal mechanisms, they will be represented as several individual elements. This might be the case if we have to predict the speed of a cyclist while taking tailwind and physical strength into ac- count. These cues would be expected to affect the criterion independently. In short, we hypothesize that causal knowledge might allow decision makers to constrain the countless number of combinations of cues that appear in a particular environment to a subset for which it is more likely that configurations have a high predictive value. In the following sections, we review some experiments conducted to test whether causal knowledge helps people to select a subset of reliable (indi- vidual or compound) cues and aids learning of cue validities. Causal knowledge helps us select cues in the environment In two experiments, Garcia-Retamero et al. (2006) examined the impact of pro- viding causal information about cue criterion relationships on decision-making processes. More specifically, these authors analyzed whether causal knowledge about the cues in the environment had an effect on the selection of a subset of cues that were used to make decisions, and whether that knowledge made the compu- tation of cue validities easier. Specifically, we tested the following predictions. Based on our assumption that causal knowledge can help us identify highly valid cues in the environment, we hypothesized that participants would look up cues that were causally connected to the criterion (in short, causal cues) earlier than non- causal cues, even when they had the same validity.2 They would also rely more on causal cues than on non-causal cues in their decisions, and they would be more confident and faster in their decisions when causal cues were available than when no causal cues were available. On the other hand, given that causal knowledge reduces the number of cue criterion relationships to keep track of to compute validity, we hypothesized that participants would be more exact in their validity estimates for causal than for non-causal cues and consequently would also be more accurate in their inferences. The two experiments tested these hypotheses. The first addressed the basic assumption that causal cues are preferred over non-causal cues. The second repre- sented a stronger test by allowing participants to learn the validities of the cues after they had received information about which cues were causally related to the crite- rion. The experiments were computer based and used two-alternative forced-choice tasks. On each trial, participants were presented with two alternatives (i.e., two species of insects) and had to decide which would show a higher criterion value (i.e., 2 For the sake of brevity, we just distinguished between causal and non-causal cues in these experiments, thereby ignoring alternatives such as indirect causal relationships (e.g., causal chains; see Waldmann & Hagmayer, 2001). 123 Causal knowledge and simple heuristics 371 which would do more damage to a crop). To make this decision, they could look up information on up to four cues (i.e., properties of the insects, such as the presence or absence of a particular metabolic factor), represented by little boxes on the screen that could be clicked to retrieve information (see Bröder, 2000, 2003; Newell & Shanks, 2003; Rieskamp & Hoffrage, 1999, for a similar experimental procedure). Two of these cues had, counterbalanced across participants, a high validity (.85) and the other two a low validity (.65). All four cues had a discrimination rate of .56.3 Causal knowledge was manipulated in a between-subjects design. In the causal group, participants were provided with information that causally related two of the cues to the criterion via the experimental instructions (e.g., the metabolic factor makes the insects hungry and aggressive ). For these two cues, the instructions suggested an underlying causal mechanism that went beyond the possible covaria- tion between the cue and the criterion. The remaining two cues were neutral. Here, participants received information that did not link the cue causally to the criterion (e.g., the metabolic factor leads to green and blue coloration of the insects body ). Which cues were causally linked to the criterion and which were neutral was counterbalanced across participants. In the control group, information about all four cues was neutral. A pre-test confirmed that the causal cues, but not the neutral cues, were indeed perceived as having a strong causal effect on the criterion. In Experiment 1, participants went through a decision phase in which information about whether the cues were present or absent was not automatically displayed; instead they had to look up information for one cue after another. When a cue was looked up (at the cost of 1 cent) the cue values of both alternatives were shown. After looking up one cue, participants could stop and decide on one of the alter- natives at any time. After each decision, participants received feedback about whether the decision was correct (if so, they earned 7 cents). At the end of the experiment, participants made an estimation of cue validities. In Experiment 2, before they entered the decision