Introduction
Health economists have long been interested in examining the determinants of, and potential policies for, reducing unhealthy behaviors in the population. Although a main focus in this area has historically been on issues of policy involving taxation, access restrictions, advertising, etc., a shift toward evaluating the basic social or nonmarket determinants of unhealthy behaviors has occurred in the literature. This is perhaps most obvious in the research regarding children and adolescent behaviors, wherein peer pressure is often thought to play a substantial role in determining choices such as smoking decisions. Indeed, there is now a large and growing literature in health economics that asks variants of the question, ‘‘Do peers influence an individual’s health behavior decisions?’’ The areas of interest range from substance use to eating behaviors and weight outcomes whereas the peer group definitions range from best friends to classmates, residential neighbors, and beyond. Indeed, although there have been several recent reviews of the literature examining social effects on health behaviors (Fletcher, 2010a, 2011a), these papers are being updated because of the rapid expansion in research in this area. This new research is based on previous work; moreover, quasi-experimental methods as well as new identification strategies are utilized in the research.
One reason for the increasing interest in achieving these research milestones is the policy implications of the existence of peer effects in health behaviors. Specifically, peer effects often imply a ‘social multiplier’ for interventions – if the health of one individual is increased, the effect of the intervention may be multiplied through peers. This type of social effect is seen as an ‘endogenous social effect’ in the literature of economics. In contrast, peer effects that operate through the characteristics of peers are labeled ‘exogenous social effects or contextual effects’ (see Manski, 1993). The presence of such endogenous social effects could increase the potential benefits of intervention without increasing the costs. In contrast to these benefits, the presence of peer effects could also work to spread unhealthy behaviors (such as smoking). The awareness of a social multiplier operating in determining health can also help to inform whether targeted (e.g., based on influential individuals within networks) or broad-based policy is more effective. Also, peer effects imply that the composition of a person’s neighborhood and/or school could affect his/her health behavior; because many policies can reorganize peer groups, such as school ability grouping (tracking), busing, school grade-span configuration, and residential zoning, there are a host of potentially important policy domains that plays a role in reducing poor health behavior in the presence of peer effects.
Although research regarding health decisions of socially connected individuals may continue to expand along with the multiple growth of interaction within social ties, the empirical hurdles to credibly estimating peer influence remain relatively unchanged and difficult to overcome. This article discusses some of the general empirical issues with their brief history, current controversies, and future directions.
Empirical Issues
Just as the policy and health importance of peer effects is likely to be substantial, so too is the empirical difficulty of credibly estimating causal effects. There are (at least) four standard primary empirical issues that researchers face. In many empirical settings, some are generic problems of measurement and omitted variables, whereas others are somewhat specific to peer effects research.
First, researchers must define a relevant peer group. This step seems simple, but data limitations typically force researchers to define peers on the basis of convenience rather than on theory. This has created peer group definitions that range from state-based groups to nominated best friends, and everything in between. For example, Harding (2003) uses census tracts, Evans et al. (1992) use metropolitan level data, Case and Katz (1991) use city block level data, Fletcher (2010a) uses school grades, Fletcher (2010b) uses school classrooms, Mayer and Puller (2008) use ‘Facebook Friends,’ and Sacerdote (2001) uses roommates to create relevant reference groups for the outcomes to be examined. Although there are several data sets that include nominated friends and peers, the vast majority do not. New data sets may reduce this issue over time, particularly those collecting online social network data, but this will raise the issue of whether online social contacts represent an important and relevant peer group for the determination of health decisions, and if they do, what types of health decisions are relevant when considering online peers.
A second empirical difficulty is the endogeneity of peer groups. Does a person smoke because his friend smokes or did he choose his friend for the sake of smoking? Because individuals typically have some degree of choice over their interaction with others (schoolmates, neighbors, friends, etc.), separating peer selection from peer influence is a particularly difficult empirical problem, and peer selection effects would typically inflate standard estimates of ‘peer effects.’ In fact, there seems to be a ‘relevance-endogeneity’ trade-off between the first and the second empirical difficulties (Fletcher, 2010a). As the researcher broadens the definition of the peer group (such as pertaining to the state level), the endogeneity of the peer group probably diminishes, but the relevance of the peer group may weaken. In contrast, best friends are probably a relevant definition of a peer group for many health behaviors but the endogeneity of best friend is magnified.
A third empirical difficulty in peer effect research lies in its potential nature for omitted variable bias through shared influences. For example, smoking bans may reduce tobacco use in all members of a school-based or community-based peer group. These shared factors can lead to inflated estimates of peer effects if sufficient control variables are not included.
A fourth empirical difficulty in peer effects research is the reflection problem (Manski, 1993), where the researcher may be unable to distinguish between whether Bill influences Ted or Ted influences Bill. Although it is not essential to disentangle these two influences in order to establish whether there is any social effect for determining health behaviors, it can be useful to separate these effects in order to understand the importance of the initial causal effect as against the feedback effects to further understand the processes of health spillovers. Although most researchers explicitly acknowledge each of these difficulties, they often adopt different approaches in attempting to overcome them.
Indeed, there is a two-decade-old history examining peer effects in many health behaviors, which can provide some examples of the difficulty with this research topic while outlining ways that other researchers have attempted to circumvent the empirical issues as outlined above. Typically, researchers have used neighborhood or school-based definition of ‘peers’ when examining health behaviors such as tobacco, alcohol, and drug use.