Rational Behavior in Rats
Aaron P. Blaisdell is an Associate Professor of Psychology at the University of California, Los Angeles, a member of the UCLA Brain Research Institute, and an APA Fellow, Division 6. He received his B.A. in Anthropology in 1991 at the State University of New York, Stony Brook, his M.A. in Anthropology in 1996 at Kent State University, his Ph.D. in Psychology in 1999 at the State University of New York, Binghamton under the mentorship of Ralph R. Miller, and was an NRSA postdoctoral fellow in Robert Cook’s lab at Tufts University. His research concerns the comparative analysis of learning and cognition in pigeons, rats, and humans, with primary interests in spatial and causal cognition, behavioral variability, and problem solving behavior. His work is currently funded by the National Science Foundation and the National Institute of Neurological Disorders and Stroke.
Science is dedicated to increasing our understanding of our incredibly complex universe. Modeling the universe is a daunting task given its sheer complexity; and hidden variables abound. To simplify the task, the scientist attempts to model only a tiny piece of the world at a time to discover how it works. Scientists can test their models by manipulating key variables of the model while holding the remaining variables constant—that is, they conduct controlled experiments. This approach, termed the scientific method, allows us to continue to push the boundaries of our knowledge. This approach is predicated on a causal interpretation of the world. The universe and everything in it, including our own private experiences, is made up of a series of causally-connected events (Mackie, 1974). Humans are successful agents precisely because our causal representations mirror the causal texture of the world (Blaisdell, 2008; Waldmann, Cheng, Hagmayer, & Blaisdell, 2008).
As a comparative psychologist, I’m interested in how well other species approximate natural causal agents and by what mechanisms they do so. Before addressing my own research on these questions I first briefly review contemporary research on the processes of human causal cognition.
In his penetrating analysis, David Hume (1748/1977) concluded that we do not have direct access to the causal powers underlying our world. Instead, he claimed that—although the causality of the world is real—we can only infer causal knowledge from patterns of observed relationships between events. Many contemporary learning theorists have adopted Hume’s empirical approach to causal learning. Associations derived from observed contingencies, such as through Pavlovian and instrumental conditioning, serve as the basis for causal predictions (e.g., Allan, 1993; Shanks & Dickinson, 1987). There is no need to resort to concepts such as causal power or mechanism in accounts of causal learning in humans or other animals.
Recent challenges from philosophy, statistics, and psychology argue that we can go beyond the information given (i.e., contingency) by entertaining hypothetical causal models that we can use to dissect cause-effect relationships using our own actions (i.e., interventions) on the world (see Cheng, 1997; Glymour, 2001; Gopnik et al., 2004; Pearl, 2000; Sloman, 2005; Spirtes, Glymour, & Scheines, 1993; Waldmann & Holyoak, 1992; Woodward, 2003). Just as physicists can test models of how the world works, we can use causal models to make predictions about possible causal relationships which can be tested experimentally through interventions (Buehner & Cheng, 2005; Buehner, Cheng, & Clifford, 2003; Liljeholm & Cheng, 2007; Waldmann, 1996, 2000, 2001; Waldmann, Cheng et al., 2008; Waldmann & Hagmayer, 2005; Wu & Cheng, 1999). Even young children appear to naturally go beyond contingency information and construct hypothetical causal models which they systematically test and refine through manipulation and observation (Gopnik et al, 2004; Schulz, Gopnik, & Glymour, 2007). Thus, causal model theory provides a top-down approach to causal induction (i.e., learning) which can be contrasted with bottom-up approaches (Waldmann & Holyoak, 1992). There is mounting empirical support that causal reasoning in humans goes beyond mere associative learning and instead favors causal model theory.
Interventions play a special role in deriving causal inferences. For example, suppose we observe a change in the level of a barometer. We also expect to observe a concomitant change in the weather. This is because the state of the barometer and the weather are both the direct result of changes in atmospheric pressure (Figure 1, left panel). Given this causal knowledge, we would not expect the weather to change if we intervened and tampered with the barometer, thereby artificially altering its reading (Figure 1, right panel). We realize that such a change in the barometer is attributable to our actions and not to changes in air pressure. The difference between reasoning by interventions and by observations displays the fundamental power of causal model theory (Waldmann & Hagmayer, 2005; Waldmann, Hagmayer, & Blaisdell, 2006; Waldmann, Cheng et al., 2008). An intervention renders the variable intervened on independent of its normal causes because those causes are no longer setting it – the intervention is (Pearl, 2000; Sloman & Hagmayer, 2006; Woodward, 2003).