phase, participants went through a learning phase in which the values of the four cues were automatically provided to allow participants to learn cue validities. Note that the actual cue validities were not provided. Otherwise, Experiment 2 was identical to Experiment 1. In line with our hypothesis, the results of Experiment 1 show that when causal information about some of the cues was available, without a separate phase for learning cue validities before decision making, participants preferred to start searching with a causal cue, regardless of its validity. They also followed the causal cues more often than the neutral ones and were faster and more confident in their decisions than when they did not have causal cues to rely on. Finally, participants were also more precise in estimating the validities of the causal cues than those of the neutral ones. Note that, in this experiment, they could check which cues were reliable predictors of the criterion throughout the decision-making phase of the experiment, but they still preferred to rely just on causal cues. When participants had the opportunity to learn about cue validities before the actual decision-making phase (i.e., in Experiment 2), their search processes were influenced by both causal information and validity. Specifically, participants in the causal group preferred to start by looking up the causal high-validity cue over the rest of the cues. Furthermore, these participants became more accurate in their 3 The discrimination rate of a cue is the proportion of pair comparisons in which the value of that cue in the two decision alternatives differs (Gigerenzer & Goldstein, 1996). 123 372 R. Garcia-Retamero, U. Hoffrage decisions. They also achieved, on average (i.e., across all cues), a higher precision in estimating cue validities. Overall, the higher frugality and accuracy in the causal group led to a higher final payoff than in the control group. These results could be interpreted as evidence that participants perceived causal information as indicative of highly valid cues in the environment, as they gave priority to causal cues in both the search and the decision processes. Furthermore, causal knowledge seems to facilitate cue validity learning, as participants in the causal group were more precise in their validity estimations. In sum, our experi- mental results show that information about the causal relations between cues and criterion could be used as a meta-cue for identifying and selecting relevant cues in the environment, helping participants target a manageable subset of cues and, consequently, leading to more accurate learning of cue validities. Causal knowledge helps us process cues as configurations Garcia-Retamero et al. (in press b; see also Garcia-Retamero, Hoffrage, & Dieck- mann, in press a) went one step further and analyzed whether causal knowledge might also guide people in the selection of the relevant compound cues in the environment in order to improve the accuracy of their inferences. Note that fast and frugal heuristics do not take compound cues into account to predict a criterion. However, as we said above, this is sometimes required to make accurate predictions. For example, according to the diathesis-stress model, hereditary predisposition and a current stressor (e.g., a traumatic experience) are assumed to be necessary for the onset of certain diseases such as schizophrenia (Walker & Diforio, 1997) or depression (Abela & Seligman, 2000). Previous studies have shown that people can and do process compound cues in some problems (Edgell, 1993; Shanks, Charles, Darby, & Azmi, 1998; Williams & Braker, 1999). However, none of the fast and frugal heuristics proposed so far by Gigerenzer et al. (1999) can adequately deal with such problems. With this in mind, Garcia-Retamero et al. (in press b) raised the question of whether the repertoire of fast and frugal heuristics could include a strategy that processes compound cues. In particular, they proposed the Take The Best Configural heuristic (TTB Configural for short). Like TTB, TTB Configural is a heuristic in which cue search is ordered according to the cues validities. In contrast to TTB, however, TTB Configural also includes compound cues in the cue search ordering under some conditions. TTB Configural begins searching with the cue that has the highest validity be it an elemental or a compound cue. If the cue search ordering requires that a compound cue be looked up, TTB Configural looks up the cues that jointly constitute that compound. Correspondingly, the stopping rule entails that as soon as a cue (either individual or compound) discriminates, search for further information is stopped. As far as the decision rule is concerned, TTB Configural decides in favor of the alternative to which the cue (be it an individual or a com- pound cue) points. We posit that the use of TTB Configural is supported by causal knowledge be- cause such knowledge may help to identify the relevant compound cues in the environment. Specifically, when the decision maker expects that several cues act through a common causal mechanism in bringing about the criterion, such cues will 123 Causal knowledge and simple