Peer Effects, Social Networks, and Healthcare Demand

Overview

The notion that social influences are important in both the development of numerous health outcomes and decisions related to healthcare use has great intuitive appeal. Health economists have modeled many potential mechanisms through which social influences affect outcomes ranging from practice patterns among providers to the choice to engage in risky behaviors. Effectively understanding the nature and scope of these effects is critical; if such influences are ignored estimates of the impact of policy interventions will in many cases be biased because they neglect the indirect pathway that occurs due to spillovers or what is known as the social multiplier effects. There have been many recent theoretical and empirical developments in this area on this topic. For example, studies on animal populations have shown that there is an association between social status and increased odds of specific diseases due to biochemical responses to low status affecting a creature’s immune system. More abstractly, economic theorists are increasingly directly modeling concerns for social status in the specification of utility function because this is as important an aspect of human decision-making potentially including the choice of medicine as one’s occupation.

Research on social interactions first appeared in the economics literature with Veblen’s analysis of conspicuous consumption, where consumption levels are used to signal wealth. Formal analysis of the influence of social groups in economic models developed following Schelling (1971) who demonstrated how the existence of these interactions may result in the formation of ghettos and segregation of individuals across neighborhoods, even in situations where most individuals prefer living in an integrated neighborhood. Formal models generally include social interactions by allowing for strategic complementarities, which occur when the marginal utility to one person of undertaking an action is increasing with the average amount of action taken by one’s peers. Although the initial developments in this literature were primarily made by theorists, there has been both a growing body of empirical work and policies proposed to take advantage of social interactions in health-related behaviors. This research is summarized in the following text.

Definitions Of Peer Effects, Social Learning, And Social Network Effects

As with many topics which claim interdisciplinary roots, the concept of social network effects as it relates to behavior can fall prey to semantic differences that may confuse many readers. Here, first the concepts of peer effects and social learning are defined which, in addition to being common topics in the economic literature, are the key foundation of social network models and effects. This is then followed by the definitions of the concept of social network models, which represent social ties through which peer effects and social learning occur across defined communities.

Peer Effects And Social Learning

Peer effects are commonly studied in economics and studies take a broad view of what constitutes peer influence. For example, one study examined decisions by school children in Kenya of whether to take drugs that kill intestinal worms already in the body. These drugs directly helped the individual who takes them, and generate positive externalities by breaking the transmission cycle, a pathway commonly referred to as the social multiplier. Researchers often try to make a distinction of whether the peer’s behavior versus their peer’s characteristics influenced one’s decision because only the former pathway would lead to a social multiplier.

For policy purposes, a key issue is to understand the channels through which peer effects operate. In this example, do children take these drugs because of information sharing from communicating with those who took the drug earlier, social learning by observing how others’ behavior and subsequent outcomes, reduced stigma or identity/image concerns may be lessened because others have taken the drug, or is it simply imitation? Although understanding the pathway the social interactions operate is important as discussed in the next subsection, even identifying these effects is challenging.

Part of the challenge researchers must face is how to properly define an individual’s peer group. In many papers researchers make ad hoc assumptions on the structure of peer groups. For example, only people who work in the same department within an organization or all individuals in a certain geographic area are considered. Researchers thus rely on the use of aggregates – such as the average characteristics or lagged behaviors of all classmates – to proxy for the social network. Such a setup is constraining: It means that individuals within a group must interact with each other, rather than with individuals outside the group. This can be a strong assumption if, for example, groups are formed by the researcher at the grade level, which suggests that students only interact with kids in the same grade but not with kids in different grades at the same school.

Social Network Models

If the concept of a peer group is not defined correctly, measurement error can be introduced by potentially omitting relevant peers, and when those neglected channels are not considered the underestimation of actual information flows can occur. As a result, more recent research is trying to directly model the formation of social networks and utilizing methods from graph theory to consider the microstructure of interaction among individuals within a community. As Figure 1 outlines, while direct peer effects (panel (a)) and social norms shared throughout a community (panel (b)), each capture potential avenues of social learning and influence, there is a third way of conceptualizing this pathway, where the structure of individual social ties throughout a community can capture these community effects (panel (c)). Generally, social networks are defined as a set of actors and relationships (or ties) linking the actors. Networks can be egocentric in which case the network is built out from a subset of a population, or sociocentric, where a full population of individuals is included. Most social networks measured are egocentric due to the challenges related to data collection, though sociocentric networks are more empirically appealing to study as there are no assumptions required about missing data from unobserved individuals and social ties biasing results. Social network analysis can be used to study the structure of a social organization and how this structure influences the behavior of individual actors. As such, this kind of analysis extends the study of peer effects and social learning beyond a given actor’s social ties.

