Advances in Consumer Research Volume 10, 1983 Pages 585-589
EXPERTISE AND THE STRUCTURE OF FREE RECALL
J. Wesley Hutchinson, University of Florida
The hypothesis that product expertise results in changes in memory structure was tested by comparing the free recall of brand names of nonprescription cold remedies for pharmacy students and marketing students. A computer algorithm (NETSCAL) was employed to construct network representations of associative memory for the two groups. As hypothesized, the pharmacy students exhibited more ingredient related structure.
One hypothesis about memory that is of particular importance to consumer researchers is that learning is not merely incremental, but can result in the "restructuring" of memory. That is, as a consumer acquires more information about products, she does not simply file it away, she uses it to organize her knowledge about those products. The effectiveness of advertising that is intended to introduce or reposition brands is critically dependent on this hypothesis, as are virtually all consumer education programs. Decision processes will be affected by both the resulting changes in the relative salience of product attributes and the nature of free recall when the consumer generates a choice set for a particular need (see Bettman, 1979 and Lynch & Srull, 1982). This paper reports the results of a study designed to test the restructuring hypothesis by observing structural aspects of unaided brand name recall for groups that differed in terms of product expertise.
Traditionally, free recall tasks have been separated according to whether the items to be recalled were constrained or unconstrained. Typically, a constrained item set is a list of words that is presented to an individual at an experimentally controlled time and at controlled frequencies. The instructions to the individual in the constrained task are to recall as many of the words from the list as possible. In an unconstrained task the item set is typically a semantic category (e.g., foods). No experimental presentation of the items is made. Rather, the individual is simply to draw on his knowledge of the world and recall as many instances of the category as possible. Many aspects of the two paradigms are not directly comparable (e.g., the serial position of items in the presentation order). However, the temporal aspects of recall (i.e., what items are recalled when) are comparable and have proved to be a valuable source of information about memory processes. Two phenomena have been particularly important.
First, when the cumulative number of items recalled is plotted against the time elapsed since recall began a characteristic shape is invariably found. These curves are fit extremely well by concave, asymptotic functions, and much research has been devoted to determining the exact form of this function and the factors that influence its parameters (e.g., Bousefield & Sedgewick, 1944, 1954; Gruenwald & Lockhead, 1980; Indow & Togano, 1970). The importance of this finding is that it suggest that retrieval is not a simple dump of memory akin to emptying books from a bookshelf one at a time. This type of process implies a linear function. If, instead of a dump there is a search of memory in which items are tagged as they are recalled, then occasionally items that have already been recalled will be retrieved. This will happen increasingly often as more words are recalled. Thus, the time between new words will increase leading to a concave function. Simply put, the first words recalled have an "inhibitory" effect on the recall of remaining words.
The second phenomenon concerns the systematic departures from the theoretical curve that are observed in individual, but not group, data. Items tend to be recalled in semantically related bursts (e.g., Bousefield & Cohen, 1955; Gruenwald & Lockhead, 1980; Reitman & Rueter, 1980). This type of clustering in free recall is thought to be related to the structural aspects of either the memory store, or the retrieval strategy. [Recent research suggests that the principal effects concern retrieval strategy. In the present context this is not a pivotal distinction, and it can be side-stepped by acknowledging the these results pertain only to the "operational" organization of memory, regardless of its exact mechanism.] It is assumed that related items are more strongly associated in memory and therefore they are recalled together. Moreover, because they are more strongly associated, or somehow "chunked" together in memory, more sets composed of clusters of items are more easily remembered than sets of unrelated items. This relationship between temporal proximity in recall and associative strength has been used successfully to examine the clustering that occurs "naturally" in memory. Multidimensional scaling and cluster analysis of temporal proximity data have provided tentative validation of this approach to mapping out the structure of memory (see Friendly, 1977, 1979). This is the approach that will be taken in this paper to demonstrate differences in memory structure that are related to expertise about the product class.