Figure 1. Observing an effect (left) versus intervening in an effect (right) of a common cause: While an observation of an effect allows inferring the presence of its cause, an intervention in the same variable renders this variable independent of its cause. See text for details.
As a comparative psychologist, I wondered to what extent other species engage in similar types of causal inferences. I review evidence from my laboratory suggesting that rats do engage in some forms of causal reasoning that appear to go beyond contemporary associative accounts and closely approximate the predictions of causal model theory (see also Sawa, 2009; Waldmann, Hagmayer, & Blaisdell, 2006; Waldmann, Cheng, et al., 2008).
Seeing versus Doing in Rats
Our first experiment asked whether, like humans, rats can also draw causal inferences from their own interventions (Blaisdell, Sawa, Leising, & Waldmann, 2006). Rats were placed in conditioning chambers where they were trained on various causal models between events (audiovisual cues and food). Rats received Pavlovian pairings of a Light with Tone (LightTone) and Light with Food (LightFood) to teach the rats a common-cause model in which a Light was a common cause of both Tone and Food (left panel of Figure 2 – analogous to how air pressure is a common cause to both changes in the barometer and changes in the weather, Figure 1). All rats also received separate Pavlovian pairings of a Click with Food to teach them that the Click was a direct cause of Food. After training, some rats were tested on the Tone presented alone. Rats in the Observe test condition received presentations of the Tone periodically during each test session. Figure 2 shows that these rats had a high expectation of food during the Tone as indicated by the high rate of nose poking during the Tone at test. This behavior is consistent with the view that the rats accessed a common-cause model to infer from one effect (Tone), through the Light, to the other (Food) (Blaisdell, 2009).
Figure 2. Left Panel: Causal model used in Blaisdell, Sawa, Leising, & Waldmann (2006), Experiment 1. Light is the common cause of Tone and Food. Click is a direct cause of Food. Arrows indicate directionality from Cause to Effect. Right Panel: Mean nose pokes during test trials with the Tone (Common-Cause) or Click (Direct-Cause) when the stimulus was preceded by a lever press (Intervene) or not (Observe), from Blaisdell et al. (2006), Experiment 1.
For rats receiving the Intervene test condition, a novel lever was inserted into the cage during each test session. The rats had never seen this lever before, and had no prior training on the lever. Rats are exploratory beasts and most of them pressed the lever at least a few times in each test session. For rats receiving an Intervene test condition, the Tone was presented every time the rat pressed the lever (rats in the Observe test session had a lever, too, but the lever was non-functional—i.e., pressing the lever did not cause the Tone to come on for the Observe test condition). Thus, rats in the Intervene test condition intervened on the Tone. The nose-poke data from the rats that Intervene on the Tone provide the crucial test of causal model theory. Causal model theory predicts that subjects should be sensitive to whether the event was merely observed or was produced by an intervention. Thus, a novel intervention onto an event should act to separate that event from its previously established causes, such as when tampering with a barometer results in no prediction that the weather will also change (Sloman & Lagnado, 2005; Waldmann & Hagmayer, 2005). Thus, if the rats understood that their lever-press intervention (and not the Light) produced the Tone on the Intervene test, then they should NOT have expected Food to be available. This hypothesis was confirmed: rats that turned on the Tone through their intervention on the lever nose poked much less than rats receiving the Observation test (Figure 2). The statistical relationship between Tone and Food experienced during training was identical in the Intervene and Observe test conditions, and thus an associative account would predict equally high (or low) rates of nose poking in both conditions. Our results, therefore, are consistent with causal model theory but not with an associative account.