Social Network Effects And The Challenge Of Identification

Both intuition and previous work suggest that social influences and, by extension, social networks, are important when driving economic behavior. Assuming that these factors are significant, there remains the difficult question of assessing the magnitude of such an effect. There are three main identification challenges facing an empiricist. The first was termed the ‘reflection problem’ by Manski (1993) and is an issue that mimics a simultaneity problem. For example, in studying cigarette smoking a reflection problem arises when student and peer smoking are determined simultaneously, which inherently convolutes the measure of peers’ influence.

The next challenge for the analyst is more complicated because individuals generally choose their friends in part based on the characteristics they favor. This second challenge is a form of selection bias and leads to a correlated unobservables problem. Social networks are not created in a vacuum, as the result of a random stochastic shock. Rather, they are formed based on the preference of individual actors (nodes) in the vast majority of cases. Therefore, it is entirely conceivable that sorting into networks (also known as homophily) may occur based on traits and behaviors that are linked with and/or signals of future preference. Thus, one usually does not know with certainty if an individual may be influenced to choose a specific diet or health plan because of the influence of their friend, or they chose their friend because of the friend’s revealed preference in the first place. In the latter case, it may be possible to observe what appear to be examples of social influence or contagion within a network, but are actually, primarily, artifacts of prior selection.

The third and final challenge involves the possible presence of unobserved group-level characteristics that affect both individuals and their peers. That is, some third confounding factor is responsible for the observed association between one’s behavior and that of their peers. Taken together, if an individual is a member of some group, can the analyst distinguish a role for the characteristics and behaviors of others in the group in influencing that individual’s choices?

More formally, much of the empirical literature on social economics has involved variations of a general linear model, dubbed by Manski the linear-in-means model.

where Y_i is the health outcome under study for individual i who is a member of peer group g at time t, X_igt is a vector of individual exogenous controls, X_-igt is a vector of contextual controls common to all members of group g since the -i notation indicates everyone but person i, Y_-igt is the mean peer choice in group g, G_g reflects common environmental influences affecting all members of group g, and ε_igt is a random error term with mean 0. β₁ expresses individual effects, β₃ contextual effects, β₄ correlated effects, and β₂ endogenous social effects. The linear-in-means model thereby provides a formal expression to three hypotheses often advanced to explain the common observation that individuals belonging to the same group tend to behave similarly.

The ‘reflection problem’ occurs if the peer variable measures peer group members at time t, which is obtained at the same time one’s own health outcome is measured. It is called the reflection problem because it is similar to the problem of interpreting the almost simultaneous movements of a person and her reflection in a mirror. To overcome this challenge and identify the endogenous social effect, researchers generally ensure that all the regressors are known (predetermined) at the time of regression, which in theory avoids simultaneity problems. That is, the peer variable is constructed using earlier behaviors that were hopefully measured immediately before any interactions among the group g. Manski notes that if the transmission of peer effects really follows this temporal pattern, the identification problem is alleviated.

Empirical Approaches In The Estimation Of Peer Effects In A Social Network

Given the challenges outlined earlier, how can a researcher infer true parameter estimates for social influence along networks? Although this is a thorny (some might argue intractable) challenge, there are in fact a number of experimental and analytical approaches that can be used to control for selection bias among other identification issues that economists are quite familiar with.

Overcoming Selection Bias

Researchers have attempted to overcome the selection bias in one of three main ways. Studies have used insights from randomized experiments to induce credible exogenous variation into aspects of social networks in an effort to identify their impacts. This research design can be seen in a number of papers, including the work of Duflo and Saez (2003) who explored not only the existence but also the mechanism underlying peer effects in the context of demand for benefits at the workplace. Specifically, they examined the role of social learning on the choice of employer-sponsored retirement plans, using individual data on employees of a large university a random sample of employees and focused on the question of whether people are influenced by the decisions of other employees in the same department. In a subset of departments some individuals were encouraged by an offer of a financial incentive to attend a benefits information fair organized by the employer. Not all departments were treated and the authors compared both benefits fair attendance and retirement plan enrollment decisions across departments and also looked within departments comparing outcomes of those who receive the treatment with their untreated coworkers. Receiving the letter led to a large increase in the likelihood of attendance and untreated individuals within departments where some individuals treated also had higher odds of attending the fair and plan enrollment 5 and 11 months after the fair. This presents convincing evidence that peer effects likely influence demand for benefit plan decisions.

