Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Social selection and influence of alcohol & marijuana implicit cognitions and behaviors: a longitudinal investigation of peer social network dynamics
(USC Thesis Other)
Social selection and influence of alcohol & marijuana implicit cognitions and behaviors: a longitudinal investigation of peer social network dynamics
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
SOCIAL SELECTION AND INFLUENCE OF ALCOHOL & MARIJUANA IMPLICIT COGNITIONS AND BEHAVIORS: A LONGITUDINAL INVESTIGATION OF PEER SOCIAL NETWORK DYNAMICS by Kathryn Coronges A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (PREVENTIVE MEDICINE) August 2009 Copyright 2009 Kathryn Coronges ii ACKNOWLEDGMENTS Mary Ann & Chih-Ping for the shoulders Alan for the soft nudges and dependable loyalty Peter for asking “but where’s the network,” and for the quiet nods of encouragement throughout my search Tom for the open door, ear, & mind; and for the road maps, compass and unmistakable cheering from the other side Mom for showing me how to create mysteries from soil and build houses out of sand Dad for showing me how to create simplicity from chaos, and purpose out of uncertainty Janie for holding up a mirror with raw honesty and genuine adoration, and for sharing countless life lessons over tripe, tori-sashi and arepas Joe for giving me the fairytale, and for the unwavering patience, perspective, and integrity iii TABLE OF CONTENTS ACKNOWLEDGMENTS.......................................................................................................... ii LIST OF TABLES...................................................................................................................... v LIST OF FIGURES.................................................................................................................. vii ABSTRACT...............................................................................................................................viii CHAPTER 1 INTRODUCTION............................................................................................. 1 1.1 Social Structural-Implicit Cognitive Paradigm........................................................... 2 1.2. Specific Aims.............................................................................................................. 5 1.3 Substance Use – Model Behavior for a Social-Cognitive Analysis........................... 7 1.4 Relevance of Proposed Aims..................................................................................... 8 CHAPTER 2 REVIEW OF SOCIAL NETWORK ANALYSIS RESEARCH............... 10 2.1 Social Network Analysis .......................................................................................... 11 2.2 Network Theories of Contagion & Diffusion............................................................ 13 2.3 Social Network Approach to Behaviors..................................................................... 16 CHAPTER 3 REVIEW OF COGNITIVE ASSOCIATION RESEARCH....................... 18 3.1 Explicit and Implicit Cognitions: Two Distinct Memory Systems........................... 19 3.2 Associative Memory................................................................................................. 20 3.3 Effects of Social Features on Associative Memory.................................................. 21 3.4 Implicit Cognition Approach to Behaviors............................................................... 23 3.5 Measures of Associative Memory............................................................................ 25 CHAPTER 4 INTEGRATION OF SOCIAL & COGNITIVE DOMAINS...................... 28 4.1 Methods & Measurements Across Social and Cognitive Domains.......................... 30 4.2 Contagion & Diffusion Models of Implicit Cognition.............................................. 31 4.3 Social Selection and Behavioral Influence................................................................ 32 4.4 Network Selection and Social Influence of Drug Behaviors .................................. 32 4.5 Contagion of Implicit Attitudes................................................................................. 36 4.6 Cognitive Influence Through Associative Memory .................................................38 4.7 Associative Memory Drives Social Selection .......................................................... 39 4.8 Summary of Cognitive Contagion Research............................................................. 40 CHAPTER 5 THEORETICAL MODELS & HYPOTHESES......................................... 41 5.1 Study Plan................................................................................................................. 41 5.2 Analytic Plan............................................................................................................. 41 5.3 Hypothesis Testing.................................................................................................... 42 5.4 Comparison of Network & Regression Approaches................................................. 49 CHAPTER 6 STUDY METHODS ....................................................................................... 51 6.1 Study Methods........................................................................................................... 51 6.2 Survey Instruments.................................................................................................... 53 iv CHAPTER 7 LOGISTIC REGRESSION ANALYTIC PROCEDURE........................... 66 7.1 Multilevel Mixed-Effects.......................................................................................... 67 CHAPTER 8 LOGISTIC REGRESSION RESULTS........................................................ 70 8.1 Descriptive Statistics................................................................................................. 73 8.2 Cognitive and Behavioral Influence Results............................................................. 70 8.3 Logistic Regression of Alcohol Behaviors............................................................... 75 8.4 Logistic Regression of Marijuana Behaviors............................................................ 79 8.5 Logistic Regression of Alcohol Cognitions.............................................................. 81 8.6 Logistic Regression of Marijuana Cognitions........................................................... 83 8.7 Summary of Behavioral Influence Logistic Regression Analysis............................. 85 CHAPTER 9 SIENA NETWORK MODELING ANALYTIC PROCEDURE.............. 92 9.1 SIENA Modeling Procedures.................................................................................... 93 9.2 Rate and Objective Functions................................................................................... 95 9.3 SIENA Modeling Assumptions.............................................................................. 102 9.4 Hypothesis Testing.................................................................................................. 103 CHAPTER 10 SIENA NETWORK MODELING RESULTS........................................ 105 10.1 Descriptive Statistics ........................................................................................... 105 10.2 Model Selection Procedures................................................................................ 110 10.3 Dynamic Network Selection Modeling............................................................... 117 10.4 Summary of Network Selection Models.............................................................. 127 10.5 Network & Behavioral Coevolution Models....................................................... 127 10.6 Summary of Network Behavioral Coevolution Results...................................... 136 CHAPTER 11 REVIEW OF FINDINGS ACROSS MODELS, ACROSS METHODS ..140 11.1 Comparison of Network Analysis Results (Studies 2 and 3) ...................................140 11.2 Comparison of Logistic and Network Findings from Studies 2 & 3 ........................147 CHAPTER 12 DISCUSSION OF FINDINGS................................................................... 151 12.1 Methodological Considerations........................................................................... 152 12.2 Review of Behavioral and Cognitive Influence Findings.................................... 157 12.3 Intervention Implementation Effects.................................................................... 161 12.4 Applications to Intervention Strategies................................................................ 162 12.5 Limitations of Study............................................................................................. 166 12.6 Conclusions.......................................................................................................... 173 REFERENCES............................................................................................................................... 175 APPENDICES........................................................................................................................ 186 Appendix A: Assumptions of SIENA Simulation Procedures....................................... 186 Appendix B: Assumptions of Theoretical Model.......................................................... 188 v LIST OF TABLES Table 5.1 Summary of SIENA Network Modeling Hypotheses and Their Implications...... 48 Table 5.2 Analytic Strategies to Measure Network Effects in Dissertation Studies.................. 50 Table 6.1 Alcohol and Marijuana Cues for Associative Memory Tasks.................................. 58 Table 6.2 Means of Alcohol & Marijuana Implicit Cognition Responses Across Cognitions Tasks.............................................................................................. 60 Table 6.3 Comparison of Alcohol and Marijuana Use Predictions of Across Cognition Tasks............................................................................................... 62 Table 8.1 Demographics of Study Participants........................................................................... 71 Table 8.2 Cognitive and Behavioral Change Counts From Baseline to Follow-up................... 73 Table 8.3 Logistic Regression Results for Alcohol Behaviors on Ego and Alter Effects in Continuation HS Students......................................................................... 76 Table 8.4 Logistic Regression Results for Marijuana Behaviors on Ego and Alter Effects in Continuation HS Students.......................................................................... 80 Table 8.5 Logistic Regression Results for Alcohol Cognitions on Ego and Alter Effects in Continuation HS Students.......................................................................... 82 Table 8.6 Logistic Regression Results for Marijuana Cognitions on Ego and Alter Effects in Continuation HS Students......................................................................... 84 Table 9.1a Endogenous Network Selection Parameters in Actor-oriented Modeling................ 99 Table 9.1b Exogenous Network Selection Parameters in Actor-oriented Modeling.................. 99 Table 9.1c Behavioral Influence Parameters in Actor-oriented Modeling............................. 100 Table 10.1 Population Demographics & Attributes for SIENA Analysis .............................. 105 Table 10.2 Score Tests for Endogenous and Exogenous Network Parameters in High School Standard Condition Classes............................................................ 113 Table 10.3 Score Tests for Endogenous and Exogenous Network Parameters in High School Experimental Condition Classes.................................................... 114 Table 10.4 Score Tests for Endogenous and Exogenous Network Parameters in High School Control Condition Classes.............................................................. 115 Table 10.5 Friendship Selection Models in Standard Intervention Classrooms....................... 121 vi Table 10.6 Friendship Selection Models in Experimental Intervention Classrooms................ 124 Table 10.7 Friendship Selection Models Control Classrooms.................................................. 126 Table 10.8 Selection and Influence Co-Evolution Models of Alcohol and Marijuana Cognitions and Behaviors in Standard Intervention Classes................. 129 Table 10.9 Selection and Influence Co-Evolution Models of Alcohol and Marijuana Cognitions and Behaviors in Experimental Intervention Classes......... 132 Table 10.10 Selection and Influence Co-Evolution Models of Alcohol and Marijuana Cognitions and Behaviors in Control Condition Classes...................... 135 Table 10.11 Co-evolution Selection& Influence Parameter Results by Condition.................... 138 Table 11.1 Summary of Homophily & Assimilation Parameters for Drug Cognitions and Behaviors............................................................................... 143 Table 11.2 Summary of Network Modeling and Coevolution Findings from Continuation HS Intervention................................................................................... 145 Table 11.3 Comparison of Regression and Network Analyses Findings on Social Influence by Outcome................................................................................... 148 Table 12.1 Review of Behavioral Influence and Network Selection Dissertation Findings................................................................................................ 158 vii LIST OF FIGURES Figure 2.1 Types of Network Data..................................................................................... 11 Figure 5.1a,b Influence of Friends Marijuana & Alcohol Activities on (a) Behaviors, and (b) Cognitions: Longitudinal Regression Models of Cognition and Behavior Exposure................................................................ 43 Figure 5.2 Coevolution of Network Selection & Behavioral Influence of Alcohol & Marijuana Cognitions & Behaviors................................................ 45 Figure 6.1 Example of Word Association Task Instructions............................................. 56 Figure 6.2 Example of Social Network Survey Item......................................................... 64 Figure 6.3 Illustration of Network Exposure (Alter Effects) Calculation.......................... 65 Figure 7.1 Logistic Regression Equation for Behavioral Influence Models...................... 67 Figures 8.1a-d Significant Alter and Ego Effects on Alcohol and Marijuana Activities.......... 89 Figure 9.1 Objective Function Equations for Selection & Influence ERG Models............... 97 Figure 9.2 Dyadic Behavioral Assimilation Configurations................................................. 101 Figure 10.1 Illustration of Friendship Networks & Alcohol Activities in Experimental Condition Classes........................................................................... 107 Figure 10.1 Illustration of Friendship Networks & Marijuana Activities in Standard Condition Classes.................................................................................. 108 Figure 10.3 SIENA Actor-Oriented Evaluation Functions for Gender Selection................... 119 Figure 12.1 Proposed Alcohol Using Dynamics for Standard and Experimental Conditions.............................................................................................................. 165 Figure 12.2 Attribute Data Transposed onto Transitive Tie Formation................................. 172 viii ABSTRACT Individuals are embedded in social contexts - their cognitions, attitudes, resources, and normative references are meaningful to the extent that they define individuals relative to others. Yet, statistical models are routinely applied to behavioral theories, which assume that individuals act independently from one another. Even day-to-day decisions about private activities require cognitive pathways that depend on social meanings and contexts. Implicit cognitions, characterized by being non-conscious and automatic, are stored in memory with physical and social information. Evidence suggests that implicit attitudes can be communicated and spread through social exchanges, and can influence behaviors and drive social structures. Social network analysis (SNA) is a set of tools and concepts that assess how relationships between people influence their behaviors (rather than teasing out interdependence as a nuisance). A relatively new area of SNA allows statistical inference about behavioral and network evolution using stochastic actor-oriented modeling. In these dissertation studies, the role of classroom friendships on alcohol and marijuana implicit cognitions and behaviors are examined by regression and SNA techniques. Data are from an evidence-based drug intervention delivered to High school students over 3- months. Classes were randomly assigned to either the: standard program, (led by health educators), experimental program, (facilitated by peer leaders), or control condition (no intervention). Regression findings show that individuals adopt their friends' behaviors and cognitions regardless of program condition. Network findings show that in the standard condition, friends' alcohol behaviors influenced one’s own behaviors, and there was a preference for alcohol using friends. In the experimental condition, findings show preferences for friends who are non- drinkers and who have similar alcohol behaviors, and there was an overall reduction in alcohol cognitions. An explanation for this may be that using existing social relationships to disseminate ix educational material facilitates gradual shifts in social norms towards healthy alternatives to alcohol and marijuana use. Interventions that target both (overt) behavioral and implicit cognitive changes at both the individual and group levels may be more effective at creating a cultural setting that supports healthy behaviors. Harnessing the spread of unconscious cognitive tendencies through friendship structures may be an effective strategy for creating behavioral change. 1 CHAPTER 1 INTRODUCTION Each person carries with him his cognitive field as a map of the world. He responds not the world, but to the map... Since the effects of communication depend on the manner in which it meshes with an existing cognitive map, we might entertain the notion that these effects will take place more readily when the cognitive maps of the communicators are similar in structure. --Runkel, 1956 (pp178-9) Almost all instances of decision-making are influenced by interactions with the people around us. Social exchanges play an important role in every level of our lives, whether they are transient encounters, intimate networks of relationships, formal and informal group affiliations, or broader interactions within one’s culture. Individuals are embedded in social contexts - their cognitions, attitudes, resources, support systems, and normative references are formed as a result of shared experiences with others, are relevant to the extent that they define the individual relative to others. Despite this, theoretical models routinely apply social influence and selection mechanisms to data that focus on individuals’ perceptions of the world without taking into account larger social contexts within which the person is embedded. Even day-to-day decisions involving entirely individualistic activities require associative cognitive pathways that integrate subjective meanings, which are entirely embedded within social meaning and contexts. Furthermore, the random sample correlative approach typically used in behavioral research is grounded in assumptions about individuals acting independently from one another. Consequently, many behavior modeling strategies have been limited because they do not address various levels of social dependencies. Social network analysis is a particularly useful set of techniques in mapping social dependencies while accounting for multiple levels of social exchanges. Social network theories attempt to explain individual and group activities by how people are positioned in their social 2 community and the specific distribution of behaviors in that community. Unlike classic statistical techniques that correlate individuals to their attributes, attitudes, and behaviors, social network analysis methods focus on capturing dependencies that exist between individuals and other social entities. Information about the individual can be overlaid with information about who is connected to whom, revealing unique information about social exchanges and subsequent behavioral changes (Scott, 2000; Wasserman & Faust, 1994). Relational analyses can answer questions about how behaviors emerge from relationships between people, and how they change over time. Constructs characteristic of many psychological & sociological paradigms, such as social support, social norms, cognitive dissonance, and observational learning, can be supplemented with information about how an individual’s behaviors and attitudes relates to the attitudes and behaviors of their contacts. In addition, longitudinal data can reveal how individualistic activities spread through groups, what kinds of people are more influential and what kinds of people are more susceptible to these influences. Social network methods provide unique tools to account for dependencies between behaviors and social structures. 1.1 Social Structural - Implicit Cognitive Paradigm A social structural perspective of cognition is proposed as the basis for the studies presented in chapter 5. Cognitions in these studies refer to implicit memory, which are now considered to be a major cognitive system (distinct from rational, conscious processes) that governs problem behaviors. Implicit cognitions are characterized by their extremely quick activation, which is experienced as spontaneous and with relatively little awareness. Cognitions generated from implicit paths draw from associations in memory, rendering these cognitions to be heavily biased on past experiences and personal meanings. Behavioral models that only address direct measures typically rely on the notion that rational processes determine one’s actions. Yet, in real life 3 settings, many behaviors are experienced as automatic, where there is little opportunity to weigh pros and cons of a decision Unlike explicit perceptions of attitudes and norms, implicit attitudes can be measured indirectly revealing pools of perceptions and knowledge that are unavailable to conscious introspection (Ames, Franken & Coronges, 2006, Kahneman, 2003, Roediger, 1990). The social structural - implicit cognitive paradigm proposed here stem from two main lines of established research findings. First, social network analysis inquiries find that behaviors are largely a result of relations between people. From this perspective, behaviors and innovations can be tracked as they diffuse through a network via social connections, such as friendships and work affiliations. Typically, research is focused on explicit or tangible behaviors and practices. However, the second line of research supported by several domains of psychological and cognitive studies suggests that implicit evaluations are also exchanged from person to person. Research supports the notion that implicit attitudes are a cornerstone in how individuals relate to one another. Therefore, it is proposed here that implicit cognitions, much like explicit attitudes, can be communicated and spread through social networks, and that implicit exchanges can influence behavioral decisions and drive social structures. Research is reviewed that describes the importance of social context in implicit cognitive processes. Cognitive associations are found to be intrinsically embedded within memories of physical and social settings, where the very presence of other individuals can trigger or inhibit particular cognitive pathways. Implicit processes can be disrupted, altered and sometimes converge when individuals are given group tasks (Asch, 1956; Loftus, Miller & Burns, 1978; Roediger & McDermott, 1995; Roediger, Meade & Bergman, 2001; Valente, Unger & Johnson, 2005). Past research also supports that implicit attitudes can be conveyed and processed implicitly through forms of symbolic language where both the sender and receiver are unaware that such an exchange has occurred (Carley, 1997; Carley & Palmquist, 1991; Choudhury, 2004; DiMaggio, 4 1997; Pentland, 1986). Examining how social networks facilitate (or hinder) accessibility of specific information may be an important aspect of social learning that remains almost entirely untapped. Cognitive association data, therefore, may provide a much understudied aspect of social influence and selection, where friends may be highly integrated in each other’s associative memory structures underlying many behavioral choices. Both social structural data and implicit cognition assessments provides many challenges for researchers. Implicit associations require indirect measures, making contagion of implicit information much harder to detect as compared to the spread of explicit attitudes. Secondly, social network analysis requires relational data from an entire network to link individuals to one other and software capable of calculating meaningful indices from relational data represented as matrices. Until recently, network analysis has been confined to exploratory, qualitative, or visual research. However, new advances in network modeling have enabled statistical assessments of network configurations, radically increasing the kinds of questions that can be answered with network data (Contractor, Wasserman & Faust, 2006; Monge & Contractor, 2003; Snijders, 2005; Snijders, Steglich & Schweinberger, 2007). A burgeoning network approach uses simulation techniques to generate estimates and standard errors for network and behavioral preferences. Identifying the evolution of social network preferences over time employ what have been called stochastic actor-oriented models (Burk, Steglich & Snijders, 2007; Snijders, 2001; Snijders et al., 2007). These can be used to assess what behaviors drive social network evolution, and whether social relationships influence behavioral changes. This technique is essentially a form of agent-based modeling, with the exception that the resulting models are used for statistical inference about behavior and network evolution. Behaviors of one’s friends have been widely acknowledged to correspond to changes in one’s own behaviors, but only an analytic framework that can model these processes as co- 5 evolving systems can capture the complexity of social-behavioral dynamics. In addition, unlike regression models, which assume independence across participants, actor-based network analysis statistically accounts for the effects that friends have on one another (at the dyadic and triadic levels). Expanding social network analysis from exploratory or descriptive strategies to statistical ones is relatively new among social network analysts, but one that is almost entirely unexplored in the broader academic community. A major advantage of statistical network analysis is that it bases the estimates for friendship patterns and behavioral changes on social dependencies rather than mathematically teasing them out. Network models of social selection and behavioral influence, for example, can assess the extent to which people are drawn to others based on similarities of traits (e.g., gender, drug use) while accounting for specific influences of their existing friendships (Robins, Elliott & Pattison, 2001; Steglich et al., 2007). Of course, network analysis modeling requires new assumptions that are also imperfect representations of reality. Comparisons of simulation modeling network statistics to traditional correlative regression techniques can provide new insights into the advantages and disadvantages to these methods in regards to appropriateness of estimation assumptions, ease with which real field data can be managed, and conceptual operationalization of mechanics of social and behavioral changes. 1.2 Specific Aims This dissertation focuses on combining social network and memory association data to explore hypotheses about how friendship structures are related to cognitive and behavioral changes. These mechanisms might be particularly relevant for behaviors that are heavily reliant on social encounters, such as marijuana and alcohol use. Objectives of the studies presented address the exchange of drug-related cognitions and behaviors through classroom friendship ties, in the context of how a drug education program influences these social exchanges and behaviors. 6 Social selection and behavioral influences are modeled separately (as described below as specific aims 1 and 2) and simultaneously (specific aim 3). Hypotheses are based on assumptions about the contagiousness of implicit cognitions, the role of implicit attitudes on friendship choices, and the interaction between cognitive tendencies and susceptibility to social influences. Three specific aims are described below: (1) Assess behavioral and cognitive changes that occur as a result of social relationships over the course of the drug intervention. This research will determine whether friends’ cognitive associations and behaviors around drug use at baseline influence one’s own cognitions and behaviors at post-test. Analyses will address the extent to which decisions are driven by individual (ego) attributes and to what extent they are driven by the influences of one’s friends. (2) Evaluate whether selection of friends are driven by drug cognitions and behaviors. This research will examine whether friendship selection preferences can be explained by drug cognitions and behaviors. Analysis will be conducted to address whether friendships more likely to form when individuals have similar cognitions and behaviors (homophily), and whether selection-preferences are encouraged or depressed by the different interventions. (3) Test social selection and behavioral influence models simultaneously to parse out the integrated but distinct social processes. This research will assess the co-evolving processes of friendship selection based on alcohol and marijuana activities and behavioral changes as a result of friends’ influences. Assessments will test whether selection and influence processes occur concurrently, or if one mechanism shows more significant patterns after controlling for the other. 7 1.3 Substance Use – Model Behavior for a Social-Cognitive Analysis While the social structural-implicit cognitive paradigm is presented here as a general framework for socially relevant behaviors, these relationships are unlikely to exist for all behaviors and for all socially connected groups. Features of some behaviors make them susceptible to these dependencies more so than others. Alcohol and marijuana use have a number of features that make them particularly appropriate for this combined social network-implicit memory approach. First, drug use is very relevant in social contexts. Typically people know with relative accuracy whether their close friends drink alcohol or smoke marijuana, and in fact, these activities are important defining characteristics of adolescent friendships and peer group identifications (Sussman, Pokhrel, Ashmore & Brown, 2007; Sussman, Simon, Stacy, Dent, Ritt, Kipke, Montgomery, Burton & Flay, 1999). Second, drug use is thought to be largely governed by associative memory processes, and in some studies implicit evaluations have been found to be stronger predictors of alcohol and marijuana use than explicit attitudes (McCusker, 2001; Stacy, 1997). Third, drug use is typically a highly visible activity, which allows for exchange of implicit attitudes. Implicit communication around more isolative drug activities, such as heroin injection, would be less susceptible to fluid social influences, and therefore, less amenable to the paradigm presented here. After several decades of national attention to the issue, alcohol and marijuana use remains a significant public health problem and many large-scale interventions show only small effects. A national survey of high-school seniors reveals that about 83%of American adolescents have tried alcohol by the 12 th grade and about 28% report binge drinking in the last 30 days (CDC, 2007), with the highest rates in ages 16 to 25. Of particular concern has been a general rise in the popularity of marijuana, viewed by researchers and the public as an important "gateway" drug leading to the use of "hard drugs" such as cocaine and heroin (Kandel & Yamaguchi, 1993; Kandel, Yamaguchi & Chen, 1992; Newcomb & Bentler, 1986; Osgood, Johnston, O’Malley & Bachman, 1988). Unlike 8 so many diseases, trajectories of debilitating drug behaviors can be improved with lifestyle choices made during vulnerable years of adolescence. Questions arise as to why youth continue to use drugs when clearly undesirable consequences are laid out for them. Understanding how new drugs diffuse through youth populations, and more importantly, why are some individuals so much more susceptible to abusing mind-altering substances, while other can limit themselves to recreational use is essential for designing effective interventions. Substance use remains a complex challenge that would likely benefit from a social-cognitive approach that could address the complex cognitive and social processes in play. 1.4 Relevance of Proposed Aims In the set of studies presented in this dissertation, relationships between social structures and behavioral changes are examined through applications of distinct but broadly overlapping social dynamic and cognitive processing theories. Despite advancements in establishing relationships between associative memory and substance use, as well as between social networks and substance use, there has been little attention devoted to dependencies between these two fields. Although researchers frequently use network terms metaphorically, to date, no studies have applied social network analytic techniques to the study of implicit cognitive associations. Tools now exist to measure statistical relationships between associative memory and social networks, which will undoubtedly lead to greater articulation of social processes and inform educational strategies. This dissertation contributes new ways to conceptualize social exchanges, accounting for symbolic communication that may occur without either party’s awareness. Empirical analysis of the dynamics of friendships and behaviors allow a rare glimpse at selection and influence mechanisms at play in the trajectory of drug use from a group perspective. This dissertation incorporates the following unique features to understand communication dynamics around drug use in young adults: 9 (1) Inclusion of both implicit and explicit behavioral exchanges between individuals (2) Accounts for levels of structural dependencies among friends (3) Models selection and influence processes simultaneously (4) Applies correlative regression and network simulation techniques to the same hypotheses and data (5) Compares cognitive and behavioral diffusion processes across different intervention strategies Findings will be discussed in terms of their implications for behavioral interventions and education strategies. So far, little attention outside of social network analysis research, has suggested taking advantage of interpersonal dependence to disseminate health-related initiatives. If findings support the assertion that implicit cognitions of one’s friends influence a person’s own implicitly held associations in memory, harnessing the diffusion of unconscious cognitive tendencies may be the beginning of a new framework for designing, implementing, and evaluating educational programs. 10 CHAPTER 2 REVIEW OF SOCIAL NETWORK RESEARCH The relations among information can occur within the mind of an agent or between agents, such as "shared ideas" or the "I know that you know" linkages. Knowledge exists within and between individual agents and groups of agents…Adaptation occurs in a composite agent as the member agents change or the connections within and among them change. --Kathleen Carley, 2002 (p.7258) Social network analysis is a specific application of the mathematics of graph theory. Graph theory is a branch of discrete mathematics that deals with how space can be quantified in terms of entities (also called nodes or vertices) and their connections (also called links, relations, or edges). A set of nodes is called a graph (also called a network) and their specific configuration of links can be refereed to as the ‘realization’ of the network (Scott, 2000; Wasserman & Faust, 1994). Social network data specify relationships or affiliations between people (e.g., ‘x is friends with y’ or ‘x attended conference with y’) and enable mathematical expressions of social configurations (e.g., connectedness or centralization of the network, (Wasserman & Faust, 1994). In their most basic form, networks exist as relational data in the form of matrices. The most common network metrics are calculated from adjacency matrices, where the columns and rows list the same set of nodes, and cells where two nodes meet indicate whether they have a relation (indicated by ‘1’) or not (designated by ‘0’). Additionally, attributes of the actors in the network can be included so that social groups can be considered in terms of their shared characteristics (Scott, 2000; Wasserman & Faust 1994) Network diagrams can be drawn to visually represent the pattern of relationships between entities. Nodes are typically represented as circles and connections between them are depicted as lines (with arrows to indicate direction of relation). Attribute information is usually depicted by color, shape or size of the nodes, as shown below in Figure 2.1. 11 Figure 2.1 Types of Network Data Figure 2.1 For a hypothetical set of actors 1-4, an adjacency matrix (NxN) is used to display the actors’ relationships to one another, while the vector (1XN) is used to represent one of their attributes, gender. Relational data plus attribute information can be depicted in a network diagram. Social network analysis techniques have been established in order to empirically quantify influences from direct social ties. Instead of describing behaviors as a function of an individual’s attributes, such as perceived social support or behavioral norms, the network approach focuses on how relations between people determine individual behaviors. The broad network perspective is that behaviors are not self-contained, but rather embedded within networks of social influences and social meaning (Burt, 1987; Carley, 1991; Friedkin, 2004; Granovetter, 1978; Valente, 1996). Behaviors are viewed as a consequence of who a person is connected to, and to whom their affiliates are connected (i.e., “friends of friends’”). 2.1 Social Network Analysis The social network analysis offers multiple levels of analysis from local to global effects. Actors can be described in terms of their network position, such as their centrality (number of ties an individual has), density (the extent to which one’s friends are friends with each other), or alter effects (how many friends engage in a specific activity). Global level measures describe the 1 2 3 4 1 1 0 0 1 2 0 0 1 1 3 0 1 0 1 4 0 1 0 1 M/F 1 M 2 F 3 F 4 M 2 4 3 1 Relational Data Attribute Data Network Diagram 12 network in its entirety, such as centralization (the extent to which linkages are focused around one or a few actors in the network) or clustering (the extent to which subgroups exist in a network). Beyond descriptive network metrics, the network can be characterized at various levels of its structure. For example, analysis of the dyadic level examines pairs of network members together (e.g., the extent to which communication links from actor A to actor B are reciprocated or the extent to which actors A and B share specific traits). Triadic levels reveal links between three individuals, for instance, the extent to which links are balanced or transitive among groups of three. The network perspective posits that structural levels are inherently embedded within (and dependent on) all other network structures. A long-standing problem of general statistical approaches for analyzing social processes is that independence assumptions are artificially imposed. Since network data are by definition, relational, network models can address social dynamics while controlling for multiple levels of social dependencies. Network Statistics. Recent advances in social network science have led to the development of techniques that generate statistics to assess structural tendencies of networks. Techniques are employed that can identify the tendencies for specific network configurations (e.g., reciprocity or triadic relations) to occur more so than they would by chance (Robins, Snijders, Wang, Handcock, & Pattison, 2007; Snijders, 1995). Identifying the evolution of these preferences over time employ what have been called stochastic actor-oriented models (Burk et al., 2007; Snijders, 2001; Snijders et al., 2007). These can be used to assess what behaviors drive social network evolution, and whether social relationships influence behavioral changes. This technique is essentially a form of agent-based modeling, with the exception that the resulting models are used for statistical inferences about behavior and network evolution. The most common software for these analyses is Simulation Investigation of Empirical Network Analysis (SIENA). SIENA uses maximum likelihood estimation to simulate thousands 13 of networks guided by social dependencies proposed by the researcher. Each proposed relational dependency (e.g. reciprocity, transitivity) is represented by a separate parameter in the model. Parameter estimates of the specified network and behavioral functions derived from the simulated networks are compared to the estimates found in the observed data. T-tests are conducted to determine whether the observed and simulated variables are significantly different from one another. Thus, probabilistic configurations are compared to the observed network data in order to test whether each parameter occurred in the observed data more than it would occur in randomly generated networks with the same number of actors and links. Network parameters are thought to guide network structures if they fit the data significantly more than a randomly generated network (Burk et al., 2007). Parameters with a positive parameter value have a greater probability than would occur by chance of being present in the network, while configurations with a negative parameter value appear less than they would occur by chance (Robins, Pattison, Kalish & Lusher, 2007). This analytic framework can model social and behavioral processes as co-evolving systems and statistically account for the effects that friends have on one another (at the dyadic and triadic levels). 2.2 Network Theories of Contagion and Diffusion From the network perspective, information and influence spread through relational ties between individuals. Diffusion models focus on the accumulation of new information or tools in a society, where contagion models focus on mechanisms by which they spread. One’s role or position within the network structure can correspond to the extent to which they have access to particular resources (e.g., diverse versus homogenous set of relations) and particular network configurations lend themselves to facilitated exchanges while other constrain the network (e.g., well-distributed versus dense areas with no access to the rest of network). Models that focus on individual and network structures amenable to diffusion often identify basic network features such as centrality, density, 14 clustering, and individual adoption behaviors, such as thresholds (Valente, 1995). Patterns of social network ties can also be characterized as potential exchange points for individual resources and for behavioral influences. Models that focus on mechanisms of social exchange typically address the kinds of relationships that facilitate the spread of resources and behaviors, such as whether the actors are directly connected or if they hold similar structural positions, or whether the relationship is reciprocated or exists within a triad, or whether the tie is based on friendship, collaboration or competition. Diffusion of innovation studies focus on mapping the rate of adoptions and the distribution of an innovation in a society. Diffusion rates are dependent on the (1) the characteristics of the innovation, (2) the intrinsic qualities of network members (e.g., willingness to try something new) and, (3) the density and overall interconnectivity of the network (Rogers, 2003; Valente, 1995, 2005). A number of studies have identified characteristics of an innovation that influence its adoption: relative advantage over what it is replacing, compatibility to a person’s values and needs, complexity of its use, trialability without commitment, communicability and its observability. Individual susceptibilities to innovations also influence whether diffusion occurs. Thresholds are usually discussed in terms of an individual’s innovativeness or their willingness or adopt a new behavior or attitude. Threshold scores are calculated as the proportion of one’s personal network that have adopted a behavior before the individual has done so (Rogers, 2003). Threshold studies show that some people are quick to change their attitudes towards a new behavior, even if few or none of their friends have endorsed it, while another person may require most of their contacts to be engaged in the behavior before they will adopt it. People’s behaviors are influenced by local social networks to the extent that their threshold of susceptibility is reached (Granovetter, 1978; Valente, 1995). 