Our results cannot be explained by simple response competition by which lever pressing and nose poking are incompatible responses. Rats that were tested on the Click which was a direct cause of food, rather than the Tone, showed no disruption of nose poking during the Click that was turned on by a lever press intervention (Figure 2, right-hand bars). If pressing the lever merely attenuated nose-poke responding during the Tone due to interference or response competition, then we should expect lever pressing to equally disrupt nose poking during a Tone or a Click. The disruptive effect of the lever-press intervention on nose poking during the Tone but not during the Click supports causal model theory. Interventions should lead to discounting of the causes of the intervened-on event, but not of its effects. If the Click is represented by the rat as a cause of the Food, then intervening on the Click should lead to the expectation of Food. Thus, the pattern of results of Experiment 1 of Blaisdell et al. (2006) is consistent with the predictions of causal model theory but not with those of an associative account.
We also compared the causal inferences drawn from an Intervention on a common-cause model to that of a causal chain. Rats learned a causal chain consisting of Tone-->Light-->Food (Figure 3). When compared to rats that received common-cause training, the lever press intervention had no effect on nose-poke responding during test trials with the Tone when the Tone was a distal cause of Food in the causal chain (Figure 3; Blaisdell et al., 2006, Exp. 2). These results add support to the view that the rats understood the causal relationship between their action and an outcome, because for the Chain groups the Tone was a cause of the Light, which was a cause of Food. It should not have mattered whether the Tone was merely observed or caused by an intervention; the Food should have been expected in either case.
Figure 3. Left Panel: Causal model used in Blaisdell, Sawa, Leising, & Waldmann (2006), Experiment 2. Common Cause: Light is the common cause of Tone and Food. Causal Chain: Tone is a cause of Light which is a cause of Food. Arrows indicate directionality from Cause to Effect. Right Panel: Mean nose pokes during test trials with the Tone after Common-Cause training, Causal Chain training, or when Light was unpaired with Food (Unpaired) when the test Tone was preceded by a lever press (Intervene) or not (Observe), from Blaisdell et al. (2006), Experiment 2 (a and b).
What’s so Special about Actions?
The experiments discussed above provide compelling evidence for causal reasoning in rats that associative theories fail to explain. One key prediction of causal model theory is that an event should serve as an intervention to the extent that is viewed as deterministic and independent of other events (Waldmann et al., 2006). Deterministic causes—such as between switches and lights—are more readily perceived as being causal than are probabilistic causes—such as between sugar consumption and Type 2 diabetes. Likewise, causal status can be most readily determined by causes that act independently of the causal system on which they act (Woodward, 2003). Unlike arbitrary observed events, one’s own actions typically are seen as deterministic and independent, which allows the actor to infer causality after very few learning trials (Blaisdell, 2008). This is why experimental outcomes are always preferable to epidemiological correlations in establishing cause-effect relationships in science and medicine.
Like humans, rats appear to treat their own actions as special (Leising, Wong, Waldmann, & Blaisdell, 2008). We compared the efficacy of an action to that of a salient exogenous cue to serve as an intervention. All rats first received common-cause training as described above. Then rats were allocated to one of three test conditions: Intervene on the Tone, Observe the Tone, or observe a novel Click followed by the Tone (Exogenous Cue condition). Consistent with the predictions of causal model theory, we (Leising et al., 2008, Exps. 1 & 2) found a lever press but not an exogenous cue (the novel Click) to serve as an effective intervention and reduce expectation of Food (left panel of Figure 4). Thus, a lever press serves as a powerful intervention from which the rat can derive causal inferences.
Figure 4. Left Panel: Mean nose pokes during test trials with the Tone after Common-Cause training when the test Tone was preceded by a lever press (Intervene), or not (Observe), or by a novel Click (Exogenous Cue) from Leising, Wong, Waldmann, & Blaisdell (2008), Experiment 1. Right Panel: Mean nose pokes during the first test trial with the Tone after Common-Cause training when the test Tone was preceded by a lever press (Intervene) or not (Observe) from the meta-analysis reported by Leising et al. (2008).
