Inferring interventional predictions from observational learning data

Previous research has shown that people are capable of deriving correct predictions for previously unseen actions from passive observations of causal systems (Waldmann & Hagmayer, 2005). However, these studies were limited, since learning data were presented as tabulated data only, which may have turned the task more into a reasoning rather than a learning task. In two experiments, we therefore presented learners with trial-by-trial observational learning input referring to a complex causal model consisting of four events. To test the robustness of the capacity to derive correct observational and interventional inferences, we pitted causal order against the temporal order of learning events. The results show that people are, in principle, capable of deriving correct predictions after purely observational trial-by-trial learning, even with relatively complex causal models. However, conflicting temporal information can impair performance, particularly when the inferences require taking alternative causal pathways into account.     

Associative and causal reasoning accounts of causal induction: Symmetries and asymmetries in predictive and diagnostic inferences

Associative and causal reasoning accounts are probably the two most influential types of accounts of causal reasoning processes. Only causal reasoning accounts predict certain asymmetries between predictive (i.e., reasoning from causes to effects) and diagnostic (i.e., reasoning from effects to causes) inferences regarding cue-interaction phenomena (e.g., the overshadowing effect). In the experiments reported here, we attempted to delimit the conditions under which these asymmetries occur. The results show that unless participants perceived the relevance of causal information to solving the task, predictive and diagnostic inferences were symmetrical. Specifically, Experiments 1A and 1B showed that implicitly stressing the relevance of causal information by having participants review the instructions favored the presence of asymmetries between predictive and diagnostic situations. In addition, Experiment 2 showed that explicitly stressing the relevance of causal information by stating the importance of the causal role of events after the instructions were given also favored the asymmetry.     

The compatibility heuristic in non-categorical hypothetical reasoning: Inferences between conditionals and disjunctions

A new theory explains how people make hypothetical inferences from a premise consistent with several alternatives to a conclusion consistent with several alternatives. The key proposal is that people rely on a heuristic that identifies compatible possibilities. It is tested in 7 experiments that examine inferences between conditionals and disjunctions. Participants accepted inferences between conditionals and inclusive disjunctions when a compatible possibility was immediately available, in their binary judgments that a conclusion followed or not (Experiment 1a) and ternary judgments that included it was not possible to know (Experiment 1b). The compatibility effect was amplified when compatible possibilities were more readily available, e.g., for ‘A only if B’ conditionals (Experiment 2). It was eliminated when compatible possibilities were not available, e.g., for ‘if and only if A B’ bi-conditionals and exclusive disjunctions (Experiment 3). The compatibility heuristic occurs even for inferences based on implicit negation e.g., ‘A or B, therefore if C D’ (Experiment 4), and between universals ‘All A’s are B’s’ and disjunctions (Experiment 5a) and universals and conditionals (Experiment 5b). The implications of the results for alternative theories of the cognitive processes underlying hypothetical deductions are discussed.     

Are Causal Structure and Intervention Judgments Inextricably Linked? A Developmental Study

The application of the formal framework of causal Bayesian Networks to children’s causal learning provides the motivation to examine the link between judgments about the causal structure of a system, and the ability to make inferences about interventions on components of the system. Three experiments examined whether children are able to make correct inferences about interventions on different causal structures. The first two experiments examined whether children’s causal structure and intervention judgments were consistent with one another. In Experiment 1, children aged between 4 and 8 years made causal structure judgments on a three‐component causal system followed by counterfactual intervention judgments. In Experiment 2, children’s causal structure judgments were followed by intervention judgments phrased as future hypotheticals. In Experiment 3, we explicitly told children what the correct causal structure was and asked them to make intervention judgments. The results of the three experiments suggest that the representations that support causal structure judgments do not easily support simple judgments about interventions in children. We discuss our findings in light of strong interventionist claims that the two types of judgments should be closely linked.     

