Newer Approaches And Extension Of Outcomes
Although the more traditional literature examining peer effects and health behaviors has focused primarily on substance use outcomes and has used a range of empirical approaches, the more recent literature in this area has broadened research designs and has dramatically expanded the range of health outcomes under study – especially, weight and mental health outcomes.
Apart from the literature in health economics, the set of studies that has received the most media attention is from Nicholas Christakis, James Fowler, and a set of coauthors. Their first study has brought a new outcome of interest to the literature by examining whether obesity is ‘socially contagious.’ Specifically, the authors have found that the chances of an individual becoming obese increased by more than 50% when his/her friend has become obese. The authors have used the Framingham Heart Study data, which contain up to 32 years of longitudinal measures of BMI for individuals in one area of Massachusetts. To these data, the authors have matched information from the original respondents’ records, on which respondents have individually been asked to name a person who can be contacted in case the survey team does not reach them directly at follow up; this contact person is treated by Christakis and Fowler (2007) as a ‘friend.’ Thus, the first issue with this research is whether the contact person is truly a peer. The authors estimate regressions using the following parsimonious empirical model:
where, the health (obesity) of person i is linked to person j and δ is the coefficient of interest – the endogenous social effect or ‘social multiplier.’ A positive estimate on d suggests that an intervention which reduces the chances of an individual becoming obese will also reduce the chances of obesity in his/her peer.
To overcome endogeneity, Christakis and Fowler (2007) have assumed that lagged health outcomes for the friend (healthjt-1) is a sufficient control, that is, after controlling for lagged obese status of a friend, they have assumed that there is no additional issue of friendship selection. Unfortunately, to the extent that this control variable does not completely eliminate selection effects, the estimated coefficient of interest (δ) will probably be upwardly biased. The authors have controlled for own-lagged health in order to control for aspects of the individual’s genetic disposition or other time-invariant characteristics. The second issue is confounding due to shared influences. Without explicitly controlling for shared environmental factors, the authors have appealed to a comparison between mutually nominated friends and nonmutual friends (with unreciprocated nomination), arguing that directionality of nominations does not matter if environmental confounding is the primary explanation. Finally, the authors neither discussed nor attempted to overcome the empirical complications from the reflection problem. Unfortunately, each of those empirical issues listed above would probably lead to upwardly biased estimates of peer effects. So, what proportion of the 50% estimated peer effect is due to bias and what proportion is an actual peer effect? To address these empirical concerns, Cohen-Cole and Fletcher (2008a) have provided an examination focusing on one of the empirical issues in peer effects models – shared environmental factors that may bias upward the estimates. The authors have used the National Longitudinal Study of Adolescent Health (Add Health), which includes the nationwide longitudinal data on adolescents in the US over approximately seven years. Although the Framingham study has a much longer time horizon and focuses on adults, the Add Health data contain information on actual ‘best friends’ who are named by the respondent; this is arguably a more appropriate peer than the contact person in the Framingham data. Cohen-Cole and Fletcher have first estimated eqn [5] on the basis of the Add Health data in order to replicate the baseline findings of Christakis and Fowler (2007) that are based on the Framingham data. Interestingly, both papers, using different data of different age groups, arrive at point estimates for d from eqn [1] for the ‘peer effect’ of BMI of 0.05, meaning that a one unit increase in a friend’s BMI over time is correlated with a 0.05 unit increase in one’s own BMI. However, when Cohen-Cole and Fletcher controlled for shared environmental factors such as school-fixed effects, the coefficient fell by approximately 40%, no longer being statistically significant. Thus, current evidence suggests that the empirical problems described above are problematic enough to reduce confidence in any peer effects in obesity resulting from this specific model.
In an attempt to further explore the potential upward bias in the Christakis/Fowler empirical model, Cohen-Cole and Fletcher (2008b) took an alternative approach. The authors asked the question: ‘Is the empirical model [5] so weak that it would produce estimates of peer effects in behaviors where the true peer effect should be zero?’, that is, the authors conducted a falsification test of the empirical model by showing that estimating eqn [5] with the Add Health data would also produce results suggesting ‘social contagion’ in outcomes that are unlikely to be contagious: acne, headaches, and height. Indeed, the estimates for peer effects in these health behaviors are in some case larger than the Christakis/Fowler estimates of peer effects in obesity. The results of the falsification exercise strongly suggest that the model is insufficiently specific to distinguish between true social effects and the alternative hypotheses as discussed above (e.g., endogeneity of friendships and exposure to shared environmental factors). As in previous work, Cohen-Cole and Fletcher (2008b) have shown that the magnitudes of the fictional social network effects are reduced and when shared environmental influences are controlled, these effects largely disappear.