15 Network contagion paradigms address the channels and dynamics of influence that emerge from network ties. Social cohesion refers to contagion processes that occur through direct contact between individuals, whereas structurally equivalence contagion refers to exchanges among those who share similar positions in their network. For cohesion to occur, contact between individuals serves as the mechanism through which people are exposed to attitudes and information. Cohesion studies reveal that behavioral influence is stronger between individuals that interact more frequently (Carley, 1986b; Romney & Faust, 1982), those that are valued more (Burt, 1987), those with more than one social function (friend and co-worker, for example), and those who are mutually invested in the relationship (Monge & Contractor, 2003). Unlike cohesion, structural equivalence does not require that individuals have a direct relation in order for social and behavioral influence to occur (Burt, 1987). Mechanisms of contagion and diffusion can also be viewed as a question of behavioral influence (are friends’ behaviors contagious?) or social selection (are friends selected because of their behaviors?). The tendency to select others with similar attributes and behaviors has been well established by social scientists as a consistent organizing principle of human relations. Selection and influence are usually framed in the context of homophily, the tendency for people to cluster and associate with others who share similar attitudes and behaviors. Dozens of studies have found that a person is much more likely to endorse the same behaviors as their friends. Homophily has been studied on the basis of similarity in age, gender, education, prestige, social class, tenure, and occupation (e.g., Ibarra, 1992; Kandel, 1978; McPherson, Smith-Lovin & Cook, 2001). Homophilic tendencies are attributed to social influence mechanisms when observable patterns of clustered attributes occur as a result of the friendship or relations. Social selection mechanisms are thought to be in play when clustered attributes are the very qualities that drew the individuals together 16 (McPherson et al., 2001). In its various forms, diffusion and contagion models can be used as tools to conceptually and mathematically describe the role of social networks in behavioral changes. 2.3 Social Network Approach to Behaviors Exposure of attitudes and behaviors through social relations increases the likelihood that the individual will hold that same belief or practice that behavior (Jaccard, Blanton & Dodge, 2005; Kandel, 1978; Strang & Tuma, 1993; Rice, Donohew & Clayton, 2003; Valente, 1995). A wide range of health-related behaviors, such as needle sharing, alcohol, tobacco & marijuana consumption, condom and family planning practices, and alternative medical therapies have been shown to be highly predicted by the behaviors of one’s direct social ties (Latkin & Knowlton, 2005; Rosenfield, Asavasena & Mikhanorn, 1973; Wellman, 1988). Social Networks & Drug Behaviors. Drug use is a highly social activity and appears to be initiated with friends and maintained among groups of friends (Conrad, Flay & Hill, 1992; Kandel, 1985; Kobus, 2003; Leventhal & Cleary, 1980; McAlister, Perry & Maccoby, 1979; Pearson, Sweeting, West, Young, Gordon & Turner, 2006; Sussman, Dent, Stacy, Burton & Flay,1995; Tyas & Pederson, 1998). Exposure to friends who use tobacco, alcohol, and marijuana has been firmly established as one of the most important determinants of adolescent drug use (Marcos, Bahr, & Johnson, 1986; Ritt-Olson, Unger ,Valente, Nezami, Chou, Trinidad, Milam, Earleywine, Tan & Johnson, 2005; Valente, Gallaher & Mouttapa, 2004). Empirical evaluations of the number of friends who are using tobacco or marijuana are directly correlated to one’s own current and future use (Alexander, Piazza, Mekos, Valente, 2001; Rice et al., 2003). In addition, drug using has been correlated to specific network positions. For example, popularity (network centrality) is associated to higher drug use (Pearson, Steglich & Snijders, 2006), particularly in schools where prevalence of use is relatively high (Burt, 1987; Carley, 1991; Friedkin, 2004; Valente et al., 2005). Overall, 17 network models reveal aspects of local and macro-level social environments that influence tobacco, alcohol and marijuana use. 18 CHAPTER 3 REVIEW OF COGNITIVE ASSOCIATION RESEARCH The human mind is susceptible to cultural representations, in the way the human organism is susceptible to disease. -- Sperber, 1985 (p74) Cognitive theories that have been applied to behavior models often include measures of beliefs, knowledge, intentions, expectancies, and perceptions of social norms (e.g., Theory of Planned Behavior - Ajzen & Fishbein, 1980; Social Cognition Theory - Bandura, 1977). Cognitive theories are based on the notion that behavioral decisions are governed in large part by the individual’s internalized responses to their environment, and cannot be predicted by the external environment alone. Cognitions in these models typically refer to direct assessments about one’s attitudes or perceptions, which rely on deliberative processes where pros and cons of an action can be consciously weighed. Models of social influence that incorporate deliberative cognition assessments often ask participants about their perceptions of others. For example, in the Theory of Reasoned Action, participants are asked how someone important to them would view a specific behavior, such ask “how favorably do you think your friends feel about using condoms” and a social norms assessment might be “how many of your friends support this behavior” (Fishbein, Hennessy, Kamb, Bolan, Hoxworth, Iatesta, Rhodes &, Zenilman, 2001). These cognitions assessments assume that social influence occurs through conscious explicit knowledge of the activities of others within one’s social network: “awareness does not imply influence, but it is a necessary condition for it” (Marsden & Friedkin, 1996, p20). Focused on less frequently in behavioral models are automatic cognitive processes that circumvent conscious deliberation. Yet, it is well established that during any cognitive encoding 19 or retrieving of information, conscious decision-making is activated simultaneously with automatic or implicit pathways, which occur without any conscious awareness (Kahneman, 2003). Implicit pathways depend on associative memory links, which are formed as a result of concepts or features occurring together (or consecutively) in real life. One of the features of implicit cognitive activity is that decisions can be made and carried out with the conscious awareness of making the decision or the rationale underlying it. Research focused on explicit cognitions only may miss important social influences that occur outside of the conscious realm. 3.1 Explicit and Implicit Cognitions: Two Distinct Memory Systems In general, memory processes can be categorized as either explicit or implicit. Explicit processes involve memories that we are conscious of, can verbalize, and can deliberately bring forth to make decisions. Explicit memories are described as autobiographical memories associated with context, space, and time. Implicit memories are characterized as thoughts whose source is not explicitly recalled and which do not usually reach consciousness (Greenwald & Banaji, 1995; Roediger, 1990). Mechanisms of implicit processes include classical conditioning (where exposure to two stimuli sequentially links them in memory, so that subsequent exposure to one leads to memory activation of the other), procedural memory (motor skills and habits, e.g., tying your shoe, hitting a ball), and priming (facilitation of perception of stimuli due to pre- exposure to them without awareness that the stimuli had ever been perceived). Conditioning, priming, and procedural memory processes rely on the individual’s ability to (implicitly) ‘remember’ previous experiences without conscious recollection or awareness of that experience. Evidence for two distinct memory systems over the last 20 years has lead to widespread acceptance of what was once highly contested by researchers (Shimamura, 1986). Neurobiological and behavioral evidence support that there are, in fact, two highly integrated, but distinct memory processes that work simultaneously. Some brain regions that are activated during 20 recall memory tasks are not amenable to conscious reflection, and further, synaptic response to stimuli is found to occur milliseconds before conscious thought can even be perceived and before action is executed (Phillips et al., 2003). The most profound evidence comes from research in amnesiacs, who are entirely unable to form new long-term memories, yet they perform perfectly normally on implicit tasks (Schacter, 1990). They can apparently build new associative links in memory during the task, even though they have no conscious recollection of the exposure that lead to the new memories. Research clearly demonstrates that implicitly activated memories function independently from explicit processes. 3.2 Associative Memory Theoretical perspectives that consider non-deliberate, automatic activation of thoughts are based on how memories are linked together through lexical, semantic or experiential associations. Associative memory links result in cognitive chains, where activation of one memory automatically triggers memories to which it is linked. Associated concepts become readily available in order to prepare the individual for related concepts should they, too, appear in the environment (Masson, 1995). There are a number of cognitive models that describe associative processes with different sets of assumptions about how features of memories become activated (e.g., multiple trace memory model - Hintzman, 1990; spreading activation - Collins & Loftus, 1975; activation-at-a-distance - Nelson, McEvoy & Pointer, 2003; neural networks – Hopfield, 1982). The simplest case of associative memory formation is when a person is exposed to two stimuli concurrently or sequentially creating a connection between the two events (or features of the events) in memory. For example, a person who obtains a sensation of calm while smoking a cigarette will likely form an associative link between ‘calm’ and ‘cigarette,’ and will subsequently think of cigarettes when seeking calmness without having any awareness that such a 21 strategy is underway. Associative processes become much more complex when applied to social nuances and complex emotional relations or opinions. When encountering a new person, for example, their physical or social presence may unconsciously (and automatically) stimulate memories of an incident or person from one’s past. The encounter will cause a clear sense of positive or negative affect, but neither the specific memory nor the awareness of the affect needs to be identified in order to influence decisions. 3.3 Effects of Social Features on Associative Memory An important aspect of implicit cognition processes is that associative memories are vulnerable to modification when they are retrieved. In fact, memory researchers describe remembering as a process of reconstruction, where the memory is created each time it is retrieved (LeDoux, 1996). Memory retrieval in a social context can be interpreted, transformed, or strengthened by associations that are relevant to that specific social situation (Duvarci & Nader, 2004). This perspective differs from localized or fixed models, which assume the memory is static and unchanged during retrieval. These models are appropriate for describing explicit knowledge, while implicit memories are stored with associated information, which can be changed or strengthened instantaneously during the activation process. A consequence of the vulnerable nature of implicit memory is that social exchanges and complex social dynamics can alter the memory itself. Basic science research provides support for the notion that cognition is dependent on social context. In particular, associative memory is considered to be a constructive process, whereby the specific associative links that are accessible in memory will vary depending on the social context within which they are activated. Several lines of research have shown different aspects of how social relations influence implicit cognition. Memory retrieval studies show that simply altering the social environment and the specific social setting within which the memory is retrieved, the set of associative linkages that are readily 22 accessible are significantly altered (Mitchell, Nosek & Banaji, 2003; Wittenbrink, Judd & Park, 2001; Wyer & Srull, 1989). For example, Wyer and Srull (1989) found that different associates were elicited in connection with the concept ‘self’ after participants were primed with various social settings (e.g., ‘at home,’ ‘as a teacher,’ ‘at the last office party’). Other memory studies show that when asked to recall a specific prototype, for example a person of a specific gender or racial group, stereotypical features of the prototype are more easily activated in subsequent tasks (Blair, Ma & Lenton, 2001; Greenwald & Banaji, 1995; Mitchell et al., 2003). Using the Implicit Association Test, Mitchell et al. (2003) found that individuals categorize Black athletes faster when category constraints focus on ‘occupation,’ and slower when categorizing white politicians in the context of ‘race’. These findings imply that associations between Blacks and athleticism were stronger than between Blacks and political; whereas categorizing Whites by occupation was more cognitively supported than categorizing them by race. Group dynamic studies also show that individuals’ cognitions can form and evolve as a result of being part of a group. An individual’s memories are found to change over time to converge with other people’s, where qualities and values held by one’s friends and close companions are sometimes recalled as their own (Decety & Sommerville 2003; Roediger et al., 2001). Other studies suggest that remembering can become a ‘group process’, where corroboration of artificial stories creates false memories, conformity of perceptual judgments occur over time, and groups can disrupt the individual’s performance on free recall tasks (false memories – Roediger et al., 2001; conformity – Asch, 1956; disrupting strategies - Decety & Sommerville, 2003). Group free recall studies show that participants perform better as a group of aggregated scores, than if they were alone. For example, Roediger et al. (2001) found that during a free recall task involving 24 household objects, up to 30% of inaccurately recalled items were due to those suggested by a confederate compared to when they were measured individually. 23 As a collection of studies, findings in memory recall and group memory studies strongly support the need to consider social context when investigating social influences. In addition, social influences appear to act on the implicit cognitions system by altering availability of associative memories. Investigations of social dynamics should consider an individual’s implicit cognitions as potentially transmittable through social encounters. 3.4 Implicit Cognition Approach to Behaviors Behavioral models that do not incorporate implicit processes assume that people generally engage in rational decisions. Implied in this assumption is that if people know about a consequence, then this knowledge is available and called upon during decision-making. However, automatic memory pathways account for up to 30% of variances in health-related decisions, independent of explicit cognitions (Greenwald & Banaji, 1995; McCusker, 2001). In some cases, indirect measures of attitudes correlate with behaviors more so than explicit attitudes (McCusker, 2001; Rudman & Click, 2002; Swanson, Rudman & Greenwald, 2001). Contradictions between explicitly stated views and automatic assessments are known as dual attitudes (Dovidio, Kawakami & Gaertner, 2002; Wilson, Lindsey & Schooler, 2000) and are found to be particularly disparate around stigmatized beliefs, such as drug use or racial stereotypes. Implicit Cognitions & Substance Use. Substance abuse is largely governed by implicit processes, where there is little impact of rational thought or weighing of consequences. In many cases, individuals begin smoking or using illicit substances even though they are well aware of their dangers. Addicts are found to endorse both negative and positive aspects of drug use in surveys, but they tend to draw on the positive features more quickly during implicit recall tasks (McCusker, 2001). Further, availability of positive associations are directly associated with rates of actual use. 24 In addition, associations between positive affects and illicit drugs result in unconscious cognitive biases that make it harder to avert attention away from drug-related stimuli (Ames et al., 2006). Attentional biases towards drugs can be detected by measuring reactions to images or words that are presented for a duration that is thought to be too brief for conscious processing (approximately 250 milliseconds, see Jones et al., 2002). For example, in a change-detection task, two versions of a visual scene are quickly presented on a monitor (objects are deleted or rotated from the original scene) and detection of this change is assessed (Jones, Jones, Blundell & Bruce, 2002). B. T. Jones et al. (2003) demonstrate that participants who detected subtle alterations in marijuana images reported higher levels of marijuana use than participants who were equally likely to detect a neutral change (Jones, Jones, Smith & Copley, 2003). Attentional biases have been shown for alcohol, marijuana, cocaine, and opiate users, and are stronger for users who report higher cravings or obsessive thoughts (Field, Mogg, Zetteler & Bradley, 2004; Franken, Kroon & Hendriks, 2000; B.C. Jones et al., 2002; B.T. Jones et al., 2003; Lubman, Peters, Mogg, Bradley & Deakin, 2000). In word association studies, drug use is predicted by responses to simple survey items that ask participants to give the first word that comes to mind, given drug-related cue stimuli (Stacy, 1995; Stacy, Ames, Sussman & Dent, 1996; Ames, Sussman & Stacy, 2005). In a prospective study of substance use among college students, Stacy (1997) found that associative strengths between positive outcomes (‘relaxing’ or ‘feeling good’) were better predictors of subsequent alcohol and marijuana use than outcome expectancies, sensation-seeking, acculturation, and gender. In other word association studies, high and low risk populations are found to differ markedly in their associative responses to words with more than one common meaning, one of which is drug-related (i.e., double entendres, such as bud, draft, pitcher, shot, hit; Stacy 1995, 1997). For example, the ambiguous cue ‘bud’ was more likely to elicit drug-related responses 25 such as ‘marijuana’ or ‘Budweiser’ in high-risk youth, while low-risk youth were more likely to give responses such as ‘rose’ or ‘friend.’ Further, changes in cognitive associative features of drug-related concepts are found to correspond to changes in drug using behaviors (Bovasso, Szalay, Biase & Stanford, 1993; Szalay, Bovasso, Vilov & Williams, 1992). One study used an association task measured before and after an intensive drug treatment program to create a matrix of word associations (Bovasso et al., 1993). Individuals who completed the treatment had an associative structure that was larger, more centralized, and less dense compared to persons entering the program. In a network analysis of word association data, Coronges et al. (2007) found that as exposure to drug activities increase, associations around drug-related cues becomes more integrated in memory structures. For example, the word “beer” is considerably more accessible and central in associative chains in college students (and is elicited often regardless of the cue) as compared to seventh graders. Taken together, the collection of studies shows that word association instruments capture cognitive associations that are important in predicting drug use. 3.5 Measures of Associative Memory Associative strength is a test of how strongly concepts are linked together in memory, or how easily one concept activates another (Yin & Knowlton, 2005). Associative memory is expressed by facilitated performance on tasks that specifically do not require conscious processing (Roediger, 1990; Schacter & Graf, 1986). One of the simplest implicit assessments is the free word association task. Exposure of a single word has been shown to prime associative pathways, facilitating access to related concepts (e.g., Neely, 1991). From an associative memory perspective, distributions of associative connections and their relative strengths around a concept is what gives meaning to information and memories (Deese, 1965). 26 Word association assessments use cues words to stimulate associative pathways and elicit automatic responses from memory. Outcome-behavior tasks evaluate the strength of the association between affects or outcomes and specific behaviors. Cues are designed to target specific behaviors. Word association responses indicate that a connection between the cue and the response exists in memory and is activated automatically (e.g., target response might be ‘beer’ if stimuli cue is ‘bud’). The frequency of the response, given a set of cues, estimates the connection strength of the target and cue stimuli. Extensions to simple word cues include the use of phrases, such as “feeling good” or “Friday night” (Ames, 1998; Stacy, 1997; Stacy et al., 1996a) and ‘continued free associations’, where cues are repeated multiple times (e.g., fun:____, fun:____ fun:____). Consistently, results from word association tasks find that responses predict a range of behavioral outcomes, including alcohol and marijuana use (Ames, 1998; Stacy, 1997; Stacy et al., 1996a), memory biases in depressive patients (Watkins, Martin & Stern, 2000), risky sexual behavior (Stacy, Newcomb & Ames, 2000), and general health-enhancing behaviors such as seat belt use and exercising (Benthin, Slovic, Moran, Severson, Mertz & Gerrard, 1995). Other commonly used associative strength assessments are reaction time tasks, such as the Implicit Association Test (IAT), the visual dot probe, and Stroop tasks (Franken, Kroon & Hendriks, 2000; Franken, Stam, Hendriks & Brink, 2003; Greenwald & Banaji, 1995; MacLeod & MacDonald, 2000; Watson & Humphreys, 2000). These measures are based on the fundamental finding that individuals react faster when categorizing strongly associated or compatible concepts and slower when categorizing concepts that are not associated or not compatible in memory. Participants are asked to quickly pair words or images that are presented quickly on a screen. Faster reaction times are found to correlate with associative strength between the two concepts. In general, cognitive processes that are challenged during associative memory 27 tasks may be critical for understanding the impetus underlying many behaviors and social contextual effects on these behaviors. 28 CHAPTER 4 INTEGRATION OF SOCIAL & COGNITIVE DOMAINS [A new] view of culture is...the interaction of shared cognitive structures and supra-individual cultural phenomena that activate those structures to varying degrees --DiMaggio, 1997 (p264) Models used to direct persuasive campaigns, such as drug prevention programs, typically assume behavioral changes are a result of both internal cognitive capacity and external environmental factors. Social behavioral paradigms typically acknowledge that cognitive tendencies act as filters through which social information is encoded and retrieved, and cognitive approaches usually recognize the fundamental role of social context in memory organization and retrieval. There is also a wide range of models that have addressed their interaction. Support for cognitive dissonance theory, for example, finds that people choose to be with others who are like themselves in order to decrease exposure to inconsistent ideas (Festinger, 1957). Computational models often specify rules of cognitive processing and social interactions within pre-existing social and cognitive landscapes, including the individual’s internal representations of the world (Carley & Prietula, 1994). The conceptual elaboration and support for social-cognitive models of behavior is tremendous and diverse -from social anthropology (Johnson, 1996); to behavioral ecology (Dunbar, 1998), neuroscience (LeDoux, 1996), cognitive neuroscience (Decety & Sommerville, 2003), computational simulation social psychology (Bandura, 1989), cognitive psychology (DiMaggio, 1997), social cognition (Howard, 1994), cognitive science (DiMaggio, 1997), computational science (Carley, 1990, Clancey, Sierhuis, Damer & Brodsky, 2004), and artificial intelligence (Collins & Quillian, 1969). These fields specify that human behavior is a result of cognitive processes that are meaningful as they relate to their social contexts. As a collection of social and biological sciences, human behavior is described as arising out of specific cognitive processes that determine how individuals 29 interpret and attach meaning to their experiences in conjunction with what types of environmental resources they can access to respond to these experiences. However, even with the widespread recognition, when implemented in the field, studies tend to emphasize either social structure and interactions, or cognitive organization and processes. In some cases, unexplained variances in social or cognitive models may underscore the importance of considering these two paradigms together. For example, dozens of studies find that exposures to drug behaviors through direct social ties correlate with an individual’s own drug use, but that thresholds of social triggers vary from person to person (Granovetter, 1978; Valente, 1995). That is, some people are intrinsically more susceptible to social influences than others. Differences in thresholds of behavior change may be due, in part, to pre-existing cognitive associations. Cognitive aspects, particularly at the implicit or non-conscious level tend to be overlooked in these kinds of studies. Convergence of social network analysis concepts and cognitive network contagion theories can broaden the perspectives of how social information is exchanged and internalized (Carley, 1986a; Krackhardt, 1987; Marsden, 1996). While cognitive and social constructs are typically acknowledged as part of any coherent behavioral model, there has been little attention devoted to empirical assessments of dependencies between social interactions and cognitions. The dozen or more theoretical models that do integrate social and cognitive perspectives, have not yet attempted to integrate major streams of research in social network analysis and theories of implicit cognition, and no studies have been conducted that measure the spread of implicit attitudes through social interactions (Gazzaniga & Heatherton, 2004). These models will need to account for the fact that the information one has available to them at any given moment is highly dependent on the social and physical setting in which a person is situated. Recent advancements in social network analyses and implicit cognitive studies call for more thorough investigation into empirical socio-cognitive approaches. An integrated approach combining 30 networks theories and cognition theories, might offer more complete pictures of social mechanisms and behavioral change processes. 4.1 Methods & Measurements Across Social and Cognitive Domains One major constraint in attempts to integrate cognitive and social perspectives has been the limited number of valid and robust assessment tools and analytic techniques that cross from one discipline to the next. Studies outside of cognitive psychology use what might be considered distal measures of specific cognitive activity, such as co-occurrence of words or concepts in texts or documents (Monge & Contractor, 2003, chapter 6), semantic differentials (reactions to concepts on bipolar scales (Osgood, 1964; Runkel, 1956), cognitive social structures or knowledge structures (perceptions of who knows whom and what they know; Krackhardt, 1987). These kinds of measures may not capture the dual nature of cognition, which involve both conscious and unconscious processes. On the other hand, methods to measure social influence employed by many behavior and intervention researchers often consist of self-report assessments of social support, such as marital status, participation in community events, and contact with friends (Barrera, 1986), but often do not include empirical evaluations of direct social interactions, such as sociometric data. Studies that incorporate social network effects typically use correlative regression analyses (e.g., logistic and linear regression), which assume independence across observations. While social network models are more appropriate for explorations of dependencies between social structure and behaviors, studies that apply social network analysis to health and psychology are usually published in one of three key social network journals (Social Networks, Connections, and the Journal of Social Structure). Segregation of findings into distinct fields might lead to a lack of diffusion of valid, reliable and robust measurements and methods. Overall, we often find weaknesses in methodology or conceptual frameworks of social or cognitive factors, and findings 31 are usually reported in journals specific to one field because measures or analytic approaches are rarely recognized across fields. 4.2 Contagion & Diffusion Models of Implicit Cognition Contagion of implicit evaluations requires that implicit social cues are exchanged during social interactions and contains information that is not expressed through explicit means. As with explicit attitudes, exposure to implicit attitudes can also be operationalized as the extent to which one’s direct contacts possess particular cognitions. While implicit evaluations are also thought to spread through symbolic communication, it is suggested that direct exposure to individuals is still required for contagion of implicit evaluations. Measurements of social exposure to implicit cognitions can be determined in much the same way as are explicit behaviors or overt attitudes. In this case, dense networks would create social settings where implicit cues are ubiquitous, increasing the salience of social exposure. Exchanges of implicit information may function differently, and in some cases, more readily than explicit exchanges. Implicit evaluations might be particularly important for attitudes whose concealment would be advantageous. Revealing deviant behaviors, such as drug use or risky sexual behaviors, might be wrought with uncertainty and risk. On the other hand, implicit cognitive tendencies towards drug use are relatively risk-free, and are not subject to rational concerns around uncertain consequences. Therefore, there are some cases where implicit evaluations might be diffused more easily than knowledge. Success of diffusion is also dependent on the extent to which the innovation is compatible with one’s own views or needs (Valente, 1995). Compatibility of implicit evaluations to one’s own associative structure appears to be crucial in successful transmission of implicit information (Levesque, Wilson & Wholey, 2001; Palazzolo, 2005; Palazzolo, Serb, She, Su & Contractor, 2006; Runkel, 1956). In many cases, implicit evaluations differ from explicit ones. Scenarios where friend’s implicit evaluations are 32 compatible with one’s own, while their explicit attitudes vary, could lead to greater implicit diffusion compared to explicit diffusion. 4.3 Social Selection and Behavioral Influence of Drug Activities As discussed in chapters 2 and 3, both social network influences and cognitive tendencies are independently predictive of drug use and other health-related behaviors. Models of behavioral influence go hand-in-hand with social selection theories. Social selection and social influence models are often pitted against one another in order to explore whether people’s relationships influence their behaviors or whether behaviors determine who has relationships with whom (Hall & Valente, 2007; Kandel, 1978; Sieving, Perry & Williams, 2000; Snijders & Baerveldt, 2003; Steglich, Snijders & Pearson, 2007; Thornberry, Lizotte, Krohn, Farnworth & Jang, 1994). Changes in network structure can be equated to ‘social selection’, while changes in individual behaviors are identified as ‘social influence’ mechanisms (Friedkin, 1998). 4.4 Network Selection and Social Influence of Drug Behaviors Delinquent behaviors are apparently encouraged by one’s friends, but the spread of these behaviors among friends may also occur as a simple consequence of a greater number of opportunities to engage in the behavior as a result of the social activities (Akers, 1992; Matsueda & Anderson, 1998). Selection of friends who engage in similar delinquents activities may be a strategy to reduce cognitive dissonance (i.e., “if my friends do it, it’s okay for me to do it”) but also may result from a straightforward relationships where shared subcultures that lead to friendships simply coincides with shared delinquent activities. Causality can be determined by sophisticated examination of these relationships. Social network and behavioral research examining selection and influence mechanisms in the context of drug using behaviors suggest that these tendencies do not result from shared social activities or subcultures, but are instead the result of complex social dynamics (Reed & Rountree; 33 Thornberry et al., 1994). A number of longitudinal studies find support for both selection and influence in adolescent drug use (Ennett & Bauman, 1994; Fisher & Bauman, 1988; Hall & Valente, 2007; Kandel, 1978). A person with friends who smoke and use alcohol are more likely to initiate those same behaviors (Ennett & Bauman, 1994; Hall & Valente, 2007; Sieving et al., 2000). Homophilic smoking and/or alcohol use are also found to be predictive of subsequent friendship selections (Fisher & Bauman, 1988; Wills & Cleary, 1999). Other studies find that both selection and influence mechanisms occur concurrently (Ennett & Bauman, 1994; Hall & Valente, 2007). Statistical Social Network Analysis. Because of the complexity involved in the investigation of relationship formations and dissolutions concurrently with behavioral adoption and cessation, the task of teasing out mechanisms of selection versus influence is empirically burdensome. As discussed in previous sections, one of the major advances in network science is recently developed software that can model the co-evolution of behaviors and social ties. Statistical tests are performed to separate effects from social influence (assimilation of friends’ behaviors) and effects from selection (the behavior leads to selection of friends). The number of studies that have compared social influence with social selection processes while statistically accounting for social dependencies are few because the actor-oriented network modeling approach is relatively new. The following section outlines some of the studies conducted using these network-behavioral co- evolution models to explore selection and influence around drug use or delinquent behaviors in adolescents. Snijders and Baerveldt (2003) did a study using the Dutch Social Behavioral study, a two-wave survey in classrooms from 1994 –5 of 1528 students ages 13 to 18 years. Their study focused on modeling value differences between creating and breaking friendship ties. They found that similarity of the level of delinquency leads to the formation of friendship ties, but interestingly similarities also caused greater likeliness of friendship dissolution. Burk et al. (2007) looked at minor delinquency 34 (e.g., vandalism, petty theft) in Swedish 6 th graders, and found that while delinquency was quite low, in general, both selection and influence were significant factors in friendship and behavioral changes. Selection for friends with a similar degree of delinquency was equally strong for reciprocated friendships as for unidirectional ties, whereas influence was significantly stronger for reciprocated friendships. In another study, three waves of data were collected from 137 adolescents, in Glassgow, Scotland in the Teenage Friends and Lifestyle Study. Friendship data, substance use, and other lifestyle variables were collected from 137 students, from 1995-1997 when they were 13 years old. A number of statistical network analysis studies have been published with these data. In one analysis, Snijders (2008) found that people with either low or high alcohol use had tendencies towards others with similar levels of alcohol use, while those who reported medium values of alcohol use did not select friendships based on alcohol consumption. In this study, homophily based on the number of nominations sent, number of nominations received (popularity) and sex were also included in the models, indicating that the significant selection effects were based on alcohol use alone. They also found support for behavioral influence (the tendency to change towards the behavior of one’s friends). In another analysis using these same data, Steglich et al. (2007) found different results when they included smoking in the analysis. That is, when alcohol and smoking are modeled together, selection based on behavioral similarities was no longer significant for either smoking or alcohol. However, influence effects were significant for smoking, and even more so for alcohol. Social selection was found to be driven by homophily of students reported “pocket money,” their gender, and if they shared a romantic relation. Because these other attributes were not all included in the original analysis by Snijders (2008), smoking selection effects that he reported appear to be accounted for by romance and money homophily. 35 Another analysis of the Scottish data focused on sport activities, where models were run separately for sports and tobacco, sports and alcohol, and sports and marijuana (Pearson, Steglich, & Snijders, 2006). Again, based on the different set of variables included in the model, findings differed. When sport activities were included in the models, friend selections were found to be significantly driven by alcohol use, and people also tended to assimilate to their friends level of alcohol use. That is to say, individuals adopted their friends’ alcohol using behaviors, and selection of individuals also selected friends based on similarities in alcohol use. Marijuana use was dominated by behavioral influence effects, and tobacco showed significant selection effects. In addition, interaction effects were found between marijuana and tobacco, where marijuana users were found to be more likely to be cigarette smokers. Summary of Drug Selection and Influence Studies. Overall, there is support for both selection and influence effects for substance use and more generally, delinquent behaviors. A handful of studies have used a structural network approach, and only a few have considered these mechanisms as co-evolving processes in a statistical network framework. A summary of findings shows support for delinquency homophily, but in one study, these similarities also lead to greater chance of breaking off the friendship. This may simply be an artifact that results as adolescents form and dissolve friendships: if there are more friendships with “like” others, there will be more opportunities for these relationships to end. In other studies, friends are found to have a greater influence when the friendship is reciprocated, while selection patterns were found regardless of reciprocity. Non-reciprocated ties from smokers to non-smokers appear to be protective of future smoking, where perhaps the deliberate non-reciprocation represents a rejection of influence by the specific person or an aversion to the behavior (Hall & Valente, 2007). Co- evolution studies also showed support for both selection and influence mechanisms. More complex patterns are found when multiple substances are included in models, probably due to correlations of the behaviors. 36 When tobacco use is modeled along with alcohol, for example, selection patterns based on homophily become insignificant for both smoking and alcohol, while influence effects remain significant for both substances. While these studies vary in their findings, what is clear is that both selection and influence patterns are at play, and their effects are most likely interconnected. Explanations in the literature commonly couple the mechanisms, suggesting for example, that drug use is driven by a reinforcing environment of peer networks, which in turn, leads to a greater selection of delinquent friends (Kirke, 2004; Thornberry et al., 1994; Urberg, Degirmencioglu & Tolson, 1998). In Kandel’s 1978 seminal paper, she suggests that “[h]omophily between friends...results from two complementary processes whereby adolescents who share certain prior attributes in common tend to associate with each other and then to influence each other as the result of the continued association” (Kandel, 1978). 4.5 Contagion of Implicit Cognitions There is theoretical and empirical evidence suggesting that unconscious cognitive biases that accompany (and in many cases precede the conscious attitudes towards) behaviors can also be conveyed between individuals. Like homophilic preferences around behaviors, the studies presented here propose that associative memory structures are also susceptible to social selection and influence processes. It is plausible that cognitions are spread like behaviors through social exchanges. Some cognitive and behavioral research provides evidence to suggest that implicit cognitions are communicated through symbolic or non-explicit channels. While there are no studies that specifically assess the spread of associative memory features through social networks, a number of studies from a wide range of research domains support the notion that unconscious exchange of cognitions occurs through social encounters, and further, that memory structures are altered by these encounters and tend towards cognitive assimilation. 37 Non-Verbal Communication. Research in non-verbal communication studies show that emotional (e.g., empathy) and social status evaluations (e.g., centrality) are communicated through symbolic language with little or no conscious awareness by the communicator or the recipient. In fact, approximately one-third of the variance in social outcomes (e.g., marriage success, jury decisions, speed dating attraction) can be explained by quantifying non-verbal social interactions such as body movement, facial expressions, and tone of voice (Ambady & Rosenthal, 1992, 1993; Pentland, 2004; Philpott, 1983). Smiling, close physical proximity, increased eye gaze, direct body orientation have all been correlated to perceptions of affection, involvement, non-dominance, intimacy and attraction (for review, see Walther, 1996; Walther, Loh & Granka, 2005). For example, remarkable findings are presented in the dissertation study by (Choudhury, 2004), where he found that involuntary expressions and unconscious conversing strategies (vocal signaling and dialogue turn-taking dynamics) observed in controlled settings were associated with centrality in a person’s social network. He suggests that social signals are part of continual, ‘implicit negotiation’ between members of a social network. Associative Memory Studies. Associative memory studies also suggest that implicit information can be exchanged. In particular, groups of individuals who share a common culture often have common associative memory linkages around relevant concepts, raising the possibility that associative information is spread through cultural symbols. In addition, people are more likely to have common associative structures if they interact with greater frequency (Ward & Reingen, 1990) and if their relationship functions at multiple levels (i.e., “multiplexity,” for example, friends & colleagues; Reingen, Foster, Brown & Seidman, 1984). Szalay has contributed a number of word association studies with findings that people from the same culture are more likely to have similar cognitive associates around culturally salient concepts. In one study, Szalay and Maday (1983) found that Blacks and Whites elicited similar associative 38 responses to the cue “ancestor,” whereas responses to cues “society” and “problems” were more similar between Hispanics and Blacks. Findings are inconclusive, but might indicate shared cultural meanings lead to common associative structures. Similarly, (Manguno-Mire & Geer, 1998) found that gender and sexual orientation lead to differences in the number of associative links within and between sexual and emotional concept clusters. In another study, Johnson and Boster (1990) found that experts in animal identification had associative links around specific fish based on form or function, while novices linked information about fish based on morphology. 