UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Transfer of acquired rule to conditional reasoning as a function of content similarity, attribute dimension… Vavrik, John 1995

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


831-ubc_1995-060772.pdf [ 2.08MB ]
JSON: 831-1.0088360.json
JSON-LD: 831-1.0088360-ld.json
RDF/XML (Pretty): 831-1.0088360-rdf.xml
RDF/JSON: 831-1.0088360-rdf.json
Turtle: 831-1.0088360-turtle.txt
N-Triples: 831-1.0088360-rdf-ntriples.txt
Original Record: 831-1.0088360-source.json
Full Text

Full Text

TRANSFER OF ACQUIRED RULE TO CONDITIONAL REASONING AS A FUNCTION OF CONTENT SIIv11LARITY, ATTRIBUTE DIMENSION SIZE, AN]) ACQUISITION MODE by JOHN VAVRIK The University of British Columbia, 1979 B.Sc. M.A. The University of British Columbia, 1987 A THESIS SUBMITTED N PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE STUDIES (Department of Educational Psychology and Special Education) We accept this thesis as conforming to the required standard  THE UNWERSITY OF BRITISH COLUIVIBIA July, 1995 ©John Vavrik,  1995  in partial fulfilment of the requirements for degree at the University of British Columbia, I agree that the Library sion freely available for reference and study. I further agree that permis copying of this thesis for scholarly purposes may be granted by the In presenting this thesis  department  or  by  his  or  her  representatives.  It  is  understood  that  an advanced shall make it for extensive head of my copying  or  my written publication of this thesis for financial gain shall not be allowed without permission.  (Signature)  Department of  Z’’1T/O-&L  The University of British Columbia Vancouver, Canada Date  DE-6 (2)88)  1ttct,  /‘Z /5S  t)Oj-  7  EP’’ 1 (/17/O-€,  ABSTRACT  In this study, deductive reasoning with conditional statements of the form IF x THEN y was investigated in a transfer of training paradigm. Subjects induced conditional rules under different training conditions and then solved a series of deductive syllogisms based on conditional rules the content of which was either related (near-transfer) or unrelated (far-transfer) to that of the previously induced rules. The conditions under which the conditional rules were induced varied in terms of the acquisition mode (prediction/diagnosis) and in terms of the size of the attribute dimensions from which rule instances were drawn (binary/trinary). The deductive reasoning task, as the criterion task for transfer, included all four types of conditional syllogisms (Modus Ponens, MP; Denying the Antecedent, DA, Affirming the Consequent, AC, and Modus Tollens, MT). The accuracy of deductive performance on near-transfer was significantly higher than response accuracy on far-transfer on the DA & AC but not on the MP & MT argument forms. Transfer of conditional rule knowledge induced in a predictive mode was significantly better compared to a diagnostic mode on the AC form. Overall, near as well as far-transfer performance was similar for the binary and trinary conditions. In addition, neither the binary nor the trinary condition resulted in improved far-transfer performance compared to controls. The results provide evidence against purely instance-based as well as purely rule-based models of deductive reasoning. Instead, they suggest that conditional knowledge is represented at an intermediate level of abstraction, and that conditional reasoning is adaptive to the form of the deductive argument. Different types of deductive argument types may trigger different reasoning strategies.  11  TABLE OF CONTENTS  Abstract Table of Contents  iii  List of Tables  v  List of Figures  vi  Acknowledgement  vii  I. ANALYSIS OF RESEARCH PROBLEMS AND ISSUES  1  1 A.INm0DUCTION 3 B. RATIONALE FOR THE STUDY 6 C. RESEARCH ON INDUCTIVE RULE ACQuIsITIoN 11 D. RESEARCH ON DEDUCTIVE REASONING WITH A CONDITIONAL RuLE 1. ANALYSIS OF RESEARCH IssuEs RELATED TO RULE TRANSFER AS A FUNCTION OF 13 CONTENT SIMiLARITY. 2. ANALYSIS OF RESEARCH ISSUES RELATED TO THE EFFECTS OF ArrEIBuTE DIMENSION 15 SIZE ON FAR-TRANSFER OF AN AcQuIRED RULE. INTERACTION EFFECTS BETwEEN 3. ANALYSIS OF RESEARCH ISSUES RELATED TO THE ATTRIBuTE DIMENSION SIZE AND DEDUCTIVE ARGUMENT TYPES IN NEAR-TRANSFER OF AN 17 AcQUIRED RULE. IssUES RELATED TO THE EFFECTS OF INDUCTIVE AcQUISmON 4. ANALYSIS OF RESEARCH 21 MODE IN NEAR-TRANSFER OF AN ACQUIRED RULE. 25 E. SUMMARY OF HYPOTHESES AND EXPERIMENTAL PREDICTIONS  28  II. METHOD A. SUBJECTS AND DESIGN B. ExPERIMENTAL TAsKS AN]) MATERIALS C. APPARATUS D. PROCEDURE  28 31  35 35 38  III. RESULTS  A. ANALYSIS OF INDUCTiVE RULE TRAINING PERFORMANCE B. ANALYSIS OF DEDUCTIVE REAsoNING PERFORMANCE C. ADDITIONAL ANALYSIS OF TRANSFER DATA  38 46 57  61  IV. DISCUSSION  111  A. SUMMAT1VE OVERVIEW B. IMPLICATIONS FOR THEORIES OF CoNDITIoNAL REASONING C. PROPOSED ACcoUNT OF CONDITIONAL REASONING D. IMPLICATIONS FOR EDUCATION E. LIMITATIONS OF THE STUDY AND SuGGESTIoNs FOR FURTHER RESEARCH  61 62 70 74 75  REFERENCES  80  APPENDICES  85  APPENDIX A: AN ILLUSTRATION OF RULE TRANSFER BASED ON THE PROPOSED hYPOTHESES APPENDIX B: INSTRUCTIONS TO SUBJECTS  iv  86 90  LIST OF TABLES  Table 1: Means and standard deviations of number of instances and total processing processing time for differenr inductive rule acquisition conditions  39  Table 2: Category frequencies for reported inductive learning strategies  42  Table 3: Response frequencies for each option on a rule assessment item  44  Table 4a: Target rule selection by reported strategy for binary group  45  Table 4b: Target rule selection by reported strategy for trinary group  45  Table 5: Scoring key for eight types of conditional syllogisms  47  Table 6: Mean scores and standard deviations on deductive reasoning performance for the binary group  48  Table 7: Mean scores and standard deviations on deductive reasoning performance for the binary group  49  Table 8: Mean scores and standard deviations on deductive reasoning performance for the binary group  50  Table 9: Mean scores and standard deviations for overall deductive reasoning performance  52  Table 10: Proportion of subjects following either a conditional or biconditional interpretation  58  Table 11: Transfer scores summed over all argument types by rule selection score  59  V  LIST OF FIGURES  Figure 1: Experimental Design Layout  30  Figure 2: Causes and effects under different rule truth-table classes  31  Figure 3: Types of content used in the inductive learning task  32  Figure 4: Task sequences used for each group  36  Figure 5: Number of instances for rule acquisition  39  Figure 6: Processing time for rule acquisition  40  Figure 7: Overall transfer performance  53  Figure 8: Far transfer by dimension size  54  Figure 9: Near transfer by dimension size  55  Figure 10: Near transfer by acquisition mode  57  Figure 11: Transfer performance by rule score  60  vi  ACKNOWLEDGMENTS  This document is dedicated to myself in keeping with the inescapable conclusion that the pursuit of a Doctoral Dissertation is essentially a selfish process of symbolic self completion during which I alienated my friends, taxed the patience and goodwill of my supervisor, and neglected to hug my wife each time she endured yet another thesis induced mood swing. My dedication is also a celebration of the fact that this is the only paragraph I got to write without someone editing it.  vii  I. ANALYSIS OF RESEARCH PROBLEMS AND ISSUES  A.Introduction  The overall aim of this study is to gain a better understanding of how people reason with logical rules or propositions, particularly those that can be expressed in the aform “IF X THEN Y”. Such statements are called conditional statements, and reasoning with them is referred to as conditional reasoning. Below are some illustrative examples of conditional reasoning in different contexts: A physics student has learned that if two particles have opposite electric charges then they will attract each other. At some later time she observes two particles that attract each other. What can she conclude about the electric properties of the particles? A school psychologist remembers from his graduate training that damage to the frontal lobes results in poor performance on problems requiring metacognitive monitoring. He is presented with a student who does poorly on problems requiring metacognitive monitoring (such as the Towers of Hanoi problem). What can he conclude about the neurological cause of the disability? An intern on an emergency ward remembers from her medical training that a certain toxic agent Tx causes skin to turn blue. She observes that a patient suspected of poisoning has pink colored skin. What can she conclude about the cause of the poisoning?  These examples point to the ubiquity of conditional reasoning in decision making, problem solving, and scientific reasoning. (See also Braine & O’Brien, 1991; Lee, 1985; Piburn, 1990; Popper, 1959; Rips, 1990; Sternberg, 1986). Yet, despite their importance, most people reason poorly with conditionals, and the cognitive process underlying this fundamental form of reasoning is not well understood (Evans, 1982; Evans, Barston & Pollard, 1983; Marcus & Rips, 1979; Markovits & Nantel, 1989).  1  As the examples illustrate, conditional reasoning typically involves two distinct situations: an acquisition phase, where the “IF X THEN Y” rule is learned, and an application phase where the rule is used in conjunction with other information to make deductive inferences leading to some general conclusion. Both situations occur naturally in academic, professional, and everyday reasoning contexts, and both have been studied intensely, but mostly independently, for over three decades (Lee, 1985). One reason for the limited number of studies investigating inductive rule acquisition together with deductive rule application may be the fact that the inductive learning phase typically involves considerable time commitment by subjects, and often requires detailed and time consuming design and development of appropriate instructional materials (Pollard & Evans, 1983; and Markovits 1988).  Most of the work on rule acquisition has been traditionally carried out within the concept-learning research paradigm. Typically, these studies investigated learning of concepts which were defined by simple logical rules from a set of positive and negative instances of the concept. In other words, these studies have been concerned with inductive reasoning (e.g. Bourne, 1970; Bruner, Goodnow, & Austin, 1956;).  Research on the application, or utilization, of a conditional rule, on the other hand, has been the focus of another research tradition dealing with the other mode of human reasoning, namely deductive reasoning (e.g. Evans, 1982; Johnson-Laird, Byrne, & Schaeken, 1992). Within this research tradition, the interest has been on observing reasoning performance on deductive reasoning tasks such as the Wason Selection Task, and various forms of syllogistic arguments.  A key assumption motivating this study is that we cannot fully understand and improve conditional reasoning unless we consider both the acquisition of the rule 2  (inductive learning), and the application of the rule (deductive reasoning) together (Lee, 1985). It is well established now that many types of declarative and procedural knowledge are often not used (accessed) in situations where they clearly ought to be used because f differences between acquisition and application conditions (e.g. Bereiter & Scardamalia, 1985; Morris, Bransford, & Franks, 1977; Vosniadou & Ortony, 1989). It is reasonable to assume that the knowledge structures underlying logical rules, including the conditional, are no exception. Deductive use of a conditional rule may, therefore, depend on how the rule was inductively acquired in the first place.  The purpose of this study, therefore, is to determine whether the conditions under which a conditional rule is learned (induced) influence transfer of the rule to conditional (deductive) reasoning with the rule.  The following section contains a discussion of current research problems and issues pertaining to the basic purpose of the study outlined above. The discussion has two aims: (a) to help explain the rationale for this study, and (b) to develop a conceptual framework from which research hypotheses can be articulated. Since each of these aims requires a slightly different treatment of the issues, the rationale will be presented first, followed by a more detailed analysis of the research issues needed to develop the conceptual framework for specific hypotheses.  B. Rationale for the Study  One reason for studying transfer of an acquired rule is that it would advance current understanding of human reasoning on the basis of past learning. To see why this may be of interest, consider that despite a long history of research on deductive reasoning,  3  there is not an agreed upon position on the cognitive mechanism underlying deductive reasoning in general, and conditional reasoning in particular. Currently, there are at least three different views. According to one prominent perspective, people make deductions by systematically applying logical rules of inference to argument premises and conclusions (Inhelder & Piaget, 1958; Newell and Simon, 1972). The opposing view holds that reasoning is carried out by retrieving specific instances from memory (Griggs & Cox, 1982; Polard & Evans, 1981). According to the third view, some combination of logical rules and specific experience-based instances might be involved in propositional reasoning (Rips, 1990). Currently, the hybrid view appears most promising since consistent evidence for purely content-free rules or specific instances has not been forthcoming (Smith, Langston, & Nisbett 1992). However, the current hybrid models of propositional reasoning have not been articulated in enough detail to allow the formulation of specific predictions that can be empirically tested, or to inform educators interested in designing learning conditions conducive to the acquisition of conditional knowledge (Lee & Lee, 1992).  If differential reasoning performance was found with rules induced from two different sets of instances, then a more detailed specification of the role of rules and instances in hybrid models would be possible. An example of such a situation would be two sets of instances that both have a flill conditional relational structure but differ in some other ways, such as in terms of the number of unique positive and negative instances of the rule in each set. A concrete example would be the number of distinct cases of symptoms in the presence or absence of various disease agents studied by medical students to learn diagnostic rules of inference.  Another reason for investigating how rule acquisition influences subsequent reasoning is that it may inform educational practice in terms of instructional strategies for 4  enhancing reasoning performance. As was pointed out above, despite the central role of conditional reasoning in many cognitive tasks ranging from inferences about causes and effects in scientific research to problem solving and decision making in applied fields such as medicine, law, and engineering, conditional reasoning performance is generally poor, and inconsistent (e.g. Evans, 1982; Evans, Barston & Pollard, 1983; Markovits et al., 1989; Wason, 1966). While the reasons for this vary depending on which of the three theoretical perspectives one holds, it is generally agreed that in order to reason effectively with a conditional rule, one must have an appropriate rule structure in memory (Braine & O’Brien, 1991; Lee, 1985).  Just how conditional rule knowledge is acquired is not clear. What is clear is that teaching the logical form of a conditional rule by direct instruction alone is not effective (Cheng, Holyoak, Nisbett, & Oliver, 1986). On the other hand, Lee (1985) was able to significantly improve conditional reasoning performance through training that contained, as a key component, inductive rule learning from a set of instances with a conditional structure. However, Lee did not manipulate the parameters of the inductive rule learning task and therefore it is unclear whether different inductive learning conditions are equally effective in building an appropriate conditional rule structure. If it could be shown that different rule learning conditions do result in differential reasoning performance, it would offer educators more options for designing learning conditions to enhance this fundamental reasoning process. The rationale for this study is therefore based on both theoretical, as well as practical, instructional considerations.  Acquiring and Reasoning With Conditional Rules  5  An overview of research issues and problems:-As was afready alluded to in the introduction, virtually all previous research on reasoning with logical rules in general, and conditional (IF.. THEN..) rules in particular, with the exception of Lee (1985), has been carried out independently either from the perspective of rule acquisition (inductive reasoning research tradition), or rule application (deductive reasoning research tradition). Before developing hypotheses about transfer between the two modes of reasoning in the case of the conditional, a review of relevant findings from each of the two research traditions is necessary.  Whenever any findings from the separate traditions imply something about transfer from rule acquisition to rule application, such implications will be pointed out. In addition, relevant assumptions underlying cognitive processing and knowledge representation in the learning and usage of logical rules will be identified. It will then be possible to propose specific hypotheses about the influence of different acquisition conditions on transfer of an acquired rule to conditional reasoning. This chapter will conclude with a summary of the hypotheses along with a set of testable predictions derived from them.  C. Research On Inductive Rule Acquisition  Rule acquisition as concept learning: Concept learning studies have traditionally investigated two learning processes involved in the acquisition of concepts defined by logical rules: attribute learning and rule learning (Bourne 1970; Haygood & Bourne, 1965; Lee, 1984). Within this research paradigm, the assumption has been that most concrete as well as most abstract concepts can be defined in terms of criterial attributes (or features)  and some rule-like relationship between them: C=R(x,y...)  6  where C stands for a concept, R is a rule, and x, y... are the defining attributes (Bourne, 1970). The kinds of rules that have been investigated in this research tradition have been rules derived from propositional calculus of formal logic. The conditional rule is one of  four basic relationships (rules) that represent the full calculus. The others are the conjunction (and), disjunction (or), and the biconditional (1ff). Experimental tasks used in this concept learning paradigm include: 1. Attribute Identzjlcation:-Under this paradigm, subjects are given a rule (e.g. a conjunction of two attributes), and their task is to learn the relevant attributes (e.g. red, circle) from a set of instances containing both relevant and irrelevant attributes.  2. Rule Learning:-In rule learning subjects are told what the relevant attributes are and their task is to learn the relationship (rule) between them. 3. Rule Identiflcation:-Within this paradigm subjects are informed what the relevant attributes are, and previously they have learned two or more rules. Their task is to identify how the previously learned rules applies to the present set of instances.  4. Complete Learning:-In this learning paradigm subjects are simply introduced to various attributes (colors, shapes, types of events), are told that the concept involves a certain number of these, and their task is to learn both the relevant attributes as well as the appropriate rule.  