The effect of causal knowledge on judgments of the likelihood of unknown features

People frequently infer unknown aspects of an entity based on their knowledge about that entity. The current study reports a novel phenomenon, an inductive bias people have in making such inferences. Upon learning that one symptom causes another in a person, both undergraduate students (Experiment 1) and clinicians (Experiment 2) judged that an unknown feature associated with the cause-symptom was more likely to be present in that person than an unknown feature associated with the effect-symptom. Thus, these findings suggest a specific mechanism in which causal explanations influence one’s representation of and inferences about an entity. Implications for clinical reasoning and associative models of conceptual knowledge are discussed.     

Inferring causal networks from observations and interventions

Information about the structure of a causal system can come in the form of observational data—random samples of the system's autonomous behavior—or interventional data—samples conditioned on the particular values of one or more variables that have been experimentally manipulated. Here we study people's ability to infer causal structure from both observation and intervention, and to choose informative interventions on the basis of observational data. In three causal inference tasks, participants were to some degree capable of distinguishing between competing causal hypotheses on the basis of purely observational data. Performance improved substantially when participants were allowed to observe the effects of interventions that they performed on the systems. We develop computational models of how people infer causal structure from data and how they plan intervention experiments, based on the representational framework of causal graphical models and the inferential principles of optimal Bayesian decision‐making and maximizing expected information gain. These analyses suggest that people can make rational causal inferences, subject to psychologically reasonable representational assumptions and computationally reasonable processing constraints.     

Independence and dependence in human causal reasoning

Causal graphical models (CGMs) are a popular formalism used to model human causal reasoning and learning. The key property of CGMs is the causal Markov condition, which stipulates patterns of independence and dependence among causally related variables. Five experiments found that while adult’s causal inferences exhibited aspects of veridical causal reasoning, they also exhibited a small but tenacious tendency to violate the Markov condition. They also failed to exhibit robust discounting in which the presence of one cause as an explanation of an effect makes the presence of another less likely. Instead, subjects often reasoned “associatively,” that is, assumed that the presence of one variable implied the presence of other, causally related variables, even those that were (according to the Markov condition) conditionally independent. This tendency was unaffected by manipulations (e.g., response deadlines) known to influence fast and intuitive reasoning processes, suggesting that an associative response to a causal reasoning question is sometimes the product of careful and deliberate thinking. That about 60% of the erroneous associative inferences were made by about a quarter of the subjects suggests the presence of substantial individual differences in this tendency. There was also evidence that inferences were influenced by subjects’ assumptions about factors that disable causal relations and their use of a conjunctive reasoning strategy. Theories that strive to provide high fidelity accounts of human causal reasoning will need to relax the independence constraints imposed by CGMs.     

Individual differences in causal learning and decision making

In judgment and decision making tasks, people tend to neglect the overall frequency of base-rates when they estimate the probability of an event; this is known as the base-rate fallacy. In causal learning, despite people's accuracy at judging causal strength according to one or other normative model (i.e., Power PC, DeltaP), they tend to misperceive base-rate information (e.g., the cause density effect). The present study investigates the relationship between causal learning and decision making by asking whether people weight base-rate information in the same way when estimating causal strength and when making judgments or inferences about the likelihood of an event. The results suggest that people differ according to the weight they place on base-rate information, but the way individuals do this is consistent across causal and decision making tasks. We interpret the results as reflecting a tendency to differentially weight base-rate information which generalizes to a variety of tasks. Additionally, this study provides evidence that causal learning and decision making share some component processes.     

Blickets and babies: the development of causal reasoning in toddlers and infants

Previous research has suggested that preschoolers possess a cognitive system that allows them to construct an abstract, coherent representation of causal relations among events. Such a system lets children reason retrospectively when they observe ambiguous data in a rational manner (e.g., D. M. Sobel, J. B. Tenenbaum, & A. Gopnik, 2004). However, there is little evidence that demonstrates whether younger children possess similar inferential abilities. In Experiment 1, the authors extended previous findings with older children to examine 19- and 24-month-olds' causal inferences. Twenty-four-month-olds' inferences were similar to those of preschoolers, but younger children lacked the ability to make retrospective causal inferences, perhaps because of performance limitations. In Experiment 2, the authors designed an eye-tracking paradigm to test younger participants that eliminated various manual search demands. Eight-month-olds' anticipatory eye movements, in response to retrospective data, revealed inferences similar to those of 24-month-olds in Experiment 1 and preschoolers in previous research. These data are discussed in terms of associative reasoning and causal inference.     

