Heterogeneity in Place Effects on Health: The Case of Time Preferences and Adolescent Obesity

Abstract

We leverage a natural experiment in combination with data on adolescents’ time preferences to assess whether there is heterogeneity in place effects on adolescent obesity. We exploit the plausibly exogenous assignment of military servicemembers, and consequently their children, to different installations to identify place effects. Adolescents’ time preferences are measured by a validated survey scale. Using the obesity rate in the assigned installation county as a summary measure of its obesity-related environments, we show that exposure to counties with higher obesity rates increases the likelihood of obesity among less patient adolescents but not among their more patient counterparts.

1. Introduction

Efforts to reign in the obesity epidemic in the U.S. have had limited success. Obesity rates among adults and children in the U.S. are at an all-time high at 42% among adults and 19% among children ages 2-18 years (Fryar, Carroll, and Afful 2020; Hales et al. 2020). The thinking behind what drives obesity, and consequently how to address it, has evolved considerably over the last four decades. Obesity prevention efforts have shifted from their initial focus on individual-level factors, such as lifestyle choices, to place-based environmental factors. The shift was due, in part, to evidence of limited effectiveness of behavior modification interventions focused at the individual level (Minkler 1999). Since then, a large literature has studied the role of place-based factors, such as built, social, economic and policy environments, on obesity and related behaviors (Diez Roux and Mair 2010; Diez Roux 2001; Arcaya et al. 2016; Tseng et al. 2018; Lam et al. 2021; Kim, Cubbin, and Oh 2019). However, this literature too has found limited effectiveness even among efforts to alter the built, policy, and economic environments through, for example, the introduction of supermarkets and green spaces, taxation of sugary beverages, or through other community- and school-based interventions (Dubowitz et al. 2015; Bleich et al. 2013; Lam et al. 2021; Mayne, Auchincloss, and Michael 2015; Fletcher, Frisvold, and Tefft 2010a; Fletcher, Frisvold, and Tefft 2010b).

The apparent limited success of obesity prevention efforts may be attributed, at least in part, to two empirical challenges. First, our understanding of the causal effects of place-based environmental factors on obesity remains limited due to the methodological challenges resulting from self-selection of individuals into places. It is difficult to disentangle correlation from causation in observational data – an issue that has been highlighted most recently in surveys of the literature on place effects on health (Couillard et al. 2021; Deryugina and Molitor 2021; Chyn and Katz 2021). These surveys have highlighted a subset of studies that leverage movers to assess the causal effects of place, although they focus primarily on adults and non-health outcomes. Moreover, with the exception of the Moving to Opportunity housing mobility experiment, the concern regarding potential endogeneity of moving itself, and of moving to specific places, remains a hurdle in this literature.

The second challenge is that the focus on “one size fits all” strategies ignores potential heterogeneity in place effects on obesity. In other words, place-based factors may not influence all individuals equally and this heterogeneity may explain the small or null effects observed at the population level. Correlational studies suggest heterogeneity with respect to sociodemographic characteristics such as gender, race-ethnicity, and socioeconomic status. These studies often find stronger associations of place with obesity in females versus males but find few consistent patterns with respect to other socio-demographic characteristics (Tcymbal et al. 2020; Duncan et al. 2012; Kranjac et al. 2021; Kranjac et al. 2019; Daniels et al. 2021; Galvez, Pearl, and Yen 2010; Kim, Cubbin, and Oh 2019; Jia et al. 2019). However, efforts to assess heterogeneity are absent from causal studies, and consequently, we have little understanding of how the effects of place might vary by individual. Identifying the sources of heterogeneity within a causal framework can offer insights into developing tailored interventions for obesity prevention instead of a one-size-fits-all approach.

In this paper, we assess whether time preferences represent an important source of heterogeneity in understanding the effect of place on obesity. Healthy behaviors typically involve intertemporal tradeoffs. For example, eating healthy often means forgoing tempting, unhealthy foods in lieu of more healthy options today, but the health benefits of such choices often accrue later in life. Indeed, research suggests that individuals who are more patient are more likely to engage in healthy behaviors and have better health outcomes, such as lower body mass index and/or obesity (Golsteyn, Grönqvist, and Lindahl 2014; Seeyave et al. 2009; Courtemanche, Heutel, and McAlvanah 2015; Sirois 2004). The intuition is that less patient individuals, who place relatively greater weight on immediate gratification, may be less likely to engage in the healthy choices offered in their environments because those choices provide largely delayed benefits. If time preferences influence how individuals respond to their environments, this may help explain more generally why interventions, programs and policies designed to increase healthy behaviors may not appear to be successful at the population level.

We use a natural experiment in combination with data on adolescents’ time preferences to assess heterogeneity in place effects on adolescent obesity. To our knowledge, this paper provides the first quasi-experimental evidence on this issue. Adolescence represents a particularly compelling age to study this question. It is a crucial stage for developing a sense of self and identity and their preferences, goals, motivations, and behaviors evolve towards independence and autonomy (Becht et al. 2016; Meeus et al. 2005). Further, adolescent obesity rates have quadrupled during the past thirty years, making the study of how to reduce adolescent obesity particularly policy-relevant (Ogden et al. 2014; NCHS 2012).