The process of rule acquisition:-How do people acquire concepts defined by logical rules? Bout-ne (1970) proposed that people utilize “an intuitive version of the logical truth table”. A formal truth table for the conditional “IF X THEN Y” (usually written as “X ——> Y”) is shown below:  x  y  T T F F  T F T F  + -  +  +  According to Bourne, rule acquisition involves building a mental representation that reflects this kind of truth table structure. This is accomplished by two component processes referred to as attribute coding and rule mapping (Lee, 1984). During attribute 7  coding, the relevant stimulus features, (e.g. X and Y), are collapsed into the four truth table classes XY, —XY, X-Y, and -X--Y. This is followed by mapping each of these classes onto the two response categories of the rule (+ or -), that is, positive or negative instances.  One source of evidence cited in support of this assumed mechanism is the observation that with increased experience on a concept learning task, subjects begin to categorize all instances of a particular truth table class in the same manner (Bourne, 1970; 1974). Lee (1984) subsequently established the sufficiency of the two component processes by demonstrating that training subjects on each component can lead to improved rule acquisition. With respect to attribute coding, Lee argued that since this process comprises affirmation/negation and combinatory operations, attribute coding difficulty should in turn depend on the difficulty of these two operations. But since the number of combinatory operations is the same for all truth table classes, attribute coding difficulty should be a ifinction of the negation operations that are involved or, in Lee’s words, “the number of attributes requiring negations”. What Lee is referring to here is the number of implicit references to the negation of the target attributes which is essentially -  the notion of contrast sets to be discussed later in the section reviewing research on deductive reasoning. With binary dimensions, the number of negations is 0,1,1,1 for TT,TF,FT, and FF classes respectively, while with trinary dimensions the numbers are 0,2,2,4. Indeed Lee (1984) found that error patterns for the four classes, during rule acquisition, were consistent with these theoretical difficulty indices.  The encoding issue is relevant to this study. It suggests that transfer of a rule may be influenced by the number of instances from which the rule is acquired. Because a small number of dimensions is easier to encode, transfer to deductive reasoning that is based on 8  recall of specific instances should be facilitated by sets of instances with few dimensions. Situations where access to specific instances in deductive reasoning is possible are situations where the rule content is the same during acquisition and application. “Rule content” refers to the specific semantic propositions, the “X’s” and the “Y’s”, referenced by the rule “IF X THEN Y”. This situation is referred to as “near-transfer” condition.  In contrast to instance-based models of reasoning, rule-based accounts would not predict differences in near-transfer as a function of dimension size. However, according to rule-based accounts, deductive reasoning with content that is different from the acquisition phase (referred to as “far-transfer” condition) should be better with rules induced from sets of instances with widely varying attribute dimensions because the increased variability of training set is likely to induce a more generalizable, content-independent knowledge structure. Studies of far-transfer have shown that increasing the number of different training examples from which new knowledge (such as a problem schema, for example) is induced, makes that knowledge more applicable (i.e. generalizable) to new situations (e.g. Gick & Holyoak, 1987; Holyoak & Koh, 1989). The implications regarding near and fartransfer are developed more systematically later; they were introduced here only because they arise naturally from the discussion of rule acquisition and concept learning.  Pragmatic constraints in rule acquisition:-As Lee (1985) noted, we cannot assume that a conditional rule structure, induced in a traditional concept learning task, is automatically associated with a conditional proposition “IF X THEN Y”. This is because the set of instances (e.g. cause-effect pairs such as cl-el, c2-el, c2-e2,...) corresponding to a target rule (e.g. “IF cause = ci, THEN effect  =  el”) warrants many valid propositions  such as “when ci occurs then ci follows”, or “when c2 occurs then ci follows”, etc. Clearly, the reasoner must rely on some other principle or method to arrive at the target rule expressed in the appropriate causal language. 9  One way to accomplish this is to employ a prediction task using explicit linguistic expressions of the form “IF cause = ci THEN effect = ?“ and guiding the learner, through prompts and appropriate feedback, toward the correct formulation of the conditional rule (Lee, 1985). However, this may not be the only, or the best method available, particularly in naturalistic settings where such explicit and structured guidance is unlikely to occur. What other principle might be invoked?  Recent work on inductive rule-based mechanisms points to pragmatic considerations as likely candidates (Holland, Holyoak, Nisbett & Thagard, 1986). Holland et al. note that in natural contexts such as scientific or medical research, pragmatic considerations, typically the goals of the researchers, can constrain inductive inferences. In this context, this means that some pragmatic considerations may constrain the number of possible propositions that can be made about a set of observed relations.  Specifically, one pragmatic and fi.indamental goal of research is to derive parsimonious and general statements about a set of observations. In fact, such statements are often elevated to the status of scientific laws. If the pragmatic principle of generality and parsimony is applied to a set of observations with a conditional structure, it constrains the possible valid propositions to one that optimally summarizes the observed set relations. That proposition is precisely the material conditional: “Whenever cause ci occurs, effect el follows”, or equivalently, “IF cause = ci, THEN effect = el”. In other words, pragmatic constraints such as goals enable the construction of appropriate lexical entries to the underlying rules. Furthermore, if the traditional concept learning task incorporates a pragmatic context such as a predictive or diagnostic context, which utilizes the IF.. THEN.. language naturally, not only the target rule, but also its associated causal language should be readily acquired. 10  To conclude then, the evidence presented suggests that conditional knowledge may be induced from appropriately selected sets of instances using traditional concept learning tasks. Furthermore, pragmatic constraints may be utilized to associate a conditional (IF.. THEN...) statement with the induced knowledge representation.  D. Research on Deductive Reasoning With a Conditional Rule  The specific objectives of the present study are to assess transfer effects of rule acquisition variables on deductive reasoning performances. Since the transfer criterion task is a deductive reasoning task, it is useful to first describe the two most commonly used tasks in deductive reasoning research. An overview of these tasks may also facilitate the interpretation of research findings reviewed in this section.  The Wason Selection Task:-Most conditional reasoning studies have utilized some variants of the now classic Wason’s Selection Task (Wason, 1966). In a typical version of this task, a subject is shown a pack of cards and is told that each card has a letter on one side and a number on the other. The subject is then told that the letters and numbers are related by a rule which states that:  ‘Tf There is a vowel on one side, then there is an even  number on the other side “. The subject is then presented with four cards laid down in front of her (e.g. A, K, 2, 7). The subject’s task is to select those cards, and only those cards, that would be necessary to turn over in order to determine if the stated rule does indeed hold. The correct selection is Card A and Card 7. In other words, the task involves the verification of a falsifying claim with a set of data (instances). The problem with this task is that it assesses only one of several forms of deductive inferences. A more complete test of conditional reasoning is the syllogism task (e.g. Aristotle; Marcus & Rips, 1979;  11  Lee & Lee, 1985, 1992; Woodworth & Sells, 1935). This is the task used in the present study.  The Conditional Syllogism Task: This task requires subjects to either derive or validate conclusions to arguments which have a conditional rule as their first premise. There are actually four distinct forms of deductive arguments with conditional propositions: Modus Ponens (MP)  Modus Tollens (MT)  Denying the Antecedent (DA)  Affirming the Conseiuent (AC)  First premise: Second premise:  ifC Then E C  Conclusion:  E  First premise: Second premise:  not E  Conclusion:  not C  First premise: Second premise:  if C Then E not C  Conclusion:  Indeterminate  First premise: Second premise:  if C Then E E  Conclusion:  Indeterminate  If C  Then E  Virtually all data on conditional reasoning have been obtained using some form of these two tasks. Each of the following four sections presents critical analyses of the research findings and identifies research issues that are investigated in this study. The sections conclude with statements of specific research hypotheses.  12  1. Analysis of Research Issues Related to Rule Transfer as a Function of Content Similarity.  Effect of “familiar” rule content (semantics) on conditional reasoning:-Both the Wason task and the syllogism task have been tried with abstract rules such as “IF A THEN 2”, as well as with rules that refer to familiar world knowledge such as “IF a metal bar is heated THEN it expands”. Reasoning performance with familiar rule content is always different from abstract content, and often it is better, particularly on the Wason task (Pollard & Evans, 1987; Evans, 1982; Griggs & Cox, 1982; Markovits, 1986; Wason & Shapiro, 1971). It should be noted here that rule content means the objects or events referenced by the conditional rule.  In the syllogism tasks, familiar content gives rise to what has been called the “BeliefBias” effect (Evans, Barston, & Pollard, 1983). It is manifested by subjects tending to confirm conclusions which they believe to be true, and discon±irming conclusions believed to be false, regardless of the argument structure from which the conclusion was deduced.  Clement & Falmagne (1986) have shown that reasoning with conditional rules improves when the “semantic relatedness” between the antecedent and the consequent is high, or when the “imagery value” of the conditional statement is high. This suggest that reasoning may be facilitated by access to particular instances or events in memory  -  instances that constitute relevant information to the inferencing process.  The above research findings suggest that access to specific instances of antecedent-consequent relations stored in memory may influence conditional reasoning  13  performance. More specifically, better access to rule instances should lead to improved reasoning performance with the rule.  Content familiarity and rule transfer:-Consider the following two transfer situations. In the first situation, a conditional rule used in a deductive task references the same entities or events that the reasoner encountered when initially learning the rule. A typical example is the case of medical training where students learn the relationship between specific disease agents and disease symptoms. The agents and symptoms are typically the same ones the students will encounter later in their practice. So although the reasoning process changes from induction (learning the relationships) to deduction (inferring a diagnosis based on presented symptoms and applying the learned rule), the actual entities involved in the reasoning (the specific symptoms and agents) are the same. This is what is meant by the term near-transfer. In a second type of transfer, the entities and events encountered during rule acquisition are different from the entities or events contained in the premises of deductive arguments. This situation represents afar-transfer condition. Here the availability and access to specific instances will not be helpful and deductive reasoning performance may suffer. On the basis of the effects of content familiarity reviewed above, it is therefore hypothesized that content similarity inherent in near-transfer conditions would facilitate more effectively access to the conditional rule structure, and to apply it to deductive inference, than that prevailing in far-transfer conditions.  14  2. Analysis of Research Issues Related to the Effects of Attribute Dimension Size on FarTransfer of an Acquired Rule.  The basic question posed here is whether rules induced from a wider variety of instances are more generalizable, and thus more applicable to a wider range of deductive reasoning situations than smaller sets in inductive rule acquisition. It is a ubiquitous observation that when limited human memory capacity is overloaded, one has to bring into order (pattern) an array of complex stimuli (contents). Similarly in the reasoning task situation. The reasoner must rely on more general reasoning rule structures to reach correct conclusions. These are discussed next.  The Role of Content-free Inference Schemas in Deductive Reasoning:-In addition to the evidence of content effects on deductive reasoning reviewed in the preceding section, there is also evidence that some forms of deductive reasoning are not affected to any significant degree by rule content (Braine & O’Brien, 1991; Braine, Reiser, & Rumain, 1984). For example, when subjects are confronted with the following argument: Major premise: Either P Minor premise: Not P  or Q  Conclusion?  most correctly conclude that  “Q” must be the case, regardless of what P and Q stand for.  In other words, in this case, reasoning performance does not depend on specific propositional content in the argument premises.  15  The example illustrates what Braine et al. call an “inference schema”. An inference schema refers to an abstract form of the rule structure concerned, and determines what conclusion can be inferred from the general form of a given argument or a rule, rather than their particular content. According to Braine et al., inference schemas play an important role in human deductive reasoning because they can be used to derive new inferences from a given set of premises.  Examples such as the above have been cited in defense of the view that formal logical rules underlie much of human reasoning. According to this view, people can acquire generalizable rule structures, such as inference schemas, which they apply in their deductive inferencing. Because these rule structures are sensitive to the form rather than the particular content of rules, they can be applied across a wide range of reasoning situations (Rips, 1990).  Rule transfer and schema induction:-The question of where generalized inference schemas originate has received little attention. In fact, as Lee (1985) pointed out, the existence of such logic structures is usually just assumed. It was noted earlier that using an inductive rule learning paradigm, Lee (1985) was able to induce a generalizable, contentfree conditional logic structure in a significant proportion of his subjects. However, in his study, Lee did not vary the number of instances (examples) from which the logic structure was induced, although trinary dimensions were used.  Researchers interested in transfer of cognitive skill (e.g. Singley and Anderson, 1989) have repeatedly found that the generalizability of a rule or a cluster of rules (a schema) increases with exposure to a wider variety of examples (e.g. Catrambone & Holyoak, 1989; Gick & Holyoak, 1987). Their research suggests that the degree to which a conditional rule is independent of particular content (i.e. is more generalizable) may be a 16  function of the number of rule instances provided in the acquisition phase. More instances should lead to the induction of more content-free rules.  The examination of related literature clearly indicates the merit of investigating the issue surrounding the usefulness of content-free rules. More specifically, it is hypothesized that under the conditions where the rule content does not match the acquisition content, and cueing of specific relevant instances is unlikely, generalizable, content-free rules, if available, need to be activated to meet the task demand of deductive inferences. In this context, such rules are hypothesized to be more readily acquired from exposure to large rather than small sets of exemplars.  3. Analysis of Research Issues Related to the Interaction Effects Between Attribute Dimension Size and Deductive Argument Types in Near-Transfer of an Acquired Rule.  Awareness of alternatives in conditional reasoning:-Although a conditional rule such as ‘IF cause THEN effect” explicitly refers to only one cause and one effect, it implicitly references other, alternative causes and effects (e.g. “IF we eat candy THEN we get cavities” but drinking Coke will also do the job). These alternatives are in fact essential components of the conditional logic structure and they play a key role in conditional reasoning. Consider the conditional rule “IF cause = ci, THEN effect = el”. A number of researchers have shown that awareness of or familiarity with alternative causes (c2,c3) of a given effect (el) improves reasoning with indeterminate forms of the conditional syllogism, namely the DA and AC forms (Cummins, Lubart, Alksnis & Rist, 1991; Markovits, 1984). The central idea here is the availability ofalternatives along a given attribute dimension (e.g. the number of alternative causes of a particular disease symptom). This notion is captured in the concept of contrast set which is discussed below.  17  Several studies investigated reasoning with arguments and rules containing negated propositions (e.g. ‘IF NOT A THEN 2”) and found that performance is more error-prone and time-consuming when negations are present (Evans, 1982; Oaksford & Stenning, 1992). Lee et al.(1992) ranked the four basic syllogisms according to a hypothesized difficulty level where one of the difficulty parameters was the number of negations that needed processing. For example, the Modus Tollens (MT) type was one level above the Modus Ponens (MP) type because it involved two more negation operations. Overall, the results supported the contention that higher levels were more difficult to process.  Why is reasoning with negated propositions more error-prone? According to one influential account, the reason is that our natural language-understanding mechanism contains a “NOT-heuristic” which focuses attention only on the named proposition that is negated, and not on all the propositions implied by the negation. Since the implied information is often critical for correct deductions, more errors occur (Evans, 1982, 1984). Another hypothesis is that negations increase processing load because they involve an extra mental operation, similar to a NOT operator in a computer implementation of Boolean logic (e.g. Lee et al., 1992).  However, neither of these accounts can explain why some negated propositions are more error-prone than others, as was discovered in a recent study that investigated the effect of negated constituents more systematically than had previously been done (Oaksford et al., 1992). These authors claim that negated propositions which are easier to process are ones that facilitate mental construction of contrast sets.  Contrast Sets and Conditional Reasoning:-According to Oaksford et al., premises that contain negated arguments cannot be directly represented as specific instances in 18  memory because it is not clear what instances the negated argument actually represents. To see why, consider the following deductive task: Major premise: Minor premise:  If X is on one side of a card, then Y is on the other side. Xis not on one side.  Conclusion:  Clearly, the minor premise can represent an infinite variety of potential instances (any letter, any number, in fact anything at all). As a result, a reasoning program will be difficult to implement.  