The advantage of timely intervention

Can people learn causal structure more effectively through intervention rather than observation? Four studies used a trial-based learning paradigm in which participants obtained probabilistic data about a causal chain through either observation or intervention and then selected the causal model most likely to have generated the data. Experiment 1 demonstrated that interveners made more correct model choices than did observers, and Experiments 2 and 3 ruled out explanations for this advantage in terms of informational differences between the 2 conditions. Experiment 4 tested the hypothesis that the advantage was driven by a temporal signal; interveners may exploit the cue that their interventions are the most likely causes of any subsequent changes. Results supported this temporal cue hypothesis.     

Inductive reasoning about causally transmitted properties

Different intuitive theories constrain and guide inferences in different contexts. Formalizing simple intuitive theories as probabilistic processes operating over structured representations, we present a new computational model of category-based induction about causally transmitted properties. A first experiment demonstrates undergraduates’ context-sensitive use of taxonomic and food web knowledge to guide reasoning about causal transmission and shows good qualitative agreement between model predictions and human inferences. A second experiment demonstrates strong quantitative and qualitative fits to inferences about a more complex artificial food web. A third experiment investigates human reasoning about complex novel food webs where species have known taxonomic relations. Results demonstrate a double-dissociation between the predictions of our causal model and a related taxonomic model [Kemp, C., & Tenenbaum, J. B. (2003). Learning domain structures. In Proceedings of the 25th annual conference of the cognitive science society]: the causal model predicts human inferences about diseases but not genes, while the taxonomic model predicts human inferences about genes but not diseases. We contrast our framework with previous models of category-based induction and previous formal instantiations of intuitive theories, and outline challenges in developing a complete model of context-sensitive reasoning.     

Seeing versus doing: two modes of accessing causal knowledge

The ability to derive predictions for the outcomes of potential actions from observational data is one of the hallmarks of true causal reasoning. We present four learning experiments with deterministic and probabilistic data showing that people indeed make different predictions from causal models, whose parameters were learned in a purely observational learning phase, depending on whether learners believe that an event within the model has been merely observed ("seeing") or was actively manipulated ("doing"). The predictions reflect sensitivity both to the structure of the causal models and to the size of their parameters. This competency is remarkable because the predictions for potential interventions were very different from the patterns that had actually been observed. Whereas associative and probabilistic theories fail, recent developments of causal Bayes net theories provide tools for modeling this competency.     

Causal learning about tolerance and sensitization

We introduce two abstract, causal schemata used during causal learning. (1) Tolerance is when an effect diminishes over time, as an entity is repeatedly exposed to the cause (e.g., a person becoming tolerant to caffeine). (2) Sensitization is when an effect intensifies over time, as an entity is repeatedly exposed to the cause (e.g., an antidepressant becoming more effective through repeated use). In Experiment 1, participants observed either of these cause—effect data patterns unfolding over time and exhibiting the tolerance or sensitization schemata. Participants inferred stronger causal efficacy and made more confident and more extreme predictions about novel cases than in a condition with the same data appearing in a random order over time. In Experiment 2, the same tolerance/sensitization scenarios occurred either within one entity or across many entities. In the manyentity conditions, when the schemata were violated, participants made much weaker inferences. Implications for causal learning are discussed.     

Sorting out categories: Incremental learning of category structure

Two experiments examine how inferences might promote unsupervised and incremental category learning. Many categories have members related through overall similarity (e.g., a family resemblance structure) rather than by a defining feature. However, when people are asked to sort category members in a category construction task, they often do so by partitioning on a single feature. Starting from an earlier result showing that pairwise inferences increase family resemblance sorting (Lassaline & Murphy, 1996), we examine how these inferences lead to learning the family resemblance structure. Results show that the category structure is learned incrementally. The pairwise inferences influence participants’ weightings of feature pairs that were specifically asked about, which in turn affects their sorting. The sorting then allows further learning of the categorical structure. Thus, the inferences do not directly lead learners to the family resemblance structure, but they do provide a foundation to build on as the participants make additional judgments.     