4.6 Cognitive Influence Through Associative Memory One of the interesting benefits of congruent associative memory structures is that it enables individuals to have more effective communication and in turn, may facilitate influence among cognitively similar pairs. A small number of studies have investigated relationships between people and their effectiveness of communicating associated information. In semantic differential studies, Runkel (1956) showed that the extent to which college students agree on domain specific associations with their instructors, the more likely they were to have higher exam scores. In another study, participants were asked to convey single words (pertaining to specified domains) to a partner, by giving one-word clues with no visual contact (Padget & Wollosin, 1980). Pairs of individuals who displayed similar domain-specific cognitive associations had more successful trials in eliciting domain-relevant words than dissimilar pairs. Group-task studies also show that congruent cognitive structures lead to more effective & efficient communication, collaboration, problem-solving and coordination of activities (Levesque et al. 2001; Palazzolo, 2005; Palazzolo et al., 2006). Research in transactive memory, for example, show that the extent to which the organization of memories converge, coordination of activities becomes easier and more efficient (Levesque et al., 2001). Group performance is found to improve when individuals are able to establish mutual conceptions of group-level goals and organization of 39 the resources needed to achieve those goals (see Palazzolo, 2005, for discussion). Social information processing theory (Salanck & Pfeffer, 1978) usually discussed within the context of businesses, asserts that people use information from others to learn what aspects of the environment to focus on and how to interpret events, in the context of achieving shared goals (as outlined by Umphress, Labianca, Brass, Kass & Scholten, 2003). More recently, Bandura (1989) extended social cognitive theory to collective agency, which suggests that collective power is the shared belief in the group’s ability to produce desired results. 4.7 Associative Memory Drives Social Selection Implicit processes may also influence cognitive homophilic preferences through their role in allocating attentional resources at an unconscious level. Associative memory linkages bias perceptions, which both draws attention to specific stimuli and increases susceptibility to other stimuli (Ames et al., 2006; Franken et al., 2003). Attentional bias theories are suggested in this dissertation as an individualistic cognitive strategy that potentially drives the selection of like- minded others. Attentional bias’ studies show that despite attempts to concentrate on assigned tasks, unconscious attention towards personally relevant stimuli correlate with salient memories. In a social context, representations of friends and relations become associated with behaviors, attributes, and attitudes, such that friends can act as associative cues for specific behaviors, enabling particular memories to be more easily recalled in the company of specific individuals. Potentially, social exchanges with cognitively congruent others will activate more salient implicit pathways as a result of their shared implicit evaluations. Positive associations around drug use, for example, may lead to an unconscious attentional bias towards people who trigger drug associations in memory. Homophily is suggested to be a reward-seeking strategy, whereby people seek out others like themselves to reduce cognitive dissonance. Exchanges with others who share attitudes and behaviors may stimulate anticipation of positive appraisals. Biased memory strategies might play a 40 role in the activation of reward systems, where exchanges with others who share implicit association in memory may trigger salient positive associations and stimulate anticipation of positive appraisals (Field et al., 2004; Franken et al., 2000; Jones et al., 2002). It may be beneficial to the individual to assign greater cognitive resources to like-minded others, increasing the chance of friendship ties. Therefore, tendencies towards like-minded others may occur, in part, through non-deliberate implicit cognition strategies. 4.8 Summary of Cognitive Contagion Research These branches of research show that selection and influence mechanisms may be occurring below the threshold of consciousness. Interactions of implicit cognitive frameworks among friends might be important in how friends enable or discourage behaviors and why certain individuals seem to be susceptible to social influences more so than others. First, evidence suggests that there is, in fact, effective communication of implicit attitudes through unconscious non-verbal social exchanges, such as symbolic facial gestures and body language. Second, individuals with similar memory structures are effective at conveying information, and thus, very likely have a greater behavioral influence on one another. Third, implicit processes may play a role in cognitive homophilic preferences through their role in allocating attentional resources, such that people are drawn to others with similar cognitive associations. The collection of research suggests that behavioral changes associated with social relationships might function, in part, through implicit exchanges. The investigation of whether there are homophilic tendencies towards common implicit cognitions through either selection or influence is well founded and may be crucial in promoting behavioral change. 41 CHAPTER 5 THEORECTICAL MODELS & HYPOTHESES Creating a new theory is not like destroying an old barn and erecting a skyscraper in its place. It is rather like climbing a mountain, gaining new and wider views, discovering unexpected connections between our starting point and its rich environment. -- Albert Einstein 5.1 Study Plan Three specific aims will be addressed in three separate studies: (1) study 1 will address the first specific aim using regression models to test whether social influence is occurring as a result of exposure to friends’ cognitions and behaviors, (2) study 2 will focus on the second specific aim using dynamic network models to test whether selection of friends is explained by tendencies towards cognitively and behaviorally similar classmates, and (3) study 3 will address the third specific aim using dynamic network behavioral co-evolution modeling to test selection and influence simultaneously. Intervention condition will be accounted for in each of the studies, and models will be run separately for marijuana and alcohol 5.2 Analytic Plan Two different methodological approaches are employed, revealing different aspects of these processes. Analytic strategies include multiple regression analysis but the dissertation is focused on statistical network analysis. Multiple regression analyses will be used to assess social influence at baseline to explain the variances of behaviors at post-test (study 1). Egocentric friendship evaluations are used to create alter effects variables, where alters are defined by the student’s identification of friends without corroboration from those classmates. Cognitions and behaviors of friends are aggregated into scores that represent the amount of exposure the student has to drug-related 42 cognitions and behaviors through these friendships. Alcohol and marijuana use at post-test will be regressed on exposure to corresponding cognitions and behaviors at baseline. Statistical network analysis will be employed to estimate the extent to changes in friendships from baseline to post-test are due to social selection effects (study 2), and the relative effects of selection and behavioral influence when modeled together (study 3). Dyadic and triadic structural dependencies are controlled for in the models as endogenous effects, which means that they represent the effects of the network structure on the network structure. Cognitions, behaviors, and other demographic variables (age, gender, and ethnicity) are included in the models as exogenous effects, which means that they are variables that are external to the network. Selection will be assessed with parameters estimating friendship formation based on homophily, while influence will be assessed with parameters estimating assimilation of friends’ behaviors. 5.3 Hypothesis Testing Specific Aim 1. The first specific aim will be addressed in study 1, using regression analyses to test whether friends’ alcohol and marijuana cognitions & behaviors influences one’s own cognitions behaviors over time (see Figure 5.1a,b). Social influence is tested using regression methods, where cognitions and behaviors are regressed on friends’ cognitions and behaviors. Longitudinal data is expected to show that high drug cognitive and behavioral exposures at baseline predict an increase in an individuals own drug use after completing the 12-week drug intervention (controlling for one’s own cognitive biases and intervention condition). In addition, this study will model cognitions at follow-up as predicted by friends’ cognitions and behaviors. 43 Figure 5.1a,b Influence of Friends Marijuana & Alcohol Activities on (a) Behaviors, and (b) Cognitions: Longitudinal Regression Models of Cognition and Behavior Exposure Figure 5a. Logistic Regression model of friends’ drug activities on alcohol & marijuana behaviors. Figure 5b. Logistic Regression model of friends’ drug activities on alcohol & marijuana cognitions. Age Gender Ethnicity SES In-degree Baseline alcohol/ marijuana use Exposure to implicit cognitions (baseline, follow-up) Exposure to drug behaviors (baseline, follow-up) Alcohol & Marijuana Behaviors (follow-up) Age Gender Ethnicity SES In-degree Baseline alcohol/ marijuana use Exposure to implicit cognitions (baseline, follow-up) Exposure to drug behaviors (baseline, follow-up) Alcohol & Marijuana Cognitions (follow-up) 44 The following hypotheses are tested with separate models for alcohol and marijuana: Study 1 Hypotheses Behavioral Influence Alter-Ego Influence s of Drug-Activities H1: Exposure to alcohol and marijuana behaviors through friends’ is predictive of alcohol and marijuana behaviors longitudinally, controlling for drug use at baseline, drug cognitions, intervention condition, gender, ethnicity, and age. H2: Exposure to alcohol and marijuana cognitions through friends’ is predictive of alcohol and marijuana cognitions longitudinally, controlling for drug cognitions at baseline, drug behaviors, intervention condition, gender, ethnicity, and age. Alter-Ego Cognition-Behavior Interactions H3: Exposure to alcohol and marijuana cognitions through friends’ is predictive of alcohol and marijuana behaviors longitudinally, controlling for drug use at baseline, cognitions, intervention condition, gender, ethnicity, and age. H4: Exposure to alcohol and marijuana behaviors through friends’ is predictive of alcohol and marijuana cognitions longitudinally, controlling for drug cognitions at baseline, drug behaviors, intervention condition, gender, ethnicity, and age. Specific Aim 2: To explore the second specific aim, dynamic network analyses will be conducted in Study 2, where pairs and higher-order friendship structures predict changes in alcohol and marijuana use (see Figure 5.2). Analyses assess whether individuals select friends based on (1) their own behaviors (ego attributes), (2) the behaviors of those selected (alter attributes), and/or (3) the similarities between the ego and the alter. Ego, alter and similarity 45 parameters determine how relations change as a function of behaviors and attributes, while controlling for levels of structural dependencies. Theories regarding selection of ‘like-others’ guide hypotheses that students will increase connections with others who share their beliefs and activities. In this study, individuals are hypothesized to select friends based on common cognitive and behavioral frameworks. Probabilistic models will assess how likely it is that the observed network configuration is guided by the proposed drug-related homophilic selection mechanisms more so than they would by chance alone. Figure 5.2 Coevolution of Network Selection & Behavioral Influence of Alcohol & Marijuana Cognitions & Behaviors The following hypotheses will be tested for alcohol and marijuana separately and each of the three program condition groups will be run as separate models (control, standard, network). Similar Cognitions Similar Behaviors Age similarity Ethnicity similarity SES similarity # Best friends in school Endogenous Controls Density (outdegree) Dyadic dependence (Reciprocity) Triadic dependence (Transitivity, Balance) Exogenous Friendship Behavioral Influence via assimilation Social Selection via homophily Exogenous Mechanisms 46 Study 2 Hypotheses Social Selection Effects H5: Drug Use Selection (Homophily) Effect- Students are more likely to become friends with other students who have the same drug using status than would be expected by chance, controlling for dyadic and triadic dependencies, and selection effects of age, gender, and ethnicity. H6: Drug Cognition Selection (Homophily) Effect - Students are more likely to be friends with other students with the same level of drug cognitions than would be expected by chance, controlling for dyadic and triadic dependencies, and selection effects of age, gender, and ethnicity. Specific Aim 3. The third specific aim will be assessed from a dynamic network perspective in Study 3 with models that simultaneously account for changes in network links due to behaviors and behavioral changes due to friendships (see Figure 5.3). Similarity parameters test selection mechanisms through the assessment of friendship changes relative to cognitive & behavioral changes, while influence mechanisms are detected with parameters that measure assimilation of cognitions & behaviors towards friend’s cognitions & behaviors. 47 Study 3 Hypotheses Network Selection Coevolving Effects H7: Drug Use Selection (Homophily) Effect- Students are more likely to become friends with other students who have the same drug using status than would be expected by chance, controlling for dyadic and triadic dependencies, and selection effects of age, gender, and ethnicity, and drug-related influence effects. H8: Drug Cognition Selection (Homophily) Effect - Students are more likely to be friends with other students with the same level of drug cognitions than would be expected by chance, controlling for dyadic and triadic dependencies, and selection effects of age, gender, and ethnicity, and drug-related influence effects. Behavioral Influence Coevolving Effects H9: Drug Use Influence (Assimilation) Effect - Individual’s own substance use will become more similar to their friends’ use, while accounting for dyadic and triadic dependencies, and selection effects of drug cognitions, drug behaviors, age, gender, and ethnicity. H10: Drug Cognition Influence (Assimilation) Effect - Individual’s cognitions will become more similar to their friends’ use, while accounting for dyadic and triadic dependencies, and selection effects of drug cognitions, drug behaviors, age, gender, and ethnicity. Table 5.1 shows a summary of the network modeling hypotheses and implication if they were found to hold true in these data. 48 Table 5.1 Summary of SIENA Network Modeling Hypotheses and Their Implications Study 2: Network Selection of Alcohol/ Marijuana Cognitions & Behaviors Hypotheses Implications H5: Drug-Use Similarity Effect, students are more likely to become friends with other students with similar drug behavior status than would be expected by chance. Support for this parameter would indicate that friendship structures have a tendency to occur as a result of shared drug behaviors, suggesting that friendships are selected based, in part, on drug activities. H6: Drug-Cognition Similarity Effect, students are more likely to be friends with other students with the same level of drug cognitions than would occur by chance alone. Support for this parameter would indicate that friendship structures have a tendency to occur as a result of shared drug cognitions, suggesting that friendships are selected based, in part, on unconscious drug associations. Study 3: Co-Evolving Influence & Selection of Alcohol/ Marijuana Cognitions & Behaviors Hypotheses Implications H7: Selection: Drug-Use Similarity Effect, students are more likely to become friends with other students with the same drug behavior status (homophily effect of drug use) than would be expected by chance, while accounting for influence effects. Support for this parameter indicates that friendship structures have a tendency to occur as a result of shared drug behaviors even after controlling for influence effects, suggesting that friendships are selected based, in part, on drug activities. H8: Selection: Drug-Cognition Similarity Effect, students are more likely to be friends with other students with the same drug cognition status than would be expected by chance, while accounting for influence effects Support for this parameter indicates that friendship structures have a tendency to occur as a result of shared drug cognitions even after controlling for influence effects, suggesting that friendships are selected based, in part, on unconscious drug associations. H9: Influence: Drug Use Assimilation Effect, students own substance use will become more similar to their friends’ average use, while accounting for selection effects. Support for this parameter indicates that individuals who are in dyadic and triadic friendship structures become similar in their drug use over time even after controlling for selection effects, suggesting drug behaviors of friends persuades one’s own behaviors H10: Influence: Drug Cognition Assimilation Effect, student’s own cognitions will become more similar to their friends’ cognitions, while accounting for selection effects. Support for this parameter indicates that individuals who are in dyadic and triadic friendship structures become similar in their cognitions over time even after controlling for selection effects, suggesting that friends’ drug cognitions persuades one’s own cognitions 49 5.4 Comparison of Network and Regression Approaches Dissertation studies to be presented are summarized in Table 5.2. The main features of each study are listed with attention to the differences across statistical methods. In pictorial depictions, circles represent individuals, the lines represent their friendships, and the shade of the circle represents the behavior where a dark circle indicates that the individual engages in the behavior while no color indicates that they abstain from the behavior. Diagrams in gray shading represent intermediate stages between observed time points, which are estimated in SIENA modeling. The descriptions of each study underscore differences in the analytic perspectives. In study one, friends’ activities are correlated to the ego’s behavior at post-test (pictorially represented as changes in the shading of the circle for ego-networks). Study two uses behaviors and relations at baseline to model the tendencies for friendship ties to be formed or dissolved at post-test accounting for all network structural changes (depicted as added or deleted links within the entire network). Study three uses behaviors and relations to model the tendencies for (1) friendship ties to be formed or dissolved (depicted as added or deleted relational links) and, (2) behaviors to be initiated or discontinued (depicted as changes in the shading of the circle) while accounting for all network structural changes. Network modeling strategies assume that network and behavioral dynamics are constant, and that in between the observed time points, there are continuous micro-steps of behavioral and network changes. 50 Table 5.2 Analytic Strategies to Measure Network Effects in Dissertation Studies Study Time frame Analysis type Construct Rate of Change Description Network and Behavioral Changes Detected by Analysis Study 1: Social Influence Models of Alcohol & Marijuana: . Longitudinal Regression Influence Static Odds ratios indicates the influence of friends’ drug activities on ego’s drug activities. Analyses will determine the extent to which individuals behaviors change as a result of classroom friendships. Intervention will be controlled for in the models. Study 2: Social Selection Network Models of Alcohol & Marijuana Longitudinal Statistical Network Analysis Selection Dynamic Selection parameters indicate whether friends are selected on the basis of alcohol and marijuana cognitions and behaviors. Analyses assess the tendency for drug homophily, while controlling for structural dependencies. Study 3: Dynamic Network Selection & Behavioral Influence Models of Alcohol & Marijuana in Friendship Networks Longitudinal Statistical Network Analysis Selection & Influence Dynamic Coevolution Selection and influence parameters indicate the relative strength of network and behavioral changes. Analyses assess selection based on drug homophily and influence of friends, while controlling for other structural dependencies. Baseline Post-test Baseline Post-test Post-test Baseline 51 CHAPTER 6 STUDY METHODS 6.1 Study Methods Participants. Data were collected during a school based drug education program as part of the University of Southern California's Transdisciplinary Drug Abuse Prevention Research Center (TPRC). The participants attended Continuation high schools, which are alternative school setting that focus on technical and life skills. These students are characterized as students who had difficulties in regular public high schools, and tend to be high-risk for drug and criminal behaviors. Schools were randomly selected from a list of all Continuation high schools in Los Angeles County to be recruited for the study. Some 1493 students were invited to participate with 980 (65.6%) providing valid consent forms. Of the 980, 29 students were not interviewed at baseline because they were absent on multiple visits. Data were collected from 938 students in 75 classrooms from 14 schools. Of the 938 eligible students, 894 (80%) completed the network items survey, and of those 572 (63.9%) of those individuals provided network and cognition data on the post-test survey. Of the final 572 participants included in the analysis, 40 are missing baseline sociometric nominations and 85 are missing data at post-test. Therefore, not all participants have data for both baseline and follow-up exposures, but are included in the analysis to contribute to estimates for patterns of alter/ego influences. The final data set included 572 participants: 160 (28%) from the control condition, 191 (33%) in the standard intervention condition, and 221 (39%) in the experimental intervention condition. Inclusion criteria. Missing data is particularly restrictive when using a statistical network analytic approach because the unit of measurement is the network (in this case, the classroom). 52 Therefore, incomplete classroom data correspond to only partial observation of each measurement unit. Statistical network analytic software is still relatively young and research on permissible proportions of missing data are not yet available. The rule of thumbs discussed when using SIENA software are that networks must contain at least 20 actors, and have no more than 20% of the network missing. However, these requirements are extremely constraining for data used in these analyses, with only 5 classrooms fitting these criteria. Preliminary analyses of these five classes showed that the models were reliable - in that they consistently “converged” and standard errors for the parameter estimates were within reasonable ranges. However, the lack of significant results suggested there was a power issue, thus the expansion of the analysis to 21 classrooms. The final classrooms were included in studies 2 and 3 if they met the following criteria: (1) classes had to have 10 or more students with baseline and post-test data, and (2) classes had to have at least 60% non- missing from entire roster list. It should be noted that rosters often include students who are no longer in the class, and therefore, considering the network to be made up of the entire roster list gives conservative estimates of the percent non-missing. Study Procedures. Participants were asked to fill out paper surveys in the classroom during their normal health class period. The general purpose of the study was described to students and their parents in informed consent letters. The intervention was an evidence-based substance use prevention program called ‘Towards No Drug Abuse’ (TND). Previous trials using this curriculum have been found to be effective in reducing tobacco, alcohol, marijuana, and hard drug use at a one-year follow-up in high-risk youth attending alternative high schools (Sussman, Dent & Stacy, 2002; Sussman, Earleywine, Wills, Cody, Biglan, Dent & Newcomb, 2004). The curriculum included 12 sessions of 1-hour classroom lessons that incorporate role-playing along with modules aimed at reducing misconceptions around drug use (Sussman, 1995; Valente, Ritt- Olson, Stacy, Unger, Okamoto & Sussman, 2007). Subjects completed the baseline survey in the 53 spring before receiving the program, and they were asked to complete a post-test survey upon completion of the program. Friendship networks, drug cognitions and drug behaviors were assessed at baseline and post-test. A classroom randomized controlled trial design was implemented to compare control classes with those receiving either one of two forms of the TND curriculum. The three conditions were: (1) standard intervention, where students received TND as it was designed in a typical classroom format delivered primarily by health educators, (2) TND experimental condition, a peer-led interactive version of TND, where students were encouraged to work in peer groups lead primarily by nominated peer leaders, and (3) control condition, where students received no intervention at all. In a recent study, Valente et al. (2007) showed that the experimental condition (termed network condition) accelerates peer influence of drug use, both in the negative and positive directions. There were three forms of the cognition task. The study’s assessment design at baseline was randomized by type of cognition assessment. The three forms of the implicit association task were: (1) affective outcomes, which consist of positive drug affects, such as “feeling good” and “having fun”, (2) locations, consisting of high-risk drug using locations, such as “in my friends bedroom” or “in my yard”; and (3) compound cues, where the two types of cues occur together, such as “having fun, in my bedroom” or “feeling good, in my yard”. Students completed memory association items first while scales containing explicit content about drug use appear later in the questionnaire, so as to not prime the students with drug related content. 6.2 Survey Instruments Demographics. Demographic covariates were collected directly from participants. Covariates include gender, age, grade, and ethnicity. Past research has shown peer influences on drug use may 54 vary by gender and ethnicity (Griesler & Kandel, 1998). Gender was coded ‘1’ for male, and ‘2’ for female. Ethnicity. Students were asked to select their ethnic background. The response options were listed as follows: (1) Asian or Asian American, including Chinese, Japanese, and others; (2) Black or African American; (3) Hispanic or Latino, including Mexican American, Central American, and others; (4) White, Caucasian, Anglo, European American; not Hispanic; (5) American Indian/Native American; (6) Mixed; Parents are from two or more different groups; (7) Other (write in): _______. Because the majority of the participants were Latino, all ethnicities besides Latino were aggregated. Ethnicity was coded as ‘1’ for Latino and ‘2’ for all other ethnicities. Socio-economic Status. The SES index consists of the mean of parents’ occupational and educational levels. Questions are “What is your mother/father's main job?” and “What is the highest grade completed by your mother/father? “ The highest level of occupational and educational level completed by either the students’ mother or father, which each have a maximum value of ‘1,’ are added together and divided by 2. The resulting scale spans from 0 to 1, with a value of ‘1’ indicating that the parents completed a graduate program and were employed as a professional requiring a license or degree, such as a doctor, professor, or lawyer. Age. Age is calculated from students’ reports of their date of birth. In the regression analysis, age is a continuous variable, while it is dichotomized as above or equal to 16 and below 16 years old for input into the network modeling software. # of Best Friends in School. Number of friends in school is a proportion of a person’s best friends who attend the same high school. Students are asked to list up to five of their five best friends, and then whether the friend attends their school. The number of best friends who attend the same school are added up and divided by the number of best friends the participant listed. 55 This variable represents the student’s integration into the school and is a proxy for the extent to which the friendships within the class are important for the student. Drug Behaviors. Self-reports of past 30 days use for alcohol and marijuana were assessed with an 11-item rating scale. Participants were asked how many times they have used various drugs in their lifetime and in the last month. Participants responded to a list of 11 specific drugs, two of which are alcohol and marijuana. Frequency of use is coded from 0-10. Next to the target drug category, response choices included "0" for never used, “1” as 1-10 times, “2” as 11-20 times, and other response options increased in intervals of 10, with the last category “10” as 91-100+ times. The reliability and predictive validity of these items have been previously established, ranging from 0.8 to 0.85 (Graham, Flay, Johnson, Hansen, Grossman & Sobel, 1984). Behavior scores have non- normal distributions and were therefore converted into dichotomous scales. Alcohol and marijuana scores were recoded such that ‘0’ indicates never having used in the last 30 days, and ‘1’ represents using 1 or more times in the last month. Drug Cognitions. Word association scores are generated from cue-behavior tasks, which ask participants to write the first behavior or action that come to mind given a phrase stimulus. Cues were selected as potential alcohol and marijuana associates, and have previously been found to activate drug-related responses, particularly in individuals with who are at high risk for current or future substance use. Each cue is listed two times. An example of the survey instructions is shown below. 56 Figure 6.1 Example of Word Association Task Instructions As mentioned above, there were two variations of the target cue words: affective outcomes and locational cues: (1) Affective outcome cues consist of positive anticipated consequences of alcohol or marijuana use (e.g., “feeling relaxed”; “having fun”), intermixed with 15 phrases not likely to be related to substance use (e.g., “being tolerant”; “pleasing family”). Behavioral outcomes used in this measure include those found previously to be self-generated drug associates at moderate to high frequency among college students (e.g., Stacy, 1994), high school students (Stacy, Galaif, Sussman & Dent, 1996) and a sample of 112 drunk-driving offenders (unpublished data, Stacy). (2) Location tasks consist of situations determined to be high-risk for substance use by high school youth (e.g., “my bedroom”; “parking lot”). The locations used in survey were taken from a prior study where students were asked to indicate the location where they were a few minutes before the last time they used a drug (or drove after using a drug; see Sussman et al., 1999). The drug- related cues are mixed with 15 low-risk locations (e.g., “library”, “living room”). At baseline, three forms of the survey were randomly distributed to participants. Form one contained 11 affective outcome target cues, form two contained 10 location target cues, and form three contained 7 compounded target cues, which consisted of a combination cues taken form the Write down the VERY FIRST behavior or action the words make you think of. 1. In my bedroom: eating ice cream (first behavior or action that comes to mind) In my bedroom: drinking beer (next behavior or action that comes to mind) 57 outcomes and locations tasks (e.g., ‘at my friends house, feeling good’). At post-test, all students received the same form, consisting of 11 outcome cues. Each student must have responded to at least two associative memory cue items to be counted as a completed task. Research suggests that given that the cues target drug-related associations, a cue that elicits no response, can be considered as a non-drug response. Therefore, students who only gave one response for the entire set of cues were considered to have missing cognition data. Presented below are the cues for each form that target alcohol and marijuana behaviors. Cues were only included in the analysis if they were successful in eliciting the targeted drug- related responses from at least two individuals. For example, only ten of the eleven alcohol cues were used because the cue “feeling peaceful” generated zero alcohol related responses and thus was eliminated. Similarly, “feeling attentive,” and “feeling peaceful” were excluded from the outcome alcohol scale, and “on the sidewalk” was eliminated from the location alcohol scale. See Table 6.1 for a list of cues included in alcohol and marijuana associative memory scores for the three baseline, and one post-test word association tasks. 58 Table 6.1 Alcohol and Marijuana Cues for Associative Memory Tasks Coded Memory Associations. Responses to each cue word were entered verbatim into a computer based coding system (Ames et al., 2005). Cognition scores have non-normal distributions and were therefore converted into dichotomous scales. Associative responses were coded as ‘1’ if the responses were drug-related, and ‘0’ if it was not drug-related. Responses were categorized into the following drug categories: marijuana, alcohol, tobacco, speed, hallucinogens, crack/cocaine, heroin, and ecstasy. The same cues are used for both alcohol and marijuana because many of the self-generated outcomes of these behaviors are the same across BASELINE TASKS POST-TEST TASK Outcome cues Location Cues Compound Cues Outcome cues Alcohol (#cues = 9) Alcohol (#cues = 9) Alcohol (#cues = 7) Alcohol (#cues = 10) feeling sociable in my yard in my yard, feeling relaxed feeling attentive feeling dreamy in my bathroom in my yard, feeling good feeling sociable feeling good in a field friends house, feeling relaxed feeling dreamy feeling happy in a parking lot friends house, feeling good feeling good feeling relaxed friends house my bedroom, good feeling happy having fun in the park friends house, fun feeling relaxed laughing in my bedroom my bedroom, relaxed having fun euphoria my friends bedroom laughing forgetting problems living room euphoria forgetting problems Marijuana (#cues =11) Marijuana (#cues =10) Marijuana (#cues =7) Marijuana (#cues =11) feeling attentive in my yard in my yard, feeling relaxed feeling attentive feeling sociable in my bathroom in my yard, feeling good feeling sociable feeling peaceful on the sidewalk friends house, feeling relaxed feeling peaceful feeling dreamy in a field friends house, feeling good feeling dreamy feeling good in a parking lot friends house, fun feeling good feeling happy friends house my bedroom, fun feeling happy feeling relaxed in the park my bedroom, relaxed feeling relaxed having fun in my bedroom having fun laughing my friends bedroom laughing euphoria living room euphoria forgetting problems forgetting problems 59 these two drugs. Therefore, the number of cues included in alcohol and marijuana analyses differed only by the aforementioned criterion. Categories were selected if the response made direct reference to the drug (weed, beer), or if it described a mode of taking the drug (smoke, drink, joint), or if the response described effects of a drug (buzzed, blazed, high). After all the responses were coded, the student’s score is aggregated for a raw continuous cognition score. Scores were then dichotomized, where students with at least one alcohol or marijuana-coded response are coded as ‘1,’ and those with no alcohol or marijuana-coded responses are coded as ‘0.’ Scores are generated separately for alcohol and marijuana. Two judges coded the open- ended responses. If there was a discrepancy between the two judges, a third judge made the tie- breaking decision. Coder inter-rate reliabilities had an average Cronbach’s alpha of 0.82. Combining IC Scales /Distributions. Within classrooms, students received one of either of three survey variations. The three scales will be considered as representing a single construct, so that individual scores can be compared across classrooms. Based on prior word association studies, drug-related cues will trigger a chain of associated concepts from memory. If the cues are effective drug-targets as these have shown to be, they will very likely elicit a drug-related response if the concepts are salient in memory. Scales are dichotomized, so that the relative strength of each scale is minimized. Thus, dichotomized responses to the three versions of the survey include drug-related cues that are assumed to equally capture drug-related cognitions. Distributions of drug cue-response data are analyzed to assess their equivalence. Cognition Scales. Means and standard deviations for responses generated by cues are reported for each of the three outcome-behavior tasks in Table 6.2. 60 Table 6.2 Means of Alcohol & Marijuana Implicit Cognition Responses Across Cognitions Tasks Distributions of drug cue-response data reveal the means and variances of the three word association forms significantly differ from one another. Mean alcohol cognition scores are 0.25, 0.11, and 0.27, for forms 1, 2, and 3, respectively. These differences in means are significant (F=12.87, p<0.001) and their respective variances are also significantly different (chi 2 = 37.68; p<0.001). Mean marijuana cognition scores are 0.33, 0.22, and 0.31, for forms 1, 2, and 3, respectively. These differences in means are significant (F=5.10, p<0.006) and differences in variances between forms are just below statistical significance (chi 2 = 5.50; p<0.06). Compound cues (form 3) elicited the highest number of alcohol associations and affective cues (form 1) elicited the highest number of marijuana associates. Effects of form on correlations with alcohol and marijuana use were assessed cross- sectionally and longitudinally. Form 1 (affective cue cognition task) was used as the reference against Forms 2 3 for these analyses. In addition, variables were generated that were counts of the number of form types one’s friends were given. For example, suppose student A named 5 friends, two of who completed form 1, three of whom completed form 2, and none completed form 3. Analysis of person A’s drug use has three variables that control for each of the following counts: # friends with affective cue task (=2), # friends with locations cue task (=3), and # Word Association Task Alcohol Use Marijuana Use N Mean SD Mean SD Form 1 - Affective Cues 290 0.25 0.43 0.33 0.47 Form 2 - Locations Cues 271 0.11 0.31 0.22 0.41 Form 3 - Compound Cues 273 0.27 0.45 0.31 0.46 Total 834 0.21 0.41 0.29 0.45 61 friends with affective cue task (=0). These variables are important because alters’ cognition scores, might be biased by the forms that they completed. Because form 2, on average, generates fewer drug-related responses, a person with more friends that received form 2 will be more likely to have a lower IC alter effects than a person with friends that received forms 1 or 2. The analyses presented in Table 6.3 were conducted to detect whether the forms bias results. The Cross-sectional Ego Model includes only ego attributes, where baseline behavior is regressed on the ego’s cognitions and survey form. This model tells us whether ego’s baseline cognition instrument (forms 1-3) differentially predicts their behaviors. The Cross-sectional Exposure Model includes ego and alter attributes, where baseline behavior is regressed on the ego’s cognitions, cognition survey form, egos’ behaviors, ego’s cognitions, and a count of their cognition forms. This model tells us whether alters’ baseline cognition instruments (forms 1-3) differentially predict egos behaviors. Finally, the Longitudinal Change Model includes ego’s and alters’ follow-up attributes while controlling for baseline values. The longitudinal model is the most important because it indicates whether the difference in cognition instruments bias relationships between one’s own cognition scores, their friends’ cognition scores, and one’s own drug using behaviors. 62 Table 6.3 Comparison of Alcohol and Marijuana Use Predictions Across Cognition Tasks * p<0.10; ** p<0.05 The first diagnostic is to explore the effects of form type on prediction of drug use cross- sectionally and longitudinally. Ego’s form is not found to be associated with alcohol or marijuana use in any model (see affective and compound cue task variables in ‘Ego Data’ section). This indicates that the three forms are statistically equivalent in their relationship with ego’s current alcohol and marijuana use, cross-sectionally and longitudinally. The second diagnostic is the effects of friends’ forms on behavior models. Form type is associated with alters’ cognition scores. For every friend who completed the compound cue task, students were 24% less likely to report alcohol use and 31% less likely to report marijuana use at follow-up (see values highlighted in yellow), after controlling for ego’s baseline and follow-up cognitions, their behaviors at baseline, cognition exposure scores and the forms that their alters received. The analysis shows that students who had more friends that completed the compound cue task were significantly less likely to report changes in drug use from baseline to follow-up. Alcohol Marijuana Cross- sectional Ego Model Cross- sectional Exposure Model Longitudinal Change Model Cross- sectional Ego Model Cross- sectional Exposure Model Longitudinal Change Model Ego Data ego's use baseline 9.28*** 23.85*** ego's cognitions baseline 8.77*** 8.77*** 2.58** 7.02*** 7.01*** 1.57 ego's cognitions follow-up 11.64** 7.02*** affect cue task (form 1) 0.76 0.71 1.05 0.95 0.84 1.23 compound cue task (form 3) 0.79 0.75 0.95 0.83 0.74 0.89 Alter Data alters' cognitions baseline 1.59 1.17 2.00** 1.32 alters' cognitions follow-up 2.00 1.12 # friends - affective cue task 1.04 0.89 1.12 1.12 # friends - location cue task 1.00 1.22 0.93 0.89 # friends - compound cue task 0.98 0.76* 0.97 0.69** 63 Moderation models were conducted to further investigate whether the influence of alter cognition scores derived from friends’ who completed the compound cue task (form 3) were different from those who completed forms 1 and 2. Interaction terms were created by multiplying alcohol and marijuana cognition exposures by a ‘form 3’ dummy variable (0= alter received either form 1 or 2; 1 = alter completed form 3). The bias of the form 3 scale appears to be spurious in the prediction of marijuana use. Students with more friends that completed form 3 were less likely to report marijuana use at follow-up, but the relationship between marijuana cognitions and changes in drug use from baseline to post-test were not statistically significant. However, there is a statistically significant interaction between and friends’ alcohol cognitions and form 3 cognition task scores in their effect on alcohol use (OR = 2.03, p=0.05). These findings suggest that form 3 scores must be adjusted to make the scales equivalent. Weighted IC scores can control for the biases in the cognition forms such that the weighted values represent equivalent measurements of cognitive tendencies towards alcohol and marijuana. The weighting factor is calculated by the value that when multiplied with forms 2 & 3 results in means that are equivalent to the mean of form 1. For example, form 2 registers lower than the other two forms, then a person with more friends completing form 2 will have lower alter drug cognition score and therefore, their scores should be weighted upwards. The same set of diagnostics was conducted on the weighted the cognition scores, and results showed that the biases found in the unweighted scores were eliminated. Social Networks. Social network data were collected by asking students to write the names of up to 5 other classmates that they considered friends. Network data allows for sociometric maps to be drawn linking students to their five best friends in their class. These data are asymmetrical, in that some students may nominate ‘friends’ in the class while those students do not reciprocate that “friendship.” Classic network analysis accounts for these directional influences among pairs, and 64 statistical network analysis can also account for these patterns in triads. An example of the survey instructions for friend nominations is shown below. Who are your five BEST FRIENDS in this class? Write their names on the lines below starting with your best friend in this class. After you write their name look at the list of names on the roster that has been provided. Match the name to the number and write the number in the boxes. For example, your best friend’s name may be John Angeles. Then you would write his name and then look up his number, which is 1 2 and then write that in the boxes. It is written in as an example below. FIRST NAME LAST NAME Roster Number Example John Angeles 1 2 1 2 3 4 5 Figure 6.2 Example of Social Network Survey Item Indegree and outdegree are network indices considered to be proxies for a person’s connectedness in their social environment. A person’s social integration may affect the amount of information they are exposed to as a result of their social environment. Indegree. Indegree is the number of friend nominations received by classmates divided by the number of students in the class. Indegree is considered to be a proxy for popularity or connectedness within the social environment. A person’s social integration is thought to affect the amount of information they are exposed to as a result of their social environment. 