Dynamics of adolescent friendship networks and smoking behavior

Abstract

The mutual influence of smoking behavior and friendships in adolescence is studied. It is attempted to disentangle influence and selection processes in reciprocal and non-reciprocal friendships. An actor-based model is described for the co-evolution of friendship networks and smoking behavior. This model considers alternative selection and influence mechanisms, and models continuous-time changes in network and behavior. The data consists of a longitudinal sample of 1326 Finnish adolescents in 11 high schools. Findings suggest that selection as well as influence processes play an important role in adolescent smoking behavior. Selection had a relatively stronger role than influence, in particular when selecting non-reciprocal friends. The strength of both influence and selection processes decreased over time.

Introduction

One of the main preventable causes of cancer, heart disease, and premature death is cigarette smoking (Ezzati and Lopez, 2003, Office for National Statistics, 1997, US Department of Health and Human Services, 1994, Warren et al., 2006). Many youngsters experiment with smoking, which often results in becoming a regular smoker in adulthood (Chassin et al., 1996). During adolescence, smoking behavior tends to be similar among friends (Bauman et al., 1984, Eiser et al., 1991, Ennett et al., 1994, Sussman et al., 1990). This similarity in smoking behavior, which can be regarded as network autocorrelation (Doreian, 1989), could be caused by selection of similar others as friends as well as by influence processes where friends adjust their smoking behavior to each other, or by a combination of these. This article will demonstrate the use of stochastic actor-based models (Snijders et al., 2007a, Steglich et al., submitted for publication) capable of disentangling influence and selection processes by simultaneously representing changes in friendship network structure and changes in smoking behavior among adolescents. In particular, the impact of friendship reciprocity on selection and influence processes will be explored. Reciprocal friendships may offer higher friendship quality, which in turn could result in more opportunities for influence processes leading to smoking behavior similarities among friends (Parker and Asher, 1993, Urberg et al., 2003). Pearson and Michell (2000) examined non-reciprocal and reciprocal friendships and concluded that adolescents in the periphery of peer groups were the most important targets of influence.

Several studies attempted to disentangle selection and influence processes in the context of smoking behavior and suggested that selection of friends based on smoking behavior may be just as important as influence processes to explain similarity among friends, or even more important (Cohen, 1977, De Vries et al., 2006a, Ennett and Bauman, 1994, Fisher and Bauman, 1988, Kandel et al., 1978, Mercken et al., 2007). Studies that considered friendship reciprocity showed mixed results. Several studies found stronger support for influence within reciprocal compared to non-reciprocal friendships (Mercken et al., 2007, Parker and Asher, 1993, Urberg et al., 2003), other researchers did find strong support for influence of non-reciprocal or desired friends (Aloise-Young et al., 1994).

Disentangling selection and influence processes, as well as the role of friendship reciprocity in these processes, is difficult, due to the dynamic interdependent nature of friendship networks and smoking behavior. Previous studies have three main shortcomings. First, although most previous studies did include important alternative influence processes such as the influence of parental and sibling smoking (Avenevoli and Merikangas, 2003), they did not control adequately for alternative explanatory selection mechanisms. A smoking adolescent, for example, might choose a smoker as a friend because this individual already indicated the adolescent as his friend (reciprocity) or because this particular person was already a friend of the adolescent's other friends (transitivity). Further, the selection of this smoking friend might be based not on similarities in smoking behavior but on similarities in age, gender, alcohol consumption, school achievement, etc. Support for these alternative causes of tie formation was found by previous researchers (Burk et al., 2007, McPherson et al., 2001, Snijders and Baerveldt, 2003). Failing to control for alternative mechanisms might result in an overestimation of the strength of smoking-based selection processes. Second, researchers did not consider the continuous changes of network structure and smoking behavior over time happening between observations. Longitudinal data is mostly gathered at only a few discrete moments, which makes it impossible to unequivocally identify the processes responsible for a network or behavioral change. In between two observation moments, changes will occur in friendships and smoking behavior, and a change may even be followed by a change back to the original value before the next observation moment. Fig. 1 demonstrates influence and smoking-based selection processes that are likely to be diagnosed incorrectly on the basis of discrete observations if change between the observations is not accounted for. Consecutive observations are denoted here by T1 and T2, and some sequences of changes that may have occurred between observations are also indicated. Analysis techniques that are based on classifying observed changes as being due to influence or selection without accounting for the possibility of other intervening changes (De Vries et al., 2006a, Ennett and Bauman, 1994, Mercken et al., 2009) may be misleading, and it is preferable to use a technique that does take this possibility into account.