Now, according to this account, before a premise containing a negated argument can be processed, the reasoner must first determine the contrast set implied by the proposition. A contrast set is a set of all objects or events implicitly referenced by a negated proposition. The difficulty of some negated propositions, like the abstract one above, is that they reference large, ambiguous contrast sets, whereas other propositions like “NOT AN EVEN NUMBER” imply a smaller, better defined contrast set, namely the set of “ODD NUMBERS”. According to Oaksford et al., ambiguous contrast sets increase cognitive load and consequently reasoning performance suffers. The problem with the Oaksford et al. study, however, is that the relationship between negated propositions and their implied contrast sets was only assumed. A more direct empirical test of this hypothesis is needed.  Contrast set size and rule transfer:-How might contrast set size influence transfer of an acquired rule to conditional reasoning? The findings suggest that the size of the contrast set might interact with argument form to influence reasoning performance. Large contrast sets increase cognitive load because more elements need to be processed.  19  However, large contrast sets also contain more alternative antecedents and consequents associated with a rule. As was discussed earlier, awareness of these alternatives improves reasoning on indeterminate argument types, namely the Denying the Antecedent (DA) and the Accepting the Consequent (AC) type. It seems reasonable to propose, therefore, that reasoning with indeterminate argument forms should be more accurate with rules inductively acquired from large contrast sets. On the other hand, reasoning with deterministic argument forms that also contain a negation, namely the MT type, should be more difficult with rules acquired from large contrast sets. This is because in this case awareness of alternatives is irrelevant, and so the only impact of the large set size is to increase cognitive load, which will hinder the reasoning process.  To illustrate this line of reasoning more clearly, consider a conditional rule such as “IF cause  =  ci THEN effect  =  e 1” learned from exposure to either one of two sets of  instances: one having binary attribute dimensions (e.g. two causes, ci & c2 and two effects e i & e2), and another having trinary dimensions (e.g. three causes ci, c2 & c3 and three effects el, e2 & e3).  Consider the DA argument type first. In this type of syllogism, the minor premise “NOT ci” references the contrast set {c2} in the case of binary dimensions, and {c2, c3 } in the case of trinary attribute dimensions. Since each cause is presumed to be associated with an effect, ei, e2 and el, e2, e3 become activated in memory in the case of the binary and trinary dimensions, respectively. Since a greater number of different effects is activated with trinary dimensions, correspondingly larger number of conclusions is possible. As a result, an indeterminate response, which is the correct response for the DA type, becomes more probable with the larger dimension size.  20  Next consider the AC type. Here, affirming the consequent proposition “ci” activates { ci, c2) and { ci, c2, c3 } for the binary and trinary dimensions, respectively. Since more alternative causes become activated in the trinary case, the correct indeterminate conclusion should be reached more likely in this case, as compared to the smaller dimension case.  Finally, consider the MT type. Here the negated consequent “NOT e 1” references {e2}, and {e2, e3 } in the binary and trinary dimensions, respectively. The associated causes that become activated are {c2} for the binary, and {c2, c3 } for the trinary case. Both of these activated sets represent contrast sets for the proposition “NOT ci”, which happens to be the correct conclusion for this argument type. However, since the contrast set for the binary condition contains only one element { c2 }, it should be processed more easily than the larger set { c2, c3 }. Therefore, the smaller dimension size should lead to better performance on the MT argument type.  The foregoing analysis leads to a hypothesis that increased availability of alternative antecedents and consequents associated with larger sets of induced rule exemplars should be particularly helpful in reasoning on the two indeterminate argument types, DA and AC. On the other hand, a determinate type that requires processing of a negated proposition (MT) should be easier with rules acquired from smaller sets because cognitive load is reduced, and because the availability of alternatives is not a factor with determinate forms.  4. Analysis of Research Issues Related to the Effects of Inductive Acquisition Mode in Near-Transfer of an Acquired Rule.  21  Another question concerning inductive rule transfer is whether the rule was induced in predictive or diagnostic contexts. A brief review and an analysis of relevant research findings, and an articulation of relevant theoretical assumptions is in order.  The central idea in this regard is the ability of a reasoner to access instances of a rule during deductive reasoning. The argument put forward here is that if instance-based accounts of reasoning are correct, then conditional reasoning should be facilitated by access to specific instances of an acquired rule. In turn, memory access should be facilitated by transfer-appropriate processing. Two processing modes explored here are the predictive and the diagnostic modes. What follows is an attempt to apply the principle of “Transfer Appropriate Processing” to the transfer of a conditional rule, where “processing” refers to the directionality of reasoning involved in predictive and diagnostic modes of reasoning.  The problem is to identif,’ specific parameters of the inductive rule learning task that might affect the ease with which rule instances can be accessed during the deductive reasoning phase. First, it is necessary to state some assumptions about representational and processing aspects of memory implied by instance-based accounts of reasoning. Assumed representation of instances in memory:-What might a memory representation of rule instance look like? Anderson & Bower (1973) and Hinton and Anderson (1981) proposed a simple model of verbal associative memory where objects or events, which may include conditional rule knowledge of antecedents and consequents, are connected by bidirectional associational links where activation can spread from antecedents to consequents or visa versa. An important feature of this representation is  22  that the strength of the activation is not assumed to be symmetrical. In other words, the strength of association between X and Y is not necessarily the same as the strength of association between Y and X. Below is one example of such a knowledge representation acquired from exposure to a set of cause-effect relations with a conditional structure: ci c2 c2  >>> >>> >>>  ci el e2  It should be noted that all factors found to influence conditional reasoning performance, including clause relatedness, imagery, and generation of alternatives, are all factors that motivated the development of associative models of memory. Put another way, these content effects are consistent with memory models based on spreading of activation and strength of association between objects or events stored in memory (Ross &Bower, 1981).  Instance-based accounts suggest that conditional rule knowledge may be acquired from exposure to specific instances (e.g. causes and their associated effects) and that the instances are associated via spread of activation along asymmetrical links from one to the other (e.g. links from causes to effects, or from effects to causes). Given this type of associational memory structure, the next question concerns the retrieval of specific instances of antecedents and consequents.  Retrieval of instances from memory and transfer-appropriate processing: According to Tulving’s Principle of Encoding Specificity, “remembering of events always depends on the interaction between encoding and retrieval conditions” (Tulving, 1984, p.242). A similar position is expressed in the principle of Transfer Appropriate Processing put forward by Morris, Bransford, & Franks (1977). They stress that “the value of  23  particular types of acquisition activities must be defined relative to the types of activities to be performed at the time of the test” (p.53 1). The principles of Transfer Appropriate Processing and Encoding Specificity do not restrict the nature or the types of processes involved during acquisition and retrieval. The principles only require that the processes be “similar”. What is meant by the term “similar”, at least within the present study, will be addressed shortly. What implications do these principles have for a study of rule transfer?  According to an instance-based account, conditional reasoning should be facilitated by access to rule-related instances in memory, where memory access in turn depends on the similarity between encoding conditions and retrieval conditions. If the type of processing during rule acquisition is in some respect similar to the type of processing required during rule application, then access to rule instances should be facilitated and reasoning performance should improve. This is a general statement of the last hypothesis. To develop it further it is necessary to explain what is meant by “similar type of processing”.  Predictive and diagnostic inductive reasoning in rule transfer:-There are two principal modes of reasoning associated with conditional rules of the form IF X THEN Y: predictive and diagnostic. Predictions typically require reasoning from causes to effects. Diagnoses, on the other hand, involve reasoning in the opposite direction, from effects to causes. In a naturalistic setting, a conditional (or any other) relationship between causes and effects may be learned in either one of these modes by being exposed to causes linked to effects, or effects linked to causes (Bjorkman & Nilsson, 1982; Waldman & Holyoak, 1992).  Predictive and diagnostic modes of reasoning are important not only in learning logical rules but also in using the rules in deductive reasoning tasks (Evans & Beck, 1981). 24  To see why, consider the conditional syllogisms presented earlier. Note that the only difference between the four types of syllogisms can be found in the minor premise and in the conclusion (the major premise is the same for all of them). In the MP and DA forms, the minor premise refers to an antecedent (a cause), while the conclusion refers to a consequent (an effect). In other words, here a cause is given and an inference has to be made regarding an effect. This is precisely the direction of reasoning required when making a prediction from cause to effect. In contrast, the MT and AC argument forms -  contain a minor premise that references a consequent (an effect), and a conclusion that references an antecedent (a cause). Here the effect is given, and an inference has to be made regarding a cause. The direction of reasoning in these two cases thus corresponds to making a diagnosis from an effect to a cause. -  On the basis of the foregoing analyses, the last hypothesis can now be stated more specifically. It is hypothesized that when the reasoning mode (predictive / diagnostic) under which a conditional rule is acquired matches the reasoning mode required in the deductive task, memory access to specific rule instances alternative causes and effects  -  -  particularly  is facilitated, resulting in improved reasoning  performance as compared to conditions where the modes of reasoning do not match.  E. Summary of Hypotheses and Experimental Predictions  First, it is hypothesized that familiarity (in terms of lexical attributes) enhances the application of the acquired rule. In other words, familiarity with rule attributes facilitates transfer of rule knowledge. Consequently, it is predicted that near-transfer deductive  25  performance (i.e. where the rule content in the deductive task matches the content encountered during inductive rule acquisition) should be better than deductive performance on far-transfer (i.e. where the rule content on the deductive task is different from the content encountered during rule acquisition). Specifically, it is predicted that there will be fewer reasoning errors on the deductive transfer task with rules that have content which matches the content used in the inductive rule acquisition task, as compared to rules with content that does not match the inductive task content.  Second, some situations can arise where the rule content does not match the acquisition content, and cueing of specific relevant instances is unlikely. When that happens, it is hypothesized that accurate deductive inferencing presumably relies on generalized, content-free rules. Such rules are more readily acquired from exposure to large rather than small sets of exemplars. It is predicted, therefore, that rules acquired from larger sets of training examples should transfer more readily to inferencing situations with different propositional content as compared to rules acquired from small exemplar sets. If this is the case, there will be fewer reasoning errors with content that is different from the acquisition content with rules induced from instances with trinary, as compared to binary attribute dimensions.  Third, it is hypothesized that increased availability of alternative antecedents and consequents associated with larger sets of rule exemplars during inductive training can be particularly helpful in reasoning on the two indeterminate argument types, DA and AC. In contrast, a determinate type that requires processing of a negated proposition (MT) should be easier with rules acquired from smaller sets because cognitive load is reduced, and because the availability of alternatives is not a factor with deterministic types.  26  More specifically, as compared to rules acquired from sets of instances with binary attribute dimensions, there will be fewer reasoning errors with rules acquired from sets of instances with trinary attribute dimensions on the DA and AC forms. Also, as compared to rules acquired from sets of instances with trinary attribute dimensions, there will be fewer reasoning errors with rules acquired from sets of instances with binary attribute dimensions on the MT argument form. Finally, when the mode of reasoning (predictive vs. diagnostic) under which a conditional rule is acquired matches the reasoning mode involved in the deductive task (MP, DA vs. AC, MT), it is hypothesized that memory access to specific rule instances, particularly alternative causes and effects, is facilitated, resulting in improved reasoning performance as compared to conditions where the modes of reasoning do not match. More specifically, compared to rules acquired in a diagnostic context (AC, MT), there will be fewer deductive reasoning errors with rules acquired in a predictive context on the IVIP and DA argument types.  27  II. METHOD  A. Subjects and Design  The subjects were 120 University of British Columbia undergraduates in the Faculty of Education who were registered in the Primary Teacher Training Program. Elementary student teachers were selected because their background knowledge is more homogeneous than secondary students who specialize in different subject matter contents. The aim was to minimize any potential influence of prior knowledge. The sample comprised 42 males and 78 females, reflecting the overrepresentation of female students in the primary program. All subjects were recruited by the experimenter from 15 sections of a 400-level course on Evaluation in Teaching. Participation was on a volunteer basis only. Subjects were paid $20 for their participation.  Experimental conditions:-Subjects were randomly assigned to one of three groups: binary, trinary, and a control group. Within each group, subjects were assigned to one of 24 trial sequences as required by counterbalancing the experimental tasks. (Counterbalancing is described in more detail below). A complete replication of the counterbalanced task materials and task orders required 24 subjects in each group. In order to improve the reliability of the results, and to increase the sensitivity of statistical analyses, one more full replication of each training treatment group was completed by assigning 48 (2x24) subjects to each of two treatment groups. However, the limited logistics of the research program (i.e. funds, subject availability, and time) did not permit another replication of 24 for the control group.  28  The inductive training conditions for rule acquisition were defined in terms of two factors: (a) the size of attribute dimension (i.e., binary vs. trinary) and (b) the mode of instance presentation (i.e., predictive vs. diagnostic). The former varied as a betweensubjects factor and the latter as a within-subjects factor. Since transfer relations between inductive rule learning and subsequent application of this knowledge to deductive reasoning is the primary focus of the present study, the conditions under which transfer can occur need to be specified.  Transfer type is the first within-subjects design factor. The hypotheses require the specification of two types of transfer: (a) near-transfer, where the content encountered during inductive rule learning is the same as the content in the deductive reasoning task, and (b) far-transfer, where the deductive syllogisms contain content that is unrelated to any of the content encountered during inductive rule learning. The hypotheses also require the specification of two modes of inductive rule learning, predictive and diagnostic. As a result, the near-transfer condition is fi.jrther subdivided into familiar content encountered during prediction-based rule learning, and familiar content encountered during diagnosisbased rule learning. These factorial variations of the familiarity and the mode of induction did not cross over the far-transfer condition. In addition, it was necessary to define another within-subjects factor: type of deductive argument form (MP, DA, AC, and MT).  Transfer type and syllogism type were defined as within subject factors because using a within subject design has the advantage of reducing error variance due to individual differences. Attribute dimension size was treated as a between subject factor, however, because of possible asymmetric carry-over effects from the trinary to the binary rule acquisition conditions. As was noted earlier, a trinary set has more distinct instances and as such can be expected to give rise to more generalized causal schema which is  29  hypothesized to influence subsequent learning and transfer (Gick & Holyoak, 1987). The complete experimental design is depicted in Figure 1 below.  Deductive Transfer Inductive Learning  —  Ss  Grp  Rule  induction mode  1 BIN 48 BIN  PredlDiag Diag/Pred  49 TRI 96 TRJ  Pred/Diag DiagfPred  Far  Near Prediction related content  Diagnosis related content  Unfamiliar content Trial 1  Unfamiliar content Tnal 2  MPDAACMT MPDAACMT MPDAACMT MPDAACMT  97CTRL l2 CTRL Figure 1. Experimental Design Layout  As can be seen in Figure 1, the design required three groups of subjects, two experimental and one control. Including a control group provided an opportunity to assess the overall effect of the inductive acquisition phase. The performance of the treatment groups may be similar but better than if no rule learning took place. On the other hand, the treatment groups may perform differently, but the level of performance for both groups may actually be lower as compared to a no-treatment condition (i.e. negative transfer). Using a control group enabled a more precise assessment of these potential outcomes.  The design layout also shows two trials under the far-transfer condition. This decision was justified on the grounds that having an extra replication would increase the 30  reliability of the data and, more importantly, the extra trial would make it possible to assess potentially different practice effects on far-transfer for the two treatment groups.  