65 Social Network Exposure. Exposures to specific behaviors or attributes through one’s social network are calculated by multiplying the friendship adjacency matrix (with equal columns and rows, NxN) by the behavior vector (1xN). The product represents the exposure of the specified behaviors or cognitions of those they nominated as friends. Specifically, the new vector is equal to the number of friends one has who report the specified attribute (e.g., marijuana use, marijuana cognitions) divided by total number of friends nominated. In this study, the maximum value of an alter effect is 5, assuming the participant nominated the maximum number of friends. For example, if an individual lists three friends, one of whom reports using marijuana, then her alter effects to marijuana use is 0.33. Similarly, if two out of four friends report marijuana use, she has an alter effect equal to 0.66, and so on. See figure 6.3 for an illustration. Alters’ drug behavior scores are a measure of exposure to behaviors while alters’ drug cognition scores are a measure of exposure to cognitions. Figure 6.3 Illustration of Network Exposure (Alter Effects) Calculation Program Condition. The two program conditions were made into dummy variables with the control condition as the reference. Network Exposure = 33% Network Exposure = 33% Network Exposure = 66% Network Exposure = 66% Network Exposure = 100% Network Exposure = 100% = NO Drug = NO Drug- -related Cognitions related Cognitions = Drug = Drug- -related Cognition related Cognition Network Exposure = 33% Network Exposure = 33% Network Exposure = 66% Network Exposure = 66% Network Exposure = 100% Network Exposure = 100% = NO Drug = NO Drug- -related Cognitions related Cognitions = Drug = Drug- -related Cognition related Cognition 66 CHAPTER 7 LOGISTIC REGRESSION ANALYTIC PROCEDURE Everything should be as simple as it is, but not simpler. -- Albert Einstein In the first study proposed, regression analysis will be used to test the effects of a person’s friends’ behavior and cognitions on their own behaviors. Dichotomous outcome variables for alcohol and marijuana use dictate the use of logistic regression. Logistic regression generates a probability curve that signifies the likeliness the outcome occurred as a result of the independent variables. Analysis yields an odds ratio, which is equivalent to standardized coefficients, and can be interpreted as a unit-to-unit relationship between variables. An odds ratio of ‘2’ indicates that as the independent variable increases by one unit, the probability of the dependent variable occurring also increases by one unit. Baseline measurements are identified as Time 1, denoted as T1, and post-test measurements are identified as Time 2 (T2). All analysis is conducted using STATA software, version 10.0 (StataCorp, 2007). Alcohol use (0 =never used, 1= used in the past month) at post-test (T2) is modeled as a function of implicit cognitions at baseline (T1) and friends’ cognitions and behaviors at T2. The model controls for alcohol use at T1, friends’ cognitions and behaviors at T1, number of best friends in school at T1, popularity (indegree) at T1, and demographics (age, gender, ethnicity). Figure 7.1 represents the logistic equation to be solved in the analysis. 67 Figure 7.1 Logistic Regression Equation for Behavioral Influence Models P (Alcohol Use T2 ) = + ß 1 (Alcohol Use T1 ) + ß 2 (Behavior Exposure T1 ) + ß 3 (Behavior Exposure T2 ) + ß 4 (Cognition Exposure T1 ) + ß 5 (Cognition Exposure T2 ) + ß 6 (cognition T1 ) + ß 7 (num best friends in school T1 ) + ß 8 (indegree. T1 ) + ß 9 (age) + ß 10 (gender) + ß 11 (ethnicity) + ß 12 (SES) ß 13 (Standard Condition) + ß 14 (Experimental Condition) A lagged analysis approach will be used which includes prior time period variables to predict later ones. Lagged variables at baseline are expected to account for much of the variance in the dependent variables at follow-up. Controlling for baseline effects isolates the analyses to the change in variables. With both time points included in the model, statistically significant relationships indicate that changes in the independent variable explain changes in drug use from baseline to post- test. 7.1 Multilevel Mixed-Effects Multilevel mixed-effects are characterized as containing both fixed effects and random effects. The fixed effects are analogous to standard coefficients and are estimated in the same manner as is done in standard regression models. The random effects are not directly estimated but are summarized according to their estimated variances and covariances. The random-effects portion of the model is specified by the grouping structure of the data, which in this case vary according to classes, which vary within schools. With the power constraints of the data, only one nested grouping was accounted for. Variances across schools was much greater than across classes, and therefore, models account for random effects of school only. 68 Odds Ratios. Odds ratios and standard errors are reported for all independent models as indicators of the magnitude of influence the variable has on the dependent variable. All variables are dichotomous, except for in-degree, friend importance, age, and SES, which are continuous. For dichotomous variables, the odds represents a 1 to 1 increase in x and y. For continuous variables, an increase in odds represents a smaller unit of change. For example, in- degree scores range from 0 to14, with each unit representing an additional friendship nomination from another classmate. An odds of having an additional friend nomination will show a very small impact on the dependent variable, yet when we consider an accumulation of nominations, popularity can be determined to have a strong influence on the behavior. There are two asymptotically equivalent ways to determine whether independent variables are statistically related to the dependent variable: (1) test whether the parameter estimate differs from ‘0’ in the natural space of the model (H0: b = 0), or (2) test whether the transformed parameter differs from ‘1’ (H0: exp(b) = 1). In STATA, the standard errors for the OR’s are derived from the delta rule, and the confidence intervals reported by Stata for the odds ratios are the exponentiated confidence intervals for the betas: For the simple expression of OR b , the standard error and confidence intervals are: se(OR b ) = exp(b)*se(b) CI (OR b ) = [exp(CI b -lower limit ), exp(CI b -upper limit )] Confidence intervals of OR’s (obtained by transforming confidence intervals of betas) have two advantages – first, they do not produce negative odds ratios and second, they are usually more normally distributed in the natural (beta) space of the model. Missing Data. Social network data relies on responses in a person’s network to create individual level variables. As a result, effects of missing data are compounded when generating alter scores. Alter effects scores were calculated from the total number of available friend nominations. That is, if there was no data on a nominee for a specific attribute, then their 69 nomination was not considered in the denominator. Missing individual scores, thus, skew the importance of the non-missing friendships in a person’s network. A simple test to see whether students who have a higher number of friends with no data are more or less influenced by their friend exposure scores is to consider moderation effects in the models. For each of the exposure scores, a corresponding interaction term was created for the exposure * number of friends with for whom we have no data. These interactions test whether there are differential influence effects for students with a varying number of missing friends. However, interaction effects were found to be non-significant in all cases. 70 CHAPTER 8 LOGISTIC REGRESSION RESULTS 7.1 Descriptive Statistics Descriptive data for the full data set is presented in Table 8.1 by program condition. There are 572 students, representing 74 classrooms in 40 schools. Students are mostly Latino (69%) and male (57%) and are on average, 16.2 years old. Self-reported ethnicity was 69% Hispanic, 13% White, 5% Black, and 13% other. Students range from 14 to 21 year old, and from 9 th to 12 th grades. Continuation High School students are assigned to grades according to the number of units that they have and are not separated on the basis of grade. Socio-economic status ranged from 0 to 1, with 0.5 representing parents’ education and another 0.5 representing their job. Age, gender, ethnicity, and SES, were statistical equivalent across conditions. These characteristics are typical of a California continuation high school population. 71 Table 8.1 Demographics of Study Participants Changes in Cognitions & Behaviors. In general, after the 12-week intervention, behaviors and cognitions change very modestly. Mean scores of the changes observed are presented, as well as the absolute value of the differences between baseline and follow-up. All alcohol/marijuana behaviors show a small increase, while cognitive indicators show a moderate decrease from baseline to follow- up. Since the intervention has a low effect size, the study will have low power to detect relationships between the social factors that influence behavioral changes. Alcohol behaviors increased by 0.007 (std dev =0.46; range: -1 to +1, n=604); marijuana use increased by 0.003 (std dev = 0.41; range: -1 to +1, n=606); alcohol implicit cognition scores decreased by 0.14 (std dev = 0.56; range: -2.27 to +1, n=602); marijuana implicit cognition scores decreased by 0.08 (std dev = 0.52; range: -1.54 to +1; n=603). Control Condition Standard Condition Experimental Condition Total n 160 191 221 572 # classrooms 28212574 # schools 13 14 13 40 Age 16.4 (1.4) 16.3 (1.3) 16.1 (1.4) 16.2 Gender, (%male) 0.61 0.53 0.56 0.57 Ethnicity (% Hispanic) 0.73 0.68 0.67 0.69 Socio-economic status 1.5 (.4) 1.6 (.4) 1.5 (.4) 1.5 # best friends in school 2.0 (1.9) 2.0 (2.0) 2.2 (2.2) 2.1 Alcohol use % 0.66 0.72 0.66 0.62 Marijuana use % 0.60 0.60 0.57 0.51 Alcohol implicit cognitions % 0.34 0.18 0.24 0.24 Marijuana implicit cognitions % 0.42 0.29 0.32 0.34 BASELINE DRUG COGNTIONS & BEHAVIORS DEMOGRAPHICS 72 While the behavioral and cognitive changes for the entire population are relatively small, the number of students who either increase or decrease their cognitions or behaviors from time 1 to time 2 suggest that change occurred differentially by condition. See Table 8.2 for change counts for each outcome variable. The percentage of students in the control condition who showed a decrease in alcohol use, marijuana use, alcohol cognitions, marijuana conditions from baseline to follow-up were 10, 13, 13, and 11 percent. The percentage of standard condition students that decreased behaviors and cognitions were 14, 9, 13, and 8 percent. Experimental condition students decreased behaviors and cognitions by 8, 6, 18, and 14 percent. Only non-missing values are reflected in these percentages. Overall, change counts reveal that the experimental condition reduced drug related cognitions for the greatest number of students, while the standard condition was effective in reducing alcohol & marijuana behaviors, compared to the control and standard conditions. Condition effects were not found to be statistically significant in the logistic regression analyses, but data are reported that suggest that the standard condition reduces behaviors and increases cognitions, compared to experimental and control classes. 73 Table 8.2 Cognitive and Behavioral Change Counts From Baseline to Follow-up 8.2 Cognitive and Behavioral Influence Results Logistic mixed effects regression longitudinal analyses were conducted to test whether friends’ cognitions and behaviors contribute to one’s own cognitions and behaviors. Results for four models represent predictions of each of the following variables post-intervention: alcohol use, alcohol cognitions, marijuana use, and marijuana cognitions. Tables 8.3 and 8.4 show results predicting behaviors, and Tables 8.5 and 8.6 show results for cognition predictions. Control Condition Standard Condition Experimental Condition Total (n=160) (n=191) (n=221) (n=572) decrease (-1) .10 (15) .14 (26) .08 (18) .11 (59) No change (0) .80 (124) .77 (144) .79 (169) .78 (437) increase (+1) .10 (16) .09 (17) 0.13 (28) .11 (61) Missing 5 4 6 15 decrease (-1) .13 (20) .09 (16) .06 (13) .09 (49) No change (0) .76 (119) 0.85 (158) .85 (188) .81 (465) increase (+1) .11 (17) 0.07 (13) .07 (15) .08 (45) Missing 4 4 5 13 decrease (-1) .13 (21) .13 (24) .18 (39) .15 (62) No change (0) .82 (128) .81 (153) .77 (168) .80 (450) increase (+1) .05 (7) .06 (11) .05 (10) .05 (49) Missing 4 3 4 11 decrease (-1) .11 (17) .08 (15) .14 (30) .11 (62) No change (0) .82 (128) .82 (155) .77 (167) .80 (450) increase (+1) .07 (11) .10 (18) .09 (20) .09 (49) Missing 4 3 4 11 Alcohol Cognitions Marijuana Cognition Alcohol Use Marijuana Use 74 Tables 8.3 and 8.5 show results for alcohol activities and Tables 8.4 and 8.6 show marijuana activities. A lagged regression approach requires the inclusion of baseline and follow-up scores for all behavioral and cognitive variables. Covariates include demographic variables (gender, ethnicity, age and SES) and social network indicators (indegree and # of best friends in school). Intervention conditions were included as dummy variables for the standard and experimental conditions. Hypothesis-driven variables are categorized as ego effects (individual’s cognitions and behaviors) and alter effects (friends’ cognitions and behaviors). The first 4 columns of the Tables show results for sets of variables: (1) demographics, (2) social network indicators, (3) intervention conditions, and (4) cognitions and behaviors. These sets of variables, referred to as ‘partial models,’ were considered one at a time to identify the relative importance of general and local environmental factors in their influence on drug activities. The additional three columns in the Tables present results for what are termed as ‘full’, ‘reduced’ and ‘parsimonious’ models. The full model includes all predictor variables, covariates, and controls. The reduced model includes variables that met stepwise inclusion criteria. One variable from each set had to remain in the model (e.g., one demographic variable, one of the social network variables, and so on). In addition, of the main hypothesis driven variables (behavior and cognitions of ego and alters), post-test could be eliminated, but baseline could not. The elimination process continued until either these criteria were met, or the remaining variables were significant at the 5% probability level. The parsimonious model is the result of an unconstrained stepwise elimination process. The variable with the highest p-value was removed at each iteration of the model. The final parsimonious models represent variables that were significant at the 0.05 level, and represents the best predictive model with the fewest number of variables. 75 In these models, results that are significant at the 0.10 level are noted.. This is a moderately conservative alpha to infer statistical significance.. Restricting the p-value to p<0.05 would constrain the model to a point where it would give much less information about the relationships between the independent and dependent variables. The effects of friends’ cognitions on ego’s cognitions have never been studied before; therefore, there is no precedence for determining the expected effect size or power needed to detect these relationships. Conforming too closely to arbitrary statistical cut-offs might obscure important social dynamics occurring through symbolic or non-verbal communication. For this reason, an α = 0.10 is discussed in the evaluation of findings. 8.3 Logistic Regression of Alcohol Behaviors Table 8.3 presents logistic mixed effects regression for alcohol use after completion of a three month long intervention. Models test hypotheses that friends’ alcohol behaviors and cognitions influence ego’s alcohol behaviors, after controlling for ego’s cognitions and their past behaviors. 76 Table 8.3 Logistic Regression Results for Alcohol Behaviors on Ego and Alter Effects in Continuation HS Students *p<0.10, **p<0.05 Table 8.3 Sets of explanatory variables are regressed on to self-reported alcohol use upon completion of the intervention program. Odds ratios and standard errors were generated from Mixed Method Logistic Regression analyses. Log likelihood’s, Wald Chi2, p-values, and sample size (n) are listed for each model. Demographics Social Networks Program Condition Cognition & Behavior Full Model Reduced Model Parsimonious Model n 541 560 560 387 374 389 389 Wald Chi2 5.62 7.42 1.01 92.8 89.6 93.01 98.61 Probability 0.230 0.020 0.603 0.000 0.000 0.000 0.000 Log Liklihood -354.05 -366.02 -369.43 -182.1 -175.06 -183.2 -202.75 Age 1.11 (.08)* 1.00 (0.10) Gender 1.35 (.25) 1.05 (0.29) Ethnicity 1.23 (.26) 1.24 (0.36) 1.25 (0.36) SES 0.88 (.15) 0.92 (0.25) In-degree 1.12 (.05)** 1.02 (0.07) # best friends in school 0.91 (.05)* 0.99 (0.07) 0.97 (0.07) Experimental Condition 1.18 (.27) 1.19 (0.41) Standard Condition 0.96 (.22) 0.83 (0.30) 0.78 (.21) Alcohol Use (T1) 7.88 (2.13)** 8.01 (2.26)** 8.06 (2.21)** 8.01 (2.05)** Alcohol Cognitions (T2) 10.23 (10.65)** 9.47 (9.91)** 10.16 (10.58)** 11.41 (11.84)** Alcohol Cognitions (T1) 1.92 (0.66)* 1.86 (0.65)* 1.93 (.66)** 1.95 (.65)** Friends' Alcohol Use (T2) 3.24 (1.29)** 3.54 (1.46)** 3.49 (1.40)** 3.11 (1.07)** Friends' Alcohol Use (T1) 1.21 (.54) 1.26 (0.60) 1.19 (.54) Friends' Alcohol Cognition (T2) 1.35 (0.97) 1.46 (1.07) Friends' Alcohol Cognition (T1) 0.64 (.26) 0.56 (0.25) 0.60 (.24) ALCOHOL USE PREDICTION MODELS EXPLANATORY VARIABLES EGO ATTRIBUTES EGO EFFECTS ALTER EFFECTS 77 Partial Models. In the first set of results (columns 1-4), demographics, social network, program condition and drug-related activities are each modeled as a set. Findings show that ego effects, alter effects, age, and network dependencies are significant predictors of alcohol use. Older students (OR = 1.11, p<0.1), more popular students (OR = 1.12 p<0.05), and students with fewer close friends within the school (OR = 0.91, p<0.1) are more likely to be drinkers at follow-up (hereafter, referred to as T2) compared to younger, unpopular, and students with close friends within the school. After controlling for baseline (hereafter, referred to as T1) ego and alter alcohol activities, students who have alcohol-related cognitions at T2 are about ten times more likely to be drinkers at T2, compared to those with no alcohol cognitions (OR = 10.23, p<0.05); and students who have alcohol using friends are about three times more likely to be drinkers at T2 (OR = 3.23, p<0.001), compared to those without friends who drink. Full Model. After controlling for demographics, networks, conditions, and all other drug activities at T1 and T2, only ego’s past alcohol use at T1, ego’s cognitions at T2, and friends’ alcohol use remained significant. Students who report alcohol use at T1 are about 8 times more likely to report using alcohol after the intervention compared to non-users at T1 (p<0.001). Students with strong alcohol cognitions at T2 are about 10 times more likely to be drinkers at T2 compared to those with no alcohol cognitions (p<0.001). Students who have friends’ who report using at T2 are about 3 times more likely to be drinkers at T2 compared to those without drinking friends (p<0.001). Reduced Model. In the reduced model, ego’s cognitions and behaviors, and friends’ behaviors remain significant predictors of alcohol use. While not reaching significance, the model shows that students in the standard condition are slightly less likely to be alcohol users at T2. 78 Parsimonious Model. Results from a backward stepwise elimination process show that there were only minor changes for the significant variables in the reduced to the parsimonious models. In the final parsimonious model, demographics, network dependencies, and program condition variables were eliminated, leaving only ego and alter effects (Wald Chi2 = 98.61; log likelihood = 202.5; p<0.001; n=389). Alcohol users at T1 were 8 times more likely to be users at T2 compared to non-users at T1 after controlling for all other variables remaining in the model (OR=8.01, p<0.001). Those with alcohol cognitions at T2 were ~11 times more likely to be drinkers at T2 compared to those without strong cognitions after controlling for all other variables remaining in the model (OR=11.41, p<0.05). Students with drinker friends are about 3 times more likely to be drinkers at T2, compared to those without drinking friends after controlling for all other variables remaining in the model (OR=3.11, p<0.001). As a reminder, the standard errors in parentheses presented in Tables 8.3-8.6 represent SE’s of the Odds Ratio rather than SE’s of the beta estimates. Odds ratios and their SE’s do not have the familiar relationship where significance is confirmed by the coefficient divided by its SE equaling 1.96 or higher. Instead, the SE’s represent transformed values as discussed in chapter 7. As an example, in the “full model”, the OR for alcohol cognition (T2) is 9.47 with a standard error of 9.91. Another way of reporting this would be to say that alcohol cognition has a coefficient = ln 9.47 (= 2.248) and a point estimate standard error = 1.04. Using the delta rule, the standard error for the OR then, is se(OR b ) = exp(2.248)*1.046 = 9.91 Therefore, an OR = 9.47 with an SE = 9.91 and a CI = 1.23 to 73.95 is equivalent to reporting: beta coeff = 2.25, SE b = 1.04, z=2.16; p= 0.031, CI = 0.21 and 4.3. 79 It is important to note that the 95% confidence interval for this estimate is quite large: 2.25 +/- (1.04*1.96) indicating that 95% of the time, the estimate for the true population falls somewhere between 0.21 and 4.3 (with an alpha set at 0.05). Therefore, while these point estimates are statistically significant, there is a very large variance in how alcohol cognitions correlate to alcohol behaviors. Insufficient power is usually the cause of such large variances, and the range would either contract or expand with additional population data. Without knowing how a larger ‘n’ would affect these estimates, it is better to be conservative and view the correlations between behavior and cognitions as suggestive but not conclusive. 8.4 Logistic Regression of Marijuana Behaviors Table 8.4 shows models containing sets of predictor variables as well as full, reduced and parsimonious models testing proposed hypotheses about the correlates of marijuana behavioral change. 80 Table 8.4 Logistic Regression Results for Marijuana Behaviors on Ego and Alter Effects in Continuation HS Students *p<0.10, **p<0.05 Table 8.4 Sets of explanatory variables are regressed on self-reports of marijuana use at T2. Odds ratios, standard errors are presented from Mixed Method Logistic regression analyses. Log likelihood’s, and Wald Chi2, p-values, and sample size (n) are listed for each model. Partial Models. In the partial models predicting marijuana use at T2, ethnicity, social network indicators, and previous marijuana use were statistically significant. Non-Latinos were 1.88 times more likely to be smokers at T2 compared to Latinos (OR=1.88, p<0.001). Popular Demographics Social Networks Program Condition Cognition & Behavior Full Model Reduced Model Parsimonious Model n 560 560 541 390 377 392 590 Wald Chi2 0.93 18.62 11.4 131.18 123.3 129.8 198.41 Probability 0.629 0.000 0.022 0.000 0.000 0.000 0.000 Log Liklihood -387.09 -377.51 -368.72 -150.99 -145.13 -147.65 -251.87 Age 1.03 (.06) 0.90 (.10) 0.90 (0.10) Gender 1.04 (.18) 1.22 (.39) Ethnicity 1.88 (.36)** 0.94 (.31) SES 0.95 (.15) 0.82 (.25) In-degree 1.19 (.05)** 1.18 (.10) ** 0.87 (0.07)** # best friends in school 0.87 (.04)** 0.87 (.07) 0.87 (.07)* Experimental Condition 1.14 (.24) 0.96 (.36) Standard Condition 1.03 (.22) 0.72 (.29) 0.74 (.24) Marijuana Use (T1) 17.71 (5.60)** 17.71 (5.60)** 19.09 (5.96)** 19.12 (4.32)** Marijuana Cognitions (T2) 5.64 (2.42) 5.64 (2.42)** 5.43 (2.25)** 4.76 (1.36)** Marijuana Cognitions (T1) 1.47 (.35) 1.47 (.35) 0.64 (.29) Friends' Marijuana Use (T1) 1.96 (.90) 1.96 (.90) 2.00 (.87)* Friends' Marijuana Use (T1) 1.70 (.89) 1.70 (.89) 1.63 (0.84) Friends' Marijuana Cognition (T2) 0.74 (.35) 0.74 (.35) Friends' Marijuana Cognition (T1) 0.85 (.43) 0.85 (.43) 0.64 (.29) ALTER EFFECTS EGO EFFECTS MARIJUANA USE PREDICTION MODELS EXPLANATORY VARIABLES EGO ATTRIBUTES 81 students were 1.10 more likely to be smokers at T2, compared to unpopular students (OR=1.19, p<0.001). Students with more close friends within the school were 13% less likely to be smokers at T2 (OR=0.87, p<0.05). When behavior and cognition variables are modeled alone, only baseline use is significant (OR=17.71, p<0.001). Full Model. In the full model, popularity, marijuana use at T1 and marijuana cognitions at T2 are significantly predictive of T2 marijuana use, after controlling for ego’s and alters’ baseline marijuana activities, demographics, friendship importance, and intervention condition. Reduced Model. After gender, ethnicity, SES and experimental condition effects were eliminated in the reduced model, friends’ marijuana emerges as having a moderately significant influence on marijuana use (OR=2.0, p<0.10). While not reaching significance, the model shows that students in the standard condition were slightly less likely to be marijuana users than students in the control or the experimental condition. Parsimonious Model. The parsimonious model shows that students who were using marijuana in the month prior to the program were about 19 times more likely to report smoking marijuana at T2 compared to non-smokers at baseline (OR=19.12, p=0.001), after controlling for the remaining variables. Those with marijuana cognitions at T2 were about 5 times more likely to report smoking marijuana at T2 compared to those without cognitions at T2 (OR=4.76, p=0.001). Statistics for the model are: Wald Chi2 =198.41, log likelihood = - 251.87; p<0.001; n=590. 8.5 Logistic Regression of Alcohol Cognitions Table 8.5 shows odds ratios and standard errors for models predicting alcohol cognitions at T2. Mention different order – cognition before behavior – In this table, ego and alter cognitions are listed first to correspond with the outcome, followed by behaviors. 82 Table 8.5 Logistic Regression Results for Alcohol Cognitions on Ego and Alter Effects in Continuation HS Students *p<0.10, **p<0.05 Table 8.5 Sets of explanatory variables are regressed on to alcohol cognitions at T2. Mixed Method Logistic regression odds ratios, standard errors and p-values are presented. Log likelihood’s, and Wald Chi2, corresponding p-values, and sample size (n) are listed for each model. Partial Models. Sets of explanatory variables show that significant predictors of alcohol cognitions include alcohol use at T1 and T2, friends’ alcohol cognitions at T2, and ethnicity. Interestingly, those with alcohol cognitions at baseline (T1) are not more likely to have alcohol cognitions at follow-up (T2). Demographics Social Networks Program Condition Cognitions & Behaviors Full Model Reduced Model Parsimonious Model n 540 561 561 387 374 390 390 wald chi2 9.8 2.92 0.18 22.81 32.74 31.57 31.54 probability 0.044 0.2326 0.9121 0.0018 0.01 0.0001 0.0002 Log Liklihood -182.87 -194.5 -195.84 -105.17 -96.32 -101.19 -101.19 Age 1.12 (.12) 0.96 (.14) Gender 0.86 (.25) 0.57 (.24) 0.61 (.24) Latino 2.34 (.68)** 2.70 (1.06)** 2.59 (1.01)** 2.54 (.87)** SES 0.98 (.26) 0.98 (.37) In-degree 1.09 (0.07) 1.05 (.09) # best friends in school 0.91 (.07) 1.00 (.11) 1.05 (0.08) Network Condition 1.12 (.39) 1.34 (.71) Standard Condition 1.16 (.41) 1.97 (1.10) 1.47 (.59) Alcohol Cognitions (T1) 1.41 (0.38) 1.51(.44) 1.44 (.39) Alcohol Use (T2) 12.90 (13.58)** 11.88 (12.65)** 11.52 (12.16)** 29.39 (30.13)** Alcohol Use (T1) 3.19 (2.09)* 3.58 (2.43)* 3.74 (2.53)** Friends' Alcohol Cognition (T2) 5.28 (4.04)** 5.87 (4.62)** 6.34 (4.43)*** 8.20 (5.08)** Friends' Alcohol Cognition (T1) 1.15 (.61) 0.91 (0.52) Friends' Alcohol Use (T2) 1.56 (1.02) 1.58 (1.08) Friends' Alcohol Use (T1) 0.52 (.35) 0.44 (.33) 0.51 (.35) ALCOHOL COGNITION PREDICTION MODELS EXPLANATORY VARIABLES ALTER EFFECTS EGO ATTRIBUTES EGO EFFECTS 83 Full Model. The full model shows no significant changes in the explanatory variables. After controlling for ego and alter baseline alcohol activities, demographics, network indicators, and intervention conditions, only ego’s behaviors, friends’ alcohol cognitions, and ethnicity remain significant in the model. Reduced Model. The reduced model shows these same variables remain significant with little change in magnitude. While not reaching significance, the model shows that students in the standard condition are slightly more likely to have strong alcohol cognitions. Parsimonious Model. The parsimonious model shows that students who report alcohol use at T2 are about 29 times more likely to generate alcohol associations at T2 (OR=29.39; p=0.001), compared to non-drinkers. Those who have friends with alcohol cognitions at T2 are about 68 times more likely to generate alcohol associations at T2 (OR=8.20; p=0.001), compared to students whose friends are without these cognitions. Non-Latinos are about 2.5 times more likely to generate alcohol associations at T2 (p<0.006), compared to Latinos. The statistics for the parsimonious models are as follows: Wald Chi2 =31.57; log likelihood = - 101.19; p<0.001; n=390. 8.6 Logistic Regression of Marijuana Cognitions Table 8.6 presents odds ratio and standard errors for models predicting marijuana cognitions. 84 Table 8.6 Logistic Regression Results for Marijuana Cognitions on Ego and Alter Effects in Continuation HS Students *p<0.10, **p<0.05 Table 8.6 Sets of explanatory variables are regressed on to marijuana cognitions at T2. Odds ratios, standard errors, and p-values are generated from Mixed Method Logistic regression analyses. Log likelihood’s, Wald Chi2, p-values, and sample size (n) are listed for each model. Partial Models. Results from partial models show that significant predictors of marijuana cognitions include marijuana cognitions at T1, marijuana use at T2, friends’ marijuana cognitions at T2, gender, and indegree. Demographics Social Networks Program Condition Cognition & Behavior Full Model Reduced Model Parsimonious Model n 540 561 561 390 377 391 433 wald chi2 15.38 8.62 1.68 89 86.68 90.16 95.02 probability 0.004 0.0134 0.4325 0 0 0 0 Log Liklihood -303.18 -317.47 -320.99 -151.22 -141.55 -148.69 -172.86 Age 0.96 (.07) 0.85 (.10) Gender 0.60 (.12)** 0.44 (.15)** 0.48 (.15)** 0.53 (.15)** Latino 1.84 (.4)** 1.20 (.38) SES 0.83 (.15) 0.87 (.27) In-degree 1.13 (.06)** 1.05 (.07) 1.03 (.07) # best friends in school 0.90 (.05)* 1.03 (.09) Network Condition 1.26 (.033) 0.94 (.37) Standard Condition 1.39 (.35) 1.34 (.56) 1.30 (.41) Marijuana Cognitions (T1) 4.07 (1.06)** 3.77 (1.03)** 3.85 (1.00)** 4.74 (1.11)** Marijuana Use (T2) 5.42 (2.23)** 5.27 (2.26)** 5.49 (2.31)** 3.35 (1.03)** Marijuana Use (T1) 1.08 (.45) 1.12 (.48) 1.10 (.46) Friends' Marijuana Cognition (T2) 7.33 (3.16)** 6.72 (3.08)** 6.72 (2.83)** 7.76 (2.87)** Friends' Marijuana Cognition (T1) 1.27 (.54)*** 1.12 (.51)* 1.04 (.45) Friends' Marijuana Use (T2) 0.71 (.34) 0.90 (.45) Friends' Marijuana Use (T1) 1.58 (.82) 1.53 (.84) 1.52 (.76) MARIJUANA COGNITION PREDICTION MODELS EXPLANATORY VARIABLES ALTER EFFECTS EGO ATTRIBUTES EGO EFFECTS 85 Full Model. Variables that are significant in the reduced and parsimonious model described below are also significant in the full model. Reduced Model. Significant variables remain the same as in the full model even after imposed criteria. While not reaching significance, the model shows that students in the standard condition are slightly more likely to report marijuana cognitions. Parsimonious Model. In the parsimonious model, students with marijuana cognitions at baseline are about 5 times more likely to report marijuana cognitions after the intervention (OR=4.74, p<0.001) compared to those with no cognitions at T1. Marijuana smokers are about 3 times more likely than non-smokers to report marijuana cognitions at T2 (OR=3.35, p<0.001). Girls are about 50% less likely than boys to report marijuana cognitions at T2 (0.53, p<0.05). Statistics for the parsimonious model include the following: Wald Chi2 =95.02; log likelihood = - 172.86; p<0.001; n=433. 8.7 Summary of Behavioral Influence Logistic Regression Analysis Ego Attributes. Popular students are slightly more likely than unpopular students to be marijuana users. Non-Latinos are about 3 times more likely to have strong associations around alcohol use than Latinos. Boys are significantly more likely to have marijuana cognitions than girls. Unlike reported findings in Valente et al. (2007), males were not significantly more likely to increase their alcohol consumption compared to females and in-degree did not reach significance for alcohol or marijuana in the complete models. These differences are likely due to the slightly sub-sets of the population used in this study, due to the large number of missing responses for implicit cognition tasks. Intervention conditions. Standard condition neared statistical significance in all the reduced models. Interestingly, students in the standard condition were about 25% less likely to be drinkers and smokers and about 30% more likely to have alcohol and marijuana cognitions. In contrast to 86 previous findings in these data, students in the experimental condition did not show a significant increase in marijuana use, relative to the control. Program condition was not significant in any of the models, but results suggest that standard condition was effective in reducing alcohol and marijuana use, but increased the likelihood that students develop drug-related alcohol and marijuana cognitions. Ego’s Drug-Related Activities. Typically, past behaviors are the best predictors of future behaviors. The first set of results deal with the effects of past activities on future activities. This question is explored for cognitions and behaviors in order to tease out social effects from ego effects. On average, only 20% of the study population change their behaviors or cognitions from T1 to T2. Therefore, it is not surprising that baseline behaviors are the strongest predictors of corresponding behaviors at T2, three months later. However, contrary to expectations, cognitions do not show these patterns. While those with marijuana cognitions at baseline are more likely to have cognitions three months later, the magnitude of this relationship is relatively low (OR=4.7 compared to marijuana behavior, where those with who smoked at baseline are 20.4 more likely to smoke at follow-up). Further, baseline alcohol cognitions have almost no predictive value for later alcohol cognitions. This suggests that associations around alcohol are not pre-determined by established cognitive pathways, and may be highly malleable, even during a time span of only three months. Instead, changes in alcohol cognitions (15% of students decrease and 5% increase their cognitions from T1 to T2) are explained by the individual’s alcohol behavior, suggesting that behavior drives cognitions directly. Ego Cognition-Behavior Interactions. The second set of findings address whether there are intrapersonal cross-effects of behaviors and cognitions. For example, does an individual’s alcohol cognitions lead to future alcohol use, and vice versa? The results show that there is, in fact, cross- 87 effects between behaviors and cognitions. Behaviors are consistently predictive of cognitions, and cognitions are predictive of behaviors in all four models. That is, given baseline cognitions/behaviors, there are high odds that the corresponding behaviors/cognitions will occur at follow-up. For example, those with alcohol cognitions at T2 are 11.4 times more likely to also use alcohol at T2, compared to those without cognitions and after controlling for baseline behaviors and cognitions. In general, these cross effects are greater for alcohol than for marijuana. Alters’ Drug-Related Activities. The most important relationships in these analyses are those that address as alter-ego dynamics. The extent to which the alters’ activities influence ego’s are assessed by regressing average behaviors of a person’s friends’ on to ego’s behaviors. The first set of hypotheses (hypotheses 1 and 3) investigated is the effects of alters’ activities on the corresponding ego’s activity. In each of the four models, friends’ behaviors and cognitions are predictive of corresponding behaviors and cognitions. That is, influence of alters’ alcohol cognitions on ego’s alcohol cognitions, influence of alters’ alcohol behaviors on ego’s alcohol behaviors, and so on. Marijuana behaviors are the least influenced by others’ behaviors, as the odds ratios are low in magnitude (OR = 1.9) and are significant only at the 10% probability level. In contrast, those whose friends report alcohol use are 3.1 times more likely to also report alcohol use (p<0.001). Valente et al. (2007) showed that there was an interaction between friends’ behaviors and program condition. As friends’ use increases, change in ego’s alcohol and marijuana use declines in the experimental and standard conditions. In the experimental condition, as friends’ use increases, change in ego’s alcohol and marijuana use also increases. Their findings suggest that the experimental condition was effective in facilitating social influences, and that program intentions of reducing drug use are only effective on the group-level – friends’ decrease behaviors, ego 88 decrease behaviors. These interaction terms were not tested in these models, but these findings will be considered further in the final Discussion section. Surprisingly, the influence of friends’ cognitions on ego’s cognitions were considerably stronger than influences on behaviors. Those who report alcohol cognitions at T2 are 6.4 times more likely to have friends with alcohol cognitions, and those with marijuana cognitions at follow- up are 7.8 times more likely to have friends with marijuana cognitions, as compared to odds ratio effects of 3.1 and 1.9 for alcohol and marijuana alter behaviors, respectively. Alter-Ego Cognition-Behavior Interactions. The second set of hypotheses (hypotheses 2 and 4) explore the cognition-behavior cross-effects of alters’ activities on the ego’s activities (e.g., influence of alters’ alcohol cognitions on ego’s alcohol behaviors). There were no significant cognitive-to-behavioral or behavioral-to-cognitive alter influences on the ego. However, alters’ alcohol cognitions did show an inverse influence on ego’s behaviors (OR = .60; p=0.198), These odds ratios were only significant at 20% probability level, therefore, they suggest only a very small deviation from the null hypothesis. Yet, the results suggest that there is a trend for alcohol users to have few or no friends’ with cognitive tendencies towards alcohol use. Marijuana models do not show any tendencies for alter-ego cross influences on behaviors or cognitions. Summary of Regression Findings, By Outcome. Alcohol and marijuana activities have been characterized here as unique processes. Figures 8.1a-d summarize results for each outcome. Characterizations rely on parsimonious models. 89 Figures 8.1a-d Significant Alter and Ego Effects on Alcohol and Marijuana Activities Figure 8.1a Alcohol use is influenced by past use, cognitions, and friends’ behaviors, suggesting dynamics between cognitions and behaviors, within and between individuals Figure 8.1b Marijuana use is highly dependent on internal (ego) processes rather than social processes. Students who use marijuana at baseline and those with marijuana cognitions were more likely to be smokers at follow-up, while friends’ activities had a very little effect. Figure 8.1c Alcohol cognitions are driven by behaviors (and not pre-existing cognitions) and friends’ cognitions, suggesting dynamics between cognitions and behaviors - within and between individuals. Figure 8.1d Marijuana cognitions are, to a large extent, based on cognitive processes, both internally and socially. This suggests marijuana is implicitly driven and that social influences occur through implicit (rather than overt behavioral) exchanges. Alcohol Friends’ Alcohol Use Ego’s Alcohol Use, T1 Ego’s Alcohol Cognition Alter Effects Ego Effects Alcohol Use Friends’ Alcohol Use Alter Effects Ego Effects Alcohol Cognitions Ego Effects Ego’s Alcohol Use Friends’ Alcohol Cognitions Alter Effects Alcohol Cognitions Ego Effects Ego’s Alcohol Use Friends’ Alcohol Cognitions Alter Effects Marijuana Use Ego’s Marijuana Use, T1 Ego Effects Ego’s Marijuana Cognition Marijuana Use Ego’s Marijuana Use, T1 Ego’s Marijuana Cognition Marijuana Cognitions Friend’s ’ Friends’ Marijuana Cognition Ego’s Marijuana Use Alter Effects Ego Effects Ego’s Marijuana Cognitions, T2 Marijuana Cognitions Friend’s Friends’ Marijuana Cognition Ego’s Marijuana Use Alter Effects Ego Effects Ego’s Marijuana Cognitions 90 Independent explanatory variables are classified here as either the individual’s intrinsic attributes (e.g., age, gender), the individual’s activities (i.e., behaviors and cognitions) or the individual’s friends’ activities (i.e., friends’ behaviors and cognitions). Another way of classifying these variables is to say whether they are ego-based (individual’s attributes and activities) or alter- based (friends’ activities) or whether they are cognitively-driven (ego or alter cognitions) or behaviorally-driven (ego or alter behaviors). Not surprisingly, all models predicting ego’s cognitions and behaviors show a strong dependence on the ego’s activities. That is to say, behaviors are well predicted by past behaviors as well as cognitive tendencies towards the behaviors, and cognitions are also well predicted by past cognitions and behaviors. In addition, alcohol use appears to encompass ego and alter influences, while marijuana behaviors are apparently less susceptible to social influences. In addition, findings suggest that cognitions are more influenced by social exchanges than are behaviors. As for the distinction between cognitive or behavioral motives, alcohol and marijuana are driven by both. Oddly though, alcohol cognitions are correlated to a person’s behaviors more so than by pre-existing cognitive pathways, while marijuana cognitions have a relatively small relationship to past behaviors. The first conclusion based on these classifications is that alcohol use is dependent on both intra- & inter-personal factors. Conversely, marijuana use only shows dependencies on internal (and not social) factors. This suggests that marijuana behaviors are less amenable to direct social influences, and behavioral changes may have to overcome a higher cognitive threshold. The second conclusion is that for the most part, alcohol and marijuana activities are functions of both cognitive & behavioral influences, but in the case of alcohol cognitions, behavioral influences 91 trump cognitive ones. This suggests that cognitive pathways easily adapt to alcohol approval while endorsements of marijuana use are more constrained by one’s past. 92 CHAPTER 9 SIENA NETWORK MODELING Social network statistics indicate whether the observed network exhibits the hypothesized structural features more than we would expect by chance alone. ERG models are constructed where dependent variables are relational links, and predictor variables are typically similarities (or differences) in attributes between linked individuals. Potential explanatory variables can also include: an individual’s attributes (e.g., gender or age), their alters’ attributes, and the characteristics of the relation (e.g., whether or not the tie is reciprocated). In statistical estimations of longitudinal ERG models, social selection is operationalized by the formation (and dissolution) of ties between actors as a function of their attributes (i.e., the extent to which friendship structures are driven by homophily), whereas social influence is operationalized as the assimilation of friends’ attributes (i.e., the extend to which individuals become more similar to their friends). Structural configurations (e.g., reciprocity, transitivity) act as controls in the models, allowing the researcher to be sure that any attribute-based association is above and beyond any intrinsic structural effects. These techniques allow us to test multiple levels of structural characteristics simultaneously, controlling for other levels. Homophily hypotheses can be tested by whether ties are more likely to exist between actors if they share a common attribute (e.g., drug use), after accounting for the likeliness of their relationship due to other attributes (e.g., gender, age) and inherent structural tendencies (e.g., reciprocity, transitivity). 