heuristics 373 be represented as a compound and subsequently included in the cue search order of cues according to validity.4 Gigerenzer et al. (1999) pointed out that the specific decision-making strategy that is selected to be used should be contingent on the structure of the environment (see also Rieskamp & Otto, 2006). One issue in the literature on configural strategy use that has generated special interest is whether linearly separable environments are easier to learn than non-linearly separable ones, which indeed seems to be the case (see Kimmel & Lachnit, 1991; Lachnit & Kimmel, 1993; Smith, Murray, & Minda, 1997; but see Shanks et al., 1998). Bearing these findings in mind, Garcia- Retamero et al. (in press a, in press b) hypothesized that the type of environment that participants receive in the experimental task (i.e., linearly versus non-linearly separable) would have an impact on the probability that TTB Configural is used. One non-linearly separable problem that has generated special interest is the eXclusive-OR (or XOR, for short) problem (Medin & Schwanenflugel, 1981; Rumelhart, Hinton, & Williams, 1986). In an environment in which two cues, A and B, are amalgamated into a compound that obeys the XOR logical rule, an object for which one (and only one) of these cues is present is more likely to have a higher criterion value than an object for which either both or neither of these cues is present (i.e., A1B0, A0B1 > A1B1, A0B0, where 1 denotes that the cue is present, 0 denotes that the cue is absent, and > denotes the expectation of a higher criterion value). This problem is non-linearly separable as it cannot be solved by a strategy that predicts an outcome based on a linear function of the weighted cues without an interaction term (Zurada, 1992). In contrast, the AND problem has a linearly sep- arable structure. In an environment in which two cues are amalgamated into a compound that obeys an AND logical rule, an object for which these cues are present is more likely to have a higher criterion value than an object for which only one or neither of these cues is present (i.e., A1B1 > A1B0, A0B1, A0B0). Garcia-Retamero et al. (in press a, in press b) were interested in what effect the type of environment that participants encountered had on participants strategy selection, and how the causal mental model with which they approached these environments influenced the type of strategy they used. In these experiments, the two-alternative forced-choice tasks were framed as a medical diagnostic task. Spe- cifically, participants were presented with information about two patients and had to choose the one with the higher body temperature. The information provided about the patients was whether they had ingested three different substances (referred to as A, B, and C, respectively). In the first phase of the experiment, participants were provided with information about all three cues in a given comparison at no cost. They then had to make a decision. Outcome feedback about the correct option was given in this phase to enable participants to learn the cues validities. Subsequently, they went through a decision-making phase in which the information was no longer automatically displayed; instead they had to look it up sequentially by clicking boxes on the computer screen to retrieve the information. The acquisition of cue infor- mation was costly. Again, outcome feedback was provided. In these experiments, Garcia-Retamero et al. (in press a, in press b) used both simple and complex environments, that is, environments that are best handled through TTB and TTB Configural, respectively. Two complex environments were 4 Note that, in this case, both the components and the compound cue would be included in the hierarchy of cues ordered according to their validity. 123 374 R. Garcia-Retamero, U. Hoffrage generated in which two of the three cues (i.e., A and B) were amalgamated into a highly valid compound cue that obeyed either the XOR or the AND logical rule. In these environments, the validity of the critical compound, AB, was 1.00, but the validity of each of its component cues, A and B, was .50, thus corresponding to pure guessing. The third cue, C, which was not included in the compound, had a validity of .75. Consequently, these environments allowed high separability between the strategies being studied. Whereas participants using TTB should start by looking up cue C (the only individual highly valid cue in these environments) and should decide on the basis of this cue whenever it discriminates, participants using TTB Configural should start by looking up cues A and B first and should decide on the basis of the compound AB whenever it discriminates. A simple environment was also generated. In this environment, the validity of the compound AB was .50, as was the validity of each of its component cues. Just as in the complex environments, the validity of cue C was .75. That is, this was the only individual valid cue in this environment. Thus, the difference between the complex and simple environments resided in the validity of the critical compound cue, AB. Garcia-Retamero et al. (in press a, in press b) also manipulated, in