Another method to identify the impacts of social factors on health outcomes is the use of instrumental variables to mitigate the correlation between unobservables and social network variables. This approach has been used in a large number of studies that have examined the role of peers on health behaviors such as cigarette smoking and obesity. However, these studies are frequently critiqued because the statistical properties and economic validity of these instruments are of debate. For example, some used their friends’ birth weight as an instrument for whether their friend is currently obese in influencing whether one is overweight themselves. However, peer birth weight may influence other peer outcomes besides simply weight that may also directly affect one’s own health outcomes.

Finally, several studies have attempted to use very rich data to control for unobserved confounders to identify the effects of social networks on health outcomes and show that those with greater levels of contacts with friends and neighbors have a reduced likelihood of enrolling in a Medicare-managed care plan relative to purchasing a medigap policy or having coverage through Medicare alone. Although the authors do account for a large set of unobserved confounders it remains possible that more sociable households are more risk tolerant or more optimistic than less sociable households, thus making these households more open to purchasing newer or ‘riskier’ insurance products, such as Medicare Health Maintenance Organizations (HMOs). This strategy is also used in international datasets. It has been found that social network effects are large and that both temporal and spatial proximity among household heads is the mechanism underlying this effect. However, one always can be concerned that those who lead opinions in rural villages may also have certain characteristics favoring health plan adoption decisions.

Despite these limitations, it is worth noting that in certain contexts in the educational setting an explicit rule determines assignment to different peer groups. A sharp regression discontinuity design can be utilized when there is an explicit cutoff and individuals cannot change their behavior ex ante in an effort to sort to a specific side of the cutoff. This situation mimics a randomized experiment and one can simply compare individuals who just lie on either side of the cutoff. In an education setting, several countries share competitive admissions policies to secondary school leading to very different peer groups. For example, in China and Romania the secondary school system differs markedly from that of the USA which enabled researchers take advantage of the features and institutional structure of these systems. Specifically, students compete for positions in the higher ranked secondary schools by writing a high school entrance examination at the completion of junior middle school. Administrators at each senior high school grant admission to students whose exam performance is above a cutoff score. Thus, students who just get into a higher ranked school (perhaps by scoring only a point above the cutoff) have access to much stronger peers as measured by performance on the entrance examination, than students who scored just below the cutoff and now must attend a lower ranked school. One can imagine situations in healthcare settings where individuals are assigned to different treatment centers or nursing homes on the basis of health statistics creating an opportunity to examine how peers in different centers/homes affects one’s health.

In addition to the examples listed so far, a number of other methodological approaches have been advanced in the statistical community so as to mitigate the strength of assumptions in social network models. One model is to assign all subjects into two separate groups randomly, and within them assess whether there appears to be a contagion effect from nonneighbors by using time series results in the first bin to predict results in the second. Nonzero results suggest social influence has traveled along the network over time. Another, simpler approach, relates to setting parameter bounds. In this approach, bounded parameters greater than zero would show that the entirety of the observed effect could not be due to selection (homophily) alone, though the magnitude of the effect may be significantly decreased.

Selected Application Of Social Network Models To Health And Healthcare

A number of data sources such as the Add Health Study that collects survey data on self-reported friendship networks, and a growing number of social network data from electronic sources such as Facebook, LinkedIn, and others present researchers with new opportunities to study the impact of social learning and peer effects within networks on health behaviors over time. An earlier version of this approach is provided in Christakis and Fowler (2007) who used 32 years of data on 12 000 people from the Framingham Heart Study. This study garnered substantial attention from the popular press with a conclusion that obesity appears to spread through social ties.

Fortunately, these spirited academic debates have led to a multitude of methodological improvements that should shed new and more convincing light on the role of social networks in the spread of obesity.

