Reflections on Conceptualizing and Measuring Tie Strength
We welcome this opportunity to revisit “Measuring Tie Strength” (Marsden and Campbell 1984) in light of more recent social network studies. When we wrote our article, the field of social networks was in the process of emerging as a recognized area of study, though it had many precursors (Freeman 2004). Since then, interest in network phenomena has surged, engaging analysts from many disciplines. We began by observing that the concept of tie strength had been substantively fruitful, and interest in it remains high – indeed growing: Web of Science citations to Granovetter's classic (1973) statement about tie strength accelerated rapidly in the 2000s, exceeding 500 per year from 2009 to 2011. The much lower level of attention accorded to our article on its measurement follows a similar time path. Studies invoking the concept cover a broad substantive range: they examine topics including innovation (Ruef 2002), job searches (Harvey 2008), knowledge transfer (Levin and Cross 2004), minority social integration (Vervoort 2011), political participation (Lim 2008) and volunteering (Paik and Navarre-Jackson 2011), among many others. Studies of tie strength also span locations as disparate as China (e.g., Bian 1997) and Sweden (e.g., Korpi 2001). The most notable role of weak ties in social networks is their structural significance as connectivity-generating factors: they tend to be bridges that connect distant clusters within social structures. Weak ties are less subject to the closure-producing transitivity pressures that operate on stronger ones, and hence less likely to be confined within local social environments. Weak ties thereby facilitate the interpersonal dissemination of novel phenomena, be those useful information or harmful diseases. Strong ties too have their roles, of course, serving, for example, as dependable sources of social or emotional support (Granovetter 1982). If the presence of bridges is the primary concern, one might argue that measuring tie strength itself has become less necessary. The increased availability of data on “whole” social networks and myriad advances in methods for analyzing them now allow the direct identification of bridging ties within a network. Relationships having high “edge betweenness” – those that are found within large numbers of indirect paths joining pairs of actors in a network – are prone to be bridges. A well-known algorithm for locating clusters or “communities” of actors within networks (Girvan and Newman 2002) relies on this principle, successively removing the highest-betweenness ties to reveal subsets of densely interconnected actors that are relatively isolated from one another. Relationships with high edge betweenness are likely to be weak ties according to Granovetter's argument, but if the key structural feature of bridging can be tapped directly, perhaps assessing the strength of ties is less urgent. Many analysts are centrally interested in strong ties rather than bridges, of course. Moreover, many substantive studies drawing on network concepts continue to rely on egocentric network data from surveys like those we studied in our article, rather than measuring whole networks. Egocentric network data describe the local social environments surrounding individual actors in a network – usually comprising one or more of each focal actor's direct contacts (“alters”) and certain qualities of the dyadic relationships between that actor (“ego”) and the alters (e.g., Marsden 2011). Such data collection designs do not support extensive structural analysis parallel to that possible for whole networks, though some measures of local structure exist (e.g., Burt 1992). Analytic interest in egocentric data sometimes focuses on properties of individual ties, and sometimes on features of an actor's portfolio of relationships – notably network range (Campbell, Marsden and Hurlbert 1986). In such situations, proxies for the likelihood that a tie is bridging are needed, and the question we engaged in our article – how to identify stronger and weaker ties on the basis of dyadic properties such as closeness or frequency of contact – remains important. It would, however, be of interest to assess the quality of such proxy measures by correlating them with structural measures of bridging from whole-network studies. Reflecting on our article nearly 30 years after completing the research, we begin with some thoughts about the conceptualization of tie strength as well as its measurement. We then note some limitations of our research design and conjecture about their influence on our findings. We close by highlighting a few of the findings, with attention to their implications for contemporary network studies. All measurement, as distinct from recording of data, begins with conceptualization (Bohrnstedt 2010). Our article departed from the premise that tie strength is a unidimensional latent construct ranging from weak to strong. We started with Granovetter's (1973:1361) “intuitive” definition of the concept in terms of indicators, as “a (probably linear) combination of the amount of time, the emotional intensity, the intimacy (mutual confiding), and the reciprocal services which characterize the tie.” He asserted that these indicators were “somewhat independent,” though also “obviously highly intracorrelated.” In the interim, some have made thoughtful proposals about the features of a tie that make it “strong.” Friedkin (1990), for example, draws on social psychological theory about how interpersonal attachments develop. He argues that relationships grow more elaborate in sequential stages, beginning with simple awareness and proceeding subsequently to interaction, provision of resources and assistance and eventually affective attachment. Friedkin also stresses the two-sidedness of relationships, observing that the parties in a dyad may regard their tie differently at any given moment, and emphasizing reciprocity as a key property that strengthens a tie. This implies that features such as frequency and closeness accumulate rather than substitute for one another: interaction implies awareness, resource provision implies both awareness and interaction, and so on. Cross et al. (2009) propose a similar cumulative conception for studying interagency collaboration, distinguishing five levels that increase from modest information-sharing arrangements through integrated decision-making processes that seek to realize a shared vision. We are intrigued by Friedkin's account of how ties might initially become strong, but we are less sure that it captures how strength is maintained. Ties involving emotional closeness and mutual commitment may remain strong, although dormant, enduring even after frequent contact or resource exchange cease (Levin, Walter and Murnighan 2011). Such relationships may be readily revitalized under appropriate circumstances. The accumulation of dyadic properties in Friedkin's conceptualization, then, may be more evident in longitudinal studies than in cross-sectional ones. In a marketing context, Shi et al. (2009) offer another approach to conceptualizing tie strength. While they suggest that specific survey items and measurement scales for relationship strength must suit the substantive setting for any given study, they define strength in terms of