Relating Theory and Data

Essays on Human Memory in Honor of Bennet B. Murdock

William E. Hockley, Stephan Lewandowsky

• Publisher: Psychology Press
• Year: 2014
• eBook ISBN: 9781317760122
• Topic: Psychology
• Subtopic: Experimental Psychology
• Index: Psychology

This festschrift represents the proceedings of a conference held in honor of Bennet B. Murdock, one of the foremost researchers and theoreticians on human memory and cognition. A highly renowned investigator respected for both his empirical and theoretical contributions to the field, Murdock summarized and focused a large amount of research activity with his 1974 book Human Memory: Theory and Data. This unique collection of articles addresses many of the issues discussed in his classic text. Divided into five principal sections, its coverage includes: theoretical perspectives on human memory ranging from a biological view to an exposition of the value of formal models; recent progress in the study of processes in immediate memory and recognition memory; and new developments in componential and distributed approaches to the modeling of human memory. Each section concludes with an integrative commentary provided by some of Murdock's eminent colleagues from the University of Toronto. Thus, this book offers a diversity of perspectives on contemporary topics in the discipline, and will be of interest to students and scholars in all branches of cognitive science.

I	INTRODUCTION

1	Relating Theory and Data: Towards an Integration

Stephan Lewandowsky

University of Oklahoma

William E. Hockley

Wilfrid Laurier University

INTRODUCTION

At the time of its publication in 1974, Human Memory: Theory and Data was hailed as a “classic reference that captures the last 18 years' research” and “sure to be a major contribution,” to cite two enthusiastic endorsements from the book's jacket cover. In hindsight, these expectations were proven correct; Murdock's (1974) book has indeed become a classic and provided a major contribution towards the understanding of human memory.

The book was novel in interesting ways: It was organized not by the temporal partitioning of memory popular at the time and characteristic of its predecessors, but rather according to the different types of information represented in memory. For each type of information, in turn, presentation was separated between a theory chapter on one hand and a data chapter on the other.

Item, Associative, and Serial Order Information

Murdock's (1974) distinction between three fundamental types of information in memory is relevant to every-day retrieval tasks. We can recognize Liz Taylor in a photograph (item information); using that same photograph we can retrieve the fact that she was married to Richard Burton (associative information); and we may also recall that others preceded and followed Burton (serial order information). The example also illustrates Murdock's view that item, associative, and serial order information represent increasing degrees of complexity of the underlying processes. Item information is most basic, retrieval of associative information is more difficult, and recalling the order of events is more difficult still. We easily recognize Liz Taylor on a photograph, we may be able to recall that she was twice married to Richard Burton, once to Eddie Fisher, and once to a senator whose name escapes us, but we cannot remember the order of these relationships.

Originally, the distinction between different types of information was based on the pre-theoretical assumption that different processes could underlie encoding, storage, and retrieval of item, associative, and serial order information. Murdock (1982, 1983) developed this theoretical distinction more formally in his Theory of Distributed Associative Memory (TODAM). Recent research has strongly supported the validity of Murdock's distinction (e.g., Bain & Humphreys, 1988; Gronlund & Ratcliff, 1989; Humphreys, 1976; Lewandowsky & Murdock, 1989a; Murdock & Hockley, 1989a).

The distinction between item, associative and serial order information is also amply represented in the present volume. For example, Dosher, and Humphreys and Bain consider the contributions of item and associative information in recognition decisions. The relationship between item and associative or contextual information is also reflected in Mandler's review of dual process theory, and in Nilsson and Gardiner's examination of the relationship between recall and recognition memory. The representation and retrieval of item and serial order information in short-term memory recall tasks is examined in the chapters by Baddeley, Papagno, and Norris, and Crowder and Neath, and Estes. Mewhort and Popham use TODAM to provide an account of serial recall of tachistoscopic letter strings, and Healy, Cunningham, Gesi, Till, and Bourne compare recall of item, temporal and spatial information in both children and adults.