In contrast to risky health behaviors, there have been relatively few studies that have performed true social network analyses in the context of demand for healthcare and health insurance. There have, however, been a number that have looked broadly at peer effects as they relate to purchasing decisions in health or in areas (such as insurance) closely related to health. The most convincing evidence on the role of social networks relates to health plan decisions addressing the question of whether the information one receives from their peers affects their choices; even when product quality is difficult to ascertain. This information may come from direct communication with peers who have already purchased a particular health insurance bundle. Alternatively, it may arise from the observation of peers’ purchasing decisions. This phenomenon as noted earlier is often referred to as social learning.

Whereas social learning has been extensively studied in theory, the empirical evidence is limited because social learning is difficult to identify in practice. Empirical analyses on the effects of social networks have difficultly providing direct evidence of causal relationships from consumption decisions and/or the product satisfaction of other members of one’s social networks on individual decisions related to consumption of health products due to various conceptual and data problems including selection bias. Selection bias arises because people tend to associate with others based in part on some group characteristics they favor that are unobserved by the researcher. Thus, observing that individuals in the same group make similar consumption decisions may simply reflect shared preference and not informational spillovers.

Exploring social effects in universities is a popular research design. Sorensen (2006) subsequently examined that health insurance selections are correlated across employees within the same department to multiple campuses in the University of California system. Nearly all full-time and some part-time employees are eligible to enroll in one of the health plans offered through the benefits program. He uses statistical models to examine whether individualor department-level factors influence the decision to choose specific health plans. His empirical evidence provides convincing evidence that social effects (i.e., decisions of coworkers) play a role on individuals’ choices of employer-sponsored health plans that is as large as many individual factors including age, income, and family status. The strength of the effect depends on factors such as the department’s size or the employee’s demographic distance from his coworkers. His research results have large policy implications because if all of one’s coworkers in a specific department have chosen the same plan, then the social influence may overpower any individual incentives to switch plans when the provider raises the premium.

Another study uses data from a field experiment that changed the size of work-related social networks for those who were randomly assigned the intervention. These changes in the size of social networks are not due to choices by the individuals themselves and are free from selection bias, thereby providing the authors a unique opportunity to estimate the causal impact of social networks on self-assessed measures of health. The effect of social networks on different health measures depends crucially on whether one holds a job. Those assigned to have a larger social networks report greater satisfaction with their mental and physical health when they are employed and the authors show that this effect is not due to the income channel.

In summary, the estimates (only a subset of which are discussed in the preceding paragraphs) indicate that social networks are an important determinant of the health insurance choices, health behaviors, and health as well. There is a large literature demonstrating positive associations between network size and mental health outcomes whereas another literature interestingly, finds that peers play small roles on actors in the medical system such as doctors in terms of their treatment choice and decisions to specialize. Future study of physician networks and social influence within them will benefit from new datasets being created by research teams.

Not only has there been studies investigating whether social influences exist but one may also wonder whether policies that aim to influence group dynamics subsequently shape individual health outcomes. Indeed social influences have been incorporated in many recent areas of health policy including efforts to reduce obesity. For example, weight-loss support groups have been in place for many years and use social influences in a manner similar to Alcoholics Anonymous to shape health behavior. In 2007, the academic journal Obesity devoted an entire issue to the evaluation of workplace interventions to reduce obesity. Last, several states including Arkansas have built policies around the mechanism of stigmatization as the channel of endogenous social effects, by providing students with weight report cards that provide information on their rank in the body mass index (BMI) distribution. Whereas each of these policies is built around specific social mechanisms that are hypothesized to alter obesity, their design does not appear to be based on a large body of evidence. Further, most seem to not have undergone a rigorous ex post policy evaluation of their effectiveness. Designing and evaluating policies that aim to take advantage of social multipliers are clearly areas for further research.