93 9.1 SIENA Modeling Procedures Statistical longitudinal ERG modeling is carried out by the software, Simulation Investigation for Empirical Network Analysis (SIENA). SIENA executes ERG modeling according to a specific set of assumptions and modeling processes, following the actor-oriented model presented by Snijders (1995). The model is based on two functions: the objective and rate functions. The rate function indicates the rate of change, or the rate at which an actor will change their social position (network rate) or change their behaviors (behavioral rate). Rate of changes are estimated by the number of changes in the network and behavior that would be required to get from the initial configuration to the final one. The objective function indicates the tendency or preference for the current state of the network. The aim when attempting to generate an ERG model is to find the set of objective functions that maximize the probability that any graph generated by simulation will be identical to the observed network. SIENA generates parameter estimates for network configurations and behavioral dynamics specified by the user. Like all data modeling, the parameters must be chosen based on theory or empirical evidence. The aim when attempting to generate an ERG model is to find the set of parameters that maximize the probability that any graph generated by simulated ERG model will be identical to the observed network. There are three components of ERG modeling: (1) parameters specified by the user are assigned starting values (usually zero), (2) set of graphs are simulated based on the number of nodes and links in the observed data, and (3) simulation of graph evolution is repeated and point estimates for each parameter are recalculated until adequate iterations have been completed, (4) statistical tests are carried out that calculate how closely the mean and standard deviations found in the random graphs match the observed graph. Simulation. The most advanced method to estimate the parameters used in current software is the Markov Chain Monte Carlo Maximum Likelihood Estimation (MCMC/MLE) procedure. The 94 estimation procedure starts by generating a graph from a random starting point. It estimates the probability of that tie forming (or disappearing) given the ties around it and the parameter values of the various configurations in the model. If it is likely that the tie would form based on the model hypotheses, the link is generated otherwise, it is not. For example, if the formation of the tie will complete a transitive triplet and transitivity has a positive parameter, then the routine will most likely create this tie. This process is repeated hundreds or thousands of times, each time refining parameter values until the parameter estimates stabilize (Robins et al., 2007). The specifications of this procedure are detailed in Snijders (1995, 2001, 2005). Estimation. The simulated graphs are analyzed as a distribution of possible graphs, and a mean and standard error for each configuration is generated. By comparing the graph statistics of the original observed graph with the mean and standard error of the simulated graphs, the software assesses how well the model fits the true data. Statistical values for each parameter represent the likeliness of its role in observed data patterns when compared to the simulated distribution. These values can be read like parameters and errors of standard logistic regression, where the standard error can be used as a measure of the significance of the parameter value. Significance of a parameter value is recognized if it is greater than 1.96 times larger than the standard error, significant at the p<0.05 level. However, these significance values should be viewed with a great deal of caution! Model Fitting. Two criteria have to be met to indicate that the model has been run successfully. First, t-test statistics provide information about the “convergence” of the model. These values are indicators for whether the model has enough power to produce a meaningful distribution of values from which the simulations are generated. In addition, the estimation algorithm becomes unstable if the number of effects is too great. Therefore, if t-values are close to zero, we can assume that the model is replicable and there is relatively low correlation among 95 parameters. Second, standard errors become inflated when too many non-significant parameters are included. Values become very large when too many unimportant effects are included. Therefore, if standard errors are orders of magnitudes greater than the estimate, then the model is considered unstable. Score tests are used to assess how much the model improves when parameters are included in the model, as compared to the model without that parameter. Score tests provide a Goodness-of- Fit measure for specified parameters or set of parameter. This comparison is measured with a t- statistic using a Chi-square distribution to evaluate how much better the data match the theoretical data when that parameter is included in the model. Significance test reflects the model estimates when the parameter is fixed versus the model estimates when the parameter is allowed to run freely. If a score test is significant, it means that the variable improves the model controlling for all other variables in the model. Degrees of freedom indicate the number of parameters that were tested simultaneously. 9.2 Rate and Objective Functions As noted earlier, modeling is based the objective and rate functions. The rate function indicates the rate at which an actor will change their relations or behaviors. The rate function guides the number of “micro-steps” that are modeled between the observed time points where every actor i has an opportunity to change a tie variable x ij . Rate functions are assumed to be constant across all actors. Objective functions represent an individual’s evaluation of their position in the network. For example, for any two actors, i and j, they can change their relationship in one of two ways – create a new relation or dissolve an old connection. The probability of these changes can be represented as follows: Probability ( ,x) = exp (f i ( ,x(i changes relationship with j) exp (f i ( ,x(i changes relationship with h) 96 The general form of the objective function is the strength, , of the individual parameters, s, for actor i. The objective function, f i , is specified as follows: f i ( ,x) = Σ k s ilk where is a statistical parameter, i is the actor, x is the network. k represents the strength of the set of specified parameter effects. In SIENA, the network objective function is used to model selection effects and the behavioral objective function is used to model behavioral influence effects. Network objective function. Three basic effects may be considered from the network function: (1) the attribute-ego effect (the main effect of the ego’s attribute on partner selection), (2) the attribute-alter effect (main effect of alter’s attribute on partner selection), and (3) the attribute- similarity effect (the tendency of actors to select others with similar characteristics). A positive attribute-ego effect indicates that those with higher values on the attribute have a higher number of outgoing friendship nominations. A positive attribute-alter effect indicates that those with higher values on the attribute have a higher number of incoming friendship nominations. A positive attribute-similarity effect indicates that individual’s tend to nominate others with similar values of the attribute (i.e., homophily, selection). The effects of demographic attributes can be examined in the same way where, for example, we can model the effect of gender on outdegree, indegree, and the extent to which girls and boys select others of the same gender. Behavioral objective function. The behavioral objective function corresponds how behaviors of friends influence one’s own behaviors (i.e., assimilation, social influence). The behavioral tendency parameter models the overall tendency toward high values on the behavior variable. The behavioral similarity parameters models tendencies for actors to adopt the behaviors of others. Interactions between behavioral parameters with network or other attribute variables can also be modeled. For example, an interaction between implicit cognition and behavioral similarity 97 indicates that assimilation of friends drug behaviors occurs more so for individuals with high drug cognitions. All the network and behavioral objective functions are listed in Figure 9.1. Objective Function (X) = f i ( ,x) Selection Model Out-degree (density) + Reciprocity+ Transitive triplets+ Balance effect + Σ j x ij + Σ j x ij x ji + Σ j x ij Σ h x ih x hj + Σ j x ij strsim ij + Gender (alter) + Gender (ego) + Gender (match) + Σ jGENDER x ij (z j – z) + Σ j x ij (z i – z)+ Σ j x ij sim ij Ethnicity (alter) + Ethnicity (ego) + Ethnicity (match) + Σ jETHNICITY x ij (z j – z) + Σ j x ij (z i – z)+ Σ j x ij sim ij Age (alter) + Age (ego) + Age (match) + Σ jAGE x ij (z j – z) + Σ j x ij (z i – z)+ Σ j x ij sim ij Alcohol Cognition (alter) + Alcohol Cognition (ego) + Alcohol Cognition (match) + Σ jALCOHOL COGNITION x ij (z j – z) + Σ j x ij (z i – z)+ Σ j x ij sim ij Alcohol Use (alter) + Alcohol Use (ego) + Alcohol Use (match) + Σ jALCOHOL USE x ij (z j – z) + Σ j x ij (z i – z)+ Σ j x ij sim ij Influence Model Alcohol Change Tendency (for network) + Average Alcohol use Similarity (z i – z) + Σ j (x ij sim ij )/( Σ j x ij ) Figure 9.1 Objective Functions for Selection & Influence ERG Models Statistics for each parameter are calculated by how the graph statistic for the observed graph compares to the graph statistic generated by randomly simulated graphs. Simulated networks are analyzed as a distribution of possible graphs, and a mean and standard error for each configuration is generated (Snijders & van Duijn, 1997; Robins et al., 2001). By comparing the graph statistics of the observed network with the mean and standard error of the simulated graphs 98 – it is possible to assess how well the model fits the true data. This comparison is measured using a t-statistic. t-statistic for configuration =(graph statistic of observed graph) – (mean of graph statistic for 1000 simulated graphs) standard error of graph statistic for 1000 simulated graphs Parameters indicate whether the observed data exists in certain configurations more so than it would by chance alone. A positive parameter means that the configuration has a greater than random chance of being present in our network, while negative parameter estimates mean the configuration has a lower than random chance of being present in the network. The parameters specified in the proposed dissertation studies are represented mathematically and verbally in Tables 9.1a,b, and c. 99 Table 9.1a Endogenous Network Selection Parameters in Actor-Oriented Modeling Parameter Illustration of social process Network Statistic Description Example of positive parameter Out-degree (density) T1 T2 Σ j x ij The overall tendency to have ties Students tend to become friends (friendships are beneficial) Reciprocity T1 T2 Σ j x ij x ji Tendency to have reciprocated ties Actors prefer to select others who have selected them Transitive triplets T1 T2 Σ j x ij Σ h x ih x hj Tendency towards triadic closure of the local ties Actors prefer to have ties to others who have a common ties Balance effect T1 T2 Σ j x ij strsim ij Tendency to have ties to structurally similar others (structural balance) Actors prefer to have ties to others with a similar set of friendship ties Table 9.1b Exogenous Network Selection Parameters in Actor-Oriented Modeling Parameter Illustration of social process Network Statistic Description Example of positive parameter Attribute alter T1 T2 Σ j x ij (z j – z) Main effect of alter’s behavior where attribute determines in indegree. Girls tend to receive more nominations Attribute ego T1 T2 Σ j x ij (z i – z) Main effect of ego’s property on tie preference where attribute determines outdgree Girls tend to make more nominations Attribute similarity T1 T2 Σ j x ij sim ij Tendency to have ties to similar others (homophilic selection on attribute) Girls tend to prefer ties to other girls (+) (–) (+) (–) (+) (–) (+) (–) (+) (+) (+) (+) 100 Table 9.1c Behavioral Influence Parameters in Actor-oriented Modeling Tables 9a, 9b, & 9c. Parameters for network and behavioral dynamics are listed. Changes in structure captured by the specific parameter are illustrated and described mathematically and verbally. Circles represent individuals in the network and arrows represent directed nominations. The sign over the arrow indicates whether the change would result in a positive (+) or (-) parameter estimate. The color of the circles represents behavior: red and blue circles suggest different levels of a dichotomous behavior, and black and white circles are used to represent scenarios where behavior could be either level. Statistically speaking, parameter estimate values can be read like parameters and standard errors of standard logistic regression. The parameter estimate divided by the standard error can be used as a measure of the significance of the statistic. If the parameter value is normally distributed, then 95% of the time the actual parameter value will be within 1.96 standard errors of either side of the parameter estimate. In following with standard statistical practices, we can assume that the estimate is statistically significant when it is at or below the p<0.05 level. Importantly though, these coefficients do not have units of measure and therefore, cannot be directly interpreted. As an elaboration of estimation calculations, some important points should be considered. First, of all the parameters listed, only those that have specific dyadic dependencies address selection and influence hypotheses. Attribute similarity effects test selection, that is, does the data show tendencies to select others with similar behaviors more than is expected to occur in random networks? Only cases where there is a change in network linkages are considered in these Parameter Illustration of social process Network Statistic Description Example of positive parameter Tendency (linear) (z i – z) Basic tendency to adopt behavior Students tend to be alcohol drinkers Behavioral Assimilatio n (average) Σ j (x ij sim ij ) /( Σ j x ij ) Assimilation to alters’ behaviors Students tend to become similar to the behavior of their friends (-) (+) 101 parameter estimates. A positive parameter indicates that either: (1) individuals with similar behaviors are more likely to become friends, or (2) friends with dissimilar behaviors dissolve their friendship tie. A negative parameter estimate indicates that either: (1) friends who shared behaviors at time 1 dissolved their friendship at time 2, or (2) friends with dissimilar behaviors became friends at time 2. Behavioral assimilation estimates test influence, that is, do individuals adopt the behaviors of their friends? In this case, influence assessments can only be made from pairs of individuals who are friends at time 1. These friends can either share the behavior or have different levels of the behavior (in these data, be ‘0’ or ‘1’). A positive parameter indicates that friends have similar behaviors at T2 by one of three general paths: (1) individuals are similar at T1 and remain similar at T2, (2) individuals are dissimilar at T1 and one adopts the behavior of the other at T2, and (3) individuals are similar at T1 and both individuals adopt a new behavior at T2 (see Figure 9.2). Negative parameter estimates indicate that friends either retain dissimilar behaviors or become dissimilar at T2 by either one person adopting a new behavior, or both individuals adopting new opposing behaviors. Figure 9.2 Dyadic Behavioral Assimilation Configurations Figure 9.2 Tendencies for dyadic configurations are considered as a function of attritbute homophily. Generalized dyadic configurations with attributes are shown, where red and blue represent different levels of an attribute (e.g., ‘0’ and ‘1’). Six variations are shown, three of which would lead to positive parameter estimates, and three that would result in negative parameter estimates. (+) (+) (+) (–) (–) (–) T1 T2 T1 T2 102 In these studies, the first set of models use selection (network) objective functions to test the hypothesis that there is a tendency for selection of friendships based on drug behaviors and cognitions. The second set of models use social influence (behavioral) and social selection (network) objective functions will be modeled as interacting or coevolving systems. 9.3 SIENA Modeling Assumptions There are three important assumptions about the models that have been factored into the dynamics based on theory and practicality (Snijders, 1995, 2001, 2005): (1) actors have complete control over their behaviors and relationships and with complete knowledge of everyone else’s behaviors and relations, (2) changes are modeled as stochastic processes, which means that changes in the network ties at any moment depends only on the current network configuration, not on previous configurations, and (3) change is modeled as mini-steps that are ongoing between observed states. These assumptions are outlined below, but a detailed mathematical account of the network dynamics model is given in Snijders et al. (2007) and in Steglich et al. (2007). (1) Actor-oriented models. Actor-oriented decisions are based on the individual’s point of view, where actors are assumed to have full knowledge of the network. Each actor controls their outgoing ties in response to the current network structure and the behavior of other actors in the network. Actors in the network make and carry out decisions about the formation (and dissolution) of relational ties. Individuals evaluate the gains and losses each decision costs as specified by the objective function, and they make decisions in order to optimize his or her position. Every person in the network has the same objective function, as well as a residual unknown element, modeled as a random deviation (Burk et al., 2007). (2) Stochastic model. Dynamic network models use continuous-time Markov chain models to represent changes between discrete time points. Changes between observations are imputed as a result of the current network configuration, not on previous configurations. This means that there 103 are no assumptions about how the network came to its current position, and therefore, objective functions cannot evolve. (3) Continuous model. Dynamic network models assume that continuous mini-steps occur between measured time points. The assumption is that the observed data are snapshots of constantly occurring dynamics, and that each snap shot is modeled as the cumulative result from a series of smaller related changes over time. At each mini-step, one (and only one) tie or behavioral change can occur. For example, in one mini-step, a non-user can initiate marijuana use, but they cannot dissolve their non-using friends simultaneously. The structure of the network and attributes of the actors determine which outgoing tie is changed at each mini-step. A description of many of the assumptions applied in SIENA simulation procedures and model design decisions as well as their implications on the inferences made in these studies are described in detail in Appendix A. 9.4 Hypothesis Testing Theory driven forward stepwise model selection procedure is used. Test for inclusion of higher order structures. Basic structural features of the social network are explored with parameter estimates. Endogenous determinants include density, reciprocity and two triadic configurations - transitivity and balance. Accounting for higher order social structures may be critical in modeling social exchanges. The first model-building step is to test whether higher order social dependencies explain network configurations. To justify their inclusion, score tests indicate whether the model is improved when accounting for dyadic and triadic configurations. Test for inclusion of attributes. Attributes that have been shown to strongly guide friendship and influence are included as controls in the model. Covariates include gender, age, SES and 104 ethnicity. Gender and ethnicity are controlled for their affect on in-degree (alter effect), out-degree (ego-effect) and homophily (similarity effect). Test for inclusion of behaviors. Drug behaviors are added into the model to assess whether people who have similar drug behaviors are more likely to nominate one another. Drug behavior influence tests whether people’s behaviors become more like their friends. Test for inclusion of cognitions. Drug cognitions are included in the model to assess whether implicit cognitions influence the social relational or behavioral changes. The network dynamic model is expected to show that students tend to nominate other students who have similar implicit cognitive tendencies. Missing Data. Missing data were treated in SIENA by setting values to 0, such that their influence on the estimation is minimized. Missing data was not imputed because there are many arguments against this practice, and not enough information on imputation methods in SIENA. 105 CHAPTER 10 SIENA NETWORK MODELING RESULTS 10.1 Descriptive Statistics Participants. Of the 574 participants and 75 classrooms in the main study population, only 267 students in 16 classes are included in the network analysis. The criteria for inclusion was that the classroom had to have at least 10 students and at least 60% of the original roster list had to have filled out the survey at both time points Of the 16 classroom network that et this criteria, 94 are in the standard condition, 119 are in the network condition, and 54 are in the control condition. Descriptive statistics are presented in Table 10.1. None of the variables significantly differed from the larger population. Table 10.1 Population Demographics & Attributes for SIENA Analysis Standard Condition Experimental Condition Control Condition Total n 94 119 54 267 # classrooms 6 7 3 15 average# per class 16 17 18 17 Age 16.4 (1.0) 16.3 (1.6) 17.1 (1.6) 16.6 14-16 0.43 0.53 0.28 0.48 17-21 0.57 0.47 0.72 0.52 Gender, %male 0.46 0.48 0.69 0.47 Ethnicity, %Hispanic 0.56 0.66 0.47 0.61 Socio-economic status 0.81 (.04) 0.87 (.03) 1.11 (.04) 0.93 # best friends in school 2 2.2 2.1 2.1 density 0.033 0.028 0.071 0.044 out-degree 3.8 (.2) 4.1 (.1) 4.6 (.1) 4.1 (.1) in-degree 3.2 (.2) 3.6 (.2) 3.9 (.4) 3.5 (.1) max indegree 11914 14 DEMOGRAPHICS & ATTRIBUTES NETWORK PROPERTIES 106 There were six standard and seven experimental condition classes, which are shown in Figure 10.1-10.2. Network diagrams are show friendship links at the two time points, with red lines representing ties present at time 1, blue lines representing ties at time 2 and black lines representing ties present at both times 1 and 2. These images are presented purely to give a visual sense of the networks rather than conveying network data. Figure 10.1 shows experimental classrooms with alcohol cognition and alcohol behavior data, and Figure 10.2 show standard classes with marijuana cognitions and behaviors. In the both figures, circles are users and squares are non-users, and nodes shown with ‘0’s are students who reported no drug cognitions at baseline, while those with a ‘1’ did report drug cognitions. These images show relatively dense networks and a good balance of network change to network stability from time points. Clusters of students with high cognitions and high behaviors can be explored by simply observing these networks. However, it also becomes apparent that identifying influence and selection dynamics by visualization alone would be difficult, if not impossible. 107 Figure 10.1 Illustration of Friendship Networks & Alcohol Activities in Experimental Condition Classes Figure 10.1 Graph representations represent 7 experimental condition classes with alcohol activities. Lines represent directed friendship ties at the two time points, with red lines representing ties present at time 1, blue lines representing ties at time 2 and black lines representing ties present at both times 1 and 2. Nodes represented by circles are alcohol users, squares are non-drinkers, and nodes shown with a value of ‘1’s are those who reported alcohol cognitions at baseline, while those with a ‘0’ did not report cognitions. 1 0 1 1 0 0 1 0 1 1 1 1 1 1 0 1 1 0 1 1 1 1 1 1 1 1 0 1 0 0 1 0 1 0 1 0 1 0 0 1 1 1 0 0 1 1 0 1 0 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 1 1 1 0 1 1 0 1 0 1 1 0 1 1 0 1 1 0 1 1 1 0 0 0 1 1 1 1 0 1 1 1 0 1 0 1 0 1 1 1 0 0 0 0 1 1 0 1 1 0 1 0 1 T1 T2 T1&T2 User 1 Cognition 108 Figure 10.2 Illustration of Friendship Networks & Marijuana Activities in Standard Condition Classes Figure 10.1 Graph representations represent 6 standard condition classes with marijuana activities. Red lines represent friendship ties at time 1, blue lines are ties at time 2 and black lines represent ties present at both times 1 and 2. Nodes represented by circles are marijuana users, squares are non-smokers, and nodes shown with a value of ‘1’s are those who reported marijuana cognitions at baseline, while those with a ‘0’ did not report cognitions. In order to increase power for the estimation routine, it was decided that classes would be combined. The research hypotheses are structured around program conditions, and therefore, 1 1 1 1 0 1 1 1 1 1 0 1 1 0 0 1 1 1 1 1 0 0 1 1 1 1 0 0 0 1 0 0 0 0 1 1 1 1 1 1 1 1 0 1 0 0 1 1 1 0 1 1 1 1 0 0 1 0 1 0 0 1 0 1 1 1 1 0 0 0 0 1 0 0 1 0 0 0 1 1 1 0 1 1 0 0 1 1 1 0 0 1 1 1 T1 T2 T1&T2 User 1 Cognition 109 classes within each condition are combined into a single network. Classrooms defined the networks and then classes within each program were then combined into large matrices. SIENA is able to account for the fact that no linkages are possible between classes, retaining the delineation of each network. “Structural zeros” are defined for cells where individuals in different classes meet in the matrix. Structural zeros indicate to the software the adjacency matrices will be composed of diagonal blocks consisting for each classroom while the off- diagonal cells represent impossible relations. The software knows not to include these in parameter estimates. Grouping classes by condition into separate networks does assume, though, that selection and influence mechanisms across classes receiving the same program condition will function similarly. The parameter estimates, thus, reflect patterns found across classes, and comparisons across program conditions reflect differences in these generalized patterns. This approach assumes that social mechanisms are a direct result of the program condition, and that other factors driving these exchanges can be ignored. Of course, this approach may not account for other class or school level influences, but was determined to be the best way to increase the effectiveness of the actor-oriented network analyses. The next section shows results from the stepwise model selection procedure. All calculations were done using SIENA version 3.2 (Snijders et al., 2007). SIENA models must reach convergence to be valid. Three indicators are provided by SIENA output to determine adequate convergence. First, deviations between simulated and observed values for parameter estimates should be close to zero (t-scores had to have an absolute value < 0.2). Second, parameter estimates were within the same order of magnitude as their standard errors. Third, colinearity data were examined to identify variables that represent the same data patterns. 110 Models that did not converge were adjusted by removing variables to reach convergence. All network selection and influence models presented here meet criteria for convergence. Statistical rules that are applied to regression analyses hold true for actor-oriented network analysis estimates. Two-sided t-scores are calculated by dividing the parameter estimates by standard errors, which correspond to the standard deviations from the mean. Parameter estimates are normally distributed and therefore, critical values were +/- corresponding to an α of 0.05 and +/-1.645, corresponding to an α of 0.10. It should be noted that actor-oriented modeling is presented as trends over time, and is often not limited only by statistically significant findings. Network modeling simulates artificial data, which acts as comparisons to observed data in order to identify whether the observed data has configurations more so than do random networks. The parameter estimates reveal whether there are statistically significant tendencies towards dynamic pattern, but additionally, indicate whether the objective functions reliably capture data. Therefore, convergence of a model in SIENA alone suggests tendencies in the dynamic configurations of the observed data. When reporting results, the focus is on statistically significant findings, but for the central hypotheses, the direction of the estimate (whether it’s negative or positive) is also considered. 10.2 Model Selection Procedures Model selection was conducted with a stepwise elimination process. Endogenous and exogenous covariates were modeled with either alcohol or marijuana variables. Variables are tested for their significance, along with cognition and behavioral variables so that only the important network selection controls are included in the complete model. Actor-based modeling is limited by its complexity, and it is recommended that only social features that are shown to 111 drive relational configurations are included (Snijders, Spreen & Zwaagstra, 1995; Steglich Snijders & Pearson, 2006). First, higher order structural effects are modeled to identify whether dyadic and triadic dependencies are present in the data. If they are, a network approach rather than a correlative (non-dependency framework) is justified. Outdegree (density) effects are controlled for in these models but are not reported because density (the proportion of links to total possible links based on number of nodes) acts as the basic constraint of node to linkage proportions. To control for other natural structural dependencies, reciprocity, balance, and transitive triplets are accounted for. Next, a forward stepwise model selection procedure is conducted to identify important exogenous parameters (Snijders et al., 2007; van Duijn et al., 2003). Demographics (gender, age, ethnicity, and SES) and an indicator of friendship importance (the number of one’s best friends who attend the school) are tested for their impact on friendship formation. For model selection procedures, alcohol and marijuana behaviors are also included in these models, but their values are not reported because the purpose is to find attributes that account for network dynamics that are drug activities cannot explain. Sets of three parameters are used to identify different aspects of network selection processes for each variable of interest: (1) ego-attribute (the degree to which network ties form on the basis of the ego’s attribute), (2) alter-attribute (the extent to which network ties form on the basis of the alters’ attribute), and (3) similarity-attribute (the extent to which ties form as a result of the ego and alter possessing a common attribute). If the set of parameters representing the attribute was found to be significant, all three parameters remained in the model (even if only one of the parameters was significant). If a set was not found to play a significant role, all three parameters were eliminated from the model. Each set of variables was tested for statistical significance while all other variables were included in the model. Score-type tests were generated for the omitted 112 parameters (by constraining them to zero) to check whether the parameters, in fact, improved the model. The significance of the parameters was confirmed again after removing each non- significant parameter one-by-one in a backward selection procedure. Tables 10.2 – 10.4 show the model selection results for the three program conditions (standard intervention, experimental intervention and control condition). These c-score statistics measure the significance of the set of parameters (ego, alter, similarity) corresponding to a single attribute (e.g., gender) on network selection. C-scores are generated from score tests where the indicated variable is held constant and tested for its contribution to the model. Degrees of freedom for the score tests are 1 unless indicated in parantheses. P-values that are significant at the 5% level are highlighted in yellow, and those significant at the 10% significance level are highlighted in orange. Variables in bold encompass the sum of the alter, ego and similarity parameters as a group. 113 Table 10.2 Score Tests for Endogenous and Exogenous Network Parameters in HS Standard Condition Classes Table 10.2 Model selection identifies attributes that drive friendship selection in the standard condition classes. Data are presented for 5 classroom networks, with a total of 95 students. C-scores, p-values, and degrees of freedom show that dyadic, triadic, ethnicity alter, friendship importance ego, and similarity parameters were significant at the 5% significance level for the standard condition. Age was moderately significant, while gender and SES were not significant. c-score p-value c-score p-value Dyadic 24.8 <0.0001 27.7 <0.0001 reciprocity 24.8 <0.0001 27.7 <0.0001 Triadic 31.1 (2) <0.0001 25.7 (2) <0.0001 transitive triplets 8.8 0.0029 6.3 0.012 balance 30.3 <0.0001 24.4 <0.0001 Gender 2.1 (3) 0.5501 1.2 (3) 0.7378 gender alter 1.2 0.2713 0.7 0.4136 gender ego 1.2 0.2832 0.7 0.4161 gender similarity 0.0 0.8392 0 0.9874 Ethnicity 5.4 (3) 0.1442 6.1 (3) 0.1052 ethnicity alter 4.5 0.0345 4.6 0.0322 ethnicity ego 0.2 0.6918 0.1 0.7442 ethnicity similarity 0.2 0.6846 0.3 0.5952 Age 6.4 (3) 0.0941 6.3 (3) 0.0994 age alter 0.9 0.3306 1.7 0.1979 age ego 3.3 0.0674 3.6 0.0571 age similarity 1.9 0.1659 1.3 0.2543 Socio-economics 0.6 (3) 0.9041 0.96 (3) 0.8105 ses alter 0.2 0.6652 0.1 0.7473 ses ego 0.3 0.5774 0.42 0.5132 ses similarity 0.0 0.9254 0.3 0.5607 Friendship Importance 16.7 0.0008 16.8 (3) 0.0008 friend importance alter 1.5 0.2182 1.1 0.298 friend importance ego 5.0 0.0257 4.5 0.0339 friend importance similarity 9.3 0.0022 9.2 0.0024 Endogenous Structural Dependencies Exogenous Structural Dependencies Alcohol Marijuana 114 Table 10.3 Score Tests for Endogenous and Exogenous Network Parameters in HS Experimental Intervention Classes (n=119) Table 10.3 Model selection identifies attributes that drive friendship selection in the experimental condition classes. Data are presented for 7 classroom networks, with a total of 119 students. C-scores, p-values, and degrees of freedom show that dyadic, triadic, gender similarity, and ethnicity alter parameters were significant at the 5% significance level for the experimental condition. Ethnicity similarity was moderately significant for marijuana, and SES similarity was for both alcohol and marijuana. Friendship importance and age were not significant predictors of friendship selection. c-score p-value c-score p-value Dyadic 59.1 <0.0001 61.1 <0.0001 reciprocity 59.1 <0.0001 61.1 <0.0001 Triadic 90.6 (2) <0.0001 80.0 (2) <0.0001 transitive triplets 57.4 <0.0001 51.2 <0.0001 balance 59.9 <0.0001 53.1 <0.0001 Gender 18.3 (3) 0.0004 16.1 (3) 0.0011 gender alter 1.4 0.2313 2.0 0.1614 gender ego 1.1 0.300 0.8 0.3599 gender similarity 17.1 <0.0001 15.4 <0.0001 Ethnicity 5.3 (3) 0.1509 8.0 (3) 0.0452 ethnicity alter 4.2 0.0415 4.9 0.0263 ethnicity ego 0.2 0.6222 0.0 0.8455 ethnicity similarity 1.7 0.1964 2.8 0.0968 Age 2.2 (3) 0.5242 3.1 (3) 0.3717 age alter 0.5 0.4704 0.2 0.6509 age ego 1.7 0.1876 2.5 0.1134 age similarity 0.6 0.4433 0.7 0.4071 Socio-economics 6.1 (3) 0.1079 6.7 (3) 0.0813 ses alter 0.1 0.757 0.7 0.3912 ses ego 1.5 0.2209 1.7 0.1968 ses similarity 3.1 0.076 3.1 0.0787 Friendship Importance 2.9 (3) 0.4043 2.0 (3) 0.5671 friend importance alter 0.4 0.5248 0.4 0.5456 friend importance ego 0.8 0.3702 0.2 0.6128 friend importance similarity 1.1 0.2978 1.1 0.2935 Exogenous Structural Dependencies Alcohol Marijuana Endogenous Structural Dependencies 115 Table 10.4 Score Tests for Endogenous and Exogenous Network Parameters in HS Control Condition Classes (n=53) Table 10.4 Model selection identifies attributes that drive friendship selection in the control condition classes. Data are presented for 3 classroom networks, with a total of 53 students. C-scores, p-values, and degrees of freedom show that dyadic, triadic, age alter, gender alter and friend importance alter were significant at the 10% significance level for both marijuana and alcohol. Ethnicity and SES were not significant predictors of friendship selection. c-score p-value c-score p-value Dyadic 12.4 0.0004 10.2 0.0014 reciprocity 12.4 0.0004 10.2 0.0014 Triadic 30.45 (2) <0.0001 26.8 (2) <0.0001 transitive triplets 22.9 <0.0001 20.8 <0.0001 balance 12.2 0.0005 11.2 0.0008 Gender 5.4 (3) 0.1425 7.0 (3) 0.0734 gender alter 3.5 0.0625 4.1 0.0441 gender ego 1.8 0.1774 3.4 0.0644 gender similarity 0.8 0.3619 1.0 0.3192 Ethnicity 2.8 (3) 0.4201 3.0 (3) 0.3906 ethnicity alter 1.8 0.1853 1.9 0.1702 ethnicity ego 0.2 0.6843 0.4 0.5254 ethnicity similarity 2.2 0.1413 2.0 0.1554 Age 7.9 (3) 0.0488 0.4 (3) 0.0378 age alter 5.7 0.0165 6.2 0.013 age ego 2.8 0.0926 2.6 0.1047 age similarity 0.2 0.6191 0.1 0.7617 Socio-economics 0.8 (3) 0.855 0.9 (3) 0.8295 ses alter 0.4 0.5285 0.4 0.5135 ses ego 0.3 0.5967 0.3 0.5845 ses similarity 0.0 0.908 0.0 0.8651 Friendship Importance 3.7 (3) 0.2948 4.8 (3) 0.189 friend importance alter 3.5 0.0628 4.7 0.0308 friend importance ego 0.0 0.923 0.0 0.9459 friend importance similarity 0.3 0.5851 0.1 0.7093 Alcohol Marijuana Exogenous Structural Dependencies Endogenous Structural Dependencies 116 As expected, endogenous structural dependencies are found in all the networks. There are strong tendencies towards dyadic (reciprocal) and triadic (transitive triplet and balance) formations. Networks do differ in their exogenous structural dependencies. Standard Condition. There were strong tendencies towards reciprocity, balance, and transitive triples in the standard condition. Ethnicity (alter) was the only significant demographic variable and age (ego) was moderately significant. Significance was confirmed with score tests for ethnicity (p=0.035 and p=0.032, for alcohol and marijuana, respectively) and for age (p=0.094 and p=0.099 for alcohol and marijuana, respectively). When SES and gender were removed from the model, an additional model was run to check whether age became significant at the 0.05 level upon dropping other effects from the model. Age remains to be only moderately significant, and was removed from the model. Best fitting models were those that included ethnicity and friendship importance, while gender, age, and SES were removed as they were determined to be non-significant contributors to the network structure. Experimental Condition. There were strong tendencies towards reciprocity, balance, and transitive triplets in the experimental condition. Gender (similarity; p=0.0004 and p<0.0001, for alcohol and marijuana, respectively) and ethnicity (alter; p=0.0415 and p=0.0263 for alcohol and marijuana, respectively) were significant parameters. SES was significant using an alpha of 0.10 (similarity effect; p =0.076 and p=0.079 for alcohol and marijuana, respectively). Additional models were run to check whether SES became significant upon dropping non-significant effects from the model. SES remains to be only moderately significant, and was removed from the model. Age, SES, and friendship were determined to be non-significant contributors to the network and were eliminated from the model. Control Condition. For the control condition, there were strong tendencies towards reciprocity, balance, and transitive triplets. Age (alter) was the only consistently significant 117 attribute across alcohol and marijuana models (p = 0.0165 and p = 0.013, respectively). Gender and friendship were significant only for one outcome or the other. After removing non-significant parameters, gender became significant for both marijuana and alcohol models. SES, ethnicity, and friendship were determined to be non-significant contributors to the model. Best fitting model includes gender and age as the network selection control variables. 10.3 Dynamic Network Selection Modeling Dynamic network selection models include parameters testing attributes and behaviors that drive friendship selection. Complete models are run, which include significant covariates (based on model selection phase described in the previous section), and alcohol or marijuana behaviors and cognitions to investigate the core hypotheses. In addition, tendencies towards reciprocity, balance, and transitive triplets were controlled for in all models. Controlling for endogenous effects demonstrates with statistical probability that cognitive or behavioral dependencies identified in the data cannot be “explained away” by tendencies for friendship patterns associated with other attributes. Statistical significance of the parameters is determined by dividing the estimates by their standard errors. A value of 1.96 or higher approximately corresponds to significance at an α level of 0.05, while a value of 1.645 corresponds to significance at an α level of 0.10. A direct comparison of the size of parameter estimates is possible, given that they occur in the same linear combination in the objective function, but it should be kept in mind that these are unstandardized coefficients. The basic rate parameters are the expected frequencies between any pair of successive waves within which actors have the opportunity to change a network tie. For these parameters, no p-values are given in the tables, as testing that they are zero is meaningless (if rate would be zero, there would be no change at all). Jaccard values imply that change between waves 118 is relatively low, but not too low for the applicability of the model. Tables 10.5-10.7 display results examining social selection processes for each intervention condition. Objective Functions. As already discussed, SIENA modeling requires that we transform variables into three parameters termed alter-attribute, ego-attribute and similarity-attribute. Parameter estimates provide information about the tendency for a network configuration to occur as defined by the objective function for that parameter. In addition to the magnitude of the tendency and whether the tendency occurs more than would be expected to by chance, comparisons between parameters provide an overall evaluation score for specific actor decisions. Objective functions for ego, alter, and similarity can be thought of as three features of a single decision that when combined reflect the extent to which the actor prefers that choice relative to the other choices. In this way, homophilic tendencies versus heterophilic can be easily demonstrated for dichotomous decisions. As an example, gender effects can be evaluated by multiplying the average values for gender by parameter estimates for gender-alter, -ego, and -similarity. The sum of these effects reflect the role of gender in friendship selection: Evaluations for homophilic (males choose males, females choose females) and heterophilc (males choose females, females choose males) decisions can be calculated. For the experimental condition in which gender was statistically significant, homophilic choices were evaluated. Females choosing female friends was evaluated at = 0.42, and male choosing male friends was evaluated at 0.53. Males selecting females was evaluated much lower at -0.3, while females selecting males was assessed at 0.35. In other words, students prefer friendships with others of the same sex, particularly for boys, who perceive negative benefits by adding a female as a friend. Calculations for these relationships are presented as an example from the experimental condition in Figure 10.3. 