durability: “the extent to which the partners are bound … the ability of the relationship to resist both internal and external challenges” (Shi et al. 2009:665). They distinguish three constituent aspects of strength: affective (emotional attachment), cognitive (belief that the relationship is and will remain rewarding) and “conative” (commitment to sustaining the relationship even in the face of incentives to discontinue it). Their emphasis on the resilience of stronger ties echoes Haythornthwaite's (2002) discussion of the interplay between new media and tie strength. She contends that strong ties are robust in that parties linked by stronger ties will adapt multiple media to the requirements of their relationship. A convenient or organizationally prescribed media platform, on the other hand, may encourage the passive emergence of weaker ties. Once established, such ties may strengthen, but they also are more vulnerable to dissolution if the media substrate on which they rest disappears or is disrupted. If robustness or resilience is a defining feature of a strong tie, then we wonder how well indicators like frequency of contact or emotional closeness capture it. Certainly the hypothesis that closer, higher frequency and reciprocated contacts are at lower risk of dissolution is reasonable. Some findings reported in the growing literature on longitudinal studies of social networks support this hypothesis; see, for example, Schaefer et al. (2010) or Raeder et al. (2011). Future longitudinal studies may point to additional characteristics of ties that promote persistence. Finally, we raise the question of whether a measure of tie strength should reflect actual or potential resource flows. This broaches the subject of network activation that has drawn increased emphasis in theorizing about both social capital (e.g., Lin 2001) and several job search studies that invoke social capital. Put simply, that a tie has the capacity to convey resources or information of value does not assure that it will transmit them. Smith (2005) and Marin (2012) recount considerations that actors weigh when they decide whether to convey job information to their contacts. Among these are concern for maintaining one's reputation and a reluctance to appear intrusive by offering information without the assurance that it is sought – especially when dealing with less well-known contacts. We think it reasonable that a measure of tie strength should tap potential rather than realized resource flows, but the activation literature suggests that actual flows may differ systematically across ties that vary in strength. Ideally, selection of indicators and scales would follow rather than precede conceptualization. Ours was a secondary analysis, though; we made the best we could of available data. Relying on survey data about friendships, we selected indicators corresponding to Granovetter's conception of tie strength and analyzed them using multiple-indicator models for measurement developed in the 1970s and 1980s. We sought to isolate an unobserved tie strength variable defined by direct indicators of the concept (closeness, duration, frequency, breadth of discussion topics and mutuality) and validate this using predictor characteristics presumably associated with tie strength, including both relationship foci (e.g., kinship or shared neighborhoods; Feld 1982) and heterophily measures (e.g., Reagans 2011). In principle, it would be desirable to validate using outcomes as well as predictors of strength, but dyad-level outcome indicators (such as exchanges of resources within dyads) were not available to us. Our three samples represented broad ranges of adults in the communities where they were drawn, except that one comprised only native-born white male respondents. The unit of analysis in our study was the dyad, though, and the available data referred to a set of dyads selected from the stronger end of a tie strength continuum – the three closest friendships of each respondent. We noted this in 1984 (p. 485), but in retrospect we would make more of this limitation. We expect that the associations we found among tie properties were lower than they would be within some broader population of dyads – for instance, all relationships between two adults (the vast majority of which would be null) or all relationships between adults who know one another. Likewise, it probably reduced the associations between indicators and the latent strength variable, thereby understating indictor quality. We note that Friedkin's (1990) study, based on all dyads in several whole-network data sets, was more successful in isolating a unidimensional measure of strength than we were. Even if generalizations are limited to relatively strong ties, however, some of our findings bear reiteration and comment. Some were replicated in other studies; see Mitchell (1987) and Mathews et al. (1998). The several indicators and predictors of tie strength proved to be more weakly correlated than Granovetter or we had anticipated, suggesting that the concept might not be unidimensional. Indeed, in one sample that included indicators of discussion content and confiding, we identified separate but correlated “time” and “depth” aspects of strength. The prospect that tie strength might be multidimensional is broadly consistent with Wellman and Wortley's later (1990) study reporting that many close ties provide specialized forms of social support, rather than being multipurpose channels. We found that frequency of contact was weakly related to the other indicators of strength and actually negatively associated with duration of acquaintance in all three samples. This finding merits attention because many large databases of transaction records that are now used in network studies (e.g., Raeder et al. 2011) best capture the relative frequency of contact between pairs of actors. If frequency taps an element of strength that is associated with different outcomes than are other properties like closeness or relational content, such databases will depict only one aspect of the linkages among actors. Finally, we found that predictors of strength such as shared foci (family, neighborhood, workplace) or dissimilarity in occupational standing or education were rather modestly associated with strength. This raises caution as to the practice of measuring strength via relationship type, deeming kinship relations to be strong, friendships to be intermediate, and neighbor, coworker or other shared-affiliation relations to be weak. Although some relationship types may tend to be stronger ties than others, a good deal of variation in tie strength exists within types, and we continue to think it best that measures attempt to capture strength directly. Not long after our article appeared, Rogers (1987) took stock of the state of network research, highlighting both accomplishments and challenges, and remarked that the study of networks was then characterized by “weak data, strong analysis” (298). Our efforts were oriented to improving that situation by assessing the quality of survey measurement for one of the field's most widely used concepts. We are gratified that others have found our efforts worthwhile. Given the recent outpouring of interest in networks, the importance of developing high-quality measures of tie strength has surely risen since 1984.