Final Comment

There is substantial evidence that individuals’ beliefs, actions, and choices in the health sector are impacted by beliefs, actions, and choices of their peers. Insights from behavioral economics on framing and social influences are increasingly being used by the healthcare industry to influence individual choice in regards to specific products. Although substantial progress has been made on the econometrics of identification of social interactions, more careful work is needed to understand the mechanisms driving these social effects as well as whether there are moderators. The authors believe that this can be accomplished using field experiments and credible research designs. Further, in the education literature there is growing evidence that peer impacts can be distributed unevenly and more work in health economics is needed to understand the consequence of heterogeneity among peers. In the sciences, research conducted with animal populations demonstrates that social influences operating within the environment an individual engages in also directly affect gene expression, suggesting biological mechanisms underlying the social interaction effects. Findings from further empirical studies on the form and mechanisms of peer interactions are needed to guide further theoretical work. Naturally, the converse holds and despite a burgeoning literature over the past 15 years, an incomplete understanding of such effects remains. When studies are performed and results are interpreted with an appropriate amount of care and caution, one believes that there can be a great deal inferred from peer and social network models of influence to researchers in the health economics community and both health policymakers and the healthcare industry.

Bibliography:

1. Christakis, N. A. and Fowler, J. H. (2007). The spread of obesity in a large social network over 32 years. New England Journal of Medicine 357, 370–379.
2. Duflo, E. and Saez, E. (2003). The role of information and social interactions in retirement plan decisions: Evidence from a randomized experiment. Quarterly Journal of Economics 118(3), 815–842.
3. Manski, C. (1993). Identification of endogenous social effects: The reflection problem. Review of Economic Studies 60(3), 531–542.
4. Schelling, T. C. (1971). Dynamic models of segregation. Journal of Mathematical Sociology 1, 143–186.
5. Sorensen, A. T. (2006). Social learning and health plan choice. RAND Journal of Economics 37(4), 929–945.
6. Beisetov, E., Kubik, J. and Moran, J. (2004). Social interaction and health insurance choice of elderly: Evidence from health and retirement survey. Mimeo: Syracuse University, Maxwell School for Citizenship and Public Affairs, Center for Policy Research, Working Paper.
7. Blume, L., Brock, W., Durlauf, S. and Ioannides, Y. (2010). Identification of social interactions. In Benhabib, J., Bisin, A. and Jackson, M. (eds.) Handbook of social economics, pp. 853–964. Amsterdam: Elsevier.
8. Ding, W. and Lehrer, S. F. (2007). Do peers affect student achievement in China’s secondary schools? Review of Economics and Statistics 89(2), 300–312.
9. Evans, W., Oates, W. and Schwab, R. (1992). Measuring peer group effects: A study of teenage behavior. Journal of Political Economy 100(5), 84–117.
10. Fernald, R. D. and Maruska, K. P. (2012). Social information changes the brain. Proceedings of the National Academy of Sciences of the USA 109(supplement 2), 17194–17199.
11. Ferschtman, C., Murphy, K. M. and Weiss, Y. (1996). Social status, education, and growth. Journal of Political Economy 104, 108–132.
12. Jackson, M. O. (2008). Social and economic networks. Princeton, NJ: Princeton University Press.
13. Jackson, M. O. (2011). An overview of social networks and economic applications. In Benhabib, J., Bisin, A. and Jackson, M. O. (eds.) The handbook of social economics, vol. 1A, pp. 511–579. Amsterdam: Elsevier Science.
14. Manski, C. (2000). Economic analysis of social interactions. Journal of Economic Perspectives 14(3), 115–136.
15. McClellan, M., McNeil, B. J. and Newhouse, J. P. (1994). Does more intensive treatment of acute myocardial infarction in the elderly reduce mortality? Analysis using instrumental variables. Journal of the American Medical Association 272(11), 859–866.
16. Miguel, E. and Kremer, M. (2004). Worms: Identifying impacts on education and health in the presence of treatment externalities. Econometrica 72, 159–217.
17. Moffitt, R. (2001). Policy interventions, low-level equilibria, and social interactions. In Durlauf, S. and Young, P. (eds.) Social dynamics, pp. 45–82. Cambridge: Brookings Institution Press and MIT Press.
18. Newman, M. (2010). Networks: An introduction. Oxford: Oxford University Press.
19. Soetevent, A. (2006). Empirics of the identification of social interactions; An evaluation of the approaches and their results. Journal of Economic Surveys 20(2), 193–228.
20. Trogdon, J., Nonnemaker, J. and Pais, J. (2008). Peer effects in adolescent overweight. Journal of Health Economics 27(5), 1388–1399.
21. Veblen, T. B. (1934). Essays on our changing order (edited by Leon Ardzrooni), pp. xviii, 472. New York: The Viking Press.