PLOT OF UNCONSTRAINED FREE RECALL FOR A SINGLE SUBJECT (UNLABELED: COMPLETE PROTOCOL)
Figure 1 illustrates the first phenomenon with data from an individual from the study reported below in which subjects recalled brand names of products used in treating the common cold. The second phenomenon is observable rom a closer look at this same data (Figure 2). Notice how brands that have similar benefits are recalled together and that all salient bursts (i.e., items recalled during the same time interval) are composed of such items
PLOT OF UNCONSTRAINED FREE RECALL FOR A SINGLE SUBJECT (LABELED; FIRST FIVE MINUTES)
FREE RECALL OF COLD REMEDY BRAND NAMES
Nonprescription cold remedies is a product category for which consumer knowledge about benefits is fairly high, while knowledge about the mechanisms of those benefits (i.e., the ingredients) is fairly low. Thus, unconstrained free recall of brand names from this category would not be expected to exhibit much ingredient-based structure (to be defined operationally below), other than what is implicit in the benefit structure. The restructuring hypothesis predicts that groups who have been exposed to information about the ingredients in such drugs should exhibit more ingredient-based structure. If the groups are chosen to be similar on other dimensions (e.g., age, education, etc.) then nonstructural aspects or recall should not differ between the groups. The study reported here provides an initial test of this hypothesis.
Two groups of University of Florida undergraduates were sampled: pharmacy majors who had recently taken a course on nonprescription drugs (N=35) and non-pharmacy majors who were enrolled in an introductory marketing course (N=36). The pharmacy students were hypothesized to have two sources of product expertise. First, most or these students had some work experience in pharmacies and therefore more direct product experience than typical undergraduates. Moreover, their career choice presumably sensitized them to information about nonprescription drugs. Second, the recent course had exposed them to particular ways of thinking about nonprescription drugs. In particular, the course heavily emphasized matching specific ingredients to specific symptoms in making recommendations about nonprescription drugs.
Each groups responded to a four part questionnaire about colds and cold remedies which required approximately one hour to complete. Part 1 of the questionnaire asked respondents to recall as many symptoms of the common cold as possible (five minute time limit). Part 2 of the questionnaire asked them to recall as many treatments of those symptoms as possible (five minute time limit). Part 3 asked respondents to recall brand names of products used in treating the symptoms of the common cold (ten minute time limit). For this part, they were also instructed to record the time of each response. This part of the questionnaire is the principal focus of the present analysis. Part 4 was a brand name recognition test for an extensive list of products taken from the Handbook of Nonprescription Drugs (1979).
Responses to Part 3 of the questionnaire were submitted to three types of analysis. First, a simple content analysis was performed. Second, a recall function was fitted individually to each respondents data. And finally, NETSCAL solutions for pooled inter-response times for a subset of brand names were computed for each of the two groups. [NETSCAL (Hutchinson, 1981) is a computer algorithm similar to multidimensional scaling, except the underlying model is a directed network instead of a multidimensional space. A solution consists or a list of which items are connected to which other items. The connections need not be reciprocal.]
Based on the content analysis, a reduced set of brand names was selected that (1) spanned the ingredient combinations available in nonprescription cold remedies and (2) occurred with sufficient frequency to permit pooling across respondents.
COLD REMEDY BRANDS SELECTED FOR DETAILED ANALYSIS
Correlational analysis of these brands revealed two trends that are pertinent to the issue of product expertise. First, there was a substantial amount of commonality in response frequencies for the two groups (r=.62). Second, only response frequencies for the marketing students were systematically related to 1981 advertising expenditures (r=.75 for marketing students, r=.27 for pharmacy students). Finally, the evaluation given to the brands in course lectures, as rated by the instructor, were significantly correlated with the differences between the two groups in brand recall (r=.41). Positively evaluated brands were recalled more often by the pharmacy students and negatively evaluated brands were recalled less often.
The particular recall function that was fit to each respondent's data is expressed in Equation 1.
n(t) = aqt + k (1)
a + qt
This equation expresses the cumulative number of words recalled, n(t), as a three parameter function of the time, t, elapsed since recall began.
Each of the three parameters can be given a natural interpretation. The asymptote of the function, a, is an estimate of vocabulary size. The rate parameter, q may be thought of as the rate at which memory items are sampled and the additive constant, k, may be thought of as buffer size (i.e., the number of items initially in short term memory due to the delay between reading instructions and beginning the task). The restructuring hypothesis predicts a difference between the two groups only for the asymptote. Presumably, pharmacy students know more names. Vocabulary size should also be reflected in the amounts of recall and recognition if the groups to not differ on the process parameters (i.e., search rate and buffer size). This prediction is supported by the data although the effect is not large (see Table 2).