B. Experimental Tasks and Materials  Rule acquisition task:-This task was a modification of the traditional concept learning paradigm where subjects learn a target rule by observing positive and negative instances of the rule. The task was to learn a rule that reflected the relationship between dimensional attributes of interest, which in this study were specified as causes and effects. The rationale for this is based on the fact that reasoning about causes and effects is a ubiquitous cognitive activity which typically requires both inductive and deductive thought, and often involves conditional relations (e.g. Sternberg, 1986; Carison & Schneider, 1989; Cheng & Novick, 1992). In the binary-dimension set, two different causes and two different effects appeared. In the trinary set, three different causes and effects were presented. The relations between causes and effects under each rule truth table class are summarized in Figure 2. A listing of examples of causes and effects used in this study appear in Figure 3 below.  Rule Truth table Class  Set Size  ++  +—  —+  Binary  cl—el  cl—e2  c2—el  c2—e2  Trinary  cl—el  cl—e2  c2—el  c2—e2  cl—el  cl—e3  c3—el  c2—e3 c3—e2 c3—e3  Figure 2. Causes and effects under different rule truth-table classes  31  Content A. Drugs and side effects Target Rule:  IF the drug Mixolin is administered TEEN purple skin rash develops  Causes and effects used for rule acquisition: cl=Mixolin  el=Purple rash  c2=Phoresin  e2=Blue rash  c3=Benuvin  e3=Green rash  Content B. Radiation sources and plant mutations  Target rule:  IF a plant is exposed to Gamma radiation TEEN it grows square leaves  Causes and effects used for rule acquisition: cl=Gamma radiation  el=Square leaves  c2=Beta radiation  e2=Triangular leaves  c3=Alpha radiation  e3=Circular leaves  Content C: Chemical agents and fish Target rule:  IF agent TX is released in the water TEEN fish develop brittle scales  Causes and effects used for rule acquisition: cl=Agent Tx  el=Brittle fish scales  c2=Agent Ty  e2=Enlarged fish scales  c3=Agent Tz  e3=Loose fish scales  Content D: Lasers in eye surgery and retinal damage Target rule:  IF an Argon laser is used TEEN checkered spots appear on the retina  Causes and effects used for rule acquisition cl=Argon laser  el=Checkered spots on a retina  c2=Ruby laser  e2=Striped spots on a retina  c3=krypton laser  e3=Dotted spots on a retina  Figure 3. Types of content used in the inductive learning task  Each training set contained 36 instances comprising nine, randomly ordered instances of each truth table class. In the first phase of the rule acquisition task, subjects were presented with a series of predictions or diagnoses, one by one. They were asked to look at each case (i.e. a rule instance) and decide whether the stated prediction or diagnosis was true or false. Each case consisted of a cause followed by an effect (for the prediction condition), or an effect followed by a cause (for the diagnostic condition). Subjects received corrective feedback after each response, and then a new case was presented. Each case was selected by the computer at random. The process continued until a mastery criterion was reached. This criterion was set at two replications of nine instances, yielding a total of 18 consecutive correct responses.  32  During this phase of the inductive rule acquisition task, subjects were able to record previously presented instances along with their status (true/false). The rationale for the simultaneous availability of previously presented instances was two-fold: (a) it eliminated unnecessary working memory load which is desirable since the focus of the study was on cognitive processes independent of working memory load; and (b) it maximized the likelihood of task completion by reducing the time to learn the conditional rule.  In the prediction-context condition, each trial involved a brief description of a causal agent being investigated by a research team. The subject was told that prior to conducting their experiment, the researchers predicted that the particular causal agent will result in some specific effect. The subject was then asked to indicate whether the stated prediction was correct or incorrect.  In the diagnostic-context condition, each trial started with a presentation of some effect (or symptom). This was followed by a description of a diagnosis made by a team of investigators. The subject was then asked to indicate whether the stated diagnosis proved to be correct or incorrect.  Rule verbalization and selection task:-In the rule verbalization task presented in a questionnaire, subjects were asked to write down the strategy they used to decide whether a given prediction or diagnosis was correct. The multiple choice task was then introduced. It was intended to assess subjects’ ability to identif’ a rule statement expressed in an IF.. THEN. .language that best summarized the conditional relationships they encountered in the inductive phase. The task simply required subjects to select one of four options on a multiple choice item, which, in their opinion, represented the most appropriate expression of the observed relationship between causes and effects. The options are listed below. 33  1. The most parsimonious and general (target rule) statement which had the general form: IFcl THENe1  (e.g. If Mixolin is administered, then purple rash develops.) 2. A correct, but less parsimonious rule statement in the form: IF c2 THEN el or e2 (e.g. If Phoresin is administered, then either purple or blue rash develops.) 3. A rule statement based on an incorrect, reversed inference: IF el THEN ci (e.g. If purple rash develops, then Mixolin was administered.) 4. An incorrect rule statement related to a biconditional interpretation: (IFcl THENe1)AND(IFel THENc1) (e.g. If Mixolin is administered then purple rash develops, and if purple rash develops then Mixolin was administered.)  The task was also designed to help subjects learn to associate the appropriate IF.. THEN.. language with the knowledge structure they induced from the corresponding set of instances. This was accomplished by cueing pragmatic constraints of generality and parsimony discussed earlier. Accordingly, subjects were asked to choose the simplest, most concise, yet general proposition that held true for the set of cases (instances) as a whole.  Conditional syllogism reasoning as transfer task:-In the deductive task, the four syllogistic types were presented one by one in random order. Each argument form was presented with both an affirmative and negated conclusion, yielding a total of 8 distinct syllogism items.  In order to facilitate a naturally-occurring reasoning context for the subjects while not altering the traditional format of the task, subjects were told that each syllogism was a summary of an argument made by a particular student in a high school science class. Subjects were told that each student in the science class had been exposed to the same set  34  of cases (instances) as they themselves just encountered. They were asked to put themselves in the role of a teacher who is marking the validity (correctness) of each conclusion students arrived at. A validation response to a given syllogism therefore became equivalent to a decision on whether a particular “student’s conclusion” held true always, sometimes or never. No feedback was given, and each response as well as response latency were automatically recorded.  C. Apparatus  All the tasks were programmed in Turbo Pascal and were presented on an IBM 386 microcomputer. Response-time (in millisecond accuracy) was recorded for every response required in each task. Time data were collected using Assembly Language timing subroutines interfaced with Pascal code.  D. Procedure  Trial sequence counterbalancing:-Four different content areas were necessary for the conditional rule used in the deductive transfer tasks, two for near-transfer (predictionrelated, and diagnosis-related contents), and two for far-transfer (unfamiliar content trials 1 and 2). Four types of content taken four at a time yielded 4!  =  24 distinct transfer trial  sequences counterbalanced for content effects.  The inductive task required that two of the four available content types be presented in two different acquisition contexts. This yielded 4!/(4-2)!  =  12 induction  sequences counterbalanced for content. Counterbalancing to control for sequence effects on the induction tasks then resulted in a total of (2)12=24 unique induction trial  35  sequences. The complete set of 24 induction sequences followed by 24 deductive transfer sequences is shown in Figure 4. This arrangement also ensured that each type of content is equally represented in each type of transfer.  Deductive Rule Transfer Trials  Rule Induction Exp. Trials Sequence  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24  p d A B C D  = = = =  =  pA pA pA pB pB pB pC pC pC pD pD pD dA dA dA dB dB dB dC dC dC dD dD dD  prediction diagnostic Content A: Content B: Content C: Content B:  A B C D A B C D A B C B A B C D A B C D A B C B  dB dC dD dA dC dD dA dB dD dA dB dC pB pC pD pA pC pD pA pB pD pA pB pC  B A A A B A A A C C B B C C B B D D B C B B D C  C C B B B B B C B A D A B B A C B A A A C C B B  D D D C C C B B B B A C B A B A C C B S B A A A  mode mode Drugs and side effects Radiation and plant mutations Chemical agents and fish scales Lasers and retinal damage  Figure 4. Task sequences used for each group  The experiment was carried out in a laboratory setting. Two subjects were run simultaneously, each seated at one of two computer work stations located in the lab. Subjects were given an individualized 16 page booklet which explained the purpose of the study, provided an overview of the tasks, and gave step-by-step instructions on how to proceed with each task. The booklet also contained space for subjects to record any computer feedback, if they wished (see Appendix B).  36  In order to ensure that subjects read all instructions carefully, the computer keyboard was disabled between tasks and required subjects to enter a code to reactivate it. A different code was embedded at the end of each block of instructions corresponding to a particular task.  Following the completion of each inductive task, subjects were asked (in the booklet) to write down their strategy for deciding whether a given prediction or diagnosis was correct or not. They then completed the multiple choice, rule-selection task. This completed the inductive phase. Subjects then proceeded to the sequence of deductive tasks. A brief introduction to each transfer task was given. For the near-transfer tasks, it was pointed out to the subjects that the content they were about to deal with was related to the content on one of their inductive tasks. Far-transfer tasks were introduced by stating that although the content they were about to see was unfamiliar to them, their responses were nevertheless of great interest to the experimenters. Each subject completed four (2 near and 2 far transfer) sets of eight deductive problems. At the end of the computer session, subjects filled out a brief questionnaire containing basic demographic items, including gender, age, and educational background.  37  ifi. RESULTS  The data analysis plan is as follows. First, the training data is analyzed to determine whether meaningful learning did in fact occur. This is followed by a detailed analysis of the response data on the deductive transfer tasks, and tests of the relevant hypotheses.  A. Analysis of Inductive Rule Training Performance  An analysis of the inductive rule learning phase of the experiment was carried out first to establish the validity of the design by verifying: (a) that rule acquisition has taken place, and (b) that the training outcomes including the total number of instances and the total processing time required to reach criterion performance have values consistent with the different training conditions.  Analysis of Number of Instances and Processing Time Required to Reach Criterion Performance:-Outliers in the observed distribution of the total number of instances and processing time were identified using Studentized residuals obtained from fitting a linear model to the four training variables which included the total time and total instances for the prediction and diagnostic conditions. A residual value of 2.5 was used for cutoff to identifying the outliers (Beisley, Kuh & Welch, 1980). Four outlier values (two in each group) were identified by this method and were replaced by the corresponding group means based on the remaining cases. The resulting means and standard deviations of the total processing time and total number of instances required to reach criterion performance in each training condition appear in Table 1 (see also Figures 5 and 6 for a  38  graphical representation). It should be noted that in subsequent analyses of this data, the degrees of freedom were reduced by the number of observations for which the mean values were substituted.  Acquisition Mode Training Variable  Prediction  Diagnosis  Binary Group (n=48) No. of instances 27.27 26.23 6.07 8.87  N SD  Total Time M SD  456.02 255.39  593.33 352.50  Trinary Group (n=48) No. of instances 31.78 30.33 7.71 10.77  M  Total Time 577.42 231.83  M SD  615.53 234.58  Table 1. Means and standard deviations of number of instances and total processing time for different inductive rule acquisition conditions  Mean No. of Instances Required for Rule Acquisition 4O 35  EL  30 Z  25  • Binary Group rrinary Group  20 Diagnosis  Prediction Acquisition Mode  Figure 5. Number of instances for rule acquisition  39  Mean Processing Time Required for Rule Acquisition 700  -!.  600  • Binary Group UTrinary Group  500  400 Diagnosis  Prediction  Acquisition Mode  Figure 6. Processing time for rule acquisition  Both the instance and response time variables were transformed prior to analysis to stabilize any departures from a normal distribution. The instance variable was transformed with the square root transformation and the response time variable with the log transform (Box & Cox, 1964). The correlation between the total number of instances and the total processing time was moderate, with Pearson r =.32, (<.OO1) in the prediction condition, and  .49, (p<.000) in the diagnostic condition.  The response time and instance data were analyzed together using multivariate repeated measures analysis of variance (MANOVA). The analysis of the response measures revealed that the total number of instances and total processing time required to reach the mastery criterion of 18 consecutive correct responses by the trinary group were significantly greater and longer than those required by the binary group, H-L Trace=.25, (s=1, m=-2, n=2),  p<.000; F’s(1,94)=14.16, (MSE=1.05), and 6.32, (MSE=.60), p’s<.OOO  40  and .014 respectively, as expected. The corresponding effect sizes (‘s) were .38 and .26 for the instance and time measures, respectively. The diagnostic mode of rule induction did not require significantly more instances but it did require significantly more processing time, as compared to the prediction mode, ‘s(1,94)=.74 and 5.91, p’s<.391 and .017, for the instance and time measures, respectively. The corresponding effect sizes were .18 and .50 for the instance and time measures, respectively.  These results are consistent with the expectation that rule induction from a set of diverse instances should require more processing time compared to a set of instances that are less variable. On the average, each new instance from a trinary set provides more unique information about the underlying relationship than an instance from a binary set. Processing the trinary information should therefore require more time but not necessarily more instances, which is what the results have shown.  Subjects’ Introspection Reports of Their Inductive Learning Strategies:-In order to learn more about subjects’ inductive learning strategies, at the end of each inductive task subjects were asked how they were able to tell whether a particular prediction or diagnosis was correct or not. All responses were coded, yielding the following response categories: Category 1: No Articulated Strategy (Trial & Error) This category contains responses which show no evidence ofany systematic strategy. A strategy may have been used but a subject was not able to verbalize it. A typical response was: “I don’t know, it was basically trial and error  Category 2: Memory Processing Strategies This category includes responses characterized as memory-based enumeration strategies since they contained lists of various cause and effect relationships that subjects  41  encountered The most typical response was: “I recalled that Mixolin causes only purple rash, but Phoresin causes purple rash, and it can also cause blue rash”.  Category 3: Coding Strategies Responses in this category revealed afocus on particular attributes within the set of instances. The attributes subjectsfocused on included either relevant or irrelevant ones. The key characteristic of this category ofresponses is the reported attention to specfic attributes. Typical responses were: “I basicallyfocused on which diagnoses were incorrect”; “Ipaid attention to the cities where the experiments were done “.  Category 4: Rule-Related Strategies Responses in this category reflected a systematic search for a rule. These subjects typically drew a two-way or a three-way table where they systematically recorded the observed relationships between causes and effects. They also produced a concise, general statement which summarized the underlying relationship. Typical response was: “The basic pattern was that Phoresin causes dfferent rashes but Mixolin causes only purple rash “.  Frequencies of reported strategy categories appear in Table 2. Acquisition Mode  Self—Report Strategy  Prediction  Diagnosis  Binary (n=48) 1. 2. 3. 4.  6 28 10 4  Trial & error Memory coding Rule—related  7 27 9 5 Trinary (n=48)  3 32 4 9  1. Trial & error 2. Memory 3. coding 4. Rule—related  3 35 2 8  Table 2. Category frequencies for reported inductive learning strategies  42  An inspection of the data in Table 2 clearly reveals that the most frequently reported strategy under prediction as well as under diagnosis was a memory-based strategy. Explicit rule-based strategies were infrequently reported, particularly by the binary group. Trial and error strategies were also infrequent, particularly in the trinary group. One thing that is clearly evident in the verbalized reports is that the subjects do not engage in demanding cognitive operations, as shown in the frequency data (i.e. category 1 and 2 vs. category 3 and 4), although the trinary group showed a little more effort in searching some rule or pattern.  Assessment of the Propositional Form of an Acquired Rule:-Although all subjects reached the criterion performance level on the inductive task, their ability to identify the appropriate rule when expressed in the standard conditional language of the form ]F x THEN y was also assessed using a four-option, multiple choice question. The distribution of responses for each option is given in Table 3.  43  Acquisition Mode Option*  Prediction  Diagnosis  Binary (n=48) 1 2 3 4  1 2 30 15  1 2 3 4  0 4 6 38  1 3 29 15 Trinary (n=48) 3 8 8 29  Item; Based on what you have observed in the set of examples just presented, which of the following statements represents a CONCISE, GENERAL rule that BEST SUMS UP the relationship between the type of  (cause)  and the type of (effect)?  *Options: 1.  If  (effectl)  2.  If  (causel)  and if 3.  If  4.  If  then then  (effectl)  (cause2)  (effectl)  or  (causel)  (causel) )effectl) then  (causel)  then either )effect2) then  (effecti)  Table 3. Response frequencies for each option on a rule assessment item  As can be seen in Table 3, the vast majority of subjects in both groups selected one of the two propositional statements which correctly described the conditional relationships used for training (options 3 and 4). A test of the proportion of subjects in each group who selected one of these two options revealed that subjects in the trinary group selected the correct statement for the target rule (option 4) significantly more often than those in the binary group, under the prediction mode as well as under the diagnosis mode, Chi Sguare(1, j 96)  22.28,  p<.000, and Chi-Sguare(1,  =  96)  =  8.22, p<.OO1, for the  prediction and diagnosis mode, respectively.  