a between- subjects design, participants causal mental model of the environment through instructions. Particularly, we induced either a configural, an elemental, or a neutral causal model. In the configural causal model, the instructions emphasized that cues act through the same causal mechanism in bringing about the criterion. We assumed that these instructions would then lead participants to search for valid compound cues in the environment. In the elemental causal model condition, the instructions emphasized that the cues acted through different causal mechanisms to bring about the criterion. We assumed that these instructions would lead participants to process cues as individual elements in the environment. In the third condition, the neutral causal model, participants did not receive any information about the possible causal mechanisms through which the cues acted. Therefore, this condition would show how participants searched for information and how they decided spontaneously. In short, having in mind a configural, an elemental, or a neutral causal model, partic- ipants received an XOR, an AND, or a simple environment (see Table 1). Garcia-Retamero et al. (in press a, in press b) hypothesized that the configural causal model would facilitate the detection of the highly valid compound cue in the environment. Therefore, we expected that a high percentage of participants would use the TTB Configural heuristic when there was a valid compound cue in the environment and their causal knowledge about the cues hinted toward a common causal mechanism. However, for the elemental causal model condition (i.e., when participants were told that cues acted through different causal mechanisms), we hypothesized that most of the participants would behave elementally using the TTB heuristic. We expected this to be the case even if there was a perfectly valid Table 1 Design and experimental hypotheses of the experiment conducted by Garcia-Retamero et al. (in press a, in press b). TTB denotes Take The Best; TTB Configural denotes Take The Best Configural Configural causal model Elemental causal model Neutral causal model XOR environment TTB Configural TTB TTB AND environment TTB Configural TTB TTB Configural Simple environment TTB TTB TTB 123 Causal knowledge and simple heuristics 375 compound cue in the environment. For the neutral causal model condition, we wondered whether we would find evidence for spontaneous TTB Configural use (however, this applies only to the linearly separable environment, because a linearly separable structure is easier to learn than a non-linearly separable one; Lachnit & Kimmel, 1993; Smith et al., 1997). Finally, when there were no highly valid com- pound cues in the environment, we expected that a high percentage of participants would behave elementally using TTB, regardless of their causal mental model. To classify participants according to a particular strategy in our experiments, we used the Bayesian method for multiple-attribute decision-making proposed by Bröder and Schiffer (2003a). In accordance with our hypotheses, results in the experiment show that a high percentage of participants decided according to the highly valid compound cue, using the TTB Configural heuristic, when they had a configural causal mental model of the environment, that is, when causal knowledge suggested that the component cues acted through a common causal mechanism (see Table 2). Interestingly, this result was found regardless of whether the component cues were amalgamated into a compound by applying a non-linearly or a linearly separable logical rule. However, we found more evidence of TTB Configural use in the linearly separable structures than in the non-linearly separable ones. Additionally, even if there was a perfectly valid compound in the environment, a high percentage of participants represented its component cues as independent elements and used the elemental TTB heuristic when (1) these cues were said to act through different causal mechanisms (in the elemental causal model condition), and also when (2) they had no causal knowledge about the cues in the environment (in the neutral causal model condition). This result was also found regardless of whether the component cues were amalgamated into a compound by applying a non- linearly or a linearly separable logical rule. Finally, when there was no highly valid compound cue in the environment, that is, in the control condition, the elemental TTB was also the most frequently used strategy. In short, results in these experiments suggest that TTB Configural was used only when the information structure in the environment and in the mind fit together, that is, when the causal knowledge about the cues induced participants to search for highly valid compound cues in the environment and a highly valid compound cue existed. When either of these requirements was not met, the elemental TTB was the best behavioral model. Consequently, these results support the hypothesis that causal knowledge could also act as a meta-cue for identifying highly valid compound cues in the environment. These results are in line with research in the categorization arena showing that providing