Perhaps the most compelling evidence that rats can reason rationally about their own behavior comes from a meta-analysis reported by Leising et al. (2008). We have a profound sense of causality when we accidentally bump into a table and thereby spill a glass of wine. We do not need multiple observations of this relationship to realize that we caused the wine to spill. It is immediately apparent on the very first instance. Do rats likewise reason similarly about their impromptu effects on the world? To answer this question, we performed a meta-analysis of first-trial test performance of nose pokes across all of the data sets from this report involving common-cause training and Intervene and Observe testing to determine if the effect of an intervention on the expectation of Food is present on the very first test trial. As the right-hand panel of Figure 4 shows, a lever press intervention had a strong affect on nose poking during the Tone the very first time the rat Intervened. Thus, rats show the ability to draw rational inferences about their novel actions prior to any experience with that action and its effects.
What does it all mean? These experiments were motivated by comparing two approaches to causal inference: bottom-up data-driven associative models, and top-down rational models—in particular causal model theory. The evidence from my lab supports causal model theory and adds to the growing evidence for rational behavior in rats and other animals (Beckers, De Houwer, Miller, & Urushihara, 2006; Hurley & Nudds, 2006; Sawa, 2009; Watanabe, Blaisdell, Huber, & Young, 2009; Wheeler, Beckers, & Miller, 2008; but see Penn, Holyoak, & Povinelli, 2008; Penn & Povinelli, 2007). These rational behaviors are very similar to those exhibited by humans, even very young children. We have offered one minimal rational model that would be required to account for all of our rat and related human data (Waldmann et al., 2008), but I’m sure other, better models will be developed.
One caveat: What this evidence explicitly does not mean is that rats are rational agents in the sense of necessarily being aware of causal properties or making conscious decisions about inferences. This may be what is going on in the mind of the rat (or human!), but it is by no means the only interpretation of our data. Actually, Morgan’s canon (Morgan, 1903) invokes us to presume, until proven otherwise, that psychological processes—including causal inferences—operate largely in the absence of explicit awareness in both rats and humans. Moreover, causal inferences must reflect the functions of a neural architecture that is associatively connected.
What our experiments do show, however, is that like humans, rats view the world causally. We are continuing to investigate the cognitive mechanisms underlying causal inference in rats and hope to begin investigating the neural basis of causal inferences as well. This research is still in its infancy and thus the experiments themselves raise many more questions than they can answer. They do, however, lay an important groundwork that will serve as a foundation for the comparative analysis of causal cognition. The fruits of this research can provide insight into the fundamental nature of the processes underlying causal cognition, which can help both to refine the existing models of causal inference and to develop new ones. This work also holds implications for cognitive science and philosophy by helping to distinguish the unique elements of human thought processes from those shared with other species. Knowledge of these similarities and differences can inform our epistemological quest to understand causality, perhaps our greatest tool for exploring our universe.
Allan, L. G. (1993). Human contingency judgments: Rule based or associative? Psychological Bulletin, 114, 435-448.
Beckers, T., De Houwer, J., Miller, R. R., Urushihara, K. (2006). Reasoning rats: Forward blocking in Pavlovian animal conditioning is sensitive to constraints of causal inference. Journal of Experimental Psychology: General, 135, 92-102.
Blaisdell, A. P. (2008). Cognitive dimension of operant learning. In I. H.L. Roediger (Ed.), Cognitive psychology of memory. Vol. 1 of Learning and memory: A comprehensive reference (pp. 173-195). Oxford: Elsevier.
Blaisdell, A. P. (2009). The role of associative processes in spatial, temporal, and causal cognition. In S. Watanabe, A. P. Blaisdell, L. Huber & A. Young (Eds.), Rational animals, irrational humans (pp. 153-172). Tokyo: Keio University Press.
Blaisdell, A. P., Sawa, K., Leising, K. J., & Waldmann, M. R. (2006). Causal reasoning in rats. Science, 311(5763), 1020-1022.
Buehner, M. J., & Cheng, P. W. (2005). Causal learning. In K. J. Holyoak & R. G. Morrison (Eds.), Cambridge handbook of thinking and reasoning (pp. 143-168). Cambridge: Cambridge University Press.
Buehner, M. J., Cheng, P. W., & Clifford, D. (2003). From covariation to causation: a test of the assumption of causal power. Journal of Experimental Psychology: Learning, Memory, & Cognition, 29, 1119-1140.