VOCABULARY SIZE AND PROCESS VARIABLE ESTIMATES
The main hypothesis that was to be tested by this research was the restructuring hypothesis. Specifically, it was hypothesized that the structural aspects of free recall for the pharmacy students would be more strongly related to ingredient composition than that of the marketing students. In order to make a meaningful test one must postulate a specific model of recall that explicitly describes the role of memory organization in retrieval processes. The model outlined below is only one of several current models that could have served this purpose (see Crowder, 1976, Chapter 10) . The goal in this paper, however, is not to differentially test models of free recall, but to operationalize the concept of memory organization with sufficient rigor to permit an unbiased test of the restructuring hypothesis.
For the purposes of the subsequent NETSCAL analysis four postulates about retrieval are assumed.
(1) Memory for brand names can be conceptualized as a network of associations that connect brand names to each other.
(2) The protocol from an unconstrained free recall task represents one path through that associative network.
(3) The path is constructed as follows:
a. after each response an association connecting that response with another brand name is selected with a probability that is proportional to associative strength:
b. If the selected brand has not been previously recalled, it will be given as a response, and the process will continue as in (3a):
c. If the selected brand name has already been recalled then no response is given, but the next association in the path is drawn from those originating from the selected brand, not the last response, and so on until an unrecalled brand is fount.
(4) The relative strength of the associations are the same for all individuals for whom data is to be pooled.
Thus, the retrieval process is analogous to a man wandering the streets of a city, taking turns according to the ease of the path, but not noticing where he is until he gets there. The experimenter, who is in a windowless room, gives the man a two-way radio and tells the man to report in every time he finds himself in a new place. The experimenter repeats this procedure for a number of individuals. The task of the experimenter is to construct a map of the city based on the travel times between locations. First, the experimenter must assume that people walk at about the same speed and are wandering the streets of the same city. Given those assumptions and a large number of protocols, the minimum reported travel time between two locations will be a good estimate of the distance between the two locations. Since people presumably traveled only on roads, a network model (i.e., NETSCAL analysis of the travel times), not a spatial model (i.e., multidimensional scaling of the travel times), is best suited to these data. The spatial model would not be a bat approximation; however, it would miss certain significant attributes of the city that the network representation would capture. For instance, one-way streets, large intersections, and the distinction between ordinary and limited access highways would not be represented in the spatial representation. All of these features, however, have simple graph theoretic implications.
Obviously, for an associative model of memory, the explicit connections represented in a network are of central importance. The restructuring hypothesis predicts that there should be more connections between brands that share ingredients for network representations based on the pharmacy students' recall protocols than for networks based on data from the marketing students. The NETSCAL solutions for all 34 brands are too large and complicated for easy visual display; therefore, two types of reductions of those solutions are presented here. First, all connections between brands can be divided into four groups: (1) connections that are present in both solutions, (2) connections that are absent in both solutions, (3) connections that are present only in the pharmacy students solution, and (4) connections that are present only in the marketing students solution. Table 3 displays the proportions of each type of connection that connected brands that: (l) had functionally equivalent ingredients, (2) shared at least one functionally equivalent ingredient, and (3) shared no functionally equivalent ingredients.
PROPORTIONS (FREQUENCIES) OF VARIOUS TYPES OF CONNECTIONS IN NETSCAL SOLUTIONS
A second reduction of the solutions is displayed in Figures 3 through 7. Each of these networks is a condensation of the corresponding NETSCAL solution in which all brands composed of functionally equivalent ingredients have been collapsed into a single node (referred to as an ingredient class). Arrowheads indicate the direction of the connection. If no arrowhead is present, then there is a connection in each direction. Solid lines represent connections between ingredient classes that share at least one ingredient. Dashed lines represent connections between classes that share no ingredients. The ingredients are abbreviated as: (A) antihistamine, (D) decongestant, (P) pain reliever, (C) cough suppressant, (E) expectorant, (Ca) caffeine and (Ab) antibacterial. It should be pointed out that al; connections between the same pair of ingredient classes are represented as a single connection. Therefore, common connections in these figures can result from distinct connections in the uncondensed networks that connect the same ingredient classes of cold remedies. This is because the common and unique connections depicted in Figures 5 through 7 are based on the condensed networks depicted in Figures 3 and 4. Thus, although these two reductions are obviously related, they offer different perspectives with respect to level of abstraction. The first reduction is more precisely related to the proposed model and is more appropriate for quantitative analyses. The second reduction is essentially a convenient way of displaying the NETSCAL solutions in order to facilitate the interpretation of the quantitative results.