participants with a hint that encouraged the additive integration of features greatly facilitated learning of linearly separable categories Table 2 Results of the experiments conducted by Garcia-Retamero et al. (in press a, in press b). The numbers in the cells denote the percentage of participants who were classified according to TTB Configural versus TTB using the Bayesian method for multiple-attribute decision-making suggested by Bröder and Schiffer (2003a) Configural causal model Elemental causal model Neutral causal model XOR environment 50% vs. 17% 8% vs. 58% 8% vs. 67% AND environment 67% vs. 25% 8% vs. 50% 17% vs. 42% Simple environment 0% vs. 67% 0% vs. 100% 0% vs. 83% 123 376 R. Garcia-Retamero, U. Hoffrage compared to non-linearly separable categories. When, in contrast, a hint induced encoding compatible with non-linearly separable categories, then these categories were easier to learn than linearly separable ones (Wattenmaker, Dewey, Murphy, & Medin, 1986; see also Waldmann et al., 1995; Wisniewski, 1995). Furthermore, in a simulation study, Hoffrage, Garcia-Retamero, and Cziens- kowski (2005) evaluated the performance of three decision-making strategies in several environments: TTB, TTB Configural, and TTB-All (i.e., a strategy that processes all possible compound cues in the environment). Interestingly, results in these simulations show that TTB-All was not competitive in cross-validation.5 TTB- Configural was the most robust strategy, even though TTB-All also processes all the relevant compounds in the environment. General conclusions The results of the experiments reviewed above (Garcia-Retamero et al., in press a, in press b). Show that knowledge about the causal structure of the environment helped people to focus on a small and manageable subset of cues. Specifically, such knowledge influenced which cues were looked up, in which order these were looked up, and which of them were used to make decisions. Causal knowledge also facili- tated the learning of cue validities not an easy task, as Juslin and Persson (2002) pointed out. Finally, causal knowledge guided participants in the selection of highly valid compound cues in the environment to improve the accuracy of their decisions. Taken together, these findings suggest that causal knowledge can effectively reduce the computational complexity inherent in even relatively simple decision-making tasks. Seen through the lens of the fast and frugal heuristics framework, causal knowledge helps people select valid cues in the environment which might be placed in a high position in the cue ordering, that is, in the hierarchy of cues that is accessed by the search process of a decision-making strategy. To the extent that the feedback about whether a decision was correct or incorrect leads to an updating of cue validities, the cue ordering might consequently be updated as well. In this sense, causal beliefs can be perceived as hypotheses to be tested and updated with empirical data (see also Koslowski, 1996; Koslowski & Masnick, 2002). For instance, some of the selected causal cues might turn out to be highly valid cues. Such cues would remain at a high position in the cue ordering. Consequently, the expected usefulness of the causal beliefs that link those cues to the criterion would be strengthened. However, other causal cues might turn out to have low validity. The expected usefulness of the causal belief that links those cues to the criterion would therefore be reduced, and those cues would end up in a lower position in the cue ordering. Briefly, causal beliefs might act as hypotheses that constrain the cues that are selected to make decisions, and these beliefs are subsequently confirmed or disconfirmed based on the experience with the selected cues in the environment. From the results of our experiments, we can infer that people sketch an incre- mentally constructed picture of the environment to make decisions. That is, unless 5 Cross-validation refers to the analysis of the accuracy of a decision strategy when, after fitting it to one part of a data set (training set), it is applied to the other part (test set). 123 Causal knowledge and simple heuristics 377 the decision maker has causal knowledge indicating she should do otherwise, she would start with a simple representation of the environment, that is, an elemental representation where cues are treated as individual elements (see also Cheng, 2000; Novick & Cheng, 2004; Waldmann & Martignon, 1998). Spontaneously, elemental strategies such as the TTB heuristic would be used. The specific strategy that is selected for decision making would be contingent on the environmental structure (see Gigerenzer et al., 1999; Rieskamp & Otto, 2006; Todd & Gigerenzer, 2000). If cues in the environment differ in validity, and accurate decisions are crucial, the decision maker could use her knowledge about the causal structure of that envi- ronment to focus on the highly valid cues and compute their validities. As the internal representation of the environment becomes more complicated, a more complex strategy that processes compound cues would be used to make decisions. This would occur only when there is a match between the