A Quantitative Review of Associative Patterns in the Recall of Persons

The order in which people freely recall a set of words, persons' names, or other items indicates how they organize those items in memory. An individual's cognitive structure of the persons with whom he or she has some particular relation can be described by noticing how he or she associates from one person to the next during recall. Using improved statistical measurement, we review in detail five studies that systematically examined associative patterns in the recall of persons and evaluate competing hypotheses about the nature of these patterns. Individuals in these studies recalled their acquaintances, coworkers, and friends. Across studies, the results consistently show that persons recalled adjacently or successively are perceived to interact more with each other than those not recalled adjacently. No other factor describes associative patterns as well as this notion of perceived social proximity. These results, along with related research, imply the influence of social networks on memory for persons and suggest a universal feature of human social cognition. The order in which people freely recall a set of words, persons' names, or other items can indicate how they organize those items in memory (Puff, 1979). Associative patterns are one important aspect of the way people recall. Associative patterns refer to the connections or relationships between adjacently recalled items. The way an individual associates from one item to the next in free recall can reveal his or her cognitive structure of those items. Associative patterns may be identified by measuring clustering in recall. Clustering of items by a particular factor occurs when successively recalled items are more likely to share some characteristic or have some relationship than items not recalled successively. Social psychologists have proposed and tested a number of different hypotheses about associative patterns in the recall of persons. Bond and Brockett (1987) postulated that memory for acquaintances is organized on two levels. At the broader level, individuals cluster acquaintances by the social contexts in which they encounter them (e.g., school, work, family, church, etc.). That is, people should tend to list acquaintances from the same social context adjacently in recall. Within social contexts, they asserted, acquaintances are organized in memory according to personality types. In their study, Bond and Brockett (1987) observed that undergraduates clustered acquaintances moderately to strongly by social context. Within social context clusters of adjacently recalled acquaintances, however, they found that subjects clustered by personality traits only weakly. Fiske (1995) posited that acquaintances belong to one of four relationship mode categories (communal sharing, authority ranking, equality matching, and market pricing) and that individuals cluster acquaintances in recall according to these modes. He found that adults clustered acquaintances in recall moderately by relationship mode. Both Bond and Brockett (1987) and Fiske's (1995) hypotheses are categorical. Acquaintances are assumed to belong to only one social context, personality type, or relationship mode category, and associative patterns are hypothesized to correspond to these categorical structures. Brewer (1995b) presented a somewhat different hypothesis about the organization of persons in memory. His hypothesis is that individuals recall and think about the people whom they know primarily in social network terms. Brewer (1995b) reviewed several studies that reported results consistent with this hypothesis. Specifically, these studies suggested that perceived social proximity (or social interaction) is the principal associative factor when people recall persons whom they know. In other words, persons recalled adjacently or successively tend to interact more with each other than those not recalled adjacently. These studies showed that subjects almost uniformly display highly statistically significant clustering by perceived social proximity. However, the most relevant studies reported no information on the magnitude of clustering by social proximity at the level of individual subjects because no methods were available for computing the standard measure of clustering with respect to noncategorical structures (e.g., perceived interaction patterns). This shortcoming prevented quantitative comparisons of the degree of clustering by different variables or with the results of other studies. The calculation of clustering by such non-categorical variables as social proximity involves a formidable computational problem. In this paper, we describe this problem and our solution to it. Then we report on reanalyses of data from two previously published studies and present results for three other studies. The last of these studies tests the social proximity, personality, and relationship mode hypotheses directly. Furthermore, we compare quantitatively the results of these studies with other studies on associative patterns in the recall of acquaintances. Finally, we evaluate the different hypotheses about associative patterns in the recall of persons based on this research and note the implications of these results on techniques for eliciting personal and social networks. Measurement of Clustering by Non-categorical Variables Roenker, Thompson, and Brown (1971) introduced the Adjusted Ratio of Clustering (ARC), now the standard measure of clustering in recall by a single variable for individual subjects. The ARC equals (o e)/(m e), where o is the observed clustering score, e is the expected (by chance) clustering score, and m is the maximum possible clustering score for a given subject. The ARC takes a value of 1 when clustering is maximal, a value of 0 when clustering is at the level expected by chance, and negative values when clustering is less than expected. The distribution