Archive

Building on previous social norms literature, the revised framework of normative influences (A. Chung & Rimal, 2016) comprehensively describes when and how social norms can influence behavior. The framework suggests that three different types of social norms predict behavior. First, descriptive norms refer to perceptions of what most other people are doing, encouraging individuals to align with perceived widespread behavior by offering a guiding principle or cue that indicates what actions are effective and suitable for a given situation (Cialdini et al., 1990). Second, injunctive norms pertain to perceptions of what others approve of, motivating individuals to adhere to the norm to avoid social penalties for non-compliance (Cialdini et al., 1990). Third, subjective norms refer to perceived expectations from significant others to either engage in or refrain from a particular behavior, driving behavior through a combination of motivations to comply with those expectations (Fishbein & Ajzen, 1975).

The revised framework (A. Chung & Rimal, 2016) also suggests that the association between norms and behavior is moderated by variables classified into three categories: behavioral, individual, and contextual factors. Behavioral variables refer to characteristics of the actions in question, such as behavioral privacy (i.e., whether behavior is enacted in a public or private setting; M. Chung & Lapinski, 2019). Individual factors pertain to personal attributes like self-efficacy (i.e., perceived ability to perform a given behavior; Bandura, 1977). Contextual attributes relate to social and environmental backgrounds, such as time constraints (i.e., whether there is sufficient time for a behavioral decision; Cone & Rand, 2014).

This study proposes that a person's social network density can moderate the normative influence on behavior, categorizing it as a contextual factor. We argue that the effects of norms are likely stronger in a network where people are highly interconnected (i.e., a dense network) than in one with more sporadic connections (i.e., a sparse network). Such a relationship has been theorized and often assumed in social network literature (Coleman, 1988, 1990; Delgado-Marquez et al., 2013; Valente, 2010), but direct empirical evidence supporting this claim has been surprisingly scarce. The following sections review the theoretical foundation for the moderating role of network density on the norm-behavior relationship.

Network density refers to the proportion of the number of actual links among people in a given network to the number of possible links (Monge & Contractor, 2003). Links generally indicate the relations between people, and communication network studies usually define them as connections between individuals, demonstrating how messages are sent, shared, or understood (Shumate & Contractor, 2013). In online social networks contexts, these links have often been operationalized as connections between friends (e.g., Centola, 2010; Meng et al., 2016).

There are at least three underlying mechanisms to explain how density can enhance normative influence or facilitate normative reinforcement. First, given that communication plays an essential role in the formation and diffusion of norms (Carcioppolo & Jensen, 2012; Kincaid, 2004; Yanovitzky & Rimal, 2006), dense networks, by definition, can provide more pathways than sparse networks along which communication about norms can flow and result in greater exposure to normative information (Cappella, 2017). Sparse networks may not provide sufficient channels for normative information to be circulated within a given network.

Second, a highly dense network means that network members are greatly interconnected. Thus, their deviant or norm-violating behaviors in such a network are more likely to be detected and punished by other members than in a sparse network. When pre-established norms exist in such a dense network, and people in the network are concerned about possible social sanctions from failing to conform to group norms, the norms arguably exert a strong influence (Coleman, 1988; Mohnen et al., 2012).

Third, an extremely dense network tends to be a closed network in which everybody has a direct relationship with everybody else. This structural characteristic reduces the likelihood of new information being introduced to the network members from outside and results in information circulating more within the network than between networks (Burt, 2000, 2005). In addition, closed networks enhance cohesion and group solidarity among members (Coleman, 1988; Shen et al., 2014), which in turn can strengthen normative influence (Rimal & Real, 2005). As a result, a network featuring thoroughly connected individuals can self-reinforce pre-existing conditions, including their normative beliefs.

Despite decades of theorization, there is little direct empirical evidence available on the moderating role of network density in the normbehavior relationship. The scarcity of evidence may pose a problem because studies grounded in a theory that lacks empirical support can yield several issues, such as the existence of possible alternative explanations for the study findings (King, 1973). An exception to this is the Wikipedia research by Piskorski and Gorbatâi (2017). Guided by Coleman's (1988) argument, their study focused on undoing others' posts without proper justification or negotiation as a norm violation, and hypothesized that such norm violations would be less likely to occur in densely interconnected contributor networks than in sparsely interconnected ones. They analyzed Wikipedia contributors' behavioral data and showed that the norm violation was observed less frequently in dense networks than in sparse networks. Moreover, the punishment for the violation (operationalized as reverting the inappropriate undo) occurred more frequently in dense networks.