Finally, independence assumptions that underlie the employed statistical methods are violated. Even more advanced statistical techniques such as structural equation modeling used for this type of data (De Vries et al., 2006a, Mercken et al., 2009) assume incorrectly that there are no dependencies caused by the network structure of an adolescent. For example, a given individual's value on smoking behavior could appear within more than one observation, e.g., as the smoking behavior dependent variable for one case, and as smoking behavior of one of the friends supplying data for the independent variables in other cases.

New social network analysis methods have recently been developed which are able to consider alternative explanatory selection mechanisms, to model continuous-time changes in smoking behavior and friendship networks, and to take dependencies into account caused by the network structure. Stochastic actor-based models (Snijders, 2001, Snijders, 2005) have been developed to include network and behavior co-evolution (Snijders et al., 2007a, Steglich et al., submitted for publication). The following section will describe such an actor-based model for network–behavior co-evolution in the context of adolescent friendship networks and smoking behavior. A more extensive introduction is given in Snijders et al. (2010).

Actor-based models for network–behavior co-evolution assume that at two or more observation moments, a directed network and one or more behavioral variables are observed for a finite set of social actors. In our study, the actors are adolescents in a school. The network is a dichotomous relational variable, in our case indicating who directs friendship ties to whom. The behavior is assumed to be a dichotomous or discrete ordinal variable, in our case smoking behavior. Adolescents can change their smoking behavior, and also their friendship ties, in response to the current friendship network structure and the smoking behavior of the other adolescents in the network. It is assumed that all actors are fully informed about the state of the network, covariates and smoking behavior of all other actors in the network. Actors are only allowed to change their own outgoing ties and their own smoking behavior; they cannot make changes in outgoing ties or smoking behavior of other actors. Each adolescent is furthermore assumed to make decisions to change friendship ties or smoking behavior by probabilistic rules depending on the current configuration of network and behavior. The functions that determine the probabilities of change are called the objective functions, and there are separate objective functions for network change and for behavior change. Probabilities of change to a particular network or behavior state are higher accordingly as the objective function is higher; see Snijders et al., 2007a, Snijders et al., 2010. One interpretation is that the changes are the result of choices to optimize the actor's position in the network according to short-term preferences and constraints combined with random disturbances, and the objective function represents these short-term preferences and constraints. Finally, all actors consider and execute network and smoking behavior changes independently, given the current state of the network and everybody's behavior. Actors may change only one friendship tie or one level of smoking behavior at any moment in time. This implies that actors may react to each other's changes in friendship ties and smoking behavior, but do not negotiate or otherwise make joint changes based on a prior agreement. Therefore a negotiation like ‘when you start smoking, I’ll become your friend’ would need to be represented as the result of two smaller changes, between which the causal link cannot be enforced: ‘you may start smoking, but whether I will become your friend remains to be seen’. Note that while an actor cannot be certain that starting to smoke will result in a friendship, if smoking similarity has a positive effect on friendship selection then the actor does know that starting to smoke will increase the probability of being considered as a friend by smoking schoolmates.

The actor-based models for co-evolving networks are modeled according to a continuous-time Markov process in which likely developmental trajectories between observation moments are imputed (continuous-time property), and changes adolescents make are assumed to depend only on the current state of affairs (Markov property).

To model the co-evolution of friendship and smoking two models are created, one for friendship network change and one for smoking behavior change. The network evolution and behavior evolution models are integrated as one internally dependent process. In this manner, the current state of the continuously changing friendship network can be a dynamic constraint for changes in smoking behavior, while simultaneously the current state of smoking behavior can act as a dynamic constraint for changes in the friendship network.

Using this approach, four main research questions are addressed in this study:

• Do adolescents select friends based on similar smoking behavior?

• Are adolescents influenced by friends to adjust to their smoking behavior?

• Does the strength of these selection and influence processes differ for non-reciprocated and reciprocated friendships?

• What is the relative contribution of selection and influence processes over time?