CONDENSED NETSCAL SOLUTION (MARKETING STUDENTS)
CONDENSED NETSCAL SOLUTION (PHARMACY STUDENTS)
UNIQUE CONNECTIONS (MARKETING STUDENTS)
UNIQUE CONNECTIONS (PHARMACY STUDENTS)
Two patterns are clear in these analyses. First, the connections that are present in both networks are predominantly ingredient related connections. These products can be grouped by their market positions into two major benefit classes: sinus relief (A, D, P, AD, AP, DP, ADP, PCa) and more throat relief (AC, DC, DCE, ADCE, Ab). As can be seen in Figure 5, most of the common, ingredient related connections are between brands in the same major benefit class. It is also the case that most sinus relief products are tablets or capsules, and more throat relief products are typically liquids or lozenges. Thus, the advertised benefits and the perceptual images of these products could underly many of these connections just as easily as the ingredient composition of the connected brands.
The second pattern that is readily apparent in these analyses is that, relative to the marketing students, the pharmacy students have proportionately more connections between functionally equivalent brands and fewer connections between brands that share no ingredients. One-tailed z-tests of the differences between these proportions are statistically significant at the p <.05 level (z=1.86 and z=1.78, respectively). Although the two networks have about the same proportions of connections between brands that share at least one ingredient, many of these connections in the marketing students network connect the same pair of ingredient classes and those are typically in the same major benefit class. Thus, in the condensed networks, the pharmacy students network has more ingredient related connections and, in particular, there are more connections between the two major benefit classes (see the lower right section of Figure 7).
The results reported above are quite encouraging. The difference between the pharmacy students and the marketing students in terms of product expertise was clearly not maximal. One would assume, for instance, that the difference between practicing pharmacists and housewives would be much greater. Nevertheless, the methods of analysis employed in this study were sufficiently sensitive to reveal systematic differences between the groups that supported the restructuring hypothesis. Specifically, when network representations of associative memory for cold remedy brand names were constructed (via NETSCAL analysis of inter-response times in free recall), pharmacy students showed more ingredient related connections than did marketing students. Moreover, the types of connections that separated the pharmacy students from the marketing students were those that were least likely to have resulted from the confounding factors of benefit class or perceptual image of the brand.
Several important areas for future research are suggested by these results. What types of communications (e.g., advertising, formal instruction, public service programming, etc.) are most likely to lead to restructuring of memory organization? What is the relation of brand name salience to structure in free recall? What, specifically, is the relationship between retrieval processes and structural factors? Can specific models be empirically tested? How do these memory variables relate to purchase and usage? The list of open questions is long indeed; however, there is good reason to believe that pursuing them will be fruitful.
American Pharmaceutical Association (1979), Handbook of Nonprescription Drugs, New York: American Pharmaceutical Association.
Bettman, J.R. (1979) "Memory Factors in Consumer Choice: A Review," Journal of Marketing, 43, 37-53.
Bousfield, W.A. & Cohen, B.H. (1955) "The Occurrence of Clustering in the Recall of Randomly Arranged Words of Different Frequencies of Usage," Journal of General Psychology, 30, 149-165.
Bousfield, W.A., Sedgewick, H.W. & Cohen, B.H. (1954) "Certain Temporal Characteristics of the Recall of Verbal Associates," American Journal of Psychology, 67, 111-118.
Crowder, R.G. (1976) Principles of Learning and Memory, Hillsdale, N.J.: Erlbaum.
Friendly, M.L. (1977) "In Search of the M-gram: The Structure of Organization in Free Recall," Cognitive Psychology, 9, 188-249.
Friendly, M.L. (1979), "Methods For Finding Graphic Representations of Associative Memory Structures," in C.R. Puff (Ed.) Memory Organization and Structure, New York: Academic Press
Gruenwald, P.J. & Lockhead, G.R. (1980), "The Free Recall of Category Examples," Journal of Experimental Psychology: Human Learning and Memory, 6, 225-240.
Hutchinson, J.W. (1981), "Network Representations of Psychological Representations," Doctoral Dissertations Stanford University.
Indow, T. & Togano, K. (1970), "On Retrieving Sequence From Long-Term Memory," Psychological Review, 77, 317-331.
This research was supported by a Research Development Award from the University of Florida Division of Sponsored Research (RDA-2 82-83) and by a summer research award from the University of Florida College of Business Administration and the Center for Econometrics and Decision Sciences.