While the Wikipedia study (Piskorski & Gorbatâi, 2017) provides valuable evidence, there are some important shortcomings. For example, their study centered on proscriptive norms that could be perceived as both descriptive (e.g., “most contributors avoid undo”) and injunctive (e.g., “most contributors disapprove of undo”) norms (Bergquist & Nilsson, 2019). Thus, it is unclear what type of norms actually contributed to the findings. This uncertainty was also, in part, due to their methodological choice to collect behavioral data unobtrusively, without measuring the contributors' perceptions of such behavioral rules. Our study is designed to overcome these shortcomings.

The literature on social norms suggests that the influence of norms on behavior can be both direct and contextual (A. Chung & Rimal, 2016). Additionally, the social network literature posits that the normative influence on behavior is likely to be more substantial in a dense network than in a sparse one (e.g., Coleman, 1988). Despite this, there has been limited reporting of empirical data concerning the interaction between norms and network density. This study aims to bridge this knowledge gap in the context of unfriending. We first plan to replicate the existing findings regarding the direct positive influence of social norms on unfriending intention (Verswijvel et al., 2019). The dependent variable was selected due to its robust association with actual behaviors as demonstrated in meta-analyses (i.e., a correlation coefficient of .82 after accounting for sampling and measurement errors, as per M.-S. Kim & Hunter, 1993), and to circumvent potential validity issues when observing the behavior experimentally. Following that, we will investigate whether the normative influence is indeed stronger in a dense network compared to a sparse one. In this study, the emphasis is specifically on descriptive norms, mainly because perceived descriptive norms are generally more susceptible to change in experiments with limited exposure to normative

messages than injunctive norms are (Lapinski et al., 2013). Similarly, subjective norms were also deemed difficult to experimentally manipulate, as those perceptions involved participants' significant others' expectations. Considering these, the following hypotheses were proposed:

H1: Perceived descriptive norms will positively influence unfriending intention.
H2: The relationship between perceived descriptive norms and unfriending intention will be moderated by network density; specifically, the influence of perceived descriptive norms on unfriending intention will be stronger in a dense network than in a sparse one.

This research employed a 2 (descriptive norms condition: high vs. low) × 2 (density condition: high vs. low) between-subjects online experiment. High-descriptive norms refer to the greater prevalence of unfriending a target person on social media, whereas low-descriptive norms mean the lesser prevalence of the same behavior. High-density networks (i.e., dense networks) refer to the greater interconnectedness among friends on social media, while low-density networks (i.e., sparse networks) indicate lower interconnectedness among them. A hypothetical scenario was used to induce experimental manipulations. The dependent variable of the study was participants’ intention to unfriend the target person online.

In total, 220 college students signed up for the study from two different research participant pools at Michigan State University (MSU), and 200 of them participated in the study. Study participants received either $2 Amazon Gift Cards or class credits as compensation depending on the pools they used to sign up for the study. After removing 17 unusable cases (i.e., participants under the age of 18 and instances of duplicate participation), the data collected from 183 participants were used in the subsequent analyses. The sample was predominantly White (76%) and female (81%), with ages ranging from 18 to 29 (M = 20.42, SD = 1.69). The demographic characteristics of the study sample are presented in Table 1. There was no statistically significant difference in the dependent variable between the participants from the two pools or across the demographics.

The experiment was listed and advertised on the university’s research participation pools as a “Social Media Attitudes Study.” Participants accessed an online experiment, and only participants who reviewed a consent form (which was presented on the first page of the experiment) and agreed to proceed were able to participate in the subsequent experiment and answer the survey questions.

Participants were randomly assigned into one of the four experimental conditions. They were first asked to provide their race and gender in the beginning of the experiment. This demographic information was used to match the hypothetical target person’s race (i.e., White, Black/African American, Asian, Hispanic/Latino, or other) and gender (i.e., male, female, or other). For participants who identified their race or gender as ‘other,’ we presented a randomly selected avatar that corresponded to either their race or gender, depending on which was indicated. If participants selected 'other' for both categories, we randomly assigned one of the eight avatars. This matching mechanism aimed to enhance the experimental design's rigor by leveraging the tendency of individuals to prefer interacting with others online who share similar demographic characteristics (Al-Natour et al., 2005, 2011; Benbasat et al., 2020), and minimizing potential biases related to demographic features.