The rule selection results suggest that rule induction from a trinary set, because of its greater variability of instances, increases the need for a more integrated, parsimonious representation of the observed set of relations. The target rule statement exemplifies such 44  a representation, and the fact that the trinary group selected it more frequently provides additional evidence for the validity of the inductive training task.  To see whether self-reported strategies were congruent with selection of the target rule statement, the number of subjects who selected the target rule statement was calculated for each self-report strategy category. The results for each group (binary and trinary) are shown in Tables 4a and 4b.  Rule Options Selected  Strategy category  Other (options l,2,or3)  Target (option 4)  1. 2. 3. 4.  Trial & error Memory coding Rule—related  Prediction 1 8 3 3  1. 2. 3. 4.  Trial & error Memory coding Rule—related  Diagnosis 1 9 4 1  (N=48) 5 20 7 1 (N=48) 6 18 5 4  Table 4a. Target rule selection by reported strategy for the binary group  Rule Options Selected  Strategy category  Target (option 4)  Other (options l,2,or3(  1. Trial & error 2. Memory 3. coding 4. Rule—related  Prediction 2 24 4 8  1. Trial & error 2. Memory 3. coding 4. Rule—related  2 20 1 6  Diagnosis  (N=48) 1 8 0 1 (N=48) 1 15 1 2  Table 4b. Target rule selection by reported strategy for the tnnary group  45  When pooled together across the two groups, the data show only a weak relationship between target rule selection and reported strategy in the prediction mode, Chi-Square (3, N  96)  =  6.53, R< . and an absence of any significant relationship 091  between these two variables in the diagnostic mode, Chi-Square (3, N = 96) = 1.37, p>.713. In the prediction mode, of the thirteen subjects who reported a rule-related strategy, eleven in fact selected the target rule statement, but in the diagnostic mode only seven out of thirteen with a rule strategy selected the target rule.  These findings suggest, although not conclusively, that subjects who relied on a rule-based strategy in the prediction mode may have induced a representation of a rule that more consistently matched the target rule statement as compared to the representation induced under the diagnostic mode. This is a reasonable explanation given that the target rule is expressed in the form of a predictive rather than a diagnostic proposition.  B. Analysis of Deductive Reasoning Performance This section describes the analysis of transfer data collected from the deductive reasoning performance. The main measure used in the analysis of deductive reasoning performance was a response accuracy score on the deductive syllogisms. Each syllogism was scored according to the normative conditional interpretation of each argument type. The scoring key is shown in Table 5.  46  Syllogism type  Second premise  First premise  Modus Ponens Modus Ponens Denying Antecedent Denying Antecedent Affirming Consequent Affirming Consequent Modus Tollens Modus Tollens  P P P P P P P P  -> -> -> -> -> -> -> —>  A S N  =  =  =  Q Q Q Q Q Q Q Q  Conclusion  P P -P -P  Q Q -Q -Q  Correct response  Q —Q Q —Q P —P P —P  A N S S S S N A  Always true Sometimes true Never true  Table 5. Scoring key for 8 types of conditional syllogisms  Applying the scoring protocol to each response yielded an accuracy score of 0 or 1 for each subject on each syllogism. Since each syllogism type was presented with both an affirmative and a negated conclusion, two scores for each syllogism type were obtained. These scores were added together, resulting in one score for each type of syllogism that ranged from 0 to 2 for each subject. The means and standard deviations of these scores are listed in Tables 6, 7, and 8 for the binary, trinary, and the control groups, respectively.  47  Far  Argument type  Transfer  Trial 1  Trial 2  Near—transfer  Prediction  Diagnosis  Modus Ponens N  1.77  1.89  1.81  1.90  SD  0.55  0.31  0.49  0.43  Denying Antecedent M  1.04  1.19  1.50  1.48  SD  0.87  0.89  0.77  0.82  Affirming Consequent M  0.94  1.02  1.38  1.08  SD  0.83  0.93  0.82  0.85  Modus Tollens N  1.68  1.52  1.69  1.50  SD  0.62  0.74  0.59  0.68  Table 6. Mean scores and standard deviations on deductive reasoning performance for the binary group (N=48)  48  Far  Argument type  Transfer  Trial 1  Trial 2  Near—transfer  Prediction  Diagnosis  Modus Ponena M  1.70  1.83  1.83  1.71  SD  0.62  0.43  0.48  0.61  Denying Antecedent N  1.08  1.23  1.29  1.27  SD  0.87  0.86  0.85  0.81  Affirming Consequent M  0.81  0.92  1.27  1.10  SD  0.91  0.92  0.89  0.97  Modus Tollens M  1.66  1.56  1.54  1.52  SD  0.61  0.58  0.62  0.65  Table 7. Mean scores and standard deviations on deductive reasoning performance for the trinary group (N=48)  49  Far  Argument type  Transfer  Trial 1  Trial 2  Near—transfer  Prediction  Diagnosis  Modus Ponens N  1.63  1.58  1.58  1.50  SD  0.71  0.65  0.66  0.72  Denying Antecedent M  1.37  1.38  1.21  1.25  SD  0.88  0.77  0.88  0.79  Affirming Consequent N  1.16  1.00  1.04  1.17  SD  0.92  0.93  0.86  0.87  Modus Tollens N  1.33  1.54  1.46  1.33  SD  0.76  0.83  0.83  0.75  Table 8. Mean scores and standard deviations on deductive reasoning performance for the control group (N=24)  It will be recalled that the experimental design specified one between-subjects factor, namely the Treatment Group (Binary, Trinary, and Control), and two withinsubjects repeated measures factors. The repeated measures factors were (a) Transfer Type which included two trials of far-transfer, and two trials of near-transfer (prediction-related / diagnosis-related), and (b) Syllogism Type (MP,DA,AC,MT). Performance scores obtained under this experimental design were analyzed with a multivariate analysis of variance (MANOVA) by 1-df linear contrasts performed on the mean vectors. It should be noted that although the basic task response unit is binary (correct/incorrect), reasoning scores are based on multiple item responses, and multiple subjects.  50  Consequently, the deductive reasoning data can be appropriately analyzed by standard multivariate methods (Collett, 1991). The analysis was carried out in two steps. First, following the General Linear Model (GLM) approach, a model was fitted to the data, model parameters were estimated, and tests of main effects and interactions were performed. An overall Type I error of .10 was used for the omnibus tests. Second, individual 1-df contrasts corresponding to the predictions derived from the hypotheses were utilized to test the specific hypotheses. To control for the pyramiding of contrast probabilities, a conventional alpha level of .05 was used on tests of the individual contrasts (Huberty & Smith, 1982).  As a first step, the overall transfer performance was analyzed. Performance on the four argument types was combined, yielding a total deductive performance score ranging from 0 to 8 for each subject in each transfer condition. The data is shown in Table 9.  Far  Trial 1  Near—transfer  Transfer  Trial 2  Prediction  Diagnosis  Binary ( N=4 8) M  5.44  5.63  6.38  5.96  SD  1.89  2.02  1.66  1.65  Trinary ( N=4 8) M  5.23  5.54  5.94  5.60  SD  1.71  1.66  1.56  1.84  Control (N=24) M  5.50  5.50  5.29  5.25  SD  2.32  2.02  2.21  2.29  Table 9. Mean scores and standard deviations for overall deductive reasoning performance  51  The analysis revealed significant interactions between training conditions and transfer types E(6,351)=1.84, (MSE=.32), p<O l. In addition there was a significant main 9 effect for transfer type E(3 ,3 51)3.37,  <. 019,  however the main effect of inductive  training conditions was not significant E(2,1 17)0.77, >.47. The results of the omnibus tests warranted more detailed testing of the individual hypotheses using 1 -df multivariate contrasts. Inspection of the data suggests that the effect of inductive training is present on the near-transfer reasoning problems but not on the far-transfer problems. The specific nature of this transfer can be further examined by detailed analyses.  Tests of Specific Hvpotheses:-The first hypothesis addressed rule transfer performance as a firnction of content similarity. The verifiable prediction proposed on the basis of this hypothesis was that the overall near-transfer performance (i.e. performance across all argument types) would be better than the overall far-transfer performance for both the binary and trinary groups. No such difference would be predicted for the control group since it did not have the benefit of the inductive rule learning. Near and far-transfer performance data are summarized in Figure 7.  52  _______  Overall Deductive Transfer Performance 8  0 0 Cl)  7  C  INearI lOFar  I  a,  4  Trinary  Binary  Control  Treatment Group  Figure 7. Overall transfer performance  The analysis was further carried out using a contrast of overall performance on near and far-transfer deductive tasks. The analysis therefore also served as an additional verification of the overall validity of the transfer of training design. If any meaningful inductive learning occurred, a difference between near and far-transfer performance should be observed for the treatment groups (but not for the control group) since the neartransfer tasks used rule content that was related to the content in the inductive learning phase.  Indeed, overall performance on near-transfer was significantly better than on far transfer, (1,1 17)  =  4.57, (MSE  =  6.58), c035, and further, the interaction between  transfer and the three treatment groups was also significant, E(2, 117)  =  3.65, p<.029.  Furthermore, this performance difference was due to the two treatment groups outperforming the control group, E(1,117)  6.38, <.013. However, the two treatment  groups did not differ from each other significantly in their overall near versus far-transfer performance, E(1, 117)  =  0.91, >.342.  53  Since the transfer types were varied as a within-subjects factor with nested definitions of reasoning modes and trials, the following hypotheses were tested with each nested factor (cf Figure 1). Thus, the second hypothesis addressed far-transfer performance as a function of the training content of varying attribute dimension size (binary vs. trinary). From this hypothesis, which was elaborated earlier, it was predicted that there would be fewer reasoning errors with rule content that is different from the acquisition content (far-transfer condition) with rules induced from instances with trinary, as compared to binary attribute dimensions.  Results pertaining to the prediction based on the second hypothesis appear in Figure 8. The prediction was tested with a contrast of performance scores of the three groups on the four syllogisms with unfamiliar content. No difference in performance was found between the binary and trinary groups F(2, 117)  =  .17, (MSE = 12.15), >.683. In  addition, the performance of these two treatment groups was not significantly different from control group performance E(1,117)=.O1, >.9l’7.  Far-Transfer Performance by Attribute Dimension Size 2 a U  0  •Binary  0  Trinary 0  ControI  0 MP  DA  MT Syllogism Type  Figure 8. Far transfer by dimension size  54  AC  These results therefore falsify the second hypothesis. The inductive training with a trinary set does not facilitate far-transfer to a greater degree than a binary set of dimensioned materials. Furthermore, the results show that the rule knowledge the two treatment groups acquired during the inductive phase did not transfer to completely new, unfamiliar rule content.  The third hypothesis addressed near-transfer of an acquired rule as a ffinction of the size of the contrast set from which the rule is induced. Two predictions arising from this hypothesis were verified: (a) as compared to rules acquired from sets of instances with binary attribute dimensions, there will be fewer reasoning errors with rules acquired from sets of instances with trinary attribute dimensions on the DA and AC forms; and (b) as compared to rules acquired from sets of instances with trinary attribute dimensions, there will be fewer reasoning errors with rules acquired from sets of instances with binary attribute dimensions on the MT argument form. The data pertaining to predictions based on this hypothesis are summarized in Figure 9.  Near-Transfer Performance by Attribute Dimension Size 2 0 U Co  • Binary  0  Trinary 0  a,  MP  DA  MT Syllogism Type  Figure 9. Near transfer by dimension size  55  AC  The predictions were verified with a contrast of near-transfer performance scores of the binary vs. the trinary groups on the four syllogisms. No significant difference in deductive performance was found between the binary and trinary group on the four syllogisms, E(1,117)  =  1.36, (MSE = 11.08), p>. . 246  The result clearly falsified the third hypothesis. Deductive reasoning performance with a conditional rule was not influenced by the attribute dimension size of the set of instances from which the rule was initially induced. This result, therefore, also casts doubt on the influence of the size of a contrast set encountered during rule induction. According to the hypothesis, larger contrast sets with more alternative dimension values should facilitate reasoning performance with indeterminate argument forms, but according to the present results, this effect of the trinary inductive training set was not realized. It might be noted at this time that this particular outcome may have been due to the binary nature of the traditional deductive syllogism task. If so, then the trinary group may have been less able to utilize their conditional knowledge as compared to the binary group. This point is discussed in more detail later.  The fourth, and final hypothesis addressed near-transfer of an acquired rule as a function of acquisition mode. Two verifiable predictions were proposed on the basis of this hypothesis: (a) as compared to rules acquired in a diagnostic context, there would be fewer deductive reasoning errors with rules acquired in a predictive context on the MP and DA argument forms; and (b) as compared to rules acquired in a prediction context, there would be fewer deductive reasoning errors with rules acquired in a diagnostic context on the AC and MT argument forms. The data pertaining to the predictions based on the fourth hypothesis are summarized in Figure 10.  56  Figure 10. Near transfer by acquisition mode  The predictions were verified with a contrast of performance on the four syllogisms with a prediction-related content versus the four syllogisms with a diagnosisrelated content. Overall deductive performance by the two treatment groups was better on prediction-related content as compared to diagnosis-related content, but this difference only approached statistical significance E(1,1 17)  =  3.78, (MSE  =  1.99), p<.055. The locus  of this difference was in performance on the AC argument type, where prediction-related content significantly facilitated reasoning performance, (1,1 17)  =  4.07, p<.O46.  Performance on the remaining argument types was not significantly different.  C. Additional Analysis of Transfer Data  After testing the hypotheses, a further analysis of transfer data was carried out to investigate additional transfer effects that might inform the interpretation of the above results. A response pattern analysis was canied out for each subject to determine whether the subject’s overall response pattern to the eight syllogism types followed a predominantly conditional or biconditional rule interpretation. The percentages of subjects following each interpretation in near and far-transfer conditions are presented in Table 10. 57  Binary (n=48)  Transfer type  Near  Far  Trinary (n=48)  Near  Control (n=24)  Far  Response pattern Conditional  .68  .45  .55  .46  .51  Biconditional  .20  .35  .30  .44  .28  Table 10. Proportion of subjects following either a conditional or a biconditional rule interpretation  The conditional response structure (pattern) was the most common response structure produced by control subjects (51%), as well as subjects in the treatment groups (68% and 55% for the binary and trinary groups, respectively, on the near-transfer problems, and 45% and 46% for the binary and trinary groups, respectively, on the fartransfer problems). The biconditional response structure represented the majority of nonnormative (i.e. non-conditional) responses (28% for the control and the binary groups, and 37% for the trinary group, averaged across all transfer problems).  Analysis of transfer performance as a function of rule selection scores:-To gain additional insight into the relationship between inductive and deductive reasoning, overall deductive performance on the four transfer tasks was analyzed as a function of rule selection scores. The original rule selection score was collapsed into a new variable, labeled RC, which was defined as follows: If the correct propositional representation of the target rule was selected in both the prediction and the diagnostic condition, a score of 3 was assigned; if it was not selected in either condition, a score of 1 was assigned; otherwise, a score of 2 was assigned. A high RC score therefore reflects a consistent identification of the appropriate propositional form of the target rule. Because deductive  58  performance did not differ between the two training treatment groups (binary and trinary), the groups were combined for the purpose of this particular analysis. Transfer scores summed over the four argument types for each rule selection score are presented in Table 11 (see also Figure 11). Rule Selection Score* (N 96)  Transfer Type  1 (N=34)  2 (N=27)  3 (N=35)  Far—Transfer (Trial 1) M SD  5.09 1.78  5.26 1.93  5.63 1.72  Far—Transfer (Trial 2) M SD  5.35 2.03  5.74 1.63  5.69 1.83  Near—Transfer (Prediction> M SD  6.27 1.71  6.07 1.44  6.11 1.67  5.62 1.67  5.59 1.80  6.09 1.79  Near—Transfer (Diagnosis) M SD  If the target rule was selected in both training modes then score=3, if the target rule was selected in one of the two modes then score=2 else score=l. *  Table 11. Transfer scores summed over all argument types by rule selection score  59  Figure 11. Transfer performance by rule score  Analysis of the data in Table 11 revealed that the overall transfer performance did not vary as a function of the rule selection score, F(2,93)  =  0.38, (MSE = 8.45), p>.684.  Apparently, subjects who selected the target rule, did not perform better than those who did not. An inspection of the data reveals a noteworthy trend in the transfer scores, namely that as the rule selection score increases, the difference between the prediction-related and diagnosis-related transfer scores appears to decrease (see Figure 11). It appears that with more reliable target rule selection, performance became more independent of the causeeffect directionality encountered during inductive training. The difference scores were regressed on the rule selection scores and the relationship was tested for each group. The relationship was clearly not significant for the trinary group, (1,94)  .32, (MSE = 2.37), j>.572, however it closely approached significance for the  binary group, E(1,94)  =  3.58, (MSE = 1.85), p<062. The results suggest that for the  binary group at least, near-transfer performance became more stable as the target rule was being selected more reliably following the inductive tasks. Thus, better recognition of rule structure as expressed in multiple choice statements appears to be related to performance.  60  W. DISCUSSION  A. Summative Overview The broad aim of this study was to gain a better understanding of deductive reasoning with conditional statements, or rules, of the form IF x THEN y. The motivation for the study was based on the broad premise that deductive reasoning with conditional rules depends on how the rules were acquired, or induced, by the reasoner in the first place. The purpose of the study, therefore, was to investigate how, if at all, the conditions under which conditional rules are induced, influence rule transfer to deductive reasoning with conditional propositions. In other words, this study was an investigation of transfer between inductive and deductive conditional reasoning.  