mind and the environment, that is, when the components of such a compound are perceived to act through a common causal mechanism to bring about the criterion and the compound is indeed highly valid in the physical world. In this way, fast and frugal heuristics can also take causally interacting cues into account and exploit an even wider range of information structures to make adaptive decisions, thereby allowing more refined and higher- order knowledge to be used for decision-making. Are our conclusions about the beneficial effect of causal knowledge restricted to the family of fast and frugal heuristics? Our intuition is that the present approach might also be extended to other strategies. Causal knowledge possibly could also help to reduce the computational complexity inherent in more demanding strategies for making decisions such as the weighted additive model (WADD) which is a compensatory strategy that uses cue validities as weights (Martignon & Hoffrage, 2002). However, contrary to fast and frugal heuristics, WADD and other compensatory strategies do not model the search process. That is, they strictly assume that all the relevant and necessary information to make decisions is provided in the task. Yet in the real world, as we mentioned above, this is not, in fact, the case. We find it difficult to see how people using such compensatory strategies could use their causal knowledge to select from the wide range of alternatives in the environment those cues (individual or compounded) that are highly valid. If the assumption of cue search and selection is dropped, how would causal knowledge aid learning of cue validities? Briefly, simplification is not an inherent feature of these decision models. Consequently, in their present form, they could not benefit from the advantages of causal knowledge we pointed out above. Causal knowledge modulates decision-making processes. Not providing such knowledge in an experiment will make decision makers appear less competent than they would be in their natural environment in which such information is frequently available. The presence of causal knowledge is vital as it directs the search for information, facilitates the learning of cue validities, and improves decision accu- racy. Acknowledgements We thank Gerd Gigerenzer and Peter Todd for their helpful discussion of our results. We are deeply indebted to Chris White for his helpful comments on early drafts of the present paper. Finally, we also thank Anita Todd for editing the manuscript. 123 378 R. Garcia-Retamero, U. Hoffrage References Abela, J. R. Z., & Seligman, M. E. P. (2000). The hopelessness theory of depression: A test of the diathesis-stress component in the interpersonal and achievement domains. Cognitive Therapy & Research, 24, 361 378. Ahn, W., & Kalish, C. W. (2000). The role of mechanism beliefs in causal reasoning. In F. C. Keil & R. A. Wilson (Eds.), Explanation and cognition (pp. 199 225). Cambridge, MA: MIT Press. Boyle, P. (1997). Cancer, cigarette smoking and premature death in Europe: A review including the recommendations of European cancer experts consensus meeting, Helsinki, October 1996. Lung Cancer, 17, 1 60. Brehmer, B. (1973). Note on the relation between single-cue probability learning and multiple-cue probability learning. Organizational Behavior and Human Performance, 9, 246 252. Bröder, A. (2000). Assessing the empirical validity of the Take-The-Best heuristic as a model of human probabilistic inference. Journal of Experimental Psychology: Learning, Memory, & Cognition, 26, 1332 1346. Bröder, A. (2003). Decision making with the Adaptive Toolbox : Influence of environmental structure, intelligence, and working memory load. Journal of Experimental Psychology: Learn- ing, Memory, & Cognition, 29, 611 625. Bröder, A., & Schiffer, S. (2003a). Bayesian strategy assessment in multi-attribute decision making. Journal of Behavioral Decision Making, 16, 193 213. Bröder, A., & Schiffer, S. (2003b). Take The Best versus simultaneous feature matching: Probabi- listic inferences from memory and effects of representation format. Journal of Experimental Psychology: General, 132, 277 293. Castellan, N. J. (1973). Multiple-cue probability learning with irrelevant cues. Organizational Behavior and Human Performance, 9, 16 29. Cheng, P. W. (1997). From covariation to causation: A causal power theory. Psychological Review, 104, 367 405. Cheng, P. W. (2000). Causality in the mind: Estimating contextual and conjunctive causal power. In F. C. Keil & R. A. Wilson (Eds.), Explanation and cognition (pp. 227 253). Cambridge, MA: MIT Press. Dieckmann, A., & Todd, P. M. (2004). Simple ways to construct search orders. Proceedings of the 26th Annual conference of the cognitive science society (pp. 309-314). Mahwah, NJ: Erlbaum. Edgell, S. E. (1993). Using configural and dimensional information. In N. J. Castellan Jr. (Ed.), Individual and group decision making processes (pp. 43 64). Hillsdale, NJ: Erlbaum Edgell, S. E., & Hennessey, J. E. (1980). Irrelevant information and utilization of event base rates in nonmetric multiple-cue probability learning. Organizational Behavior and Human Performance, 26, 1 6. Garcia-Retamero, R. (in