Noncategorical approaches to feature prediction with uncertain categories

In four experiments, we investigated how people make feature predictions about objects whose category membership is uncertain. Artificial visual categories were presented and remained in view while a novel instance with a known feature, but uncertain category membership was presented. All four experiments showed that feature predictions about the test instance were most often based on feature correlations (referred to as feature conjunction reasoning). Experiment 1 showed that feature conjunction reasoning was generally preferred to category-based induction in a feature prediction task. Experiment 2 showed that people used all available exemplars to make feature conjunction predictions. Experiments 3 and 4 showed that the preference for predictions based on feature conjunction persisted even when category-level information was made more salient and inferences involving a larger number of categories were required. Little evidence of reasoning based on the consideration of multiple categories (e.g., Anderson, (Psychological Review, 98:409–429, 1991)) or the single, most probable category (e.g., Murphy & Ross, (Cognitive Psychology, 27:148–193, 1994)) was found.     

Inferring hidden causal structure

We used a new method to assess how people can infer unobserved causal structure from patterns of observed events. Participants were taught to draw causal graphs, and then shown a pattern of associations and interventions on a novel causal system. Given minimal training and no feedback, participants in Experiment 1 used causal graph notation to spontaneously draw structures containing one observed cause, one unobserved common cause, and two unobserved independent causes, depending on the pattern of associations and interventions they saw. We replicated these findings with less-informative training (Experiments 2 and 3) and a new apparatus (Experiment 3) to show that the pattern of data leads to hidden causal inferences across a range of prior constraints on causal knowledge.     

The influence of complexity and reasoning direction on children's causal reasoning

Causal reasoning is the core and basis of cognition about the objective world. This experiment studied the development of causal reasoning in 86 3.5–4.5-year-olds using a ramp apparatus with two input holes and two output holes [Frye, D., Zelazo, P. D., & Palfai, T. (1995). Theory of mind and rule-based reasoning. Cognitive Development 10, 483–527]. Results revealed that: (1) children performed better on cause–effect inferences than on effect–cause inferences; (2) there was an effect of rule complexity such that uni-dimensional causal inferences were easier than bi-dimensional inferences which, in turn, were easier than tri-dimensional causal inferences; and (3) children's causal reasoning develops rapidly between the ages of age of 3.5 and 4 years.     

People want to see information that will help them make valid inferences in human causal learning

According to higher order reasoning accounts of human causal learning (e.g., Lovibond, Been, Mitchell, Bouton, and Frohardt, 2003; Waldmann and Walker, 2005) ceiling effects in forward blocking (i.e., smaller blocking effects when the outcome occurs with a maximal intensity on A+ and AX+ trials) are due to the fact that people are uncertain about the causal status of a blocked cue X in a forward blocking design when the outcome is always fully present on A+ and AX+ trials. This should not be the case for a reduced overshadowing cue Y (B- trials followed by BY+ trials). We tested this hypothesis by asking participants which additional information they preferred to see after seeing all learning trials. Results showed (1) that all participants preferred to see the effect of the blocked cue X over seeing the effect of the reduced overshadowing cue Y (Experiment 1), and (2) that more participants preferred to see the blocked cue X on its own when the outcome on A+ and AX+ trials was fully present than when the outcome on those trials had a submaximalintensity (Experiment 2).     

An integrated account of generalization across objects and features

Humans routinely make inductive generalizations about unobserved features of objects. Previous accounts of inductive reasoning often focus on inferences about a single object or feature: accounts of causal reasoning often focus on a single object with one or more unobserved features, and accounts of property induction often focus on a single feature that is unobserved for one or more objects. We explore problems where people must make inferences about multiple objects and features, and propose that people solve these problems by integrating knowledge about features with knowledge about objects. We evaluate three computational methods for integrating multiple systems of knowledge: the output combination approach combines the outputs produced by these systems, the distribution combination approach combines the probability distributions captured by these systems, and the structure combination approach combines a graph structure over features with a graph structure over objects. Three experiments explore problems where participants make inferences that draw on causal relationships between features and taxonomic relationships between animals, and we find that the structure combination approach provides the best account of our data.