Based on part on these findings, obesity and weight-related behaviors have been studied in several additional papers. Trogdon et al. (2008) use several empirical strategies to examine peer effects. They examine both grade-level peers, similar to the cross-cohort designs already discussed, as well as nominated friends. To control for shared environmental factors, the authors control for school-fixed effects. To address friendship selection and simultaneity bias, the authors use an instrumental variables strategy, where friend’s birthweight, weight of parents of friend, and other measures are used as instruments. The limitation with this approach is that it is unclear whether these variables are good instruments for friendship selection. It appears that the instruments have been mainly employed to reduce the importance of the simultaneity issue, though the instruments still need to be excludable from the equation determining one’s own weight. In addition to controlling for shared environmental influences, the authors use school-fixed effects to partially control for friendship selection. The implicit assumption with school-fixed effects is that within schools, friendships form randomly.
Like Trogdon et al. (2008), Renna et al. (2008) also use a single cross-section of the Add Health data to examine the correlation between own and friend’s weight outcomes; however, these two papers use different subsamples and Renna et al. focus only on nominated friends. Renna et al. use schoollevel fixed effects to control for shared environmental factors and also attempt to reduce the simultaneity issue with an instrumental variables approach. The authors also use the obesity status of parents of friends as instruments. To control for selection of friends, the authors include additional control variables and acknowledge that the estimates are likely to be biased upward. The authors find evidence for peer effects for both genders in the baseline models, whereas only females in the IV models, although the point estimates are very similar. Overall, these papers are suggestive of peer effects but are unable to control for the empirical issues necessary to make the evidence more conclusive.
However, three recent papers have attempted to overcome the methodological issues with the above papers by pursuing alternative research designs. Yakusheva et al. (2011) use the roommate design described above with females from a private Midwestern university. The authors show negative correlations between having a heavy roommate and own weight outcomes. Carrell et al. (2011) stretch the literature further into the outcome of physical fitness by using random assignment to squadrons in the US Air Force Academy in order to show that squadronmates’ level of physical fitness is highly correlated with one’s own fitness. Finally, using a new instrumental variable strategy (characteristics of friend of peer), a so-called ‘friend of friend’ instrument pioneered by Bramoulle´ et al. (2009) and Fortin and Yazbeck (2011) show some evidence of peer effects in fast-food consumption. Although these papers have considerably strengthened the research designs from past work and have extended the set of health behaviors, additional work is needed to further understand the potential for whether obesity is indeed ‘socially contagious.’ This work requires different (and hopefully more representative) samples and further replication.
In addition to weight-related outcomes, the literature examining peer influences in health outcomes has also begun to examine the realm of mental health. Although some older papers have attempted to examine social influences on suicidal behaviors, this literature is yet to incorporate newer and more rigorous research designs. Hence, the existence of peer effects is still uncertain. However, other measures of mental health have been explored recently. Eisenberg et al. (2011) have applied the roommate design to a variety of anxiety and depressive symptoms using a sample of freshman college students from two universities. The authors find no evidence of peer influence in measures of happiness. However, symptoms of anxiety appear to be correlated between roommates and there is some suggestive evidence of depressive symptoms being correlated between male roommates. See also Fletcher (2010c) for evidence that classmate mental health may reduce school performance.
Considering that research has expanded the domain of health behavior under study, new directions have been adopted in empirical methods on the basis of nonexperimental data. For example, a new direction in the study of social networks with implications for the study of health is the analysis of interdependent duration decisions. Because many health outcomes and behaviors have important time components such as smoking and drinking histories, utilizing new methods in this area could prove useful. The current state of the art includes the theoretical framework as outlined in Brock and Durlauf (2008) as well as the empirical applications of de Paula (2009) and de Paula and Honore (2010). Likewise, Fletcher and Ross (2012) have attempted to combine a control function approach with a cross-cohort design (as outlined above) to estimate the effects of best friend’s smoking and drinking behaviors on individual health choices. Work by Yves Zenou and colleagues have accumulated a set of papers that build a game theoretic model of network formation with interesting empirical implications (e.g., Calvo -Armengol et al., 2009).
In addition to these new research methods, there are also new data opportunities as well as research design opportunities emerging. Mayer and Puller (2008) leverage data from the social networking website Facebook.com in order to examine correlations between friends’ health behaviors, but they do not focus on causal inference. Mapping friendship networks through the use of cell phone usage information may also transform our ability to construct and track social networks in the future (Eagle et al., 2009). However, these new data sources will not alleviate the need to confront the difficulties of estimating empirical models of social influence.