Relationships Between Theory and Data

Murdock's Theory and Data clearly spelled out biases and pretheoretical considerations at the beginning. Perhaps the most consequential of these was the clear endorsement of explicit models, either stated mathematically or implemented in a computer simulation. This was in preference to the verbal models that predominated at the time, which Murdock considered to be “vague, imprecise, and difficult to refute” (p. 6). In his view the predictions of a theory or model “seldom become clear until the model is developed explicitly” (p. 6). Ratcliff and McKoon (this volume) present an important illustration of this view in their comparison of the predictions for ROC data of the current global memory models.

For someone with this theoretical conviction, the exposition of major developments in the field ideally would have integrated theory and data, for describing one without the other is incomplete and inadequate. However, at the time this proved impossible because contemporary theories could not embrace the large existing data base. Consequently, Murdock discussed theories separate from the relevant data, although not without qualms: ‘To segregate theory and data is not a very satisfactory procedure. They should be thoroughly and carefully interrelated. However we are not quite ready for that On the one hand, none of the theories we now have explain in any depth and with great precision even an appreciable fraction of the relevant data. On the other hand, to consider data alone would be barren and uninteresting” (p. ix).

Fortunately, this is no longer the case, in no small part due to Ben Murdock's development of a comprehensive theoretical framework for human memory (TODAM: Murdock, 1982; 1983). The time has now come that we can begin to present a more coherent picture that interrelates theory and data. Whereas in 1974 theories were incomplete and relatively few in number, today numerous theories compete to offer a most comprehensive account of a wide range of data.

In this chapter, we explore developments since 1974. Naturally, given the continuing explosion of research in cognitive psychology, any such attempt must be highly selective. This selectivity necessarily reflects our biases, which, like Murdock, we wish to state at the outset First and foremost, and not unexpectedly for two graduates of the Murdock laboratory, we endorse the pivotal role of quantitative and explicit modeling in human memory. The reader is referred to Hintzman's chapter for an extensive and convincing argument for the general utility of models in memory and cognition. For this chapter, we emphasize developments within a distributed perspective on human memory, although we readily acknowledge that many other approaches are equally valuable, as some of the following chapters will undoubtedly make clear. And we make liberal references to our own work only because much of it is connected to the person we honor in this festschrift.

In the following, we illustrate the present-day relationship between theory and data by discussing three themes which, we believe, are representative of the enormous progress enjoyed by cognitive psychology during recent years. These themes follow Murdock's taxonomy of item, associative, and serial order information. In the case of item information, new data have been of principal importance in constraining and guiding the construction and development of new theory. By way of contrast, for serial order information the challenge has been to develop new theory to account for the existing data base. Associative information, finally, presents a case in which data constrained theory as much as theory generated new data.

ITEM RECOGNITION: ABANDONING THE SEARCH

Since Sternberg's (1966) pioneering study, the assumption that the retrieval of item information consists of a serial search process, or at least includes a serial search component (e.g., Atkinson & Juola, 1973, 1974; Atkinson, Herrmann, & Wescourt, 1974), has had a powerful influence on models of item recognition. Murdock (1974) proposed the conveyor-belt model to account for the retrieval of item information in the venerable study-test paradigm. This metaphorically-named serial search model was based on the fact that the latency of recognition decisions is a linear function of study-test lag (the number of events that intervene between study and test) and output position.

While the assumption of a serial search process for item recognition was consistent with Sternberg's analysis of retrieval in short-term memory, the slope of the linear functions used to index the search rate are remarkably different in Sternberg's procedure and in the study-test paradigm. Murdock and Anderson (1975) tried to bridge this gap by invoking the notion of functional set-size. They showed that the conveyor-belt model could account for both subspan and supraspan set-size effects and confidence judgments. Murdock, Hockley, and Muter (1977) also tried to account for the effects of item repetition within this framework.