press). The influence of knowledge about causal mechanisms on compound processing. The Psychological Record. Garcia-Retamero, R., Hoffrage, U., & Dieckmann, A. (in press a). When one cue is not enough: Combining fast and frugal heuristics with compound cue processing. The Quarterly Journal of Experimental Psychology. Garcia-Retamero, R., Hoffrage, U., Dieckmann, A., & Ramos, M. (in press b). Compound cue processing within the fast and frugal heuristics approach in nonlinearly separable environments. Learning and Motivation. Garcia-Retamero, R., Takezawa, M., & Gigerenzer, G. (2006). How to learn good cue orders: When social learning benefits simple heuristics. In R. Sun, & N, Miyake (Eds.), Proceedings of the 28th annual conference of the cognitive science society (pp. 1352 1358). Mahwah, New Jersey, USA. Garcia-Retamero, R., Wallin, A., & Dieckmann, A. (2006). Does causal knowledge help us be faster and more frugal in our decisions? Manuscript submitted for publication. Gigerenzer, G., & Goldstein, D. G. (1996). Reasoning the fast and frugal way: Models of bounded rationality. Psychological Review, 103, 650 669. Gigerenzer, G., & Goldstein, D. G. (1999). Betting on one good reason: The Take The Best heuristic. In G. Gigerenzer, P. M. Todd, & the ABC Research Group (Eds.), Simple Heuristics That Make Us Smart (pp. 75 95). New York: Oxford University Press. Gigerenzer, G., Todd, P. M., & the ABC Research Group (1999). Simple Heuristics That Make Us Smart. New York: Oxford University Press. Glymour, C. (1998). Learning causes: Psychological explanations of causal explanation. Minds & Machines, 8, 39 60. 123 Causal knowledge and simple heuristics 379 Glymour, C., & Cheng, P. W. (1999). Causal mechanism and probability: A normative approach. In K. Oaksford & N. Chater (Eds.), Rational models of cognition (pp. 295 313). Oxford: Oxford University Press. Gopnik, A., Glymour, C., Sobel, D. M., Schulz, L. E., Kushnir, T., & Danks, D. (2004). A theory of causal learning in children: Causal maps and bayes nets. Psychological Review, 111, 3 32. Gopnik, A., & Schulz, L. (in press). Causal learning: Psychology, philosophy, and computation. Oxford: Oxford University Press. Harre, R., & Madden, E. H. (1975). Causal powers: A theory of natural necessity. Totowa, NJ: Rowman & Littlefield. Hoffrage, U., Garcia-Retamero, R., & Czienskowski, U. (2005). The robustness of The Take The Best Configural heuristic in linearly and nonlinearly separable environments. In B. G. Bara, L. Barsalou, & M. Bucciarelli, (Eds.), Proceedings of the 27th annual conference of the cognitive science society (pp. 971 976). Mahwah, New Jersey: Lawrence Erlbaum Associates. Hume, D. (1987). A treatise of human nature (2nd ed.). Oxford: Clarendon Press (Original work published 1739). Juslin, P., & Persson, M. (2002). PROBabilities from EXemplars (PROBEX): A lazy algorithm for probabilistic inference from generic knowledge. Cognitive Science, 26, 563 607. Kant, I. (1965). Critique of pure reason. London: Macmillan (Original work published 1781). Kehoe, E. J., & Graham, P. (1988). Summation and configuration: Stimulus compounding and negative patterning in the rabbit. Journal of Experimental Psychology: Animal Behavior Pro- cesses, 14, 320 333. Kimmel, H. D., & Lachnit, H. (1991). Acquisition of a unique cue in positive and negative pat- terning? Integrative Physiological and Behavioral Science, 26, 32 38. Koslowski, B. (1996). Theory and evidence: The development of scientific reasoning. Cambridge, MA: MIT Press. Koslowski, B., & Masnick, A. (2002). The developmental of causal reasoning. In U. Goswami (Ed.), Blackwell handbook of childhood cognitive development (pp. 257 281). Malden, MA: Blackwell. Kruschke, J. K., & Johansen, M. K. (1999). A model of probabilistic category learning. Journal of Experimental Psychology: Learning, Memory, & Cognition, 25, 1083 1119. Lachnit, H., & Kimmel, H. D. (1993). Positive and negative patterning in human classical skin conductance response conditioning. Animal Learning & Behavior, 21, 314 326. Lagnado, D. A., & Sloman, S. (2004). The advantage of timely intervention. Journal of Experimental Psychology: Learning, Memory, & Cognition, 30, 856 876. Läge, D., Hausmann, D., Christen, S., & Daub, S. (2005). Take The Best: How much do people pay for validity? Manuscript submitted for publication Martignon, L., & Hoffrage, U. (2002). Fast, frugal and fit: Simple heuristics for paired comparison. Theory and Decision, 52, 29 71. Mackie, J. L. (1974). The cement of the universe: A study of causation. Oxford, England: Clarendon Press. Medin, D. L., & Schwanenflugel, P. J. (1981). Linear separability in classification learning. Journal of Experimental Psychology: Human Learning and Memory, 7, 355 368. Newell, B. R., Rakow, T., Weston, N. J., & Shanks, D. R. (2004). Search strategies in decision making: The success of success . Journal of Behavioral Decision Making, 17, 117 137. Newell, B. R., & Shanks, D. R. (2003). Take The Best or look at the rest? Factors influencing one- reason decision making. Journal of Experimental Psychology: Learning, Memory, & Cognition, 29, 53 65. Newell, B. R., Weston, N. J., & Shanks, D. R. (2003). Empirical