Cheng, P. W. (1997). From covariation to causation: A causal power theory. Psychological Review, 104, 367-405.
Glymour, C. (2001). The mind's arrows: Bayes nets and graphical causal models in psychology. Cambridge, MA: The MIT 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.
Hume, D. (1748/1977). An enquiry concerning human understanding. Indianapolis: Hackett Publishing Company.
Hurley, S., & Nudds, M. (2006). Rational Animals? Oxford: Oxford University Press.
Leising, K. J., Wong, J., Waldmann, M. R., & Blaisdell, A. P. (2008). The special status of actions in causal reasoning in rats. Journal of Experimental Psychology: General, 137, 514-527.
Liljeholm, M., & Cheng, P. W. (2007). When is a cause the "same"? Coherent generalization across contexts. Psychological Science, 18, 1014-1021.
Mackie, J. L. (1974). The cement of the universe: A study of causation. Oxford: Clarendon Press.
Morgan, C. L. (1903). An introduction to comparative psychology, 2nd edition. London: W. Scott.
Pearl, J. (2000). Causality: Models, reasoning, and inference. Cambridge, U.K.: Cambridge University Press.
Penn, D. C., Holyoak, K. J., & Povinelli, D. J. (2008). Darwin’s mistake: Explaining the discontinuity between human and nonhuman minds. Behavioral and Brain Sciences, 31, 109-130; discussion 130-178.
Penn, D. C., & Povinelli, D. J. (2007). Causal cognition in human and nonhuman animals: A comparative, critical review. Annual Review of Psychology, 58, 97-118.
Sawa, K. (2009). Predictive behavior and causal learning in animals and humans. Japanese Psychological Research, 51, 222-233.
Schulz, L. E., Gopnik, A., & Glymour, C. (2007). Preschool children learn about causal structure from conditional interventions. Developmental Science, 10, 322-332.
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 (Vol. 21, pp. 229-261). New York: Academic Press.
Sloman, S. (2005). Causal models: How people think about the world and its alternatives. Oxford: Oxford University Press.
Sloman, S. A., & Hagmayer, Y. (2006). The causal psycho-logic of choice. Trends in Cognitive Science, 10, 407-412.
Sloman, S. A., & Lagnado, D. A. (2005). Do we "do"? Cognitive Science, 29, 5-39.
Spirtes, P., Glymour, C., & Scheines, R. (1993). Causation, prediction, and search. New York: Springer-Verlag.
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, Vol. 34: Causal learning (pp. 47-88). San Diego: Academic Press.
Waldmann, M. R. (2000). Competition among causes but not effects in predictive and diagnostic learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 53-76.
Waldmann, M. R. (2001). Predictive versus diagnostic causal learning: Evidence from an overshadowing paradigm. Psychonomic Bulletin & Review, 8, 600-608.
Waldmann, M. R., Cheng, P. W., Hagmayer, Y., & Blaisdell, A. P. (2008). Causal learning in rats and humans: A minimal rational model. In N. Chater & M. Oaksford (Eds.), The probabilistic mind: Prospects for rational models of cognition (pp. 453-484). Oxford: Oxford University Press.
Waldmann, M. R., & Hagmayer, Y. (2005). Seeing versus doing: Two modes of accessing causal knowledge. Journal of Experimental Psychology: Learning, Memory, & Cognition, 31, 216-227.
Waldmann, M. R., Hagmayer, Y., & Blaisdell, A. P. (2006). Beyond the information given: Causal models of learning and reasoning. Current Directions in Psychological Science, 15, 307-311.
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.
Watanabe, S., Blaisdell, A. P., Huber, L., & Young, A. (2009). Rational animals, irrational humans. Tokyo: Keio University.
Wheeler, D. S., Beckers, T., & Miller, R. R. (2008). The effect of subadditive pretraining on blocking: Limits on generalization. Learning & Behavior, 36, 341-351.
Woodward, J. (2003). Making things happen: A theory of causal explanation. Oxford: Oxford University Press.
Wu, M., & Cheng, P. W. (1999). Why causation need not follow from statistical association: Boundary conditions for the evaluation of generative and preventive causal powers. Psychological Science, 10, 92-97.