119 Figure 10.3 SIENA Actor-Oriented Evaluation Functions for Gender Selection Average male, female and similarity are generated in the Experimental selection model. Gender alter, ego, and similarity parameter estimates are generated for the marijuana model. These values are inserted into the objective function to generate evaluation for M-->M, M-->F, F-->M, F-->F selections, where M = male, and F = female. To derive the average value for male, subtract the average gender, 0.521, from 1 (male is coded as ‘1’ in data and female is coded as ‘2’) and average female = 2-0.521. Objective Function = Σ jGENDER x ij (z j – z) + Σ j x ij (z i – z)+ Σ j x ij sim ij Male --> Male = ego * (M) + alter * (M) + similarity * (1) = 0.41*(-0.52) + -0.30*(-0.52) + 0.96*(0.50)*(1) = 0.42 Female --> Male = ego * (F) + alter * (M) + similarity * (0) = 0.35 Male --> Female = ego * (M) + alter * (F) + similarity * (0) = -0.36 Female --> Female = ego * (F) + alter * (F) + similarity * (1) = 0.53 Figure 10.3 Evaluation functions for the role of gender in the selection of friends. The evaluation for male-male selection is found to be 0.42, while females evaluate friendships with other females at 0.53. Objective functions are also evaluated for endogenous parameters. For example, in the standard condition selection models that will be presented in the next section, the density (outdegree) function is -2.72, suggesting that on average, costs of arbitrary ties is higher than their benefits. The reciprocity effect is +1.01, suggesting adding reciprocated tie is beneficial. While adding a reciprocated tie corresponds to positive benefits, the overall cost of generating the tie is Variable Average Gender Parameter Parameter Estimate Male -0.52 Gender alter -0.30 Female 0.48 Gender ego 0.41 Similarity 0.50 Gender similarity 0.96 Depedency Estimate Observed Data 120 still negative (-2.73 + 1.01 = -1.72). Objective function scores will be reported throughout the results sections, but the calculations will not be shown. Dynamic Network Selection in Standard Condition Classes In the next section, final selection models testing hypotheses 5 and 6. Results will be presented separately for the three program conditions. Table 10.5 shows parameter estimates and standard errors for dynamic network selection within standard intervention condition classes. Data represent five classroom networks, made up of 95 students in total. Alcohol and marijuana parameters are presented in two separate models. Results show that friendship formations are driven by ethnicity (alter), friend importance (ego and similarity), and marijuana cognition (alter) effects. 121 Table 10.5 Friendship Selection Models in Standard Intervention Classes (n=95) *p<0.10; **p<0.05 Table 10.5 Parameter estimates and standard errors for dynamic network selection within the standard condition network are presented. Parameters are estimated across 5 classroom networks, with a total of 95 students. Alcohol and marijuana activities are modeled separately. Results show that friendship formations are driven by ethnicity (alter), friend importance (ego and similarity), alcohol use (alter), alcohol cognition (alter), and marijuana cognition (alter) effects. Rate of network changes. The rate function indicates that based on the estimated number of changes from T1 and T2, there are about three network steps between the first and second Parameter Estimate Standard Error Parameter Estimate Standard Error rate 2.97 2.99 Endogenous Structural Dependencies outdegree (density) -0.62 0.56 -0.61 0.61 reciprocity 1.05** 0.2 1.04** 0.20 transitive triplets -0.14** 0.06 -0.12** 0.06 balance 25.56** 5.04 24.40** 4.90 Exogenous Structural Dependencies Ethnicity ethnicity alter -0.68** 0.27 -0.64** 0.26 ethnicity ego 0.09 0.29 0.1442 0.281 ethnicity similarity 0.37 0.37 0.3301 0.3649 Friendship Importance friend importance alter 0.07 0.05 0.0733 0.0533 friend importance ego -0.13** 0.06 -0.12* 0.07 friend importance similarity 0.80** 0.32 0.81** 0.32 Drug Behavior use alter 0.64* 0.39 0.26 0.27 use ego -0.10 0.31 0.02 0.26 use similarity -0.48 0.44 -0.04 0.36 Drug Cognition cognition alter 0.39* 0.24 0.64** 0.23 cognition ego 0.13 0.33 -0.34 0.51 cognition similarity -0.03 0.38 -0.48 0.53 Alcohol Marijuana 122 observed data. This means that there were approximately three opportunities to either select a new friend, dissolve an old tie, or do nothing at all. Endogenous dependencies. There were tendencies towards reciprocity and balance. Outdegree was negative, as expected, but non-significant, which implies that on average, costs of arbitrary ties is not significantly higher than their benefits. Transitive triplet formations were negative (t=-0.14, p<0.05), indicating that friendship ties were relatively unstructured in the standard condition. Exogenous attribute-based selection. Ethnicity-alter effects indicate that non-Latinos tend to be nominated as friends more so than Latinos. For non-Latino students, the value of a friendship with another non-Latino is estimated to be +0.84, while nominating a Latino is approximated at -0.42. Similarly, Latinos prefer relations with non-Latinos (+0.49) to Latinos (+0.24). Friendship importance (ego and similarity) parameters show that students who report having strong friendships within the school tend to make more friend nominations, and are more likely to be friends with others who report this closeness. Exogenous drug-based selection. Significant alter effects indicate that alcohol users (0.65, t=1.64, p<0.10) and those with strong alcohol cognitions (0.39, t=1.64, p<0.10) and marijuana cognitions (0.64, t=2.76, p<0.05) are more likely to be nominated as friends. Homophily effects are not significant, but estimates are negative for all drug activities, indicating that there is no tendency for students to become friends as a result of similar alcohol or marijuana cognitions or behaviors. Overall, drinkers and those with alcohol and marijuana cognitions are more popular, and in general, students tend not to be drawn to similar others. Dynamic Network Selection in Experimental Condition Classes Parameter estimates and standard errors for dynamic network selection within experimental intervention conditions are presented in Table 10.6. The experimental condition 123 represents 7 classrooms and a total of 119 students. Results show that friendship formations are driven by gender (alter, ego, and similarity), ethnicity (alter), alcohol (alter, ego, and similarity), marijuana use (ego), and alcohol cognition (alter) effects. 124 Table 10.6 Friendship Selection Models in Experimental Intervention Classes (n=119) *p<0.10; **p<0.05 Table 10.6 Parameter estimates and standard errors for dynamic network selection within experimental intervention conditions. Parameters reflect estimates across the 7 classroom networks, with a total of 119 students. Alcohol and marijuana parameters are presented separately. Results show that friendship formations are driven by gender (alter, ego, and similarity), ethnicity (alter), alcohol (alter, ego, and similarity), marijuana use (ego), and alcohol cognition (alter) effects. Rate of network change. The rate function indicates that there were three to four opportunities to select a new friend, dissolve an old tie, or do nothing at all. Parameter Estimate Standard Error Parameter Estimate Standard Error Rate 3.81 3.75 Endogenous Structural Dependencies outdegree (density) -2.72** 0.54 -2.45** 0.50 reciprocity 1.01** 0.16 1.01** 0.16 transitive triplets 0.06 0.04 0.08* 0.04 balance 25.33** 4.07 24.12** 4.09 Exogenous Structural Dependencies Gender gender alter -0.37** 0.18 -0.3016 0.19 gender ego 0.38* 0.20 0.41** 0.20 gender similarity 0.98** 0.28 0.96** 0.28 Ethnicity ethnicity alter 0.18 0.21 0.35* 0.22 ethnicity ego 0.03 0.25 -0.20 0.25 ethnicity similarity 0.31 0.36 0.15 0.35 Drug Behavior use alter -0.44* 0.26 -0.11 0.24 use ego 0.45** 0.21 0.52** 0.22 use similarity 0.56* 0.31 0.27 0.30 Drug Cognition cognition alter 0.35* 0.19 -0.09 0.19 cognition ego -0.4818 0.30 -0.12 0.27 cognition similarity -0.19 0.34 -0.12 0.32 Alcohol Marijuana 125 Endogenous dependencies. The experimental condition classes showed strong tendencies towards reciprocity (dyadic relations) and balance (triadic relations). Transitive triplet effects were significant when included alone (as presented in Table 10.3 in the model selection phase) but become only moderately significant in the full models. Exogenous attribute-based selection. Models showed strong tendencies towards gender homophily and selecting boys as friends were preferable over selecting girls. Selecting same gender friends is evaluated at +0.49 for boys, and +0.50 for females, compared to opposite gender friends, evaluated at +0.37 for girls selection of boys, and -0.37 for boys selecting girls. Ethnicity becomes non-significant. Exogenous drug-based selection. In contrast to the standard condition findings, alcohol alter effects in the experimental condition show that there is a preference for non-drinkers (- 0.44, t=1.67, p<0.10). In addition, similarity effects show that students with similar drinking status are more likely to be friends (0.56, t=1.78, p<0.10). While popular students tend to be non-drinkers, surprisingly they also tend to have strong cognitions towards alcohol use (0.35, t=1.84, p<0.10). Unlike the standard condition, homophily for behaviors tended to occur (as indicated by the positive similarity parameter estimates), and consistent with the standard condition, parameters for cognitive homophily show negative tendencies. Overall, similarities in alcohol use drives friendship selection, and popular students tend to be non-drinkers, but with strong alcohol cognitions. Network Selection in Control Condition Parameter estimates and standard errors for dynamic network selection models within control condition classes are presented in Table 10.7. The control condition represents 3 classrooms with a total of 53 students. Results show that friendship formations are driven by 126 gender (alter, ego, and similarity) effects, alcohol and behavior ego effects, and alcohol cognition alter effects. Table 10.7 Friendship Selection Models for Control Classrooms (n=53) * p<0.10; ** p<0.05 Table 10.7 Parameter estimates and standard errors for dynamic network selection within control condition classes are presented. Parameters are modeled across 3 classrooms with a total of 53 students. Alcohol and marijuana parameters are presented separately. Results show that friendship formations are driven by age (alter) and gender (alter) effects. Parameter Estimate Standard Error Parameter Estimate Standard Error Rate 3.64 3.62 Endogenous Structural Dependencies outdegree (density) -1.23** 0.62 -2.10** 0.60 reciprocity 0.45** 0.23 0.46** 0.23 transitive triplets 0.10 0.06 0.08 0.06 balance 10.83** 2.69 10.73** 2.67 Exogenous Structural Dependencies Gender gender alter -0.60** 0.29 -0.55** 0.27 gender ego 0.08 0.39 -0.24 0.40 gender similarity 0.38 0.45 0.22 0.46 Age age alter 0.99** 0.47 1.21** 0.47 age ego 0.31 0.48 0.24 0.46 age similarity -0.67 0.46 -0.68 0.46 Drug Behavior use alter 0.32 0.41 -0.31 0.34 use ego -0.35 0.31 -0.29 0.32 use similarity -0.44 0.46 0.57 0.42 Drug Cognition cognition alter -0.02 0.28 0.19 0.28 cognition ego -0.15 0.36 0.28 0.33 cognition similarity -0.21 0.42 0.07 0.40 Alcohol Marijuana 127 Rate of network change. The rate function indicates that there were three to four opportunities to select a new friend, dissolve an old tie, or do nothing at all. Endogenous dependencies. The networks show significant reciprocity and balance dependencies. Transitive effects are not significant. Attribute-based selection. Males and older students tend to be nominated more than females and younger students. For males, the value of selecting a male friend is about +0.77 compared to a -0.37 value of adding a female friend. For older students, the value of selecting another older student is 0.97 while the value of befriending a younger student is -0.72. Exogenous drug-based selection. There were no significant alcohol or marijuana effects. The negative parameter estimate indicates that there are tendencies for heterophily around alcohol activities, while there are homophilic tendencies around marijuana activities. 10.4 Summary of Network Selection Models In the standard condition, drinkers and those with strong drug-related cognitions are more popular. In contrast, the experimental condition show that popular students tend to be non- drinkers, but with strong alcohol cognitions. The standard condition shows that students tend not to be drawn to similar others, while the experimental condition shows that students tended to select friends with similar behaviors, but not similar cognitions. The control condition had no significant drug-related selection effects. 10.5 Network & Behavioral Co evolution Models Network dynamic models integrate dependencies between network evolution and behavioral evolution. This technique uniquely enables researchers to statistically identify social influence (whether the friendship causes the behavior) and selection (whether the behavior causes the selection of friendships) effects separately. In these models, students can select or dissolve friendships, and behaviors can be initiated or terminated. Behavioral 128 influence is operationalized as assimilation, which is modeled as the tendency to adopt the average value of your friends’ behavior (or cognition). Tables 10.8-10.10 show coevolution modeling results for standard, experimental and control conditions. Coevolution models are identical to network selection models presented in section 10.4- 107 with the introduction of influence parameters. Covariates consist of the same variables found to be significant in the forward stepwise selection procedure. The first step of the analysis was to test influence parameters only, while omitting drug-related selection parameters. These analyses were conducted to determine whether influence effects might be significant when models do not simultaneously account for drug-related friend selection. However, the results showed no significant variations in the important (hypothesis-driven) parameter estimates compared to the models of coevolution of drug selection & influence. Therefore, only the final models, which include both selection and influence parameters are presented. The following section shows results for coevolution models for the standard, experimental and control networks. Selection & Influence Coevolution in Standard Condition Network and behavior coevolution models are presented for the standard condition (5 classroom networks, n=191) in Table 10.8. Selection and influence parameters are estimated simultaneously through modeling strategies that assume they are dynamic, interacting mechanisms. Results show that ethnicity (alter), friend importance (ego and similarity), and marijuana cognition (alter) effects determine selection processes, while assimilation effects are found for alcohol use. 129 Table 10.8 Selection and Influence Coevolution Models of Alcohol and Marijuana Cognitions and Behaviors in Standard Intervention Classes (n=191) * p<0.10; ** p<0.05 Table 10.8 Network & behavior co-evolution models are presented for the standard condition (5 classroom networks). Selection and influence parameters are presented separately for alcohol and marijuana. Results show that ethnicity (alter), friend importance (ego and similarity), and marijuana cognition (alter) effects determine selection processes, while assimilation processes are significant for alcohol behaviors. Alcohol Parameter Estimate Standard Error Parameter Estimate Standard Error NETWORK SELECTION rate 4.14 4.05 outdegree (density) -1.08 1.19 -0.56 1.99 reciprocity 1.32** 0.35 1.47** 0.30 transitive triplets -0.10 0.10 -0.04 0.07 balance 22.97** 8.35 18.43** 7.73 Exogenous Attribute Dependencies Ethnicity ethnicity alter -0.62* 0.34 -0.63** 0.32 ethnicity ego 0.23 0.34 0.28 0.29 ethnicity similarity 0.38 0.44 0.29 0.41 Friendship Importance friend importance alter 0.09 0.08 0.09 0.06 friend importance ego -0.16** 0.07 -0.15** 0.06 friend importance similarity 0.95** 0.35 0.90** 0.32 Exogenous Alcohol & Marijuana Dependencies Drug Behavior use alter 1.12 0.98 0.27 0.37 use ego 0.25 0.84 0.05 0.42 use similarity -0.66 1.49 0.23 0.54 Drug Cognition cognition alter 0.20 0.30 1.54** 0.71 cognition ego 0.15 0.38 -0.82 2.10 cognition similarity 0.34 0.50 -1.23 2.20 BEHAVIORAL INFLUENCE rate behavior 1.07 0.45 rate cognition 0.53 2.10 Drug Behavior behavior tendency -0.26 0.78 -0.91 0.97 behavior avg similarity 1.99* 1.09 1.49 1.52 Drug Cognition cognition tendency 1.36 3.38 -2.33 1.89 cognition avg similarity 5.98 14.51 -0.53 2.62 Marijuana Endogenous Structural Dependencies 130 Rate of network & behavioral changes. The rate function indicates that there were about four opportunities to select a new friend, dissolve an old tie, or do nothing at all. This is approximately one more opportunity than the results for the selection-only model show in Table 10.2. In comparison to network changes, much fewer influence behavioral assimilation changes or steps were made. Rate functions indicate that there was only about one opportunity between T1 and T2 to change behavior to become similar or dissimilar to one’s friends. There were about two opportunities for cognition changes between time points. Endogenous dependencies. Outdegree remains negative and non-significant in both alcohol and marijuana models. Transitive triple tendencies also remain negative and non- significant. The network still shows strong tendencies towards reciprocity and balance structures. Exogenous attribute-based selection. Consistent with the selection-only analysis, non- Latino students tend to be nominated more, and students who have close friends within the school tend to name more friends in the class, and tend to nominate others who have close friends in the school. Exogenous drug-based selection & influence. Consistent with the selection-only analysis, students with strong marijuana cognitions tend to be selected as friends (1.54, t=2.16, p<0.05). In contrast, alcohol related selection effects are no longer present, while tendencies of friends’ influence on ego’s alcohol behaviors are statistically significant (1.99, t=1.82, p<0.10). That is, participants tend to assimilate or adopt their friends’ alcohol using behaviors. In addition, homophily tendencies for alcohol use, which were negative in the selection only model, become positive. The change indicates that apparent preference for drinkers and those with dissimilar others in the selection-only model is better explained by assimilation processes. Consistent with selection-only findings, the coevolution model shows preferences for those 131 with strong marijuana cognitions and those with dissimilar cognitions. Overall, findings indicate that marijuana cognition drives selection of friends, while friends’ alcohol use influences behavioral choices. Selection & Influence Coevolution in Experimental Condition Network and behavior coevolution models are presented for the experimental condition (7 classroom networks, n=221) in Table 10.9. Selection and influence parameters are assumed to be dynamic, interacting processes. Results show that gender (similarity), and marijuana use (ego) effects determine selection processes, while friends influence around drug activities are not significant. 132 Table 10.9 Selection and Influence Covvolution Models of Alcohol and Marijuana Cognitions and Behaviors in Experimental Intervention Classes (n=221) * p<0.10; ** p<0.05 Table 10.9 Network and behavior co-evolution models are presented for the experimental condition (7 classroom networks). Selection and influence parameters are estimated for alcohol and marijuana separately. Results show that gender (similarity), and marijuana use (ego) effects determine selection processes, while friends influence around drug activities are not significant. Parameter Estimate Standard Error Parameter Estimate Standard Error NETWORK SELECTION rate 5.07 5.17 0.67 outdegree (density) -2.80 1.77 -2.65** 0.75 reciprocity 1.90** 0.44 1.62** 0.25 transitive triplets 0.09** 0.05 0.09** 0.04 balance 22.36** 5.39 21.47** 5.13 Exogenous Attribute Dependencies Gender gender alter -0.23 0.22 -0.21 0.22 gender ego 0.08 0.26 0.07 0.22 gender similarity 0.88** 0.31 0.83** 0.26 Ethnicity ethnicity alter 0.35 0.30 0.45** 0.24 ethnicity ego -0.16 0.36 -0.29 0.28 ethnicity similarity 0.13 0.41 0.19 0.34 Exogenous Alcohol & Marijuana Dependencies Drug Behavior use alter -0.60 0.56 -0.20 0.32 use ego 0.73 0.48 0.48* 0.28 use similarity 0.54 0.61 0.22 0.42 Drug Cognition cognition alter 1.05* 0.64 -0.13 0.37 cognition ego -1.57 1.71 -0.26 0.64 cognition similarity -0.40 1.83 -0.11 0.79 BEHAVIORAL INFLUENCE rate behavior 0.66 0.28 rate cognition 1.97 1.02 Drug Behavior behavior tendency 0.79 0.76 0.05 0.63 behavior avg similarity -1.26 1.88 1.34 1.53 Drug Cognition cognition tendency -1.63** 0.81 -1.04* 0.59 cognition avg similarity 1.76 2.20 1.25 1.54 Endogenous Structural Dependencies 133 Rate of network & behavioral changes. The rate function indicates that there were about five opportunities to select a new friend, dissolve an old tie, or do nothing at all. Similar to the standard condition, coevolution models estimate one more step than the modeling results show for the selection only model. Also consistent with the standard condition, there are much fewer influence behavioral assimilation steps. There was only about one opportunity between T1 and T2 to change behaviors, while rates for alcohol cognitions indicate two steps and marijuana cognition indicates one. Endogenous dependencies. In coevolution models of standard condition classes, outdegree remains negative and shows strong tendencies towards reciprocity, balance and transitive structures. Exogenous attribute-based selection. Consistent with the selection only analysis, students tend to select friends of the same gender. Ethnicity-alter effects show that non-Latinos tend to be selected as friends more so than Latinos. Exogenous drug-based selection & influence. As found in the selection -only model, alcohol cognition-alter effects indicate that students with alcohol cognitions tend to be chosen as friends. With the addition of influence effects, all the alcohol based-selection parameters, including ego, alter, and similarity effects that were significant in the selection-only model, become non-significant. Still, the trends around alcohol use remain the same, where non- drinkers tend to be more popular and heterophilous relationships are preferred. In addition, behavioral tendency parameters indicate that students tend to decrease their alcohol behavior and cognitions over time. No behavioral influence effects were found. Overall, students with alcohol cognitions tend to be more popular, while alcohol behavior based selection become non-significant but continues to trend towards a preference for non-drinkers, in the absence of homophily. 134 Selection & Influence Coevolution Models in Control Condition Network and behavior coevolution models are presented for the control condition (3 classroom networks, n=160) where selection and influence parameters are estimated simultaneously (Table 10.10). Results show that gender (alter) and age (alter) effects determine selection processes, while no significant effects for drug-related selection and influence are found. 135 Table 10.10 Selection and Influence Co-Evolution Models of Alcohol and Marijuana Cognitions and Behaviors in Control Condition Classes (n=160) *p<0.10; **p<0.05 Table 10.10 Network and behavior coevolution models are presented for the control condition (3 classroom networks, n=160) where selection and influence parameters are estimated simultaneously. Data are presented separately for alcohol and marijuana. Results show that gender (alter) and age (alter) effects determine selection processes, while no significant effects for drug-related selection and influence are found. Parameter Estimate Standard Error Parameter Estimate Standard Error NETWORK SELECTION rate 6.85 1.48 6.8494 1.033 outdegree (density) -1.12 2.06 -2.61** 1.01 reciprocity 1.05** 0.40 1.14** 0.43 transitive triplets 0.14* 0.08 0.13** 0.07 balance 6.35** 2.79 5.76** 2.59 Exogenous Attribute Dependencies Gender gender alter -0.44* 0.26 -0.41 0.29 gender ego 0.06 0.36 -0.17 0.40 gender similarity 0.42 0.44 0.28 0.41 Age age alter 0.78 0.54 0.91** 0.43 age ego -0.13 0.45 -0.22 0.49 age similarity -0.24 0.45 -0.25 0.36 Exogenous Alcohol & Marijuana Dependencies Drug Behavior use alter 0.13 0.84 -0.44 0.51 use ego -0.35 0.57 -0.10 0.42 use similarity -0.17 1.24 0.94 0.88 Drug Cognition cognition alter 0.32 0.69 0.38 0.56 cognition ego -1.01 1.96 -0.11 0.72 cognition similarity -0.85 2.35 0.04 1.06 BEHAVIORAL INFLUENCE rate behavior 0.80 0.52 rate cognition 1.57 1.22 Drug Behavior behavior tendency -0.05 1.65 -0.01 0.95 behavior total similarity 0.75 1.78 0.12 3.41 Drug Cognition cognition tendency -5.63 10.03 -1.08 1.25 cognition total similarity -1.24 3.10 3.46 5.72 Endogenous Structural Dependencies Alcohol Marijuana 136 Rate of network & behavioral changes. The rate function indicates that about there were six to seven opportunities to change one’s networks, and one to two to change behaviors. Endogenous dependencies. Outdegree remains negative and there are strong tendencies towards reciprocity, balance and transitive structures. Exogenous attribute-based selection. Consistent with the selection-only analysis, males and older students are preferred as friends more so than younger students and females. Exogenous drug-based selection & influence. There were no significant alcohol or marijuana effects. 10.6 Summary of Network Behavioral Co-evolution Results Coevolution models assess dependencies between network and behavioral processes. Findings related to demographics or other attributes reveal interesting features of the networks, but are only considered as control variables to tease out drug-related influence and selection processes. In the summary of results, only parameters related to the study hypotheses are discussed. Coevolution results will be discussed in greater depth in the chapter 11, particularly as they compare to findings in studies 1 and 2. In the standard condition, students tended to adopt alcohol using behaviors of their friends’. In addition, standard condition classes showed preferences for befriending students with strong marijuana cognitions. In the experimental condition, students with alcohol cognitions tended to be more popular. Additionally, there were trends for selection for non- using friends’ and those with dissimilar behaviors, but these parameters were only significant in the selection-only models (not the coevolution models). There were no significant alcohol or marijuana effects found in the control classes. Summary of Influence & Selection Coevolution Findings. Table 10.11 summarizes alter and similarity effect for alcohol and marijuana activities across conditions. To explore the 137 original hypotheses of the study, homophily (selection) and assimilation (influence) parameters are presented. In addition, alter effects are shown, as they provide context for homophily effects and are an important component of friend selection. Significant findings are described, as are trends found in the intervention groups as they compare to each other and to the control condition. The control condition is used as a baseline measure for social mechanisms, and is considered to reflect natural processes that would occur in the absence of any intervention. Thus, it is plausible to rule these processes out as possible effects due to our intervention. 138 Table 10.11 Coevolution Selection& Influence Parameter Results by Condition Table 10.11 Homophily and assimilation parameters are summarized with only the sign of their estimates shown. Parameters with p-values </=0.10 are indicated with an asterisk (‘*’). Results in red indicate that they differ from the control and the alternate intervention condition. Summary of Influence & Selection Coevolution Homophily & Assimilation Trends. Using the control condition as a baseline measure for effects found in the intervention groups provides useful information about the role of the intervention on social processes. From this perspective, neither intervention condition had any effect on marijuana behaviors or marijuana cognitions. Both interventions showed negative effects for: (1) alcohol cognition homophily, and (2) assimilation of friends’ marijuana cognitions, as compared to the control condition. PROGRAM CONDITION STANDARD EXPERIMENTAL CONTROL n 95 119 53 Alcohol Alter +* - + Alcohol Homophily -+- Alcohol Cognition Alter +* +* + Alcohol Cognition Homophily +- - Marijuana Alter +* - - Marijuana Homophily +++ Marijuana Cognition Alter +* - + Marijuana Cognition Homophily -- + Alcohol Use Assimilation +* - + Alcohol Cognition Assimilation ++ - Marijuana Use Assimilation +++ Marijuana Cognition Assimilation -+ + Alcohol & Marijuana Selection Effects Alcohol & Marijuana Influence Effects 139 In summary, results of the network analysis studies using the control condition as a baseline to measure intervention effects suggest that the standard (teacher-led) program promotes homophily around alcohol cognitions and thwarts marijuana cognition assimilation. Conversely, the experimental (peer-led) program encouraged homophily around alcohol use, and thwarted assimilation of friends’ alcohol behaviors. Both interventions decreased alcohol cognition homophily and increased assimilation of friends’ marijuana cognitions. 140 CHAPTER 11 REVIEW OF FINDINGS ACROSS MODELS, ACROSS METHODS Applying social selection and social influence mechanisms to observed data first as separate paradigms, and then, as a single complex model shows scenarios where social patterns exist, but become obscured by other processes that are occurring simultaneously. Selection mechanisms revealed in the network models mostly disappear in the coevolution models. In addition, applying different methodologies to answer the same research questions in the same populations show that dynamic models find fewer significant relationships. Together, results from different analytic models derived from different methodologies can be pieced together to uncover different aspects of complex social mechanisms. Network models focus on what drives actor’s decisions, while regression models tease out causes of the behaviors at the population level. 11.1 Comparison of Network Analysis Results (Studies 2 and 3) For the most part, coevolution models show patterns similar to those found in the selection-only models. However, the cases in which findings are altered with the addition of influence parameters may reveal the complex nature of these social phenomena. Endogenous dependencies across study 2&3. Results from studies 2 and 3 confirm several of the known properties of friendship networks in adolescent classrooms, such as reciprocity, transitivity and balance. Reciprocity is a significant structural feature for friendship ties across conditions. Tendencies towards balance ties are strong and consistent while the preference for transitive ties (closed subgroups) are statistically significant in the model selection procedure, but then lose significance when additional explanatory parameters are included. One important structural feature found in this population is a high magnitude for outdegree estimates relative to reciprocity, compared to past studies. In addition, several of the models in these 141 studies had non-significant outdegree parameters. Typically, networks show a highly significant tendency towards negative density. Negative density means that on average, costs of arbitrary ties are higher than their benefits. While networks in this study showed the expected tendency towards negative density, when density and reciprocity functions are added together, they produce an overall negative value. This indicates that students might be reluctant to form any friendship ties because the cost of any given tie may be much higher than its benefits, even when they are reciprocated. In addition, non-significant outdegree effects suggest that ties are relatively random, which could be explained by the unusual setting of continuation high schools. Schools are designed as an alternative to the typically well-defined classroom structure, where there is not typically a great sense of cohesiveness within each classroom. It is possible that classroom networks are social networks artificially delineated by the researchers, but may not reflect social dynamics that are salient to the students. Attribute-based dependencies across study 2&3. Gender, ethnicity, age, and friendship importance show varying degrees of network dependencies. In the standard classrooms, students with more close friends within the school are significantly more likely to be friends with each other. Standard and experimental networks show that ethnicity is strong factor in popularity, where non- Latino students were preferred over Latinos. Students in the experimental classes tend to evaluate same sex friendships more highly than opposite gender friendships. In the control condition, boys and older students are evaluated much more highly than female friends, and younger students. Overall, (1) number of best friends in school (homophily among those who feel close to their schoolmates) is a significant selection factor in the standard condition, (2) ethnicity (preference towards non-Latinos) is significant in the standard and experimental conditions, and (3) gender (homophily for the same gender and general preference for boys over girls) is significant in the control condition. These results suggest that ethnicity effects might be intensified by the program activities, possibly due to the 142 survey items that ask the students to identify their race and their peer group identity. At the same time, the intervention may be effective in integrating the class by encouraging students to share their opinions and own experience, giving students a basis for creating mixed gender friends. Drug-based selection & influence. Table 11.1 summarizes selection and influence effects by condition. All parameter estimates for network selection and behavioral influence tendencies are shown. Only the direction of the estimate is shown in the table. Positive homophily effects indicate that students with similar cognitions or behaviors are drawn to each other and are more likely to form friendships. A positive alter effect means that students who are users or have strong cognitions around alcohol or marijuana are more likely to be selected as friends by others. A positive assimilation effect indicates that students tend to acquire the behavior/cognitions of their friends. Parameter estimates 143 Table 11.1 Summary of Homophily & Assimilation Parameters for Drug Cognitions and Behaviors Table 11.1 Positive or negative tendencies for selection and influence are presented simply as ‘+’ or ‘-‘ Only alter and similarity selection effects and assimilation influence effects are shown. Parameter estimates that are significant at the 10% significance level are highlighted in yellow. From the direction (+ or -) of the parameter estimates, patterns of homophily and alter effects become evident. In this table, cases where intervention estimates show tendencies that differ from one another and the control condition can be easily recognized. For example, in the first row, alcohol behavior alter, the experimental condition is negative while the control and standard conditions have positive parameter estimates. This might indicate that the experimental condition influenced the social selection processes around alcohol use. Conclusions based on comparisons only might be unfounded but are considered alongside statistically significant findings. PROGRAM CONDITION n Models Selection Selection & Influence Selection Selection & Influence Selection Selection & Influence Alcohol Behavior Alter ++ - - ++ Alcohol Behavior Homophily -- + + -- Alcohol Cognition Alter ++ ++ ++ Alcohol Cognition Homophily - + -- -- Marijuana Behavior Alter + + -- -- Marijuana Behavior Homophily - + ++ ++ Marijuana Cognition Alter ++ - - ++ Marijuana Cognition Homophily -- -- + + Alcohol Behavior Assimilation +- + Alcohol Cognition Assimilation ++ - Marijuana Behavior Assimilation +++ Marijuana Cognition Assimilation -+ + STANDARD EXPERIMENTAL CONTROL Alcohol & Marijuana Selection Effects 53 Alcohol & Marijuana Influence Effects 95 119 144 Homophily for alcohol cognition is found in the standard condition. In the case of marijuana, students in the standard condition show preferences for marijuana users and those with marijuana cognitions compared to other conditions. The standard intervention had an effect on selection and influence processes around alcohol use, while neither intervention affected selection processes around marijuana use. Alcohol cognition selection effects occurred in the experimental condition, and influence of marijuana cognitions occurred the standard condition. There is a tendency to engage alcohol-using behaviors that differ from one’s friends, in the experimental condition. In Table 11.2, all the main selection and influence findings from studies 2 and 3 are reviewed. Statically significant findings are reported with corresponding probability values. P- values are presented from the coevolution model, unless indicated that it is from the selection- only model (denoted as ‘SM’). Comparative findings that were presented in Table 11.1 are also shown in Table 11.2. These are non-significant findings that differ from parameter estimates found in the control and the other intervention condition. 145 Table 11.2 Summary of Network Modeling and Coevolution Findings from Continuation HS Intervention Table 11.2 A summary of findings from the selection-only study (denoted as ‘SM’) and the coevolution study is presented. For endogenous structural dependencies and attribute-based selection findings, p-values reflect statistics, which are consistent across alcohol and marijuana models. Standard Condition Experimental Condition Exogenous Dependencies Strong tendencies for reciprocity (p<0.05) and balance (p<0.05), and inconsistent tendencies for transitive ties (ns) Strong tendencies for reciprocity (p<0.05) and balance (p<0.05), and inconsistent tendencies for transitive ties (p<0.05) Attribute- based Selection Preferences towards non-Latinos (p<0.05) Students with more close friends in the school tended to nominate one another (p<0.05) Preferences towards non-Latinos (p<0.05) Alcohol Behavior Students tend to prefer drinkers as friends (p<0.1) Students tend to assimilate friends’ alcohol behaviors (p<0.1) Students tend to prefer non-drinkers (SM, p<0.1 and compared to other conditions) Students tend to select friends with similar alcohol behaviors (SM, p<0.1 and compared to other conditions) Assimilation of friends’ alcohol behaviors is thwarted, compared to std and control conditions (ns) Marijuana Behavior Students tend to select marijuana users as friends compared to exp and control conditions (ns) No significant or comparative effects Alcohol Cognition Students tend to choose those with strong alcohol cognitions as friends, compared to exp and control conditions (ns) Students tend to choose friends with strong alcohol cognitions (p<0.10) and overall there is an overall tendency to reduce cognitions (SM, p<0.1) Marijuana Cognition Students tend to choose those with strong marijuana cognitions as friends (p<0.05) Assimilation of friends’ cognitions is thwarted, compared to exp and control conditions (ns) No significant or comparative effects 146 Standard condition. Alcohol users (p<0.1) and those with alcohol cognitions (p=ns) are more popular. When the influence mechanisms are modeled, strong tendencies for the influence of friends’ behaviors on alcohol use are found (1.99, p<0.1). In these models, selection tendencies around alcohol use become non-significant, yet, positive alter and negative similarity trends remain the same. These findings suggest that alcohol-based selection and influence processes both occurred in the standard classes, but when modeled together, influence effects are stronger. These networks show a preference for friends with strong alcohol and marijuana cognitions (p<0.1, p<0.05, respectively). In addition, there is a tendency to develop marijuana cognitions that are less similar to ones friends’ cognitions. These findings show that students with strong drug-related cognitions tend to be more popular in the standard condition, and in the case of marijuana, cognitive changes reflect divergence from friends’ cognitions. Experimental Condition. There is a tendency for homophily around alcohol behaviors, where drinkers select other drinkers, and non-drinkers select non-drinking friends (p<0.1). Interestingly, non- drinkers are more popular in these networks than drinkers (p<0.1). In contrast to homophilic friend selection around alcohol use, students tend to develop alcohol-using behaviors unlike their friends’ behaviors. There is a preference for friends who are non-drinkers but also for those who have strong alcohol cognitions. While behavioral influence effects are not significant; there is a tendency to diverge from one’s friends’ alcohol behaviors. Tendencies to befriend students who have strong alcohol cognitions remain strong, even after accounting for influence effects (p<0.1). Interestingly, while non-drinkers are more likely to be selected as friends, those with alcohol cognitions tend to be more popular. There was a strong tendency to reduce alcohol cognitions over the course of the intervention. The findings suggest that students in the experimental classes tended to select friends who have strong alcohol cognitions and yet, alcohol cognitions tended to diminish over the course of the intervention. 147 Control Condition. No significant effects for alcohol or marijuana-related dynamics are found in the control condition. Notable trends are: (1) preferences for others with similar marijuana cognitions and behaviors, and (2) preferences for friends’ with dissimilar alcohol cognitions and behaviors. 11.2 Comparison of Logistic and Network Findings from Studies 2 & 3 Findings from the three studies explore mechanisms of behavioral influence using very different analysis strategies. Specifically, study 1 uses logistic regression and studies 2 and 3 use dynamic network analysis. Table 11.3 shows assessments of friends’ cognitions and behaviors across methodologies. Findings are presented by behavioral or cognitive outcomes (i.e., alcohol use, alcohol cognition, marijuana use, marijuana cognition). 148 Table 11.3 Comparison of Regression and Network Analyses Conclusions on Social Influence by Outcome Logistic Regression Findings Statistical Network Analysis Findings Conclusions Across Methodologies Alcohol Behavior ~ Friends’ behaviors influence ego's behaviors (OR=3.1; p<0.001) ~ Also, friends' cognitions are negatively associated to ego’s behaviors (OR=0.56; p=0.23)* ~ Friends' behaviors influence ego's behaviors in the standard (2.0; p<0.05) ~ While the experimental condition shows negative assimilation (ns) Logistic Regression: ~ Friends’ alcohol behaviors influence ego’s behaviors ~ Cognitions of friends’ may negatively influence ego’s behaviors Network Dynamics ~ Friends alcohol behaviors become more similar in the standard condition, and become less similar to one another in the experimental condition Marijuana Behavior ~ Friends' behaviors had a small effect on ego's behavior (OR=1.9, p<0.1) ~ No effects were found Logistic Regression: ~ Friends' marijuana behaviors influence ego's behaviors Network Dynamics ~ Friends' marijuana behaviors tend to become more similar ~ The intervention did not differentially modify social influence processes Alcohol Cognition ~ Friends' cognitions influence ego's cognitions (OR=6.4; p<0.01) ~ No effects were found Logistic Regression ~ Friends’ cognitions strongly influence ego’s cognitions Network Dynamics ~ Friends develop similar alcohol cognitions in the intervention condition Marijuana Cognition ~ Friends' cognitions influence ego's cognitions (OR=7.8; p<0.001) ~ Students in the standard condition develop cognitions that differ from their friends Logistic Regression ~ Friends’ cognitions strongly influence ego’s cognitions Network Dynamics ~ Friends marijuana cognitions become less similar to one another in the experimental condition Table 11.3 A summary of social influence findings from regression and network analyses are presented. All log odds and p-values are given for the parsimonious model except where denoted. For the regression analyses, odds ratios and p-values are taken from the ‘full models.’ Network analysis results are reported for statistically significant parameters, and for results that differ from both the alternative condition and control conditions. 149 A comparison of influence findings across methodologies shows a number of contrasting results. Logistic regression shows that students who have friends that engage in alcohol and marijuana-related activities are much more likely to also engage in those activities. Social influences around cognitions were even greater than influences of behaviors. Coevolution network modeling results show significant influence effects only in the standard condition, and only for alcohol behaviors. In one sense, given that social influence is operationalized similarly in the two methodologies, these discordant results are discouraging. However, the disparities are easily attributed to differences in modeling strategies and in fact, raise important questions about how socially dependent data should be examined. In addition, it underscores limitations of each approach. These issues will be addressed in the discussion chapter that follows. Alcohol Behaviors. Consistent with regression findings in study 1, influence models in study 3 show that students tend to assimilate their friends’ alcohol behaviors. Regression results indicate that students with friends who drink alcohol are 3 times more likely to drink, compared to those with non-drinking friends (OR =3.1; p<0.001). Dynamic network analysis modeling shows that there is a tendency for people to choose behaviors that are similar to their friends’ behaviors. In the experimental condition, students tend to select friends based on alcohol behaviors while in the standard (p<0.10) and control conditions (ns), individuals tend to acquire friends’ alcohol behaviors. In addition, regression models show that students with friends with alcohol cognitions were actually about half as likely to become drinkers themselves (OR=0.56; p=0.23). In the network analysis, the effects of friends’ implicit cognitions on ego behaviors were not significant. Marijuana Behaviors. Regression findings indicate that those with friends’ who use marijuana are about 2 times more likely to smoke marijuana, compared to those with non- 150 smoking friends (OR = 1.9, p<0.1). In contrast, network modeling results found that there were no significant or comparative effects. Alcohol Cognitions: Regression results show that students who have friends with alcohol cognitions are 6.4 times more likely to also have alcohol cognitions (OR =6.4; p<0.01). Network analysis models do not show significant or comparative alcohol cognition influence effects. From the regression framework, alcohol cognitions are strongly driven by pre-existing behaviors (and not pre-existing cognitions). In addition, students in the standard program tend to prefer others with similar alcohol cognitions (1.05; p<0.10). Marijuana Cognitions. Regression results show that individual whose friends have strong marijuana cognitions are 7.8 times more likely to also have marijuana cognitions. Network modeling results show that there was a reduction in assimilation effects in the standard condition compared to the experimental and control conditions (ns). 151 CHAPTER 12 DISSCUSION OF FINDINGS The problems we face will not be solved by the same level of consciousness that created them. -- Albert Einstein This dissertation applies behavioral influence and social selection paradigms to implicit cognitive processes. Although cognitive researchers frequently use network terms metaphorically, to date, no studies have applied social network analytic techniques to the study of cognitive associations. Given the recent advances in network modeling techniques and the problems inherent in fitting a non-dependency statistical approach to exploring socially dependent data, influence and selection mechanisms require investigation from both a network perspective and a regression random sampling one. In addition, the program design enables us to investigate social influences around alcohol and marijuana activities as they interact with specific educational strategies. The data are from a drug education program, which has been shown in previous studies to be effective among high-risk youth attending alternative high schools (Sussman et al., 2004). The intervention was delivered to 900 Continuation high school students over the course of three months during 2004 and 2005. There were two unique features of the program. The first component was an experimental implementation strategy using social network concepts. Using an algorithm that optimizes social preferences (Valente & Davis, 1999), students in the experimental condition were assigned to groups based on a sociometric network survey item, “who would make the best leader in this class?” For students in the experimental condition, much of the curriculum was delivered to the students in an interactive, group format. The second distinctive component of the program was the assessment of implicit cognitions. Evidence strongly suggests that program 152 effects on drug using behaviors are mediated through changes in implicit associations around drug concepts. Studies presented here model marijuana and alcohol behaviors separately because different sets of word association tasks are used to target each behavior. Social network and implicit cognition domains were integrated in these studies to investigate social selection and behavioral influence mechanisms. The focus of the three studies presented in this dissertation is on social factors of behaviors and cognitions. Findings from the studies provide some evidence that selection and influence of cognitions and behaviors occur. Results across methodologies are not consistent, but closer examination reveal that confluence of findings may provide a much fuller story about the data than either approach can tell alone. It also becomes clear that the modeling strategies answer hypotheses from such distinctly different points of view that it might be reasonable not to expect the same results across methods. Therefore, the first section covers several methodological considerations that became important in making conclusions about these results. Second, behavioral and cognitive influence findings are reviewed, including specific effects of program condition on these processes. Third, the relevance of findings for prevention and educational applications are discussed. Lastly, limitations of the study are presented, with special attention paid to how the large amount of missing data might distort the social network modeling how SIENA modeling assumptions might have affected the findings. 12.1 Methodological Considerations Network modeling identifies patterns of friendship structures and behavior adoption to assess the role of social structures in behavioral change. Regression modeling identifies the contribution of social and ego variables by assessing the likelihood that the behavior or cognition is present given the set of individual versus friend attributes. Attempts to compare findings from a network (relational) approach to a regression (correlative) approach underscore the fundamentally distinct 153 assumptions underlying the two strategies. Regression analyses account for environmental or other social factors by determining their contribution to predictions of the outcome. Controlling for gender in a logistic model, for example, allows the researcher to conclude that a hypothesis driven variable (e.g., friends’ alcohol use) is predictive of the outcome (e.g., ego’s alcohol use), after eliminating any effects that friends’ behaviors shares with gender in their effects on the ego’s alcohol use. Alternatively, controlling for gender in a network statistics model allows the researcher to conclude that individuals choose friends based on their alcohol using behaviors above and beyond their selection preferences of one gender over another. From a behavioral influence perspective, controlling for gender in a network statistics model merely accounts for social clustering (at the dyadic and triadic levels) based on gender, so that we can empirically account for covariates that might contribute to the friendship. As methodological considerations are discussed in more depth, clear distinctions in the analytic strategies become more apparent. Taken together, the findings reveal the complexity of the social dynamics in play. The next section outlines three steps that grew into relevant topics over the course of conducting these studies: (1) selection of covariates across analyses, (2) determining the appropriateness of data for each of the analytic techniques, and (3) considerations for comparing results across methods. Covariate Selection. The SIENA model quickly becomes unnecessarily complex with the addition of covariates. That is why forward model selection procedures are recommended, where only significant covariates remain in the model. Researchers have used results from regression models to identify important variables in the data set for inclusion in the network model (Valente Fujimoto, Chou & Spruijt-Metz, 2009). However, this strategy proved to be ineffective for these models upon closer examination. Take for example, ethnicity, which was significant in alcohol cognition regression models’. Findings from regression analyses show that Whites and other non- 154 Latinos are 2.9 times more likely than Latinos to have alcohol cognitions, whereas network findings show that non-Latinos are more likely to be selected as friends (alter effects, p<0.10). A regression approach controls for effects of gender on the behavior so that the effects due to the hypothesis-driven independent variables can be isolated. The network controls for effects of gender on friendship formation so that any effects on friendship found due to the hypothesis driven variables are above and beyond any effects that preferences for non-Latinos has on friendship formations. Theoretically, regression dependencies do not translate into network model dependencies. Tendencies for non-Latinos to have higher cognition scores have little to do with tendencies to select non-Latinos. However, the correlation between being of non-Latino decent and having higher alcohol cognition scores may reveal that tendencies for non-Latinos are a confounder in the preferences for high cognitions. For example, people may choose those with low cognitions simply as a function of choosing non-Latinos. While equally valid considerations, the framing of the research questions in the two approaches capture very different covariate effects. Based on these methodological considerations, future studies that employ both logistic regression and network modeling should consider independent covariate selections procedures but should consider behavioral correlates when considering social dynamics. Appropriateness of Data. As already discussed in depth, network analysis modeling strategies uniquely account for social dependencies that exist in the real world. However, characterizing a network of individuals as a single entity requires sufficient information about members of that group. Only five classrooms in these data met loosely set criteria set at a minimum of 20 students, and not more than 20% missing data as described by Snijders (2008). Instead these networks contained a minimum of 10 students and surveys had to be collected from a minimum of 60% of the total class. Still, the imposed criteria reduced the sample size to only 16 classes out of 70. 155 Strict criteria for allowable missing data for SIENA actor-based modeling have not been analyzed fully at this date, but several results from these data indicate that they may be insufficient to capture important complex dynamics being modeled (discussed further in the Limitations section below). While intrinsic structural features of the data fulfill modeling criteria (i.e., convergence) and show that dyadic and triadic dependencies are significant, the interpretation of specific results might suggest that these data do not provided an adequate amount of information about network structures. For example, networks presented here show a high magnitude for outdegree parameter estimates relative to reciprocity and several of the models had non-significant outdegree parameters. Social networks typically show a highly significant tendency towards negative density (negative parameter for outdegree) and when linearly combined with density, the objective function usually reflects an overall preference for social structures (versus random relations) (Pearson, Sweeting, West, Young, Gordon & Turner, 2006; Steglich et al., 2006). However, networks presented here show that density plus reciprocity functions produce an overall negative value. This means that on average, costs of arbitrary ties are higher than their benefits, supporting the basic network premise that relationships are not random (see Figure 10.1 for calculation procedure). The non-significance of outdegree effects (as reported in all the standard condition models) implies, therefore, that ties are relatively random. The high outdegree relative to reciprocity indicates that students are reluctant to form any friendship ties because the cost of any given tie may be much higher than their benefits. The reason for these inconsistencies with past SIENA findings on classroom networks is likely due to the large amount of missing data. One possibility is that these networks may simply not have enough power to model dynamic friendship structures. Alternatively, the general lack of cohesiveness in alternative school settings 156 might make these networks less meaningful than typical high school classroom exchanges. In either case, these results might indicate that these data are not particularly appropriate for network dynamic models. While meeting the procedural criteria for network modeling (see “Model Fitting” section in 9.1), parameters should be meaningful beyond fulfillment of mathematical considerations. These models should be run on larger networks to find whether outdegree parameters remain to have relatively higher magnitude than reciprocity parameters. Comparing Results. In the studies presented here, findings across methodologies show more inconsistent findings than consistent ones. There are several points that should be considered when making comparisons across methodologies. First, network models identify behavioral influence patterns while controlling for selection processes. If both mechanisms were present, then the effect size would need to be much greater to detect both clustering tendencies caused by friendships (influence), and friendships caused by the clustered attributes (selection). Reasons for the disparate results might be as simple as the lack of power in the network modeling analyses. Findings in the network modeling show that when influence parameters are included in the models, most of the significant selection effects disappear. While influence effects reach significance only for alcohol use in the standard condition networks, modeling of behavioral influence effects for all drug activities (however small they may be) may account for much of the social exchanges that are attributed to selection (when selection is modeled alone). Thus, the power to detect two concurrent processes, where the independent variable in one is the dependent variable in the other, would have to be significantly higher than detecting either one alone. Analytic tests for network modeling power criteria have not been published. Power criteria are essential to provide guidelines for assessments of coevolution processes. A second consideration for determining appropriateness of one’s data to network modeling techniques might be the extent to which the specific behavior is reliant on social-versus-ego driven 157 processes. Logistic regression findings showed that by comparing ego versus alter effects, each of the four outcomes (i.e., alcohol and marijuana cognitions &behaviors) could be characterized as being ego or alter in nature. That is, one of the advantages of attribute-based assessments is that the relative contributions of an individual’s intrinsic qualities can be distinguished from the qualities of their friends. From this perspective, alcohol use appears to encompass ego and alter effects, whereas marijuana use is more dependent on the ego’s own (past and current) behaviors and implicit cognitions. Network models of alcohol behaviors were much more effective in detecting significant influence and selection relationships compared to marijuana models. One reason for this might be the fact that alcohol appears to be much more dependent on social relationships. Therefore, logistic regression findings might be very useful in identifying the importance of social variables on specific behaviors before network models are attempted. 12.2 Review of Behavioral and Cognitive Influence Findings Table 12.1 shows influence and selection findings for logistic and network modeling studies (dissertation studies 1, 2, and 3). Logistic regression models only test behavioral influence, while network models test selection and coevolution models test influence and selection simultaneously. 158 Table 12.1 Review of Behavioral Influence and Network Selection Dissertation Findings Hypotheses Study Domain Independent Variable Dependent Variable Mechanism Regression Findings Network SIENA Findings H1 Do friends’ drug behaviors predict ego’s behaviors? #1 Behavioral Friends' Alcohol/ Marijuana Behavior Ego's Alcohol/ Marijuana Behavior Influence Strong alcohol influence (OR = 3.1, p<0.001) Weak marijuana influence (OR = 1.9, p<0.10) See hypothesis 9 H2 Do friends’ drug cognitions predict ego’s cognitions? #1 Cognitive Friends' Alcohol/ Marijuana Cognitions Alcohol/ Marijuana Cognitions Influence Strong alcohol cognition influence (OR = 6.4, p<0.001) Strong marijuana cognition influence (OR= 7.8, p<0.001) See hypothesis 10 H3 Do friends’ drug behaviors predict ego’s cognitions? #1 Behavioral Friends' Alcohol/ Marijuana Behavior Alcohol/ Marijuana Cognitions Influence No significant effects Model was unable to converge H4 Do friends’ drug cognitions predict ego’s behaviors? #1 Cognitive Friends' Alcohol/ Marijuana Cognitions Ego's Alcohol/ Marijuana Behavior Influence Weak alcohol cognition influence to reduce behaviors (OR=0.56, p=0.23) Model was unable to converge H5 Do ego's tend to become friends with those who have similar drug behaviors? #2 Behavioral Similarity of Alcohol/ Marijuana Cognition Friendship Tie Selection Not Applicable Selection of similar behaviors occurs only for alcohol, in the exp cond (p<0.1) Non-drinkers are preferred in the exp cond (p<0.1), & preference for drinkers (p<0.1) & smokers occurs (p=ns) in the std cond. H6 Do ego's tend to become friends those who have similar drug cognitions? #2 Cognitive Similarity of Alcohol/ Marijuana Cognition Friendship Tie Selection Not Applicable Preferences for friends' with alcohol cognitions occurs in the exp condition (p<0.10); and alcohol & marijuana cognitions in the std cond (p<0.10, p<. 05) 159 Table 12.1 Review of Behavioral Influence and Network Selection Dissertation Findings, Continued Hypotheses Study Domain Independent Variable Dependent Variable Mechanism Regression Findings Network SIENA Findings H7 Do ego's tend to become friends with those who have similar drug behaviors, controlling for behavioral influence? #2 Behavioral Similarity of Alcohol/ Marijuana Behavior Friendship Tie Selection Not Applicable No selection effects, controlling for influence effects H8 Do ego's tend to become friends with those who have similar drug cognitions, controlling for cognitive influence? #3 Cognitive Similarity of Alcohol/ Marijuana Behavior Friendship Tie Selection Not Applicable Preferences for alcohol cognitions remains significant in the experimental condition (p<0.10); Preferences for marijuana cognitions remains significant in the standard condition (p<. 05) H9 Do ego's tend to assimilate friends’ behaviors, controlling for behavioral selection? #3 Behavioral Friendship Tie Similarity between Ego's Alcohol/ Marijuana Behaviors of their Friends Influence See hypothesis 1 Alcohol influence occurs in the standard condition (p<0.1); Tendency to diverge from friends' behaviors occurs in the experimental condition (compared to other conditions, p=ns) H10 Do ego's tend to assimilate friends’ average cognitions, controlling for cognitive selection? #3 Cognitive Friendship Tie Similarity between Ego's Alcohol/ Marijuana Cognitions of their Friends Influence See hypothesis 2 Assimilation of friends’ marijuana cognitions is thwarted in the standard condition (compared to other conditions, p=ns) 160 Regression Findings. Regression findings show that friends’ cognitions and behaviors influence ego’s cognitions and behaviors, with cognition influences having a greater magnitude of effects. This suggests that implicit cognitions are exchanged between friends perhaps more readily than behaviors. Adopting a behavior requires a willingness on the part of the individual, which necessitates that the activity is somewhat harmonious with attitudes and perceptions around that behavior. Implicit cognitions, on the other hand, circumvent rational conscious processes, and therefore, cognitive contagion might occur more fluidly than behaviors. The strong log odds for ego’s cognitions given their alters’ cognitions relative to relationships between ego and alters’ behaviors also might indicate that implicit associations accumulate gradually before social influences on behavior can occur. That is, cognitive tendencies might increase susceptibility to behavioral influences. Future studies should follow students for a longer period of time to test whether patterns of implicit cognition exchanges leads to behavior changes over time. While the influence of friends’ drug-using behaviors has been consistently found in past studies, implicit cognitions have never been considered to be fluid through social connections. These findings point to the importance of further research to explore how implicit cognitions are exchanged over periods of time, and the role these social interactions play in behavioral adoption. Contrary to the expected findings, data show that students who have friends with strong alcohol associations are less likely to become drinkers. Students with friends who had strong alcohol cognitions at baseline were about sixty percent as likely to be alcohol drinkers after completion of the intervention. These counter-intuitive findings may be explained by the overall decline in alcohol cognitions over time, at least in the experimental condition. Strong alcohol cognitions at baseline may gradually decline over the three months and thus, friends’ with alcohol cognitions at baseline may subsequently become friends with healthy alternative cognitions. Thus, the discrete snapshots modeled in the regression models show an inverse relationship with friends’ 161 behaviors (albeit non-significance), but may actually represent a group level reduction that occurs in unison, gradually over time. Network Modeling Findings. In contrast to the strong influence effects found in regression models, as well as previous regression studies with these data (Valente et al., 2007), network modeling shows that the influence of friends’ activities only occurred around alcohol behaviors. Using comparative effects (presented in table 11.1), findings show that in some cases friends were more likely to chose behaviors dissimilar from one’s friends (e.g., divergent marijuana cognitions among friends in the standard condition). Disparate findings in the two studies may suggest that, in reality, selection and influence processes are occurring concurrently, and that testing these mechanisms independently might lead to overestimations of their effects (type I errors). However, another explanation for the lack of significant influence effects is that the network modeling did not have the power to detect both selection and influence (discussed in more depth in ‘Limitations’ section below.) Alcohol behavior homophily was found in the experimental condition (p<0.1). Interestingly, selection for like others is accompanied with preferences for non-drinkers. No homophily is found in the standard condition, but there are preferences for drinkers and smokers. No evidence for cognition based friendship selection was found. Yet, both intervention conditions showed preferences towards individuals with strong alcohol cognitions, and preferences for marijuana cognitions in the standard condition classes. 12.3 Intervention Implementation Effects In general, trends found in logistic regression analyses suggest that students in the standard teacher-led condition are slightly more likely to increase their alcohol and marijuana behaviors and slightly less likely to increase drug-related cognitive associations, compared to the experimental and control conditions; although these effects were not statistically significant. Network analysis findings show that assimilation of friends’ alcohol behaviors and selection of 162 drinkers as friends occur in standard condition classes. Preferences for friends with strong alcohol and marijuana cognitions are also found. Assimilation trends show that students tend to develop dissimilar marijuana cognitions from their friends. When selection and influence effects are considered together, trends suggest that: (1) drinkers are more popular and, therefore, behavioral assimilation of friends’ behaviors will lead to increased alcohol use, and (2) those with marijuana cognitions are more popular, but friends tend to diverge in their cognitions, which may result in an overall decrease in marijuana cognitions. In the interactive, experimental condition, selection preferences for non-drinkers are found. While alcohol behavioral influence effects were not statistically significant, their inclusion in the models reduced effects of selection effects, indicating that influence mechanisms might, in part, explain the clustering of individuals by alcohol using status. As for cognition-related findings, the interactive program shows a significant reduction in alcohol and marijuana cognitions. While overt alcohol behaviors tended to be selected against, students showed preferences for others with strong alcohol cognitions. Taken together, findings suggest that the experimental (interactive) program encourages alcohol-related homophily, and facilitates popularity of those with alcohol cognitions while reducing the appeal of overt alcohol behaviors. 12.4 Applications to Intervention Strategies Reducing the appeal of a pro- alcohol and marijuana culture in high-risk youth remains to be a great societal challenge. When considering the additional complexity of social interactions involved in these behavioral norms and behavioral change processes, prevention strategies become even more cumbersome. Reducing substance use among youth not only requires changing attitudes, awareness, and willingness to adopt alternative behaviors, but also requires changes in the social climate and susceptibilities to deviancy training among friends. Much of the research on educational /intervention strategies is aimed at the dissemination process itself (i.e., did the session have the desired effect? Did the participants understand and retain the information?) However, 163 behavioral and social dynamics that occur during, or, as a result of the intervention remains relatively unexplored. For example, does an intervention increase social ties among deviant group members, or does the program increase susceptibilities to social influence? Alcohol using models provide the most interesting results of this dissertation. Findings suggest the standard condition promotes social exchanges that ultimately increase alcohol use. First, the standard program appears to increases alcohol behavioral influences among friends. This may occur by creating a teacher-student delineation, making students more reliant on one another, thus, more susceptible to their friends’ influences. Second, there are preferences towards choosing friends who are drinkers, making alcohol users highly influential. By amplifying susceptibility to social influences while also increasing popularity of drinkers, the standard condition program appears to increase socially related alcohol use. Findings suggest that while the standard condition promotes social exchanges that increase alcohol use, the experimental condition appears to promote social selection and influence processes that ultimately reduce alcohol use. Students who have strong alcohol cognitive tendencies are at high risk for alcohol use, but have not necessarily initiated the behavior. In the experimental condition, non-drinkers become more popular, which is presumably a result of the program’s success in changing cultural norms, making alcohol use less favorable. In addition, overall drug-related cognitions are reduced, which apparently does not occur through social selection or behavioral influences, but rather occurs as a group level process. One possible explanation is that cues that had previously elicited drug-related memories become associated alternative behaviors. Ideally, alcohol links would be redirected to healthier social behaviors, but only a closer analysis could confirm what new behaviors cognitively replaced alcohol. Reductions in drug-related associative memories are not seen in the standard condition, suggesting that programs that encourage peer interactions during course activities may provide unique opportunities for students to form new associative links. Creating new associations within 164 a relevant social context may also be strengthened by their shared associative meanings among peer groups. From an interventionist perspective, peers can act as social structures through which new positive associative memories can be embedded, supporting long lasting behavioral changes. Interventions that attempt to create cognitive and behavioral changes within the individual without accounting for the larger social context ignores deeply embedded social meanings of alcohol and marijuana use, especially for adolescents. Alcohol and marijuana are typically used in social settings and often the rituals around its use are what define one group from another. Creating behavioral change from within networks of friends provides a foundation within which the individual can gradually rebuild. Overall, findings suggest that changing cultural norms around alcohol drinking facilitates the increased popularity of non-drinking students and engenders an overall reduction in alcohol associative memories. Figure 12.1 shows the proposed mechanisms for each of the two program conditions. 165 Figure 12.1 Proposed Alcohol Using Dynamics for Standard and Experimental Conditions Standard Condition Experimental Condition Figure 12.1 In the standard condition, friendships lead to similarities in alcohol behaviors. Preferences for drinkers may lead to an overall increase in alcohol use. In the experimental condition, popularity of non- drinkers increases and group level alcohol cognitive tendencies are reduced. Cognitions may be reduced as a result of classroom cultural norms and cognitive-behavioral exchanges. It should be noted that the proposed mechanisms differ from the ones put forward by Valente et al. (2007) analyzing these same data. They showed that while the experimental condition was more effective than the standard condition in reducing substance use (represented by a composite score for tobacco, alcohol, marijuana, and cocaine), the success of the program implementation came with caveat that for students with drug-using friends, the program resulted in greater drug use. Their study concludes that for students in the standard classes, increases in friends’ using behaviors is associated with reduction of ego’s substance use, while for experimental classes, a greater number of drug-using friends correlates to an increase in ego’s use. The standard condition, therefore, appears to be protective for negative social influences. Engenders increase in alcohol use ! Preferences for drinkers Influence of friends leads to assimilation Non-drinkers are/become more popular Students choose like others Engenders a group-level reduction in cognitions Engenders non-use ! Cultural norms around alcohol use becomes less favorable Non-drinkers are/become more popular Students choose like others Engenders a group-level reduction in cognitions Engenders non-use ! Cultural norms around alcohol use becomes less favorable 166 Mechanisms suggested in these studies may differ from those presented in Valente et al (2007) due to the fact that only alcohol use is considered in these analyses compared to the substance using composite score used in Valente et al (2007). It is possible that using a composite score would show support for different influence mechanisms. However, implicit cognition measures were designed specifically to capture either alcohol or marijuana cognitions, so that a composite score would not be appropriate. Another possibility is that modeling selection effects as done in these analyses may alter influence results. For example, accounting for selection effects due to drug activities may explain some of changes attributed to deviance training in Valente et al. (2007). 12.5 Limitations of the Study Attrition and Missing Values. The attrition rate was largely due to students moving to different schools (there were no students who reportedly refused to complete the survey). This was the case for 31% (or 742-515 =227 students) of the original sample. It is possible that this uncontrollable source of attrition biased the social influence and selection results. In network analyses, students who were on the class roster but did not complete a survey are considered to be missing. Unfortunately, we have no way of knowing whether these students were no longer in the class or simply absent on the day of the survey. The unusual setting of continuation high schools allows students to move from class to class as they earn credits, and further, students vary in their coursework requirements so they are not all required to be present all day long. It is plausible that students missing at baseline and/ or follow- up represent a high-risk population and are more likely to engage in illicit drug activities. However, the focus of the study is to evaluate social interactions among classmates, and therefore, we can reasonably discount the students who are rarely present in the class. It is argued that missing students are more likely to play a lesser role in social dynamics than those who are present. 167 While missing data on individual students may not significantly influence social dynamics, incomplete networks may, however, distort social structural findings. In network modeling, where the unit of measure is the network, missing students equate to incomplete networks. . Studies have not yet been published examining the allowable missing percentage of nodes for network modeling, or how to specify the power required to capture specific network effects. Remember, of course, that a missing student may have little effect on a regression analysis, but will permeate through the classes’ network dependencies, i.e., missing students cannot reciprocate ties, contribute to triadic configurations, select similar others or assimilate friends’ behaviors. Models with many missing reciprocal and triadic relations will underestimate the value of reciprocity and triadic configurations (which is likely the case in these data). More directly, excessive missing data leads to overvaluing those students with complete data. For example, assimilation is measured as the average of friends’ behaviors. If ‘missing’ friends are more likely to be users, friends’ behavior scores will be lower than they are in real life, and thus, when an ego assimilates drug using behaviors of their “missing” friend, assimilation effects will not be captured. As for missing particular variable data, the most frequently uncompleted part of the survey was the implicit cognition measures. These tasks require the students to read phrases and quickly respond by writing the first behavior that comes to mind. This can be a taxing task, particularly for those who read more slowly or those for whom English is a second language. For example, only about 9% of the population responded to the cue, ‘having euphoria,’ presumably because students did not comprehend its meaning (this cue was removed from the analysis). Poor reading skills may reflect lower cognitive function and therefore, may also indicate a higher risk for drug behaviors. To test this, logistic regression was conducted to check for associations between the number of missing word association responses and alcohol and marijuana use. The analysis showed that of those who completed the survey at baseline, students who filled out fewer word association responses were not more likely to be alcohol or marijuana users. 168 Associative Memory. In this study, alcohol and marijuana cognitions are assessed using implicit word association tasks. These measures have been validated in past studies, and have reliably predicted drug use in similar populations (Stacy, Ames et al., 1996; Stacy, Galaif et al., 1996). Implicit assessments are unique in that they capture automatic or spontaneous memory associations, distinct from conscious, rational cognitions. Free association tasks rely on the spontaneity of the reactions in order to fully tap into implicit processes. However, the nature of large classroom surveys makes it difficult to regulate the automaticity of the responses. A second important consideration of associative memory tasks accounting for how language is affected by culture. Some concepts are more salient in particular cultural groups and the same word might have very different meanings in different cultures. Multilingual individuals store and recall concepts differently, where some concepts are more relevant in one language over another (Marian & Neisser, 2000). A word might need to be translated in memory before it can elicit a response, which will draw from various brain regions outside those targeted by implicit and explicit memory tasks. Lastly, the level of education may also affect how easily a person can process written text, and possibly affects how quickly a person can recall information. However, it is likely that the relationships to word meaning in the context of social structures should hold. Third, as a result of the social and flexible nature of associative memory recall, the reliability and contextual relevance of response data should be considered. Associative memory processes are considered to be reconstructive processes, where networks of associative links are reconfigured the instant they are accessed. Memory reconstruction is dependent on a complex confluence of external stimuli, and the social, physical, emotional context within which it is embedded. Despite this, coded word association measures have both consistently predicted behaviors and shown reliability in classroom and laboratory studies. Still, the flexibility of associative memory should be considered when proposing cognition dependencies and how results are interpreted. Responses elicited in these classroom surveys may 169 not reflect responses that would be generated in other settings. For example, when a student is at school, the phrase “in my bedroom” might be more likely to activate cognitive associates that are consistent with behaviors or attitudes that are relevant to peer relationships. On the other hand, the same phrase “in my bedroom” presented to a student while at home might elicit responses related to their role in their family, such as chores or homework. The focus of the studies presented here is on social selection and behavioral influence of classroom friendships. Therefore, associations around alcohol and marijuana that are generated in the classroom are the most relevant to these relationships. The classroom setting prepares the mind for concepts that are socially relevant within peer groups, which makes them more easily activated given relevant cue word stimuli. Drug use typically occurs among peers and therefore, the ease with which associative memories around drug use are generated among peers captures associative processes in real life scenarios. In this study, to measure whether cognitions affect social dynamics of classroom friendships, associative memory tasks are appropriately conducted within the classroom. Social Networks Measures. In these analyses, hypotheses about how friends’ influence one another are modeled, with an assumption that friends’ behaviors and ego’s behavior represent distinct constructs. Alter and ego effects, however, may instead be two indicators of a broader construct, possibly social and structural in nature. That is to say, positive associations of alters’ effects on the ego may actually be the result of a shared affiliation or cultural construct under which both ego and alter are affected (e.g., gang-related activities), instead of a causative relationship. This is a discriminant validity concern that would lead to type I errors (falsely rejecting the null hypothesis). In some ways, the actor-oriented network modeling approach deals with the discriminant validity problem by controlling for other potential social structural causes of shared behaviors. For instance, this strategy can accounts for the possibility that two individuals 170 are friends because they share a particular behaviors. However, this approach does not account for broader environmental features outside of the classroom that are not surveyed. Generalizability. Generalizing these findings to other adolescents may be somewhat limited. Students in this population represent a very high-risk group with relatively high incidence of criminal activities, drug addiction, and gang membership. Engaging in alcohol and marijuana use, while illegal, is perceived as a relatively innocuous activity. Therefore, social interactions around these activities may not be as salient as a more illicit substance, like heroin or cocaine use. In addition, compared to other populations that have been studied using dynamic network analysis (Snijders et al., 1995; Steglich et al., 2006), activities of classmates may not be as relevant as they are for typical adolescents. Continuation schools are designed as an alternative to the typically well-defined classroom structure, and accommodate students with children or full time jobs. Continuation HS classrooms are relatively fluid, and attendance is not steady. In many cases, these students are more likely to have strong social networks outside of the school. Connections within these schools may not be as relevant as compared to typical schools, where students’ social lives are centered around their classmates. These networks may not be effective in facilitating meaningful or consistent interactions among students. If relationships within the classroom are not salient, independency assumptions employed by logistic regression may be more appropriate than imposing the relational assumptions that are built into the network approach. Methodological. Already discussed at length, logistic regression analysis does not account for inherent social structural dependencies in the data. However, findings from network modeling results indicate that the data may be insufficient to fully capture network dynamics. An important finding discussed earlier is the high magnitude for outdegree estimates relative to reciprocity, compared to past studies. This finding suggests that friendship choices within the classroom are relatively random and that students may weigh the cost of any given tie higher than their benefits. Results might indicate that these data are not particularly appropriate for network dynamic models 171 due to the lack of information about the networks (small n and high percentage of missing data per classroom). Limitations of SIENA Modeling Assumptions. Implications of simulation and modeling are described in detail in Appendices A & B. One of the basic assumptions in SIENA model building procedure is that there are endogenous structural features that are inherent in social structures, and that these dyadic and triadic dependencies exist regardless of specific social processes. The argument to include reciprocity and triadic configurations in the models relies on the assumption that these configurations are basic building blocks that should be accounted for and teased out so that we can capture network dynamics that are based solely on specified attributes of the individuals. This approach also assumes that natural tendencies to reciprocate a friendship tie (or close a transitive triad) would occur regardless of other socially relevant attributes, such as drinking status. Controlling for reciprocity allows the research to conclude that a significant attribute similarity indicates that homophily between alcohol drinkers occurs in addition to the tendency for friendship ties to be reciprocated. Considering dyadic and triadic structural tendencies as endogenous features of the network leads to conservative estimates for detecting selection of alcohol and marijuana activities. In addition, this approach makes unsubstantiated assumptions about the natural state or intrinsic nature of social dependencies. An alternative approach would be to consider reciprocity and triadic configurations not as intrinsic (endogenous) features of the network, but rather as structural opportunities that facilitate selection processes (exogenous features of the network). From this perspective, the challenge would be to identify the structural features that mediate selection preferences, rather than statistically remove their effects from the model. One example of how these parameters can mediate one another can be demonstrated with transitive triplet effects found in these models. There are statistically significant preferences for transitive ties when only structural effects are included in the models, but it becomes non- 172 significant when attributes and behavioral parameters are added. This suggests that transitive structures may be explained by other variables in the models. Take for example a scenario where alcohol behaviors are transposed onto transitive structures. Persons B and C may meet at a party because they have a common friend (A), but their subsequent friendship may result from their mutual decision, say, as alcohol drinkers, to go on a beer run together. Triadic structures increase exposure, such that B and C are more likely to know about each other’s alcohol using behaviors through person A, compared to more distally or unconnected network members. Person C is likely to be more easily trusted by “B”, have indirect influence on B, and alter B’s perceptions of behavioral norms. Thus, triadic configurations might very well mediate opportunities for the selection of others with similar drug activities: Individuals who have the opportunity for a transitive tie will be more likely to create a homophilic tie compared to those who do not have such an opportunity. The analytic designs presented thus far suggest that triadic tendencies should be accounted for independently of tendencies for alcohol and marijuana homophily. However, transitivity that is blind to behavioral preferences may obscure important behavioral preferences that are embedded within dyadic and triadic tendencies. In the example above, the inclusion of alcohol parameters into the network data presented in these studies seem to have lead to insignificant transitive triplet parameters. It is unclear from the simulation algorithms whether controlling for dyadic and triadic dependencies as confounders disguises mediational relationships. Balance and reciprocity C B A C B A Figure 12.2 Attribute Data Transposed onto Transitive Tie Formation = Non-drinker = Alcohol drinker 173 parameters found to be statistically significant may obscure drug-related selection mechanisms in a similar fashion. One way to test for whether dyadic and triadic configurations confound or mediate drug-based selection processes is to consider these parameters without controlling for dyadic and triadic effects. If selection mechanisms become significant, then we can conclude that the dyadic and triadic configuration occur in conjunction with behavioral selection mechanisms. Further exploration would be important to assess the temporal sequence (i.e., causality) of these effects. Future studies should employ these kinds of investigations. A full description of modeling assumptions, implications and limitations are listed in Appendices A and B. 12.6 Conclusions In conclusion, network modeling identifies social aspects of behaviors, where behavioral change is assumed to occur through structured social interactions. On the other hand, logistic regression relies on the assumption of independence, an approach that can effectively discern social (alter-based) versus internal (ego-based) dependencies. Relational versus correlational inferences reveal very different information about social mechanisms. Comparing findings from the two analytic perspectives was particularly effective for alcohol behaviors. Assimilation of friends’ alcohol using behaviors was found to occur in the standard condition in both methodologies, while other drug activities showed inconsistent findings across techniques. Results for the experimental program implementation reveal a potential mechanism for reducing alcohol behaviors. The interactive nature of the program may have been effective in reducing the appeal of having alcohol-using friends, and through the increased popularity of non-users, alcohol cognitions of the entire class are reduced. Evidence from implicit cognition research implicates implicit processes in drug addiction, and helps explain how established drug associations can make relapse inevitable for some people. Therefore, encouraging social exchanges around substance use may be more effective in producing long term behavioral change that ultimately 174 become cognitively embedded and socially supported. The importance of implicit cognitions in social dynamics is underscored by the strong findings in logistic regression, showing that cognitive contagion is stronger than behavioral contagion. Inconsistent findings across methodologies suggest that there may be insufficient power to detect simultaneous selection and influence mechanisms in network modeling. However, network modeling results provide a context for regression correlative findings, providing rich information about social-behavioral dynamics and structural dependencies.. The findings in these studies highlight the importance of creating a partnership between social network modeling and regression approaches. Ultimately, network structures are meaningless without understanding the mechanisms that the goal-driven entities are following. 