of possible clustering scores is usually not symmetric (rather, typically skewed to the right), and therefore the ARC has no universally defined lower bound. The ARC is easily computed when the variable is binary and categorical. In this case, o is the number of adjacently recalled pairs of items that belong to the same category. The expected clustering score, e, is computed as (pw/pt) * (n 1), where pw is the number of pairs of recalled items that are in the same category, pt is the number of all recalled pairs of items, and n is the number of items recalled. The maximum clustering score, m, is the number of items recalled minus the number of categories represented in the recalled items. Somewhat different procedures are required to compute the ARC when measuring clustering by a variable that is not categorical. First, for a non-categorical variable, a square hypothesized associative structure matrix must be constructed that contains the hypothesized associative strengths (such as social proximities) between each pair of recalled items. These strengths may be binary or valued. (The hypothesized associative strengths for categorical variables, or category memberships, may also be represented in this matrix form, e.g., with cells for same category pairs containing "1" values and cells for all other pairs containing "0" values). In the case of a noncategorical variable, o is the sum of the hypothesized associative strengths for the adjacently recalled pairs of items. The expected clustering score, e, is the mean offdiagonal cell value (hypothesized associative strength) multiplied by the number of items recalled minus one. The maximum clustering score, m, is much more difficult to compute for a noncategorical variable than a categorical variable. The task of computing m actually is a case of the classic Traveling Salesman Problem (TSP) (Lawler, Lenstra, Rinnooy Kan, & Schmoys, 1985). In the TSP, the goal is to find the shortest route among a set of cities (like that for a traveling salesman) such that each city is visited once and only once. The TSP typically is discussed in graph theoretic terms, with cities referred to as nodes and the distances between pairs of cities as edges. In the clustering context, the hypothesized associative structure matrix may also be conceived of as a graph, with items as nodes and hypothesized associative strengths as edges. The goal of finding the shortest Hamiltonian path (a path in which each node is visited once and only once) in the TSP corresponds to the goal of finding the maximum clustering score, once proximity data have been appropriately converted into distance data. When the number of items recalled is few, all possible permutations of nodes (or items in a subject's recall sequence) can be enumerated and the corresponding path lengths (or clustering scores) calculated. The maximum clustering score found in this enumeration is m. When the number of nodes/items recalled is greater than 10, the enumeration approach becomes infeasible. Mathematicians and computer scientists have struggled for decades to develop algorithms to solve the TSP to optimality (Lawler et al., 1985). We used Padberg and Rinaldi's (1991) algorithm to obtain provably optimal solutions to the TSP and thus m. This algorithm produces optimal solutions even in cases with large numbers (thousands) of nodes and with nonEuclidean data on the distances between nodes. In this paper, we compare ARCs for different variables (including categorical and non-categorical variables) and different studies, focusing on ARCs for social proximity. We define social proximity broadly to include measures of interactionand/or sentiment-based social ties, such as social interaction, friendship, and knowing. In the studies we describe, the hypothesized associative strengths for the social proximity variables refer to the social ties between the recalled persons. The ARCs for two variables cannot be meaningfully compared if either is categorical. An ARC for a categorical variable can be high even when clusters defined by that variable account for few adjacently recalled pairs of persons. In such a circumstance and to the extent that subjects' recalls are patterned, some other variable(s) must underlie the associations among the other pairs of adjacently recalled persons. Moreover, even though a subject may display clustering by a categorical variable, he or she may also cluster by a second variable within clusters of the first variable. To address this latter issue, when data were sufficient and available, we measured clustering by social proximity within clusters defined by a categorical variable. Such analyses control for clustering by the categorical variable. If the categorical variable is the fundamental associative factor in a subject's recall, associations within clusters defined by that category should be essentially random with respect to other variables. If subjects cluster by social proximity within clusters of a categorical variable, then it indicates that social proximity provides a fuller, more detailed description of subjects' associative patterns and suggests that clustering by the categorical variable may be a coincidental byproduct of clustering by social proximity. For these control clustering analyses, we modified Bond and Brockett's (1987) control ARC measure. For each cluster defined by a categorical variable in a subject's recall, we computed the o, e, and m clustering scores. Then we summed the observed scores together, the expected scores together, and the maximum scores together across clusters to arrive at grand o, e, and m scores for computing a control ARC. We calculated a control ARC only when the subject had at least one cluster defined by the categorical variable with four or more persons or at least two clusters defined by the categorical variable with 3 or more persons each. Appendix A shows examples of how the ARC and control ARC are calculated.