Four specific hypotheses pertaining to the transfer of conditional knowledge were proposed. The first hypothesis addressed overall differences in near and far-transfer performance. Near-transfer referred to a situation where the content from which a conditional rule was induced matched the content on the subsequent deductive transfer task. Far-transfer referred to a situation where the inductive and deductive content did not match. The second hypothesis addressed far-transfer performance as a function of attribute dimension size (binary vs. trinary). The remaining two hypotheses concerned near-transfer performance: one addressed near-transfer as a function of attribute dimension size, and the final one focused on near-transfer as a function of acquisition mode (predictive vs. diagnostic).  61  The hypotheses and the testable predictions derived from them were developed by drawing on a combination of empirical findings and theoretical claims from two competing psychological theories of human reasoning, namely a class of theories referred to as rulebased (Braine et al., 1991; Inhelder & Piaget, 1958; Nisbett 1993), and an other class of theories labeled instance, or memory-based theories (Griggs et al., 1982; Pollard et al., 1981). Rule-based theories propose that deductive reasoning is carried out by the use of rules that are sensitive to the logical form of deductive arguments. Memory-based theories on the other hand propose that reasoning depends primarily on the cueing of relevant instances by the content of argument propositions or their context (Rips, 1990). The neartransfer hypotheses were developed on the basis of the memory view, while the fartransfer hypothesis was derived from the rule perspective. Therefore, the summary of the findings and their implications will be discussed in the context of the theoretical perspective from which each hypothesis evolved.  B. Implications for Theories of Conditional Reasoning  Implications for Rule-based Theories of Deductive Reasoning:-Any discussion of rule-based theories of human reasoning runs into a difficulty as soon as it becomes necessary to specify what rules one is talking about (Rips, 1994). The current debate over the utility of rules in reasoning includes a wide variety of inferential rules such as logical rules, elementary inference schemas, pragmatic reasoning schemas, and statistical rules including the law of large numbers (Nisbett, 1993). This study focused on logical (conditional) rules, but the results, specifically those pertaining to the second hypothesis, have implications for other types of rules as well.  62  Rule transfer studies have repeatedly found that rules are generalized to novel situations more consistently if they are induced from sets of instances with high compared to low variability (Catrambone et al., 1989; Royer, 1979; Singley & Anderson, 1989). Since the trinary training set exhibits greater variability than the binary set, rule-based theories would predict the far-transfer performance of the trinary group to be better than the binary group. This was not the case in this study, however. The groups did not differ on the far-transfer tasks, and in fact no far-transfer was observed for either training group.  The lack of far-transfer in this study was not due to poor training. All subjects reached the inductive training criterion of 18 consecutively correct predictions and diagnoses, and the vast majority of subjects in both treatment groups selected one of the two correct conditional rule statements. Furthermore, the analysis of the inductive training data confirmed that the trinary group took significantly more time, and required more instances to reach criterion compared to the binary group. Since induction of rules from variable sets of instances would be expected to require more elaborate processing, the present results are consistent with the claim that meaningful training did in fact occur. Inadequate training, therefore, does not appear to be a factor responsible for lack of far-  transfer. What may have occurred instead, is that the assumed content-free rule structure might have developed from the training materials, particularly from the trinary dimensional sets, but such rule structure might not have been properly operative. As was already noted, one reason may have been the inherent binary nature of the criterion propositional reasoning task which might have mitigated the adaptive function of the content-free rule structure.  What implications does the absence of far-transfer have for rule-based theories of conditional reasoning? To help frame the rule debate, Smith et al.(1992) proposed eight  63  criteria for deciding whether reasoning involves the use of abstract rules such as the set of logical rules which includes the conditional. The criteria are as follows: 1. 2. 3. 4. 5. 6. 7. 8.  Reasoning performance is the same on familiar and unfamiliar content. Performance is the same with abstract and concrete material. Early in acquisition, a rule may be overextended. Performance deteriorates as a fhnction of the number of rules that are required for solving a task. Performance is facilitated by prior application of the same rule. A rule is mentioned in a verbal protocol. Performance on tasks with specific content is improved by training on abstract versions of the rule. Performance on problems in a particular domain is improved as much by training on problems outside the domain as on problems within it.  Many of the criteria are not directly applicable to the present study. For example, this study was not concerned with performance differences on abstract and concrete tasks; the rule content on all the tasks was at a similar level of abstraction. However, the first and last criteria are relevant here. If abstract rules are being used, then performance on familiar (near-transfer) and novel (far-transfer) deductive tasks should be similar, and experience with one task content should facilitate performance with others types of content. This clearly did not happen in the present study. Overall scores on near-transfer tasks were significantly higher than on far-transfer tasks, and experience with one type of content did not facilitate performance with unfamiliar content since the treatment groups did not do any better than the controls on arguments with novel rule content.  The present findings indicate that subjects who underwent inductive rule acquisition clearly learned something, but not generalizable inference rules that they could apply in new contexts. Subjects did not seem to rely on inferential rules or schemas that were independent of content.  64  Implications for Memory-Based Theories of Deductive Reasoning: Memory-based transfer has been shown to be facilitated by transfer-appropriate processing (Morris et al., 1977) where encoding and retrieval conditions match (Tulving, 1984). Therefore, if deductive conditional reasoning is memory-based, it should also be facilitated by transferappropriate processing. Thus, according to the fourth hypothesis which was derived from the memory view, rules induced from examples of causes followed by effects (prediction mode) should facilitate performance on deductive argument types that involve primarily cause-to-effect reasoning, namely the MP and the DA types. On the other hand, rules acquired by exposure to examples of effects being traced backward to causes (diagnostic mode) should aid performance on argument forms that also involve inferences from effects to causes, namely the AC and MT types.  The results of this study do not support the above hypothesis. The facilitative effect of transfer-appropriate processing did not materialize. The diagnostic acquisition mode did not enhance near-transfer performance on argument forms involving reversed, diagnostic inferences. In fact, the opposite result occurred on the AC form, where the predictive mode resulted in significantly better performance compared to the diagnostic mode.  The findings cast doubt on instance-based accounts of reasoning. Reasoning on the AC form has been shown to depend on the reasoner’s awareness of alternative antecedents (Markovits, 1984), which in the context of this study means alternative causes of a given effect. According to the memory-for-instances view, the recall of alternative causes of a particular effect on the AC form should have been facilitated by the matched effect-to cause directionality encountered on the diagnostic rule induction task. The fact that this did not occur suggests that recall of specific instances may not be involved in deductive reasoning. This conclusion does not contradict the fact that awareness of alternatives 65  plays a role in conditional reasoning. It may be that reasoners benefit from simply being aware of the fact that some alternatives exist, but just what these are may not be critical for making correct inferences.  Support for the instance-based account of reasoning is also undermined by the results of the test of the third hypothesis. In contrast to the last hypothesis, which addressed access to relevant instances through transfer-appropriate processing, the third hypothesis focused on the availability of such instances. This hypothesis involved a comparison of near-transfer performance of two treatment groups. The binary group induced conditional rules from sets of instances that had binary attribute dimensions, the trinary group from trinary sets. In the context of this study, the rule instances in the binary sets were constructed from two distinct causes and two distinct effects. The trinary sets were constructed from three causes and three effects.  Since the pooi of unique alternatives (of antecedents and consequents) that was available to reasoners in the trinary group was larger than the pooi of alternatives available to the binary group, instance-based theories would predict better deductive performance for the trinary group on the DA and AC types. This is because these two argument types are indeterminate, and correct conditional inferences depend once again on the ability of the reasoner to consider alternative antecedents that can lead to a given consequent (Evans, 1982; Oaksford et al., 1992). Since the near-transfer performance did not differ between the binary and trinary groups, there appears to be no advantage of having a larger pooi of alternatives available for making deductive inferences. This finding, together with the refutation of the first hypothesis, question the validity of the claim that people recall specific instances when they reason with conditional rules.  66  Implications for Theories of Reasoning by Mental Models:-Currently the most intensely debated theory of propositional reasoning in general, and conditional reasoning in particular is the mental model theory (Johnson-Laird, Byrne, & Schaeken, 1992; 1994). Because the mental model view is so prominent in the literature, and because the results of the present study have some implications for this theory, the key elements of the theory are briefly outlined below.  The mental models approach rejects the idea that people use abstract logical rules that are sensitive only to the logical form of sentences. Instead, Johnson-Laird et al. propose that people reason in more concrete ways, by constructing and manipulating mental models of argument premises. Reasoning by mental models involves the following steps: 1. The reasoner begins by studying the premises and then forms a mental model which represents a possible state of the world in which the premises hold. 2. The reasoner then proceeds to search for a proposition which is both true in the model, and also is informative in the sense that it is not a repetition of an existing premise or a trivial inference. This derived proposition represents the reasoner’s putative conclusion. 3. In the last step the reasoner looks for counter-examples to check the validity of the putative conclusion. Counter-examples are models which are consistent with the argument premises but inconsistent with the conclusion. If no counter-examples are found, the reasoner accepts the conclusion as valid.  According to the mental model theory, errors in conditional reasoning occur when reasoners fail to represent some relevant information in their model. The more their implicit model is “fleshed out” with explicit information, the better the reasoning performance (Johnson-Laird et a!., 1992; 1994). Unfortunately, the theory does not clearly specifr what determines the extent to which a model is fleshed out. As Evans (1993)  67  points out, the theory needs to be more specific with respect to (a) the nature of the mental representations, and (b) what causes the initial representations to be expanded.  The results of this study can in fact inform these gaps in the mental model theory. The theory assumes that when people encounter an IF.. THEN.. statement they automatically construct a mental representation that has either an implicit or an explicit conditional structure. Just where people acquire the knowledge to do this is never addressed. Given the generally poor reasoning results reported in the literature, it would appear that most people do not in fact have this presumed conditional knowledge.  The issue of conditional knowledge acquisition, of course, motivated the present study. Since subjects in this study were exposed to explicit examples of conditional relationships during the rule induction task, their experience should improve their ability to “flesh out” their mental models they presumably utilize on the deductive task. The present findings provide mixed support for the fleshing out component of the model theory. The  fact that both treatment groups performed better on the near-transfer task suggests that their exposure to the exemplar sets used in the inductive task facilitated their ability to construct more explicit mental representations of the premises compared to the control group. The results also suggest that if subjects did flesh out their models, their expanded representations were probably not constructed from specific instances. If they were, then the trinary group should have outperformed the binary group because the trinary set contained a greater variety of potential model building “material”. It is more likely that people relied on representations that lie somewhere between abstract rules and concrete models built from specific instances.  68  The fact that near-transfer performance Modus Tollens did not rise above the level of controls for either treatment group also has implications for the mental model theory. Performance on MT was not facilitated by inductive training and remained generally unchanged across groups and conditions. Yet, unlike Modus Ponens, which also remained unchanged, the MT scores do not indicate a performance ceiling which would limit any potential improvement. The MT form is generally considered to be relatively difficult since it contains a negation as well as a reversed inference from consequent to antecedent (Evans, 1982). The presence of both of these factors may in fact tax cognitive resources to such an extent that subjects do not even attempt to construct even implicit mental models. The mental model theory stipulates that models are represented and manipulated in working memory (e.g. Baddeley, 1990), therefore a significant cognitive load imposed by particular argument structures may constrain reasoning with internally represented models.  Summary of Major Implications From the Findings:-The results clearly rule out a reasoning process that operates on specific instances stored in memory. The trinary group always performed at or below the level of the binary group despite the fact that the trinary group had equal access to and greater availability of specific instances that could be used to generate more counterexamples or more explicit models to aid reasoning. Instancebased processing is also ruled out by the fact that transfer-appropriate processing failed to improve performance. Previous research has shown that access to encoded instances is facilitated by matching the encoding and retrieval conditions. When this was done in the present study by matching the direction of inferences between the inductive and deductive tasks, no facilitative effect of transfer-appropriate processing was observed. Since any reasoning program that relies on access to specific instances would be expected to benefit from improved memory access, the present findings indicate that an instance-based process was probably not involved. 69  At the same time, the results also do not provide evidence for a process driven solely by abstract, general rules. If such rules were used, then far-transfer would be expected to take place, particularly in the trinary condition. The inductive training data for the trinary group was consistent with the expectation of more abstract rule learning (e.g. increased processing time, and more reliable selection of the more parsimonious and general rules summarizing the observed conditional relationships). Yet, the trinary group performance on far-transfer was no better than either the trinary or the control group, indicating that the reasoning process was probably not driven by abstract rules. It should once again be noted, however, that an abstract rule structure may have developed during training, but it may not have been operated on during the deductive task.  So far, results of this research have been used as evidence for rejecting current accounts of conditional reasoning. However, considered together, the results suggest alternative accounts of deductive conditional reasoning. One such account is proposed in the sequel.  C. Proposed Account of Conditional Reasoning  Taken together, the present findings point to a reasoning process driven by rules that are specific to a domain, but generalizable over particular instances of the domain. This explains the superiority of the near over the far-transfer performance in conjunction with the lack of observed performance facilitation due to improved access and availability of particular instances in memory. But what form do these domain-specific rules take?  The verbal reports on inductive strategies suggest that these rules are not represented in the syntactic form of IF x THEN y. The majority of subjects produced  70  statements such as “Drug x causes only symptom y, but the other drugs cause all the observed symptoms”. All such statements formed the response category labeled “memorybased”, which was also the largest category for both treatment groups. Such statements can be expressed more generally as: “A member x, from the class of X’s is associated only  with one member, y, of the class of Y’s, but any of the other X’s could be associated with any member of the set of Y’s”. A rule at this intermediate level of abstraction might be optimal in terms of reasoning efficiency. It captures the general conditional structure of observations, yet it provides enough detail for constructing and manipulating mental representations in the form of mental models. The assumed associative memory structure referenced by this linguistic expression is shown below.  Several aspects of the proposed rule representation should be noted. First, the relationship between the C’s (e.g. antecedents) and E’s (e.g. consequents) conforms to the material conditional. This is corroborated by the fact that the conditional response structure (pattern) was the most common response structure in the transfer conditions.  Second, in the proposed representation the associative link connecting Cl and El is bi-directional, and stronger than all the others. As a result, the rule representation is biased toward biconditional reasoning errors, particularly when it is operated on under increased processing load, involving trinary dimensions, for example. The present findings again support this conjecture. A biconditional response structure represented the majority  71  of non-normative (i.e. non-conditional) subject responses across all transfer problems, however it was more prevalent in the trinary group as compared to either the binaiy or the control groups.  