tests of a fast-and-frugal heuristic: Not everyone Takes-The-Best . Organizational Behavior and Human Decision Processes, 91, 82 96. Novick, L. R., & Cheng, P. W. (2004). Assessing interactive causal influence. Psychological Review, 111, 455 485. Pearl, J. (2000). Causality. New York: Oxford University Press. Reichenbach, H. (1956). The direction of time. Berkeley: University of California Press. Rieskamp, J., & Hoffrage, U. (1999). When do people use simple heuristics, and how can we tell? In G. Gigerenzer, P. M. Todd, & the ABC Research Group (Eds.), Simple Heuristics That Make Us Smart (pp. 141 167). New York: Oxford University Press. Rieskamp, J., & Otto, P. E. (2006). SSL: A theory of how people learn to select strategies. Journal of Experimental Psychology: General, 135, 207 236. 123 380 R. Garcia-Retamero, U. Hoffrage Rumelhart, D. R., Hinton, G. E., & Williams, R. J. (1986). Learning internal representations by error propagation. In D. E. Rumelhart, J. L. McClelland, & the PDP Research Group (Eds.), Parallel distributed processing (pp. 318 362), Vol 1. Cambridge, MA: MIT Press. Schlottmann, A. (1999). Seeing it happen and knowing how it works: How children understand the relation between perceptual causality and underlying mechanism. Developmental Psychology, 35, 303 317. Shanks, D. R., Charles, D., Darby, R. J., & Azmi, A. (1998). Configural processes in human asso- ciative learning. Journal of Experimental Psychology: Learning, Memory, & Cognition, 24, 1353 1378. Shanks, D. R., & Dickinson, A. (1987). Associative accounts of causality judgment. In G. H. Bower (Ed.), The psychology of learning and motivation: Advances in research and theory (pp. 229 261) Vol. 21. San Diego, CA: Academic Press. Shultz, T. R. (1982). Rules of causal attribution. Monographs of the Society for Research in Child Development, 47, 1 51. Smith, J. D., Murray, M. J., & Minda, J. P. (1997). Straight talk about linear separability. Journal of Experimental Psychology: Learning, Memory, & Cognition, 23, 659 680. Spellman, B. A. (1996). Conditionalizing causality. In D. R. Shanks, K. J. Holyoak, & D. L. Medin (Eds.), The psychology of learning and motivation (pp. 167 206) Vol 34. San Diego: Academic Press. Spirtes, P., Glymour, C., & Scheines, R. (1993). Causation, prediction, and search (Springer lecture notes in statistics). New York: Springer-Verlag. Spirtes, P., Glymour, C., & Scheines, R. (2000). Causation, prediction, and search (2nd ed.). Cam- bridge, MA: MIT Press. Steyvers, M., Tenenbaum, J. B., Wagenmakers, E. J., & Blum, B. (2003). Inferring causal networks from observations and interventions. Cognitive Science, 27, 453 489. Tenenbaum, J. B., Griffiths, T. L., & Niyogi, S. (in press). Intuitive theories as grammars for causal inference. In A. Gopnik & L. Schulz (Eds.), Causal learning: Psychology, philosophy, and computation. Oxford: Oxford University Press. Todd, P. M., & Gigerenzer, G. (2000). Précis of simple heuristics that make us smart. Behavioral & Brain Sciences, 23, 727 780. Waldmann, M. R. (1996). Knowledge-based causal induction. In D. R. Shanks, K. J. Holyoak, & D. L. Medin (Eds.), The psychology of learning and motivation (pp. 47 88) Vol. 34. San Diego, CA: Academic Press. Waldmann, M. R., & Hagmayer, Y. (2001). Estimating causal strength: The role of structural knowledge and processing effort. Cognition, 82, 27 58. Waldmann, M. R., & Hagmayer, Y. (2005). Seeing versus doing: Two models of accessing causal knowledge. Journal of Experimental Psychology: Learning, Memory, & Cognition, 31, 216 227. Waldmann, M. R., Holyoak, K. J. (1992). Predictive and diagnostic learning within causal models: Asymmetries in cue competition. Journal of Experimental Psychology: General, 121, 222 236. Waldmann, M. R., Holyoak, K. J., & Fratianne, A. (1995). Causal models and the acquisition of category structure. Journal of Experimental Psychology: General, 124, 181 206. Waldmann, M. R., & Martignon, L. (1998). A Bayesian network model of causal learning. In M. A. Gernsbacher, & S. J. Derry (Eds.), Proceedings of the 20th annual conference of the cognitive science society (pp. 1102 1107). Mahwah, NJ: Erlbaum. Walker, E. F., & Diforio, D. (1997). Schizophrenia: A neural diathesis-stress model. Psychological Review, 104, 667 685. Wallin, A., & Gärdenfors, P. (2000). Smart people who make simple heuristics work. Behavioral and Brain Sciences, 23, 765. Wattenmaker, W. D., Dewey, G. I., Murphy, T. D., & Medin, D. M. (1986). Linear separability and concept learning: Context, relational properties, and concept naturalness. Cognitive Psychology, 18, 158 194. White, P. A. (1995). Use of prior beliefs in the assignment of causal roles: Causal powers versus regularity-based accounts. Memory & Cognition, 23, 243 254. Williams, D. A., & Braker, D. S. (1999). Influence of past experience on the coding of compound stimuli. Journal of Experimental Psychology: Animal Behavior Processes, 25, 461 474. Wisniewski, E. J. (1995). Prior knowledge and functionally relevant features in concept learning. Journal of Experimental Psychology: Learning, Memory, & Cognition, 21, 449 468. Zurada, J. M. (1992). Introduction to artificial neural systems. New York: West. 123