175 REFERENCES Ajzen, I., & Fishbein, M. (1980). Understanding Attitudes and Predicting Social Behavior. Englewood Cliffs, NJ: Prentiss Hall. Akers, R. L. (1992). Drugs, Alcohol, and Society: Social Structure, Process, and Policy. Belmont, CA.:Wadsworth Publishing Company. Alexander, C., Piazza, M., Mekos, D., Valente, T. (2001). Peers, schools, and adolescent cigarette smoking. Journal of Adolescent Health, 29(1), 22-30. Ambady, N., & Rosenthal, R. (1992). Thin slices of expressive behavior as predictors of interpersonal consequences: a meta-analysis. Psychological Bulletin, 111(2), 256-274. Ambady, N., & Rosenthal, R. (1993). Half a minute: Predicting teacher evaluations from thin slices of nonverbal behavior and physical attractiveness. Journal of Personality and Social Psychology, 64(3), 431-441. Ames, S. L., Franken, I. H. A., & Coronges, K. (2006). Implicit cognition and drugs of abuse. In R. W. Wiers & A. W. Stacy (Eds.), Handbook of Implicit Cognition and Addiction (pp. 363–378). Newbury Park, CA: Sage. Ames, S. L., & Stacy, A.W. (1998). Implicit cognition in the prediction of substance use among drug offenders. Psychology of Addictive Behaviors, 12(4), 272-281. Ames, S. L., Sussman, S., Dent, C., & Stacy, A. (2005). Implicit Cognition and Dissocative Experiences as Predictors of Adolescent Substance Use. American Journal of Drug and Alcohol Abuse, 31(1), 129-163. Asch, S. E. (1956). Studies of independence and conformity: A minority of one against a unanimous majority. Psychological Monographs, 70(9), 1-70. Bandura, A. (1977). Social Learning Theory. Englewood Cliffs, NJ: Prentice Hall Bandura, A. (1989). Human agency in social cognitive theory. American Psychologist, 44(9), 1175-1184. Barrera, M. (1986). Distinctions between social support concepts, measures, and models. American Journal of Community Psychology, 14(4), 413-445. Benthin, A., Slovic, P. Moran, P., Severson, H., Mertz, C.K., & Gerrard, M. (1995). Adolescent health-threatening and health-enhancing behaviors: a study of word association and imagery. Journal Adolescent Health, 17(3), 143-52. Blair, I. V., Ma, J. E., & Lenton, A. P. (2001). Imagining stereotypes away: The moderation of implicit stereotypes through mental imagery. Journal of Personality and Social Psychology, 81, 828-841. 176 Bovasso, G., Szalay, L., Biase, V., & Stanford, M. (1993). A Graph theory model of the semantic structure of attitudes." Journal of Psycholinguistic Research, 22(4), 411-425. Burk, W. J., Steglich, C. E. G., & Snijders, T. A. B. (2007). Beyond dyadic interdependence: Actor-oriented models for co-evolving social networks and individual behaviors. International Journal of Behavioral Development, 31(4), 397-404. Burt, R. S. (1987). Social contagion and innovation: Cohesion versus structural equivalence. American Journal of Sociology, 92(6), 1287-1335. Carley, K. M. (1986a). An Approach for relating social structure to cognitive structure. Journal of Mathematical Sociology, 12(2), 137-189. Carley, K.M. (1986b). Knowledge acquisition as a social phenomenon. Instructional Science, 14(3), 381-438. Carley, K. M. (1990). Group stability: A socio-cognitive approach. Advances in Group Processes, 7, 1-44. Carley, K.M. (1991). A Theory of group stability. American Sociological Review, 56(3), 331- 354. Carley, K. M. (1997). Extracting team mental models through textual analysis. Journal of Organizational Behavior, 18, 533-558. Carley, K., & Palmquist, M. (1991). Extracting, representing, and analyzing mental models. Social Forces, 70, 601-636. Centers for Disease Control and Prevention (CDC). Youth Risk Behavior Surveillance --- Selected Steps Communities, United States, 2007. Surveillance Summaries, 2008. MMWR, 57, 1-27. Choudhury, T. (2004). Sensing and modeling human networks. (Doctoral dissertation, MIT, 2004). Unpublished. Collins, A. M., & Loftus, E. F. (1975). A Spreading-activation theory of semantic processing. Psychological Review, 82(6), 407-428. Conrad, K. M., Flay, B. R., & Hill, D. (1992). Why children start smoking cigarettes: Predictors of onset. British Journal of Addiction, 87(12), 1711-1724. Contractor, N. S., Wasserman, S., & Faust, K. (2006). Testing multitheoretical, multilevel hypotheses about organizational networks: An analytic framework and empirical example. Academy of Management Review, 31(3), 681-703. Decety, J., & Sommerville, J. A. (2003). Shared representations between self and other: a social cognitive neuroscience view. Trends in Cognitive Sciences, 7(12), 527-533. 177 Deese, J. (1965). The Structure of associations in language and thought. Baltimore: Johns Hopkins Press. DiMaggio, P. (1997). Culture and cognition. Annual Reviews in Sociology, 23(1), 263-287. Dovidio, J. F., Kawakami, K., & Gaertner, S. L. (2002). Implicit and explicit prejudice and interracial interaction. Journal of Personality and Social Psychology, 82(1), 62-8. Dunbar, R. I. M. (1998). The Social brain hypothesis. Evolutionary Anthropology, 6, 178-190. Duvarci, S., & Nader, K. (2004). Characterization of fear memory reconsolidation. Journal of Neuroscience, 24(42), 9269-9275. Ennett, S. T., & Bauman, K. E. (1994). The Contribution of influence and selection to adolescent peer group homogeneity: The case of adolescent cigarette smoking. Journal of Personality and Social Psychology, 67(4), 653-663. Field, M., Mogg, K., Zetteler, J., & Bradley, B. P. (2004). Attentional biases for alcohol cues in heavy and light social drinkers: the roles of initial orienting and maintained attention. Psychopharmacology, 176(1), 88-93. Fishbein, M., Hennessy, M., Kamb, M., Bolan, G. A., Hoxworth, T., Iatesta, M., Rhodes, F., & Zenilman, J. M. (2001). Using intervention theory to model factors influencing behavior change: Project Respect. Evaluation & the Health Professions, 24(4), 363-384. Fisher, L. A., & Bauman, K. E. (1988). Influence and Selection In the Friend-adolescent Relationship: Findings from Studies of Adolescent Smoking and Drinking. Journal of Applied Social Psychology, 18(4), 289-314. Franken, I. H. A., Kroon, L. Y., & Hendriks, V. M. (2000). Influence of individual differences in craving and obsessive cocaine thoughts on attentional processes in cocaine abuse patients. Addictive Behaviors, 25(1), 99-102. Franken, I. H. A., Stam, C. J., Hendriks, V. M., & Brink, W. (2003). Neurophysiological evidence for abnormal cognitive processing of drug cues in heroin dependence. Psychopharmacology, 170(2), 205-212. Friedkin, N. E. (1998). A Structural theory of social influence. New York: Cambridge University Press. Friedkin, N. E. (2004). Social cohesion. Annual Review of Sociology, 30(1), 409-425. Graham, J.W., Flay, B.R., Johnson, C.A., Hansen, W.B., Grossman, L., & Sobel, J.L. (1984). Reliability of self-report measures of drug use in prevention research: Evaluation of the Project SMART questionnaire via the test-retest reliability matrix. Journal of Drug Education, 14(2), 175–193. Granovetter, M. (1978). Threshold models of collective behavior. American Journal of Sociology, 83(6), 1420-1443. 178 Greenwald, A. G., & Banaji, M. R. (1995). Implicit social cognition: Attitudes, self-esteem, and stereotypes. Psychological Review, 102(1), 4-27. Hall, J. A., & Valente, T. W. (2007). Adolescent smoking networks: The effects of influence and selection on future smoking. Addictive Behaviors, 32(12), 3054-3059. Hintzman, D. L. (1990). Human learning and memory: Connections and dissociations. Annual Reviews in Psychology, 41(1), 109-139. Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences of the United States of America, 79(8), 2554-2558. Ibarra, H. (1992). Homophily and differential returns: Sex differences in network structure and access in an advertising firm. Administrative Science Quarterly, 37, 442-447. Jaccard, J., Blanton, H., & Dodge, T. (2005). Peer influences on risk behavior: an analysis of the effects of a close friend. Developmental Psychology, 41(1), 135-147. Johnson, J. C. (1996). Anthropological contributions to the study of social networks: A review. In S. Wasserman & J. Galaskiewicz (Eds.), Advances in the social and behavioral sciences from social network analysis. Thousand Oaks, CA: Sage. Jones, B. C., Jones, B. T., Blundell, L., & Bruce, G. (2002). Social users of alcohol and cannabis who detect substance-related changes in a change blindness paradigm report higher levels of use than those detecting substance-neutral changes. Psychopharmacology, 165(1), 93- 96. Jones, B. T., Jones, B. C., Smith, H., & Copley, N. (2003). A Flicker paradigm for inducing change blindness reveals alcohol and cannabis information processing biases in social users. Addiction, 98(2), 235-244. Kahneman, D. (2003). A Perspective on judgment and choice: Mapping bounded rationality. American Psychologist, 58(9), 697-720. Kandel, D. B. (1978). Homophily, Selection, and Socialization in Adolescent Friendships. American Journal of Sociology, 84(2), 427. Kandel, D. B. (1985). On processes of peer influences in adolescent drug use: A developmental perspective. Alcohol and Substance Abuse in Adolescence, 4(3-4), 139-163. Kandel, D.B., & Yamaguchi, K. (1993). From beer to crack: developmental patterns of drug involvement. American Journal of Public Health, 83(6), 851. Kandel, D. B., Yamaguchi, K., & Chen, K. (1992). Stages of progression in drug involvement from adolescence to adulthood: further evidence for the gateway theory. Journal of Studies in Alcohol, 53(5), 447-57. 179 Kirke, D. M. (2004). Chain reactions in adolescents’ cigarette, alcohol and drug use: similarity through peer influence or the patterning of ties in peer networks? Social Networks, 26(1), 3-28. Kobus, K. (2003). Peers and adolescent smoking. Addiction, 98 (Suppl1), 37-55. Krackhardt, D. (1987). Cognitive social structures. Social Networks, 9(2): 109–134. Latkin, C. A., & Knowlton, A. R. (2005). Micro-social structural approaches to HIV prevention: a social ecological perspective. AIDS Care, 17, 102-113. LeDoux, J. (1996). The Emotional brain. New York: Touchstone. Leventhal, H., & Cleary, P. D. (1980). The Smoking problem: A review of the research and theory in behavioral risk modification. Psychological Bulletin, 88(2), 370-405. Levesque, L. L., Wilson, J. M., & Wholey, D. R. (2001). Cognitive divergence and shared mental models in software development project teams. Journal of Organizational Behavior, 22(2), 135-144. Loftus, E. F., Miller, D. G., & Burns, H. J. (1978). Semantic integration of verbal information into a visual memory. Journal of Experimental Psychology, 4(1), 19-31. Lubman, D. I., Peters, L. A., Mogg, K., Bradley, B. P., & Deakin, J. F. W. (2000). Attentional bias for drug cues in opiate dependence. Psychological Medicine, 30(1), 169-175. MacLeod, C. M., & MacDonald, P. A. (2000). Interdimensional interference in the Stroop effect: uncovering the cognitive and neural anatomy of attention. Trends in Cognitive Sciences, 4(10), 383-391. Manguno-Mire, G. M., & Geer, J. H. (1998). Network knowledge organization: Do knowledge structures for sexual and emotional information reflect gender or sexual orientation? Sex Roles, 39(9), 705-729. Marcos, A. C., Bahr, S. J., & Johnson, R. E. (1986). Test of a bonding/association theory of adolescent drug use. Social Forces, 65, 135-161. Marsden, P. V., & Friedkin, N.E. (1996). Network studies of social influence. In S. Wasserman & J. Galaskiewicz (Eds.), Advances in the social and behavioral sciences from social network analysis. Thousand Oaks, CA: Sage. Masson, M. E. J. (1995). A Distributed memory model of semantic priming. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21(1), 3-23. Matsueda, R. L., & Anderson, K. (1998). The Dynamics of delinquent peers and deliquent behavior. Criminology, 36(2), 269-308. McAlister, A. L., Perry, C., & Maccoby, N. (1979). Adolescent smoking: onset and prevention. Pediatrics, 63(4), 650-658. 180 McCusker, C. G. (2001). Cognitive biases and addiction: an evolution in theory and method. Addiction, 96(1), 47. McPherson, M., Smith-Lovin, L., & Cook, J. (2001). Birds of a feather: Homophily in social networks. Annual Review of Sociology, 27, 415-444. Mitchell, J. P., Nosek, B. A., & Banaji, M. R. (2003). Contextual variations in implicit evaluation. Journal of Experimental Psychology: General, 132(3), 455-469. Monge, P. R., & Contractor, N. S. (2003). Theories of communication networks. Oxford, UK: Oxford University Press. Nelson, D., McEvoy, C. L., & Pointer, L. (2003). Spreading activation or spooky action at a distance? . Journal of Experimental Psychology: Learning, Memory, & Cognition, 29: 42-52. Newcomb, M. D., & Bentler, P. M. (1986). Substance use and ethnicity: differential impact of peer and adult models. Journal of Psychology, 120(1), 83-95. Osgood, C. E. (1964). Semantic differential technique in the comparative study of cultures. American Anthropologist, 66(3), 171-200. Osgood, D. W., Johnston L. D., O’Malley, P. M., & Bachman, J. G. (1988). The Generality of deviance in late adolescence and early adulthood. American Sociological Review, 53(1), 81-93. Palazzolo, E. T. (2005). Organizing for information retrieval in Transactive Memory Systems. Communication Research, 32(6), 726. Palazzolo, E. T., Serb, D. A., She, Y., Su, C., & Contractor, N. S. (2006). Coevolution of communication and knowledge networks in Transactive Memory Systems: Using computational models for theoretical development. Communication Theory, 16(2), 223- 250. Pearson, M., Steglich, C., & Snijders, T. (2006a). Homophily and assimilation among sport- active adolescent substance users. Connections, 27(1), 51-67. Pearson, M., Sweeting, H., West, P., Young, R., Gordon, J., & Turner, K. (2006b). Adolescent substance use in different social and peer contexts: A social network analysis. Drugs: education, prevention and policy, 13(6), 519-536. Pentland, A. P. (1986). Perceptual organization and the representation of natural form. Artificial Intelligence, 28(3), 293-331. Pentland, A.P. (2004). Social dynamics: Signals and behavior. ICDL, San Diego, CA, Oct 20-23. 181 Philpott, J. (1983). The Relative contribution to meaning of verbal and nonverbal channels of communication: A meta-analysis (Masters thesis, University of Nebraska, 1983). Unpublished. Reingen, P. H., Foster, B. L., Brown, J. J. & Seidman, S. B. (1984). Brand congruence in interpersonal relations: A social network analysis. Journal of Consumer Research, 11(3), 771. Rice, R. E., Donohew, L., & Clayton, R. (2003). Peer network, sensation seeking, and drug use among junior and senior high school students. Connections, 25(2), 32-58. Ritt-Olson, A., Unger, J., Valente, T., Nezami, E., Chou, C. P., Trinidad, D., Milam, J., Earleywine, M., Tan, S., & Johnson, C. A. (2005). Exploring peers as a mediator of the association between depression and smoking in young adolescents. Substance Use & Misuse, 40(1), 77-98. Robins, G., Elliott, P., & Pattison, P. (2001). Network models for social selection processes. Social Networks, 23(1), 1-30. Robins, G., Pattison, P., Kalish, Y., & Lusher, D. (2007). An Introduction to exponential random graph (p*) models for social networks. Social Networks, 29(2), 173-191. Robins, G., T. Snijders, T., Wang, P., Handcock, M., & Pattison, P. (2007). Recent developments in exponential random graph (p*) models for social networks. Social Networks, 29(2), 192-215. Roediger, H. L. (1990). Implicit memory. American Psychologist, 45, 1043-1056. Roediger, H. L., & McDermott, K. B. (1995). Creating false memories: Remembering words not presented in lists. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21(4), 803-814. Roediger, H. L., Meade, M. L., & Bergman, E. T. (2001). Social contagion of memory. Psychonomic Bulletin & Review, 8(2), 365-371. Rogers, E. (2003). Diffusion of innovations. New, York, NY: Free Press. Romney, A. K., & Faust, K. (1982). Predicting the structure of a communications network from recalled data. Social Networks, 4(4), 285-304. Rosenfield, A. G., Asavasena, W., & Mikhanorn, J. (1973). Person-to-person communication in Thailand. Studies in Family Planning, 4(6), 145-149. Rudman, L. A., & Click, P. (2002). Prescriptive gender stereotypes and backlash toward agentic women. Gender, Hierarchy, and Leadership, 57(4), 743-762. Runkel, P. J. (1956). Cognitive similarity in facilitating communication. Sociometry, 19, 178- 191. 182 Schacter, D. L. (1990). Toward a cognitive neuropsychology of awareness: Implicit knowledge and anosognosia. Journal of Clinical and Experimental Neuropsychology 12(1), 155-178. Schacter, D. L., & Graf, P. (1986). Preserved learning in amnesic patients: Perspectives from research on direct priming. Journal of Clinical and Experimental Neuropsychology, 8(6), 727-743. Scott, J. (2000). Social network analysis: A handbook (2nd ed.) Newberry Park, CA: Sage. Shimamura, A. P. (1986). Priming effects in amnesia: Evidence for a dissociable memory function. The Quarterly Journal of Experimental Psychology Section A, 38(4), 619-644. Sieving, R. E., Perry, C. L., & Williams, C. L. (2000). Do friendships change behaviors, or do behaviors change friendships? Examining paths of influence in young adolescents' alcohol use. Journal of Adolescent Health, 26(1), 27-35. Snijders, T. A. B. (1995). Methods for longitudinal social network data. New Trends in Probability and Statistics, 3, 211–227. Snijders, T. A. B. (2001). The Statistical evaluation of social network dynamics. Sociological Methodology, 31(1), 361-395. Snijders, T. A. B. (2005). Models for longitudinal network data. In P. Carrington, J. Scott, & S. Wasserman (Eds.), Models and methods in social network analysis (pp. 215-247). New York: Cambridge University Press. Snijders, T., Spreen, M., & Zwaagstra, R. (1995). The Use of multilevel modeling for analysing personal networks: Networks of cocaine users in an urban area. Journal of Quantitative Anthropology, 5(2): 85–105. Snijders, T. A. B., & Baerveldt, C. (2003). A Multilevel network study of the effects of delinquent behavior on friendship evolution. The Journal of Mathematical Sociology, 27(2), 123-151. Snijders, T. A. B., Steglich, C. E. G., & Schweinberger, M. (2007). Modeling the co-evolution of networks and behavior. In K. van Montfort, H. Oud, & A. Satorra (Eds.), Longitudinal models in the behavioral and related sciences (pp.41-71). Mahwah NJ, Lawrence Erlbaum. Snijders, T. A. B., & van Duijn, M. A. J. (1997). Simulation for statistical inference in dynamic network models. In R. Conte, R. Hegselmann, & P. Terna (Eds.), Simulating social phenomena (pp. 493–512). Berlin: Springer. Stacy, A. W. (1995). Memory association and ambiguous cues in models of alcohol and marijuana use. Experimental and Clinical Psychopharmacology, 3, 183–194. , Stacy, A. W. (1997). Memory activation and expectancy as prospective predictors of alcohol and marijuana use. Journal of Abnormal Psychology, 106(1), 61-73. 183 Stacy, A. W., Ames, S. L., Sussman, S., & Dent, C. W. (1996). Implicit cognition in adolescent drug use. Psychology of Addictive Behaviors, 10(3), 190-203. Stacy, A. W., Galaif, E. R. Sussman, S., & Dent, C. W. (1996). Self-generated drug outcomes in high-risk adolescents. Psychology of Addictive Behaviors, 10(1), 18–27. Stacy, A. W., Newcomb, M. D., & Ames, S. L. (2000). Implicit Cognition and HIV Risk Behavior. Journal of Behavioral Medicine, 23(5), 475-499. StataCorp (2007). Stata Statistical Software: Release 10. College Station, TX, StataCorp LP. Steglich, C., Snijders, T. A. B., & Pearson, M. (2006). Applying SIENA: An illustrative analysis of the co-evolution of adolescents’ friendship networks, taste in music and alcohol consumption. Methodology, 2(1), 48-56. Steglich, C., Snijders, T. A. B, & Pearson, M. (2007). Dynamic networks and behavior: Separating selection from influence. Manuscript submitted for publication. Strang, D., & Tuma, N. B. (1993). Spatial and temporal heterogeneity in diffusion. American Journal of Sociology, 99(3), 614-639. Sussman, S. (1995). Developing School-based Tobacco Use Prevention and Cessation Programs, Sage Publications: Thousand Oaks, CA Sussman, S., Dent, C. W., & Stacy, A. W. (2002). Project towards no drug abuse: a review of the findings and future directions, PNG Publications, 26, 354-365. Sussman, S., Dent, C. W., Stacy, A. W., Burton, D., & Flay, B. R. (1995). Psychosocial Predictors of Health Risk Factors in Adolescents. Journal of Pediatric Psychology, 20(1), 91-108. Sussman, S., Earleywine, M., Wills, T., Cody, C., Biglan, T., Dent, C. W., & Newcomb, M. D. (2004). The Motivation, skills, and decision-making model of drug abuse prevention, Substance Use & Misuse, 39(10-12), 1971-2016. Sussman, S., Pokhrel, P., Ashmore, R. D., & Brown, B. B. (2007). Adolescent peer group identification and characteristics: A review of the literature. Addictive Behaviors, 32, 1602-1627. Sussman, S., Simon, T. R., Stacy, A. W., Dent, C. W., Ritt, A., Kipke, M. D., Montgomery, S. B., Burton D., & Flay, B. R. (1999). The Association of group self-identification and adolescent drug use in three samples varying in risk. Journal of Applied Social Psychology, 29, 1555- 1581. Swanson, J. E., Rudman, L. A., & Greenwald, A. G. (2001). Using the Implicit Association Test to investigate attitude-behavior consistency for stigmatized behavior. Cognition & Emotion, 15(2), 207-230. 184 Szalay, L., Bovasso, G., Vilov, S., & Williams, R. E. (1992). Assessing treatment effects through changes in perceptions and cognitive organization. American Journal of Drug and Alcohol Abuse, 18(4), 407-428. Thornberry, T. P., Lizotte, A. J., Krohn, M. D., Farnworth, M., & Jang, S. J. (1994). Delinquent peers, beliefs and delinquent behavior: A longitudinal test of interactional theory. Criminology, 32(1), 47-83. Tyas, S. L., & Pederson, L. L. (1998). Psychosocial factors related to adolescent smoking: a critical review of the literature. British Medical Journal, 7(4), 409-420. Umphress, E. E., Labianca, G., Brass, D. J., Kass, E., & Scholten, L. (2003). The Role of instrumental and expressive social ties in employees’ perceptions of organizational justice. Organization Science, 14, 738-753. Urberg, K. A., Degirmencioglu, S. M., & Tolson, J. M. (1998). Adolescent friendship selection and termination: The role of similarity. Journal of Social and Personal Relationships, 15(5), 703-710. Valente, T. W. (1995). Network modeling of diffusion of innovations. Cresskill, NJ: Hampton Press. Valente, T. W. (1996). Network models of the diffusion of innovations. Computational & Mathematical Organization Theory, 2(2), 163-164. Valente, T. W., & Davis, R. L. (1999). Accelerating the diffusion of innovations using opinion leaders. The Annals of the American Academy of Political and Social Science, 566(1), 55-67. Valente, T. W., Fujimoto, K., Chou, C. P., & Spruijt-Metz, D. (2009). Adolescent affiliations and adiposity: A social network analysis of friendships and obesity. Journal of Adolescent Health, 45(2), 202-204. Valente, T. W., Gallaher, P., & Mouttapa, M. (2004). Using social networks to understand and prevent substance use: A transdisciplinary perspective. Substance Use & Misuse, 39(10), 1685-1712. Valente, T. W., Ritt-Olson, A., Stacy, A., Unger, J. B., Okamoto, J., & Sussman, S. (2007). Peer acceleration: effects of a social network tailored substance abuse prevention program among high-risk adolescents. Addiction, 102(11), 1804-1815. Valente, T. W., Unger, J. B., & Johnson, C. A. (2005). Do popular students smoke? The association between popularity and smoking among middle school students. Journal of Adolescent Health 37(4), 323-329. Walther, J. B. (1996). Computer-mediated communication: Impersonal, interpersonal, and hyperpersonal interaction. Communication Research, 23(1), 3-43. 185 Walther, J. B., Loh, T., & Granka, L. (2005). Let me count the ways: The interchange of verbal and nonverbal cues in computer-mediated and face-to-face affinity. Journal of Language and Social Psychology, 24(1), 36-65. Ward, J. C., & Reingen, P. H. (1990). Sociocognitive analysis of group decision making among consumers. Journal of Consumer Research, 17(3), 245-263. Wasserman, S., & Faust, K. (1994). Social network Analysis: Methods and applications, New York: Cambridge University Press. Watkins, P., Martin, C., & Stern, L. (2000). Unconscious memory bias in depression: Perceptual and conceptual processes. Journal of Abnormal Psychology, 109, 282–289. Watson, D. G., & Humphreys, G. W. (2000). Visual marking: Evidence for inhibition using a probe-dot detection paradigm. Perception & Psychophysics, 62(3), 471-481. Wellman, B., & Berkowitz, S. D. (Eds.). (1988). Social structures: A network approach. Cambridge: Cambridge University Press. Wills, T. A., & Cleary, S. D. (1999). Peer and adolescent substance use among 6th-9th graders: latent growth analyses of influence versus selection mechanisms. Health Psychology, 18(5), 453-63. Wilson, T. D., Lindsey, S., & Schooler, T. Y. (2000). A Model of dual attitudes. Psychological Review, 107(1), 101-26. Wittenbrink, B., Judd, C. M., & Park, B. (2001). Spontaneous prejudice in context: Variability in automatically activated attitudes. Journal of Personality and Social Psychology, 81(5), 815-827. Wyer, R. S., Jr., & Srull, T. K. (1986). Human cognition in its social context. Psychological Review, 93, 322-359. Yin, H. H., & Knowlton, B. J. (2005). Addiction and learning. In R.W. Wiers & A.W. Stacy (Eds.) The Handbook of implicit cognition and addiction. Thousand Oaks, CA: Sage Publications. 186 Appendix A: Assumptions of SIENA Simulation Procedures Rule Description Assumption Limitations Actor- Orientation Actors have complete control over their behaviors & relations, and have complete knowledge of everyone else’s behaviors and relations Actors know all information about network members relevant to friendship decisions and therefore the assumption is that estimates represent true preferences. It is unlikely that people aware of behaviors and relations of others besides those with whom they are directly connected or who are friends of friends’. True preferences for 'like' others may not be fully realized in a model that assumes such omnipotence. Models that weigh selection and influence mechanisms of local cliques more heavily might reflect a more accurate depiction of a network where there is limited visibility. In particular, marijuana activities may be carried out more discretely than alcohol, and may lead to an underestimation of preferences for marijuana homophily. Stochastic (Markov) Process Network changes are imputed based on the current network configuration, and ignore all previous configurations. These models assume that the immediate past of a network provides all the information necessary to predict its future. Friendships change and evolve over time. It is usually the case that friends who decide to dissolve their tie are doing so based on full information about one another and do so with deliberate and decisive action. Classroom friendships that are present at time 1 and absent at time 2 will most likely remain absent at times 3 and 4, regardless of whether there is great affinity for the behaviors of the ex-friend. Social dynamics in the natural world are learning systems, and a system that learns will remember, for instance, that persons A and B were once friends, dissolved their relation, and will very likely remain non-friends in the future. A system that “forgets" times 1 and 2 in order to estimate times 4 and 5 will not account for the fact that rejection of person B does not reflect behavioral preferences. Assumption of stochasticity could lead to conservative estimates of selection and influence effects. Continuous Change Simulations model continuous change between observed time points. Network configurations between observed time points are estimated with the assumption that network & behavioral changes occur continuously between observed states. People's relationships and their behaviors are continuously changing. Models that account for these local level negotiations show that iterations of these exchanges carried out by all actors can produce reproducible global effects. The limitations of this assumption is that the simulations are behaving as they are told to behave and, thus, simulated graphs based on faulty assumptions (e.g., actor- orientation, stochasticity) could lead to estimates that do not reflect the true data. Assumptions about multiple and continuous change could lead to inferences that are susceptible to specifications that are overly defined. Discrete Change Each interaction of change provides a discrete opportunity for each individual to make either a network or behavioral change. This assumes that people either form a tie, drop a tie, make no network changes, adopt the behavior, stop the behavior or make no behavioral changes, and that any of these changes are mutually exclusive. Changes in cognitions may be triggered simultaneously with a new friendship tie or behavioral change. For instance, the very presence of a new person in one's consciousness can alter associative links throughout the associative map. In cases where cognitive associations change concurrently with network changes, the effects would be distributed between influence and selection parameters. In the same way, behaviors and cognitions may change simultaneously. Therefore, the assumption that only discrete changes can occur might overlook mutual change steps in the cognitive- behavioral-network dynamics. 187 Appendix A: Assumptions of SIENA Simulation Procedures, Continued Rule Description Assumption Limitations Rate of Change The rate function is constant across all actors The rates of network and behavior change between observed time points are modeled the same for all actors, which assumes all actors have the same number of opportunities to change both their friendship ties and their behavioral activities. Rate assumptions mitigate individual differences and thus, cannot detect differential rates of change. By using a mean rate across all individuals, models might create either too many or too few opportunities for change resulting in inaccurate estimates of selection or influence mechanisms. If some individuals have a true rate that is lower than is modeled, network and behavioral dynamics will be underestimated. That is, given the rate of change estimate, individuals appear to frequently choose no-change, but in reality, may have had much fewer opportunities to make changes than was determined by the generalized model. On the other hand, if some individuals have more opportunities for change than is modeled, the amount of change might be less given the number of true opportunities, and therefore, dynamics might be overestimated. Objective Function Every person in the network has the same objective function Individuals evaluate their social and behavioral positions with the same strategy as represented by equivalent objective functions across all actors. Using a single objective function across the network suggests that all actors evaluate that knowledge about their network equivalently. In reality, students will vary in how they perceive their network, their preferences towards actions based on those evaluations, and their thresholds for behavioral and friendship changes. For example, those who are less socially aware or less attuned to specific behaviors may be less susceptible to social influences and selection, while those with high social awareness will be more influenced by their friends and more likely to select 'like-minded' others. Alternatively, some individuals may be innovators and prefer activities that move them further away from the norm, while others benefit from assimilation and homophily. Assuming all individuals have the same personal strategies reduces the accuracy of influence and selection parameter estimates. 188 Appendix B: Assumptions of Theoretical Model Rule Description Assumption Limitations Reciprocity Actors who receive a directional tie tend to reciprocate that tie regardless of other selection preferences. The assumption is that dyadic dependencies are intrinsic in the network, in that they occur as a result of structural necessity, rather than by design. As an endogenous network characteristic, tendencies for reciprocity must be controlled for in order to identify selection preferences based on attributes. Dyadic dependencies might mediate instead of confound relationships between hypothesis-driven attributes and network dynamics. For example, pairs of individuals who have an opportunity to reciprocate each others nominations will also have the opportunity to identify behavioral commonalities and choose to reciprocate the tie based on those shared behaviors. That is, if a directional tie exists, there is a greater likelihood that preferences for specified behaviors will be exercised compared to cases where no ties are present. In cases where reciprocity occurs in conjunction with or precedes homophily driven ties, the dyadic modeling assumption would obscure these relationships and attribute the change to dyadic dependencies rather than homophily. Transitive Triplets & Balance There are intrinsic tendencies to make a tie with a person who is within a triadic set. Triadic dependencies exist as a result of social network evolution. These relationships are intrinsic in that they occur as a result of human structural tendencies. It allows the researcher to state with statistical certainty that any tendencies to choose friends based on specific behaviors occurs in addition to the tendency of friendships to evolve in a transitive way. Triadic dependencies might mediate instead of confound relationships between hypothesis-driven attributes and network dynamics. The opportunity to close a triad will also frequently coincide with the opportunity to identify behavioral commonalities within the triad and form a network tie based on shared behaviors. Triadic configurations not only suggest one’s friends endorse the ‘third’ actor, but also express an endorsement of the third actors' behaviors. Selection of friends with similar behaviors is more likely if the tie is transitive or balanced compared to actors who are not embedded in triadic configurations. In cases where triadic dependencies facilitate attribute-based selection, the triadic modeling assumptions would generate significant parameter estimates for triadic configurations, and underestimate attribute-based selection. Modeling decisions designed to account for triadic dependencies as confounders are limited in that they do not detect triads as mediators. Ego-, Alter-, Similarity- covariate There are differential tendencies to choose friends with particular attributes, to make nominations by attribute status, and for pairs to choose each other if they share the attribute. The assumption is that outdegree, indegree and similarity effects based on demographic and other individual- level characteristics may confound the relationship between the hypothesis- driven attributes and friendship choices. In some cases, covariates may act as mediators, where those who tend to select friends based on one shared attribute are also more likely to select friends for other shared attributes. For example, two individuals are drawn to each other because of their similar drug behaviors, but the fact that they’re both girls made their friendship easier to forge. In cases where selection for the hypothesized attribute is facilitated by other shared attributes, the modeling decision to include other attributes could generate significant parameter estimates for covariate similarity, while underestimating the hypothesis driven-attribute selection effects. The result could be conservative estimates of the true selection tendencies towards hypothesized behaviors.
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
A network analysis of online and offline social influence processes in relation to adolescent smoking and alcohol use
PDF
Exploring the role of peer influence, linguistic acculturation, and social networks in substance use
PDF
Social network influences on depressive symptoms among Chinese adolescents
PDF
Investigating factors that influence peer relationships and obesity during middle childhood
PDF
Contextualizing social network influences on substance use among high risk adolescents
PDF
Adolescent social networks, smoking, and loneliness
PDF
Diffusion of a peer‐led suicide preventive intervention in secondary schools: strategies to increase effectiveness of peer‐led interventions
PDF
Influences of specific environmental domains on childhood obesity and related behaviors
PDF
The formation and influence of online health social networks on social support, self-tracking behavior and weight loss outcomes
PDF
System dynamics of in-home family eating behavior: insights from intensive longitudinal data using Ecological Momentary Assessment and wearable sensors
PDF
Modeling social and cognitive aspects of user behavior in social media
PDF
Mobile self-tracking for health: validating predictors, effects, mediator, moderator, and social influence
PDF
Friendship network position on adolescent behaviors: an examination of a broker position and the likelihood of alcohol and cigarette use
PDF
A social-cognitive approach to modeling design thinking styles
PDF
Evaluating social networks and impact of micro-influencers who promote e-cigarettes on social media
PDF
Examining the longitudinal influence of the physical and social environments on social isolation and cognitive health: contextualizing the role of technology
Asset Metadata
Creator
Coronges, Kathryn
(author)
Core Title
Social selection and influence of alcohol & marijuana implicit cognitions and behaviors: a longitudinal investigation of peer social network dynamics
School
Keck School of Medicine
Degree
Doctor of Philosophy
Degree Program
Preventive Medicine (Health Behavior)
Publication Date
08/07/2009
Defense Date
04/28/2009
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
Alcohol,implicit cognition,intervention,Marijuana,OAI-PMH Harvest,SIENA,social contagion,social network analysis
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Valente, Thomas W. (
committee chair
), Chou, Chih-Ping (
committee member
), Monge, Peter (
committee member
), Pentz, Maryann (
committee member
)
Creator Email
kate.coronges@usma.edu,kcoronges@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-m2544
Unique identifier
UC1499094
Identifier
etd-Coronges-2347 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-176870 (legacy record id),usctheses-m2544 (legacy record id)
Legacy Identifier
etd-Coronges-2347.pdf
Dmrecord
176870
Document Type
Dissertation
Rights
Coronges, Kathryn
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Repository Name
Libraries, University of Southern California
Repository Location
Los Angeles, California
Repository Email
cisadmin@lib.usc.edu
Tags
implicit cognition
intervention
SIENA
social contagion
social network analysis