The hypothetical scenario was provided with visual stimuli using avatars (see Appendices A and B). Following the experimental treatments, participants were asked to complete the survey questions. The questionnaire first measured the outcome variable of the study (i.e., intention to unfriend) then the variables for induction checks (i.e., perceived descriptive norms and perceived network density). Next, the survey measured a potential covariate (i.e., attitude toward the target person’s behavior), followed by other demographic variables (i.e., age, gender, and class standing). Upon completion of the survey, participants were thanked, debriefed regarding the purpose of the study. The experimental procedure and survey questionnaire were reviewed and exempted from the Institutional Review Board at MSU (#STUDY00000384).

The hypothetical unfriending scenario was developed for this study based on previous literature (Faucher et al., 2014; Gilbert & Karahalios, 2009; Holfeld, 2014; Schacter et al., 2016; Sibona, 2014). Participants were first asked to imagine a hypothetical social media friend (e.g., Sarah for White female participants) who was not particularly close to them. Then they learned that they and the friend originally had 32 mutual friends, and the relationships between the participants and these mutual friends were closer than with the friend. The extent to which these mutual friends were also friends with one another on social media was described but varied based on participants’ network density conditions. Following that, a context behind the unfriending situation was given: The target friend recently has left some mean and rude comments on others’ posts on social media quite often. Finally, based on participants’ descriptive norms condition, either the greater or lesser prevalence of unfriending him/her among those mutual friends was presented.

In the high-density condition, a highly interconnected network was described. For example: “29 of your 32 mutual friends with Sarah are also friends with each other on social media. In other words, about 91% of your mutual friends with Sarah are also friends with one another online.” On the other hand, in the low-density condition, a sparsely interconnected network was illustrated: “6 of your 32 mutual friends with Sarah are also friends with each other on social media. In other words, about 18% of your mutual friends with Sarah are also friends with one another online.” After this message, a brief question was presented to reinforce these experimental inductions and encourage participants to focus on the stimuli: “Approximately what percentage of your mutual friends with Sarah are also friends with each other on social media? 1. More than 90%; 2. Less than 20%.”

Consistent with the conceptual definition of descriptive norms (Park & Smith, 2007), highdescriptive norms messages showed a greater prevalence of unfriending in the participants’ friends on social media: “You recently found out that a majority of your mutual friends with Sarah on social media have unfriended her. The number of your mutual friends with Sarah dropped from 32 to only 2 friends. In other words, about 94% of your mutual friends with Sarah have now unfriended Sarah.” On the contrary, low-descriptive norms messages illustrated a lower prevalence of the behavior in the same network: “You recently found that a few of your mutual friends with Sarah on social media have unfriended her. The number of your mutual friends with Sarah dropped from 32 to 30. In other words, about 6% of your mutual friends with Sarah have now unfriended Sarah.” Again, after each message, a brief question was presented: “Approximately what percentage of your mutual friends with Sarah have now unfriended Sarah? 1. More than 90%; 2. Less than 10%.”

Most of the items used in this study were drawn from prior research, with revisions made to the wording to fit the current study context. These items were selected for their demonstrated reliability and validity in previous studies, thus ensuring the robustness of the measures employed. The study variables, unless otherwise specified, were measured on a 5-point Likert-type scale (1 = Strongly disagree, 5 = Strongly agree) in which higher scores indicated greater agreement with the statement or higher levels of the variable.

Confirmatory factor analyses (CFA) using the lessR package in R (Gerbing, 2014) were performed for the scales that contained at least four items. The internal consistency and parallelism theorems were employed to generate predicted correlations between all items of the same latent variable (Hunter & Gerbing, 1982). These predicted correlations were then subtracted from their respective obtained correlations to produce the errors/residuals (Boster, 2023). The adequacy of the measurement models was assessed by the following two criteria: ample factor loadings and small residuals. In line with previous literature, factor loadings greater than .51 were considered large, and RMSE (i.e., root mean squared error) less than .05 were considered small (Boster et al., 2011; Gerbing, 2014; Grayson-Sneed et al., 2016). When the data did not meet the criteria, items that had the weakest factor loadings or produced the largest errors of fit were removed from each measure until the data were consistent with the measurement model. When the data were consistent with the unidimensional measurement model, the items were summed and averaged to form indices. The full correlation matrix of the study variables is reported in Table 2, and the retained items and their respective factor loadings are included in Appendix C.