Critical evaluation of serial search prompted a variety of empirical tests (e.g., Baddeley & Ecob, 1973; Burrows & Okada, 1971; Corballis, 1967; Hockley, 1984; Hockley & Corballis, 1982; Okada & Burrows, 1978; Shiffrin & Schneider, 1977), alternative interpretations (e.g., Atkinson, Holmgren & Juola, 1969; Burrows & Okada, 1975; Checkosky, 1971; Corballis & Miller, 1973; Nickerson, 1972; Pike, Dalgleish, & Wright, 1977; Ratcliff, 1978; Theios, Smith, Haviland, Traupman, & Moy, 1973), considerable debate (e.g., Corballis, 1986; Ryan, 1983; Shiffrin & Schneider, 1984; Sternberg, 1975), and new findings and issues (e.g., Corballis, 1986; Corballis, Murray, & Connolly, 1989; Elkind & Corballis, 1986).

Murdock abandoned the conveyor-belt model for a number of reasons. For one, the hypothesized scanning process could not account for the pattern of results obtained in the continuous recognition paradigm (Hockley, 1982; Murdock, 1980). In addition, the changes in the shape of the underlying response latency distributions as a function of set-size or test lag cannot be easily reconciled with a simple serial comparison process (Hockley, 1984; Hockley & Corballis, 1982; Ratcliff & Murdock, 1976). Finally, the development of a distributed memory system with its intuitively appealing properties obviated the need for any type of search process.

The increasing sophistication of the chronometric approach to the study of item recognition has led to a variety of new insights into the underlying processes and has provided the impetus for a number of theoretical developments. Whereas serial search models were based on the observed changes of mean response latency, a more detailed examination of the behavior of the response latency distributions (Ratcliff & Murdock, 1976) provided the empirical basis for Ratcliff's (1978) parallel resonance retrieval theory of item recognition. While search models could account for changes in mean latency in a given task, Ratcliff's model offered a far more comprehensive description of both response time and accuracy across a wide range of recognition paradigms.

The empirical analysis of latency distributions popularized by Ratcliff and Murdock provided a powerful tool to evaluate theoretical descriptions of performance and to compare the nature of processes between different cognitive tasks. Once processes have been specified, and assumptions made about their characteristics, the predicted shape of the response latency distribution is constrained, and comparison with the data can provide a powerful test of the model.

Consider, for example, Hockley's (1984) analysis of response distributions for three separate tasks: Item recognition, visual search, and judgments of item recency. For item recognition in the Sternberg paradigm, the linear increase of mean response time as a function of set-size is often interpreted to reflect a serial search component, which mathematically would be expected to increase the leading edge component of the latency distribution. However, in Hockley's data, the increased mean response latency was largely due to an increase in the positive skewness of the latency distributions with only a modest increase in the leading edge of the distributions. In other words, while the fastest responses changed very little, the slower responses became progressively slower as set-size increased. This is difficult to reconcile with any simple version of a serial search model. In contrast, for visual search Hockley (1984) found the linear increase in response time as a function of the size of the search set to be largely an increase in the leading edge of the distributions. This finding is more consistent with a controlled search process (cf. Shiffrin & Schneider, 1977) in which each additional comparison adds to the total search time. In the recency judgment task, finally, subjects were required to determine the order of items from a short study list and mean response time was a linear function of the serial position of the most recent item. The changes in the underlying latency distributions were entirely consistent with a backward serial search process (cf. Hacker, 1980; Muter, 1979).

These results, when considered together, demonstrate that latency distributions vary in systematic and interesting ways in different tasks. Clearly, any theoretical description of performance in these tasks must be able to provide a reasonable account of these results.

TODAM and Item Recognition

In TODAM, the search notion has been abandoned completely and is replaced by a composite memory vector (Murdock, 1982). In a recognition test, the vector representing the probe is compared to the memory vector. The resulting dot product represents the degree to which information in the probe matches information in memory and serves as the...

Table of contents

• Cover
• Half Title
• Title Page
• Copyright Page
• Table of Contents
• Preface
• Part I: Introduction
• Part II: Perspectives on Human Memory
• Part III: Processes in Immediate Memory
• Part IV: Processes in Recognition
• Part V: Componential Approaches to Memory
• Part VI: Distributed Approaches to Memory
• Bibliography of Bennet B. Murdock
• Author Index
• Subject Index