Effect of large-scale social interactions on body weight

Abstract

I estimate models of endogenous social interactions in body weight at the county and state levels. The results show that dispersion in body weight across time and space in the U.S. is not clearly excessive, and that much of this variation can be attributed to observable individual and regional characteristics. Models exploiting variants of methods proposed by Glaeser et al. (2003), fixed effects, instrumental variable and split-sample instrumental variable methods to address endogeneity suggest that there are not large social multipliers on body weight outcomes. The evidence suggests there may be small multipliers on BMI, obesity, and morbid obesity. There is no evidence that underweight is subject to a social multiplier. The results are sensitive to specification.

Introduction

If the perceived costs and benefits of high or low body weight depend on the distribution of body weights of the people in one's community, body weight may be subject to a social multiplier. Such a multiplier potentially helps to explain the rapid increases in obesity in the United States and other countries over the last several decades. This paper presents econometric evidence on the causes of body weight focusing on the effect of one's neighbors’ body weights on one's own body weight. Using roughly half a million observations drawn from the Behavioral Risk Surveillance System (BRFSS) from 1997 through 2002, I investigate to what extent the obesity epidemic is actually an epidemic. The goal is to assess what can be learned about the private and social causes of body weight from standard econometric models, where ‘social’ for our purposes means social spillovers at the county and state level. In so doing, I also provide estimates of the effect of regional characteristics such as income distribution, demographics, food prices, and food store access on body weight outcomes.

In the New Social Interactions literature,¹ a person's behavior depends on the behavior of those in some reference group, such as people in their neighborhood, workplace, or classroom. When such interactions are strong enough they can lead to excess variation across reference groups, the possibility of multiple equilibria, and a “social multiplier” effect in which a small change in an exogenous incentive works both directly and through an effect on the distribution of behavior in the reference group. Apparently small changes in incentives such as food prices or the cost of physical exercise can then play out as large changes in the distribution of body weight.

Economists’ explanations of changes in obesity rates over time have largely centered on private responses to changes in such incentives. Lakdawalla and Philipson (2009) argue that changes in the price of a calorie and in the opportunity cost of burning a calorie explain changes in obesity. This explanation is echoed in subsequent work by economists.² Research in other disciplines, such as evolutionary psychology and public health, is consilient with such explanations.³ Social interactions complement these literatures by explaining features of the data, such as excess variance within regions or racial groups, which are difficult to explain in terms of root causes alone.⁴

A small but rapidly growing literature empirically examines the social determination of body weight. Costa-Font and Gil (2004) present evidence that self-image affects body weight and argue that this result suggests social influences drive obesity. Burke and Heiland (2007) present a version of the quadratic conformer model in which deviations from a desired body weight – which may depend on others’ body weights – reduce utility. The authors demonstrate that a version of the model calibrated to NHANES data is able to reproduce many of the stylized facts describing changes in the distribution of body weight over time. Christakis and Fowler (2007) report that one's friends’ weights cause one's own body weight using the Framingham Heart Study. However, Cohen-Cole and Fletcher (2008b) argue that this result is an artifact of model misspecification.⁵ Cohen-Cole and Fletcher (2008b) report estimates of standard econometric models using data similar to Christakis and Fowler (2007) and show that there is at best weak evidence of endogenous social interactions, and Cohen-Cole and Fletcher (2008a) show that methods used to estimate endogenous peer effects in friend networks fail placebo tests. Renna and Nonnemaker Pais (2008) find evidence in favor of endogenous body weight interactions among adolescent friends only for women. Halliday and Kwak (2009) show that BMI is highly correlated within peer groups and that this correlation increases with BMI, but emphasize that these correlations do not necessarily arise from endogenous social interactions. Trogdon et al. (2008) find evidence for peer effects on overweight, particularly among women and for people with high BMI. Summarizing, the empirical literature thus far has focused on estimating effects among peer groups, and has found some evidence of peer effects, particularly for female and for high BMI people.

This paper provides further evidence on the extent of social influences on body weight interactions at a coarser level of aggregation than previous empirical research. Instead of estimating the effect of one's friends’ weights on one's own weight, the paper presents estimates of the effect of the body weight of people in the county or state in which one lives on one's own weight. This approach has benefits and drawbacks relative to examining interactions at lower levels of aggregation. The primary drawback is that it is plausible that interactions are stronger within narrowly defined social groups such as friendship networks. On the other hand, the social acceptability of one's body weight and one's perception of ideal weight may be driven by norms which emerge across broad regions rather than within peer networks.⁶ Further, people may expect that their body weight will be a factor in future friendships, employment relations, and marriage markets, suggesting that the body weights of people in the community may also have large effects on weight even ignoring social norms. Following Burke and Heiland (2007), either social norms or concern over effects of relative weight in labor, marriage, and other contexts could cause people to compare their own weight to the distribution of body weight of people in their region. These large-scale effects may be easier to reveal since they may be less confounded by the severe selection effects which generate correlations between the characteristics of friends.

The primary source of data is the 1997 through 2002 waves of the Behavioral Risk Factor Surveillance System (BRFSS). These data are repeated cross-sections, which in some ways limits analysis since respondents cannot be tracked over time. More than offsetting that downside is the enormous sample size. Many of the models use sample means and other statistics calculated from the data as covariates, and the the large BRFSS samples – the estimation sample includes roughly half a million observations – allows adequate estimation those statistics, even at the county level.

I use three distinct strategies in attempting to detect the fingerprints left by social interactions. After discussing suggestive evidence for and against social interactions from descriptive statistics and variance decompositions, I implement variants suitable for repeated cross-sectional data of methods developed by Glaeser and Scheinkman (2001), Glaeser et al. (2003) and Graham and Hahn (2005), which infer the presence of social interactions from contrasts between regression coefficients at the individual and aggregate levels. The simplest, and perhaps most compelling, evidence presented compares observed variance across regions with the variance we would expect to observe from sampling noise if people were randomly distributed across regions and there are no social interactions in body weight. I then report estimates from reduced form models in which regional characteristics are added to individual-level demographic and economic covariates, avoiding attenuation bias stemming from sampling-induced measurement error in regional aggregates using a split-sample instrumental variable technique. Non-zero coefficients on these aggregate outcomes suggest the presence of social interactions, but are also of interest for other reasons. In particular, all of these models include regional mean income and a measure of the regional dispersion of income in addition to other regional characteristics, so I am able to provide evidence on the partial associations between body weight outcomes and local income distribution, a topic of interest in the literature.⁷ The final econometric approach is to instrument for aggregate body weight in conventional linear-in-means quadratic conformer models, using subsets of aggregate characteristics as excluded instruments. These strategies generates inconsistent estimates if strong auxiliary assumptions fail, so I interpret estimates as providing only suggestive evidence.

The results do not provide strong support for the hypothesis that social interactions – at this level of aggregation and over moderate time spans – are the missing piece of obesity puzzle.