The proposed rule structure also explains the counter-intuitive finding that performance on the AC argument form was better in the predictive induction condition, despite the fact that the AC form is distinguished by a diagnostic inference mode. The reason for this is that the proposed rule structure optimally represents the one-to-one and one-to-many mappings which define a conditional structure, but which only hold in a predictive context. To see why, consider the proposed rule structure from a diagnostic perspective (i.e. view it from the set of consequents, the E’ s, to the set of antecedents, the C’s). From this perspective, all the consequents have one-to-many mappings to the antecedents, and consequently subjects may infer a structure based on a symmetric mapping, such as the biconditional. When the structure is viewed in a predictive context, on the other hand, the one-to-one mapping between Cl and El is readily apparent. Since this mapping represents the key defining feature of a conditional interpretation, subjects’ reasoning is more likely to reflect this normative conditional interpretation in the predictive context. This is indeed what the results show. Performance on prediction related content was better for all argument types, although it reached significance only on the AC syllogism. This finding has instructional implications which are discussed in the next section.  Finally, it is proposed that conditional reasoning may involve more than one level of information processing: the underlying reasoning process may depend on the argument type encountered. There is no a priori reason to assume that the same cognitive process must be deployed to deal with all types of conditional reasoning arguments. In fact, the present findings suggest otherwise. The results show that near-transfer performance was 72  significantly better (compared to control levels) only on the DA and AC forms. The MP and MT levels, however, remained essentially unchanged across conditions. Why?  Past research has shown that reasoning with Modus Ponens is essentially at a ceiling level, with typical performance in the 85-100 percent range (e.g. Evans, 1982). Results of this study are no different. In fact Modus Ponens is the flagship of proponents of inference schemas (Braine et al., 1991) since performance on it is unaffected by content and is highly accurate. When considered together, the results suggest that the reasoning process on the MP form is likely to be based on compiled procedural knowledge (Anderson, 1983) and as such, any further performance improvement is not likely to be observed.  In contrast, the MT form may involve little, if any, systematic reasoning. The rationale for this supposition is based on task analysis of Modus Tollens which suggests that it is, in fact, the most difficult syllogistic form since it involves both a negation and a reversed inference (Evans, 1982). Yet performance on the MT form is generally quite good, reaching as high as 80 percent proportion correct (Evans, 1993). Again the results of this study are consistent with previous research, with MT performance being second only to MP. It is proposed that the reason for such a seemingly impressive performance on the toughest syllogism form is simply due to the well documented matching bias discussed in section Dl. (Evans, 1982). The joint presence of the two resource-taxing operations may interfere with the retrieval of the memory-based rule representation. Because of the relative strength of the biconditional link in the assumed representation, the biconditional bias inherent in this rule representation would increase, thus increasing the probability of a biconditional interpretation. Under this interpretation, the negation of the consequent compels subjects to add a matching negation to the antecedent term, which in this particular case yields the correct response. 73  D. Implications for Education  The major implication of this research for educational practice is that conditional reasoning, which is ubiquitous across academic and professional disciplines, can be enhanced by appropriately designed training. Even the brief inductive training utilized in this study resulted in significant improvement in reasoning with training-related content. Longer training periods, which were not realistic in this experimental set-up, may have yielded even better performance outcomes. It should also be noted that deductive performance may have been improved by training subjects on the deductive syllogisms directly, without the benefit of the inductive phase. However, knowledge acquired by simply substituting different content into standard logical or other rule-based forms is likely to be “inert”, and bound to specific application contexts. The bottom-up inductive approach may require more time and effort, but the knowledge so constructed may be more flexible, and adaptable to different reasoning contexts. Clearly, more research is needed to resolve this issue.  The present results may also inform the ongoing debate (e.g. Cheng et al., 1986) regarding the optimal method of teaching deductive reasoning skills. Given that subjects in this study did not make inferences on the basis of content-free, generalizable rules, induction from rule instances alone, may not be sufficient to develop such rules. Inductive training may have to be supplemented by explicitly teaching students the syntactic structure of conditional rules.  The results also indicate that at least for simple conditional arguments with binary propositions such as those used in this study, inductive training sets with binary  74  dimensions appear to be equal or superior to sets with more varied instances. Another advantage is that rule learning from binary sets also requires less time compared to larger, trinary sets. This is a usefhl result from a practical perspective of teachers who may not have the time to construct large rule-learning exemplar sets, and who want to minimize instructional time.  This research also clearly established the superiority of the predictive mode of rule acquisition over the diagnostic mode. Transfer to prediction-related content was consistently equal to or significantly better compared to the diagnostic mode, including transfer to deductive reasoning which emphasized the diagnostic mode. These results may be particularly relevant for medical education where diagnostic reasoning is, of course, critical.  E. Limitations of the Study and Suggestions for Further Research  Since the limitations of the current research often motivate additional research, the two issues are discussed together. This study has several limitations. Subjects were volunteers who were paid for their participation; they were not randomly selected from the population and they lacked any intrinsic motivation on the experimental tasks. All subjects were undergraduate students in a Teacher Training Program, which limits any generalizations to other populations. Additional research with samples from other populations is clearly needed.  Methodological limitations span both the inductive and deductive tasks. The inductive task was by necessity relatively short, averaging about one hour. This length of time may not be adequate for inducing a full blown logical rule in the form of  75  automatized, compiled, procedural knowledge that can be effortlessly applied in different settings (Anderson, 1983). Additional research is needed to determine whether more extensive rule induction training would result in transfer to deductive tasks with unfamiliar content. In addition, different inductive training sets need to be tested since there is no reason to assume that the exemplar sets used in this study were in any way exhaustive or optimal.  In order to make the duration of the inductive task reasonable, subjects were allowed to record their observations, responses, and corresponding feedback from the computer during rule learning. This procedure made the task easier but it may not have encouraged deeper processing because subjects did not have to restructure the information they encountered in order to make it more manageable for their working memory. In turn, their ability to reconstruct their conditional knowledge from memory may have been reduced. Further research comparing this experimental procedure with one that forces subjects to rely completely on their memory would be informative.  The deductive task utilized standard syllogism forms which may not represent practical reasoning situations that generally involve more information in a less structured format. Perhaps more importantly, in the standard deductive syllogism, the premises present only binary propositions. In other words, each individual proposition contains only one term or its negation (p or not p, or q or not q). This may have biased the task difficulty in favour of the binary group since, with binary content, a negation implies a small contrast set which is easier to represent and thus reduces cognitive load. Therefore, future research needs to consider non-binary deductive tasks such as, for example: “Given Ifxl then yl, together with x3 is the case, what follows?”.  76  The task content may also limit generalizations to different domains. The challenge was to develop task content that was neutral in that it did not cue many pre-existing semantic associations based on experience, while at the same time providing a sense of realism and practical relevance. Every effort has been made to satisfy these conflicting criteria, but the inevitable tradeoff may have diminished the external validity of the tasks. Future research should investigate transfer within and between “real-world” content domains.  The method of scoring the deductive performance may also have masked some effects. It will be recalled that validation scores for affirmative and negated conclusions were added together to yield an overall performance score. Validation responses are generally not identical for these two situations (e.g. Lee, 1985; Marcus & Rips, 1979) and, more importantly, the response patterns may interact with different inductive training conditions. Further research is clearly needed, which may also necessitate a different analytical approach suitable for the analysis of categorical data.  As in other experimental studies in cognitive experimental psychology, the possibility existed that subjects may not have been fully motivated to engage all their available cognitive resources while working on the tasks. Consequently, some performance differences may have been attenuated if subjects did not bother to deploy some strategy or knowledge which they may have acquired during the inductive phase. However, since the context under which the tasks were presented (i.e. evaluating student learning) was highly relevant to the subjects’ coursework, the vast majority of subjects appeared highly engaged and genuinely interested in the experimental tasks. Still, future research should investigate inductive and deductive reasoning that is motivated by realistic goals and has real consequences.  77  One important issue raised by the present research concerns the boundaries of near and far-transfer. In this study, the near vs. far distinction was based on the similarity, in terms of content semantics, between the acquisition and the application of rule knowledge. Research in other areas of cognitive psychology on the transfer of cognitive skill, most notably research on analogical reasoning is relevant here. Studies of analogical reasoning indicate that transfer is a function of the similarity between the transfer task, and the reasoner’s internal representation of a related situation encountered previously (Vosniadou & Ortony, 1989). Studies of analogical reasoning have shown that internal representations based on structural features enable far-transfer when these representations are successfully mapped onto new situations that share the same structural features (Gentner & Toupin, 1986). In this context, the structural features of conditional representations are the one-to-one and one-to-many mappings between antecedent and consequent pairs, as illustrated in the model proposed above.  Unfortunately, structural mappings unlike surface mappings based on semantic relatedness, are more difficult to access since they are often less salient. One way to facilitate access to these features is to provide contextual cues. Another way is to explicitly focus attention on relations between objects, rather than the objects themselves. In terms of the present research, it would be valuable to investigate transfer of conditional knowledge using different contextual cues between the inductive and deductive phases. In this way, the applicability of even a limited, domain-specific representation may be stretched by cueing the reasoner to focus on the structural features that are presumable already present in their internal knowledge representation. Using this approach, it would be possible to test deductive transfer to other domains, that may not share any surface features with task domains encountered during rule acquisition.  F. Final Summary and Conclusion 78  Reasoning with conditional rules represents one of the most essential and ubiquitous forms of human reasoning. Despite the proliferation of research on conditional reasoning, many questions regarding the knowledge representation and the processing involved in this form of reasoning have not been fully answered. One reason is that in virtually all studies of conditional reasoning, the existence of conditional knowledge was simply assumed, and how the assumed knowledge was acquired in the first place has not been addressed. In contrast, this study utilized a transfer-of-training paradigm which made it possible to actually manipulate the learning and the application of conditional knowledge, and consequently, to test different hypotheses about its underlying form and function. Taken together, the present findings suggest that conditional reasoning is based on a knowledge representation of intermediate level of abstraction, not unlike that of basic level categories. Instead of completely generalized schemas or specific instances, subjects appear to rely on conditional knowledge that is domain-specific and that contains the minimum number of antecedents and consequents that captures the observed conditional relations. This minimum number appears to include the target antecedent-consequent pair (i.e. ci and ci), as well as an abstract representation of the respective logical complements  { not ci, not ci). Further research is needed, however, with a non-binary version of the deductive task to assess the full extent to which induced conditional knowledge can in fact be generalized.  79  REFERENCES  Anderson, J. A. (1983). The architecture of cognition. Harvard University Press. Anderson, J.A., & Bower, G.H. (1973). Human associative memory. Hilisdale, N.J: Eribaum. Baddeley, A.D. (1990). Human memory: Theory andpractice. Mass: Allyn & Bacon. Baddeley, A.D. (1986). Working memory. Oxford: Clarendon. Beisley, D.A., Kuh, E., & Welch, R.E. (1980). Regression diagnostics: Iden4fying Influential Data and Sources of Coiinearity. New York: Wiley. Bereiter, C., & Scardamalia, M. (1985). Cognitive coping strategies and the problem of “inert’ knowledge. In S.F. Chipman, J.W. Segal, & R. Glaser (Eds.), Thinking and learning skills: Current research and open questions V2. Hilisdale, NJ: Erlbaum. Bjorkman, M., & Nilsson,R. (1983). Prediction and diagnosis:An experimental comparison. Scandinavian Journal ofPsychology, 23, 17-22. Bourne, L.E. Jr. (1974). An inference model for conceptual rule learning. In R.L. Solso (Ed.), Theories in cognitive psychology: The Loyola Symposium (pp.231-256). New York: Lawrence Erlbaum. Bourne, L.E. (1970). Knowing and using concepts. Psychological Review, 77, 546-556. Box, G.E.P., & Cox,D.R. (1964). An analysis of transformations. Journal of the Royal Statistical Society, 26, 211-243. Braine, M.D.S., & O’Brien, D.P. (1991). A theory of If: A lexical entry, reasoning program, and pragmatic principles. Psychological Review, 98, 182-203.  Braine, M.D.S., Reiser, B.J., & Rumain, B. (1984). Some empirical justification for a theory of natural propositional logic. In G. H. Bower (Ed.), The psychology of learning and motivation: Advances in research and thinking (Vol. 18, pp. 317-371). New York: Academic Press. Bruner, J.S., Goodnow, J.J.,& Austin, G.A. (1956). A study of thinking. New York: Wiley. Carlson, R.A., & Schneider, W. (1989). Acquisition context and the use of causal rules. Memory & Cognition, 17, 240-248.  80  Catrambone, R., & Holyoak, K. J. (1989). Overcoming contextual limitations on problem solving transfer. Journal ofExperimental Psychology: Learning, Memory, and Cognition, 15, 1147-1156. Cheng, P.W., & Novick, L.R. (1992). Covariation in natural induction. Psychological Review, 99, 365-382. Cheng, P.W., Holyoak,K.J, Nisbett,R.E., & Oliver,L.M. (1986). Pragmatic versus syntactic approaches to training deductive reasoning. Cognitive Psychology, 18, 293328. Cheng, P.W., Holyoak,K.J. (1985). Pragmatic reasoning schemas. Cognitive Psychology, 17, 391-416. Collett, D. (1991). Modeling binary data. Chapman & Hall. Cummins, D.D., Lubart,T., Alksnis, 0., & Rist, R. (1991). Conditional reasoning and causation. Memory & Cognition,19, 274-282. Evans, J. St.B T. (1993). The mental model theory of conditional reasoning: critical appraisal and revision. Cognition, 48, 1-20. .  Evans, J.St.B.T. (1982). The psychology of deductive reasoning. London: Routledge & Kegan Paul. Evans, J.St.B.T. (1984). Heuristic and analytic processes in reasoning. British Journal of Psychology, 75, 45 1-468. Evans, J.St.B.T., Barston, J.L., & Pollard, P. (1983). On the conflict between logic and belief in syllogistic reasoning. Memory & Cognition, 11, 295-3 06. Evans, J.St.B.T., & Beck, M.A. (1981). Directionality and temporal factors in conditional reasoning. Current Psychological Research, 1, 111-120. Gentner, D., & Toupin, C. (1986). Systematicity and surface similarity in the development of analogy. Cognitive Science, 10, 277-300. Gick, M.L., & Holyoak, K.L. (1987). The cognitive basis of knowledge transfer. In S. Cormier & J. Hagman (Eds.), Transfer of learning. San Diego, CA: Academic Press. Griggs, R.A., & Cox, J.R. (1982). The elusive thematic materials effect in Wason’s selection task. British Journal ofPsychology, 73, 403-420.  81  Haygood, R.C., & Bourne, L.E. (1965). Attribute and rule learning aspects of conceptual behavior. Psychological Review, 72, 175-195. Hinton, G., & Anderson, J.A. (1981). Parallel models ofassociative memory. Hillsdale, NJ: Eribaum. Holland, J.H., Holyoak, K.J., Nisbett, R.E., & Thagard, P.R. (1987). Induction: Processes of inference, learning, and discovery. MIT Press. Holyoak, K.J., & Koh, K. (1987). Surface and structural similarity in analogical transfer. Memory & Cognition, 15, 332-340. Huberty, C.J., & Smith, J.D. (1982). The Study of Effects in MANOVA. Multivariate Behavioural Research, 17, 417-432. Inhelder, B., & Piaget, J. (1958). The growth of logical thinking. New York: Basic Books. Jackson, S.L., & Griggs, R.A. (1990). The elusive pragmatic reasoning schemas effect. The Quarterly Journal ofExperimental Psychology, 42A, 353-373. Johnson-Laird,P.N., Byrne, R.M.J.,& Schaeken, W. (1994). Why models rather than rules give a better account of propositional reasoning: A reply to Bonatti and O”Brien, Braine, and Yang. Psychological Review, 101, 734-739. Johnson-Laird, P.N., Byrne, R.M.J.,& Schaeken, W. (1992). Propositional reasoning by model. Psychological Review, 99, 418-439. Lee, S.S., & Lee, Y-H.K. (1992). STM processing capacity and rule schema in preadolescents’ and adults’ conditional reasoning. Unpublished Manuscript. Lee, S.S. (1985). Children’s acquisition of conditional logic structure:Teachable? Contemporary Educational Psychology, 10, 14-27. Lee, S. S. (1984). Separability of coded attributes in logical rule acquisition: A twoprocess analysis. Canadian Journal ofPsychology, 38, 386-401. Marcus, S .L., & Rips, L.J. (1979). Conditional reasoning. Journal of Verbal Learning and Verbal Behavior, 18, 199-223. Marcus, S .L., & Rips, L.J. (1979). Conditional reasoning. Journal of Verbal Learning and verbal Behavior, 18, 199-223. Markovits, H. (1984). Awareness of the possible as a mediator of formal thinking in conditional reasoning problems. British Journal ofPsychology, 75, 377-391.  