A one-tailed independent samples t-test was conducted to investigate the effectiveness of the descriptive norms induction on perceived descriptive norms. The data showed a significant difference between conditions, t(181) = 12.94, p < .001, Cohen’s d = 1.91. Participants in the high descriptive norms condition, on average, reported greater prevalence perceptions of unfriending (M = 4.31, SD = 0.71) than those in the low descriptive norms condition (M = 2.66, SD = 1.00). Hence, the descriptive norms induction was considered successful.

Another one-tailed independent samples t-test was performed to examine the effects of the network density induction on perceived network density, revealing a significant difference between conditions, t(181) = 4.33, p < .001, Cohen’s d = 0.64. On average, participants in the dense network condition perceived greater network density (M = 3.91, SD = 0.92) than did participants in the sparse network condition (M = 3.37, SD = 0.86). Thus, the network density induction was also considered effective.

To test the study hypotheses, a two-way analysis of covariance (ANCOVA) was performed with the intention to unfriend as the dependent variable. Descriptive norms conditions and network density conditions (and their interaction term) were the independent variables. The attitude toward the target person’s online behavior, participants’ age, and gender were included in the analysis as covariates. The ANCOVA results are summarized in Table 3.

Hypothesis 1 predicted that perceived descriptive norms would positively influence unfriending intention. The data showed a significant main effect for descriptive norms conditions, F(1, 176) = 7.26, p = .008, partial η² = .04. Participants who were exposed to high descriptive norms messages reported significantly greater intention to unfriend the target victim (adjusted M = 3.41, SE = 0.05) than did those who were exposed to low descriptive norms messages (Adjusted M = 3.19, SE = 0.06). Thus, the data were considered consistent with H1.

Hypotheses 2 predicted that the effect of perceived descriptive norms on the intention to unfriend would be moderated by network density, such that the normative influence would be stronger in a dense than sparse network. The data yielded a significant interaction effect between descriptive norms conditions and network density conditions, F(1, 176) = 4.96, p = .027, partial η² = .03. Simple effects analysis investigating the interaction pattern showed that the normative influence on behavior was stronger in a sparse than a dense network. In the sparse network condition, participants who were exposed to high descriptive norms messages reported significantly greater intent to unfriend (M = 3.47, SD = 0.52, adjusted M = 3.44, SE = 0.08) compared to those who were exposed to low descriptive norms messages (M = 3.07, SD = 0.63, adjusted M = 3.06, SE = 0.08), F(1, 176) = 12.40, p < .001, partial η² = .07. However, in the dense network condition, there was no significant effect of descriptive norms on behavior, although those who were exposed to high descriptive norms messages reported greater behavioral intention (M = 3.37, SD = 0.49, adjusted M = 3.38, SE = 0.08) than did those who were exposed to low descriptive norms messages (M = 3.29, SD = 0.56, adjusted M = 3.34, SE = 0.08), F(1, 176) = 0.11, p = .743. Despite the significant interaction effect, the data were considered inconsistent with H2 because of the unexpected interaction pattern (see Figure 1 for interaction plot).

Figure 1.  Adjusted Group Means by Experimental Conditions.

Note. DN = Descriptive Norms. Error bars indicate the 95% confidence interval of each mean.

The unique findings of the present study must be considered in the context of several limitations. First, the gender distribution within our sample was predominantly female, representing 81% of participants. This disproportion should be acknowledged when extending our findings to the general population. Nevertheless, gender has been reported as neither a significant predictor of unfriending decisions (Barnidge et al., 2022; Kim et al., 2022) nor a significant moderator in the relationship between social norms and behavior, including behavioral intentions (Rhodes et al., 2020), in extant literature. Furthermore, this research was fundamentally concerned with examining theoretical constructs, with a primary focus on internal rather than external validity. However, it is recommended that subsequent research expanding upon this study should ensure a more balanced gender distribution to strengthen the generalizability of the findings.

Second, while this study's online experiment with a hypothetical scenario ensured controlled manipulation of key variables for robust analysis, it may have limitations in external validity. For instance, the race-gender matching, although intended to bolster the design's rigor by reflecting individuals’ tendencies online to engage with demographically similar others, might not fully represent the diversity encountered in real-world social media dynamics. Thus, questions remain about how social norms and network density interact with each other in pre-established realworld networks, where social dynamics among diverse group members have evolved over time (e.g., Lapinski et al., 2017). A fruitful approach for future research would be to base investigations on unobtrusive behavioral data collected from multiple pre-existing online networks with varying levels of network density (e.g., Jang et al., 2015), to further explore this area.