82  Markovits, H., & Nantel, G. (1989). The belief-bias effect in the production and evaluation of logical conclusions. Memory & Cognition, 17, 11-17. Markovits, H. (1985). Incorrect conditional reasoning among adults: competence or performance? British Journal ofPsychology, 76, 241-247. Markovits, H. (1986). Familiarity effects in conditional reasoning. Journal ofEducational Psychology, 78, 492-494. Markovits, H. (1988). Conditional reasoning,representation,and empirical evidence on a concrete task. The Quarterly Journal ofExperimental Psychology, 40A, 483-495. Morris, C.D., Bransford, J.D., & Franks, J.J. (1977). Levels of processing versus transfer appropriate processing. Journal of Verbal Learning and Verbal Behavior, 16 519533. Newell, A., & Simon, H.A. (1972). Human problem solving. Englewood Cliffs, NJ: Prentice Hall. Nisbett, R.E. (1993). Rulesfor reasoning. Hilisdale, N.J.: Lawrence Erlbaum. Oaksford, M., & Stenning, K. (1992). Reasoning with conditionals containing negated constituents. Journal ofExperimental Psychology: Learning, Memory, and Cognition, 18, 83 5-854. Piburn, M.D. (1990). Reasoning about logical propositions and success in science. Journal ofResearch in Science Teaching, 27, 887-900. Pollard, P., & Evans, J.St.B.T. (1981). The effects of prior beliefs in reasoning: An associational interpretation. British Journal ofPsychology, 72, 73-81. Pollard, P., & Evans, J.St.B.T. (1983). The effect of experimentally contrived experience on reasoning performance. Psychological Research, 45, 287-301. Pollard, P., & Evans, J.St.B.T. (1987). Content and context effects in reasoning. American Journal ofPsychology, 100, 41-60. Popper, K.R. (1959). The logic ofscient/Ic discovery. New York:Harper & Row. Rips, L.J. (1994). The psychology ofproof Cambridge, Mass: MIT Press. Rips, L.J. (1990). Reasoning. Annual Review ofPsychology, 41, 321-353.  83  Ross, B.H., & Bower, G.H. (1981). Comparison of models of associative recall. Memory & Cognition, 9, 1-16. Royer, J.M. (1979). Theories of the transfer of learning. Educational Psychologist, 14, 53-69. Singley, M.K., & Anderson, J. A. (1989). The transfer of cognitive skilL Harvard University Press. Smith, E.E., Langston, C., & Nisbett, R.E. (1992). The case for rules in reasoning. Cognitive Science, 16, 1-40. Sternberg, R.J. (1986). Toward a unified theory of human reasoning. Intelligence, 10, 281-314. Tulving, E. (1984). Precis of elements of episodic memory. The Behavioral and Brain Sciences, 7, 223-268. Vosniadou, S., & Ortony, A. (1989). Similarity and analogical reasoning. Cambridge University Press. Waidman, M.R., & Holyoak, K.J. (1992). Predictive and diagnostic learning within causal models: Asymmetries in cue competition. Journal ofExperimental Psychology: General, 121, 222-236. Ward, S.L., Byrnes, J.P., & Overton,, WF. (1990). Organization of knowledge and conditional reasoning. Journal ofEducational Psychology, 82, 832-837. Wason, P.C. (1966). Reasoning. In B.M.Foss (Ed.). New horizons in psychology. Harmondsworth: Penguin. Wason, P.C.,& Shapiro, D. (1971). Natural and contrived experience in a reasoning problem. Quarterly Journal ofExperimental Psychology, 23, 63-71. Woodworth, R.S., & Sells, S.B. (1935). An atmosphere effect in syllogistic reasoning. Journal ofExperimental Psychology, 18, 451-460.  84  APPENDICES  85  APPENDIX A: An illustration of rule transfer based on the proposed hypotheses  To provide a more concrete picture of the hypothetical steps involved in the transfer of an acquired rule according to the instance-based account of reasoning, this section briefly, and informally, illustrates how a conditional rule might be learned, and how it might subsequently be used in making deductive inferences.  Phase I. Acquiring a Conditional Rule (Inductive Reasoning):-Let’s assume that a researcher wants to discover what effect, if any, certain environmental agents have on the health of people who have been exposed to these agents. Let’s take a simple example where she wants to investigate only two such agents known as causal agent ci and causal agent c2. The effects she is considering are health effect labeled el, and no health effect, labeled e2. Let’s assume that this researcher administers each agent and observes the effect over several trials. Here is what she might observe if the agents and heath effects have a conditional (asymmetrical) relational structure: ci -->el c2-->el c2 --> e2  After running a number of these trials, her memory for these observations might be represented as simple associations between the causal agents and their health effects: ci >>> c2 >>>  el e2  ci >>>  ci el  c2 >>>  Note that in this notation the symbol  “>>>“  indicates the strength and direction of  the association between the cause and the effect as they are represented in memory. Her next task is to conclude something from the data. Clearly, a number of different conclusions, or propositional statements emerge from the results:  86  a. “Agent ci causes a health effect el” b. “Agent c2 causes a health effect el on some trials but no effect on others” c. “Agent ci never results in the absence of a health effect”, etc.  Given that her goal is to come up with some general, concise statement that summarizes the results (i.e. a general conclusion), she notes that no general conclusion can be stated about agent c2. In fact the most appropriate general and concise statement summarizing  the results is:”Whenever agent ci is administered, effect el occurs” or, “IF ci THEN el”. Because a statement in this general form is in effect induced from the set of observations represented in memory,, it can serve as a pointer to her memory representation for that set of observations.  In addition, the researcher’s search for a general rule (conclusion), and her attempt to formulate her conclusion in a propositional form, means that the ci  >>>  ci pair has  been processed more elaborately (more deeply) then the other pairs. This means that the associative link from ci to e I has actually been strengthened by the application of the pragmatic constraints.  87  Phase II. Transfer of the Acquired Rule to Conditional Reasoning:-For the sake of brevity, the assumed reasoning process is illustrated only in the case of the DA inference form. Step 1. Representation of the First Premise:-According to the instance view, the first premise (the acquired rule) provides lexical access to the stored representation of the instances from which the rule was induced: “IFclTHENel”  >  cl>>>el c2 >>el c2 >>e2  Step 2. Representation of the Second Premise:-The negation of the antecedent in the second premise activates the set of antecedents in the first premise (since both refer to the same category of causes). To accommodate a consistent internal representation of both premises, the inference schema  or Q not P  P  Q is instantiated yielding the following unified representation of both premises: c2>> el c2>> e2 Step 3. Generating a Conclusion:-Following the association links from the antecedents to the consequents, our reasoner arrives at the conclusion that either el or e2 occurs. In other words the reasoner concludes that ci sometimes occurs, meaning that the conclusion “ci” is sometimes true which is the correct conclusion. A complete account of the hypothesized reasoning process with rules acquired under different conditions is given below.  88  00  I(D(D(D(D I WWN)N)  —  IV IV IV I0 IN)  0I(D N)IN)  ()WN)N) VVVV VVVV WI VVVV 10000 I WN)WN)  ICD(DCDCD  I I 01  N)  0  —  I VVVV 01 VVVV W 10000 I WN)WN)  N)  0  0 I (D MIN)! I V IV I0 IN)  0 W  -  0 N)  —  0  0 W  —  0 N)  —  I—i  0  IMF-  —  I VV 01 VV N) I VV 100 IN)I-  l-  I(DCD I  CDI WI  CDI  I I  VVV VVV I VVV 1000 I  CDI  CDI WI  (Dl  (Dl  ICDCDCD  I I VVV I VVV 1000 I WN)H  ICD(DCD  0  —  VV VV 00  ICDCD  I 01 MI  0  C-)  CD  CD(DCD  000 N)N)N) VVV VVV VVV  I I  1 WN)F  000 MN)N) AAA AAA AAA CDCDCD  100 I MN) I AA CDI AA N) I AA I CDCD I1  CD  000 WWW AAA AAA AAA CDCDCD WN)F1  WMF1 WMF1  000 WWW VVV VVV VVV CD(DCD  N)  MN) VV VV VV CDCD  100  I I  CDI  —  IV IV IV lCD IF1  CDI0  IF1  lCD  IV IV IV  (D10  IV IV IV IV IV lCD  CD10  IH  IV IV IV IV IV I CD,  IF 1 l—  CDI0  N)MN)  000 0  WN)I—  000 WWW AAA AAA AAA (DCD(D  I-I  CD  00 0 MN) AA V AA V CDCD V  CDCDCD(D  0000 1 N)N)N)F AAAV AAAV AAAV  CD  VVV V VVV V VVV V CDCD(D (DCDCD V WMF1 WN)F1 V  000 WWW VVV VVV VVV  N)HV CD  VVV VVV VVV CDCD V  N)N)I-  000  I1 ‘<CD c2-  F-I CDCDCD  H  w  FDi  H  F-I  ‘-  p.)  H  F-I  CD  ‘-I  w I-i  I-I CD  0  rt F-  H  0  H 0  rF  FCD  0  CD  rF H 0  0  0  CD p.) II  rt  CD II CD  F-h  01  F-I  CD  H  Ct  0  0 CD  H  0  CD  0  CD Cl) Cl) H  0 0  F-I  0-  p.)  0  P.) rF  Ct  CD  II CD ci I-I CD CD  CD  CD CD  0 i  I-I  CD  CD  F F-I  0  p.)  CD  1 F-  I-I  P.) 1 F  F  rt  F  0  0  p.)  ct  H  0  F-I CD p.) CD  APPENDIX B: Instructions to subjects  Teacher Education Study Instruction Booklet Step 1. Please read the following section: In this study we are interested in learning more about how teachers evaluate their students’ learning. Specifically, we are exploring how teachers assess the way their students apply newly acquired knowledge about causes and effects in a variety of content areas. The study involves two separate parts: Part A. Discovery Learning No. 1 Discovery Learning No. 2 Part B. Evaluating Students: Evaluating Students: Evaluating Students: Evaluating Students:  Situation No.1 Situation No.2 Situation No.3 Situation No.4  The purpose of the discovery learning part is to let new teachers like yourself go through the same learning experience that some students have actually gone through. Generally, people take less than an hour to complete the entire study. You should expect Part A to take a little longer than part B. The evaluation situations in Part B are all very short. So sit back, relax, and have fun! Step 2. Please boot-up the computer, inset the disk into drive A (or B) and do the following: i. Make drive A (or B) the active directory by entering the command a: (or B:) ii. When you see the A> prompt, start the program by entering the command go iii. Follow the directions on the screen and please, type everything in lower case only.  90  Part A. Discovery Learning: Predicting Effects of Radiation on Plants  Step 1. Please read the following section carefully: The topic we will deal with in this part involves predicting the effects of different types of radiation on plants. A variety of plants were exposed to several radiation sources by biologists in different countries. Each group of biologists made a specific prediction concerning the effect of their particular type of radiation on their plants. You will see each prediction on the screen shortly. When you begin the next computer session, you will look at each prediction being displayed on the screen, and decide if the prediction is correct or not. After each response, you will receive feedback which you can use to make a more informed decision next time. The goal is to discover a simple rule which will help you decide whether each prediction is correct or not. The program will stop after you make 18 correct decisions in a row. Step 2. Please type the lower case letter m to begin the next computer session...  You are encouraged to record relevant information about each diagnosis in the space below: Type of Radiation  Predicted Effect  91  Prediction Correct? (yin)  Step 1. Please answer the following question:  How were you able to tell whether a particular prediction was correct or not? Please write your response on the line below:  Step 2. Please turn to the next page....  92  Step 1. Please answer the following question: Based on what you have observed in the set of examples just presented, which of the following statements represents a CONCISE, GENERAL rule that BEST SUMS UP the relationship between the type of radiation and the type of mutation? Please circle your best choice, and put an X beside your second best choice, in case you have one.  1. If a plant is exposed to Beta radiation, then it grows either square, triangular, or circular leaves. 2. If a plant is exposed to Gamma radiation, then it grows square leaves. 3. If a plant is exposed to Gamma radiation then it grows square leaves, and if a plant grows square leaves then it was exposed to Gamma radiation. 4. If a plant grows square leaves then it was exposed to Gamma radiation.  Step 2. Please turn to the next page....  93  Part A. Discovery Learning: Diagnosing Causes of Strange Fish Scales Step 1. Please read the following section careflilly:  The topic we will deal with in this part involves diagnosing the causes of unusual scales of fish that swim in polluted waters. The fish were diagnosed by fisheries officials in many different locations. Each group of fisheries officials made a specific diagnosis concerning the cause of the particular appearance of fish scales they observed at their location. You will see each diagnosis on the screen shortly. When you begin the next computer session, you will look at each diagnosis being displayed on the screen, and decide if the diagnosis is correct or not. After each response, you will receive feedback which you can use to make a more informed decision next time. The goal is to discover a simple rule which will help you decide whether each diagnosis is correct or not The program will stop after you make 18 correct decisions in a row. Step 2. Please type the lower case letter y to begin the next computer session...  You are encouraged to record relevant information about each diagnosis in the space below:  Appearance of Scales  Diagnosed Cause  94  Diagnosis Correct? (yin)  Step 1. Please answer the following question:  How were you able to tell whether a particular diagnosis was correct or not? Please write your response on the line below:  Step 2. Please turn to the next page....  95  Step 1. Please answer the following question: Based on what you have observed in the set of examples just presented, which of the following statements represents a CONCISE, GENERAL rule that BEST SUMS UP the relationship between the type of chemical agent and the type of fish scales? Please circle your best choice, and put an X beside your second best choice, in case you have one. 1. If fish have brittle scales then agent Tx was released in the water. 2. If agent Tx is released then fish develop brittle scales, and if fish have brittle scales then agent Tx was released into the water. 3. If agent Tx is released in the water, then fish develop brittle scales. 4. If agent Ty is released in the water, then fish develop either enlarged, brittle, or loose scales.  Step 2. Please turn to the next page....  96  Evaluating Students: Situation No. 1  Step 1. Please read the following section: The students you are about to evaluate went through a discovery learning exercise that was identical to the one you went through. It was the one involving radiation sources and plant mutations. These students observed the same set of examples you have seen, and from these examples they discovered a simple rule about a type of radiation and a type of plant mutation. The rule they discovered happens to be the correct rule (You will see this rule when you begin to work on the computer again.) The students used this correct rule along with some additional information to arrive at various conclusions. Your task is to evaluate each student’s conclusion. That is, you are to decide whether the conclusion a particular student arrived at is: 1. Always True, or 2. Sometimes True or Sometimes False, or 3. Always False. There are eight students that you will evaluate.  Step 2. Please type the lower case letter c to begin your evaluation...  97  Evaluating Students: Situation No. 2  Step 1. Please read the following section: The students you are about to evaluate went through a discovery learning exercise that was identical to the one you went through. It was the one involving chemical agents and fish scales. These students observed the same set of examples you have seen, and from these examples they discovered a simple rule about a type of chemical agent and a type of fish scales. The rule they discovered happens to be the correct rule (You will see this rule when you begin to work on the computer again.) The students used this correct rule along with some additional information to arrive at various conclusions. Your task is to evaluate each student’s conclusion. That is, you are to decide whether the conclusion a particular student arrived at is: 1. Always True, or 2. Sometimes True or Sometimes False, or 3. Always False. There are eight students that you will evaluate.  Step 2. Please type the lower case letter o to begin your evaluation...  98  Evaluating Students: Situation No. 3  Step 1. Please read the following section: The students you are about to evaluate were learning about new drugs and strange skin rashes. This material is completely different from any of the learning experiences you have gone through here, but we are still interested in your assessment of these students. These students discovered a relationship between a type of drug and a type of skin rash. They stated this relationship in the form of a rule which you will see shortly when you start on the computer again. Even though you are not familiar with this content area, please suppose that the rule they discovered is correct. The students used this supposedly correct rule along with some additional information to arrive at various conclusions. Your task is to evaluate each student’s conclusion. That is, you are to decide whether the conclusion a particular student arrived at is: 1. Always True, or 2. Sometimes True or Sometimes False, or 3. Always False. There are eight students that you will evaluate.  Step 2. Please type the lower case letter d to begin your evaluation...  99  Evaluating Students: Situation No. 4  Step 1. Please read the following section: The students you are about to evaluate were learning about various types of lasers used in eye surgery. This material is completely different from any of the learning experiences you have gone through here, but we are still interested in your assessment of these students. These students discovered a relationship between a type of surgical laser and some strange spots on the retina of the eye. They stated this relationship in the form of a rule which you will see shortly when you start on the computer again. Even though you are not familiar with this content area, please suppose that the rule they discovered is correct. The students used this supposedly correct rule along with some additional information to arrive at various conclusions.Your task is to evaluate each student’s conclusion. That is, you are to decide whether the conclusion a particular student arrived at is: 1. Always True, or 2. Sometimes True or Sometimes False, or 3. Always False.  There are eight students that you will evaluate.  Step 2. Please type the lower case letter e to begin your evaluation...  100  ____ ______________ ___________________ _____ ____ __________________ _____ ________________ _______________ ________________ _____ _____ ____ _____  Finally, just a few (easy) questions: 2. Gender  1. Age  M  F  3. What is your educational background? B.A. in____________ B.Sc.in  M.A. in Ph.D.in Other  4. What program of studies are you currently registered in? Major  Minor Year  5. Throughout your university career, approximately how many courses did you take in each of the following, general areas? Math  Biology Chemistry____ Physics Psychology Evaluation in teaching Philosophy Art__ Special Ed That’s it! Thanks for your time and effort. It is highly appreciated.  Please return the disk and the booklet to collect your honorarium.  101  


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items