To identify the causal effects of place on adolescent obesity, we leverage the plausibly exogenous assignment of military families to different locations. Specifically, we measure adolescents’ exposure to obesity-related environments using the obesity rate in the county where their military parent’s assigned installation is located. The county obesity rate is a summary, or realized, measure of all environmental influences on obesity in that county. We combine this natural experiment with data we collected on time preferences from adolescents using a validated survey scale. Taken together, this data provides novel empirical evidence on whether adolescents with less patient time preferences face a greater risk of obesity as a result of exposure to obesogenic environments relative to their more patient counterparts.

Our data come from the Military Teenagers Environments, Exercise and Nutrition Study (M-TEENS), a cohort study of adolescents in military families. In prior work, we analyzed baseline data from this cohort at ages 12-13 years and showed that adolescents in military families “assigned” to counties with higher obesity rates were more likely to be overweight or obese, although this effect was small (Datar and Nicosia 2018). In a related study with the same baseline data, we also found that some environmental features, specifically the neighborhood food environment, were unrelated to adolescents’ dietary behaviors or BMI (Shier, Nicosia, and Datar 2016), suggesting that food environments may not matter for obesity prevention. One explanation for these small and null effects, both in our studies and in the literature more broadly, is that average effects may mask heterogeneity. Therefore, we examine heterogeneity on the basis of time preferences using newly-collected data from the same cohort in a subsequent wave.

Our results indicate that the adverse impacts of obesogenic environments are amplified for less patient adolescents. On average, a 10 percentage point increase in the obesity rate of the installation county is associated with a 5 percentage point higher probability of being obese. As predicted, this effect is larger among adolescents with less patient time preferences despite their higher underlying rates of obesity relative to more patient adolescents. Specifically, the estimated effect is 7 percentage points for adolescents at the 25th percentile of the time preference distribution (36% increase, given 19% obesity in this group). In comparison, the corresponding estimated effect is only 2 percentage points for those at the 75th percentile of the time preference distribution (25% increase, given 8% obesity in this group). In fact, patience almost completely offsets the adverse effects of obesogenic environments for adolescents near the 90th percentile of the time preference score distribution. These findings are robust to several sensitivity analyses, including alternate measures of exposure to obesogenic environments, instrumental variables estimation, and corrections for bias in self-reports of height and weight. Finally, in exploratory analyses decomposing the overall place effects into those attributable to the built, social and economic environments separately suggest that heterogeneity based on time-preferences may be strongest for place-based features relating to the built environment.

The remainder of the paper is organized as follows. Section 2 discusses the relevant literatures, including neighborhood effects on obesity, and the link between time preferences and health. Section 3 describes the data and measures, Section 4 describes the empirical approach, which leverages our natural experiment for identification. And Sections 5 and 6 report the results and conclusions, respectively.

2. Background

Literature on Place Effects on Obesity

A large literature, predominantly in the public health and epidemiology disciplines, has examined the effects of place spanning the built, socioeconomic, and policy contextual environments on obesity.¹ This literature has been reviewed elsewhere and suggests a limited role for environments in influencing obesity (Arcaya et al. 2016; Diez Roux and Mair 2010; Ding and Gebel 2012; Feng et al. 2010; Mayne, Auchincloss, and Michael 2015; Tseng et al. 2018; Lam et al. 2021). However, the majority of this work is based on observational study designs that are limited in their ability to draw causal inferences.

Here, we briefly review a smaller economics literature that has employed quasi-experimental or experimental approaches that are more amenable to causal identification. For example, an instrumental variables approach has been used to assess effects of a specific feature of the built environment, such as access to retail food outlets such as fast food, restaurants, and supermarkets. These studies exploit variation in the expansion of such outlets (e.g. Walmart supercenters), timing of exposure to these outlets, or geographic features near these outlets (e.g. proximity to highways) to identify their impacts on obesity. Some studies find significant effects of proximity to these outlets (Currie et al. 2010; Courtemanche and Carden 2011; Dunn 2010) while others do not find any effects (Anderson and Matsa 2011; Dunn 2010; Allcott et al. 2019), which raises concerns about whether heterogeneity plays a role in the mixed findings.

Other studies have taken the approach of looking at the impact of neighborhoods as a whole instead of focusing on specific features. The most well-known among these is the Moving to Opportunity (MTO) housing experiment. MTO randomized individuals from very low-income neighborhoods to receive housing vouchers to move into higher income neighborhoods. The study’s findings indicated that moving to a higher income neighborhood was associated with improved mental and physical health, including lower risk of extreme obesity and diabetes (Ludwig et al. 2011; Sanbonmatsu et al. 2012). However, another study that further explored the mechanisms underlying the observed effects on obesity in the MTO experiment found that they were not explained by neighborhood-level factors such as food prices, restaurant and food store availability, the availability of physical activity facilities, the prevalence of crime, or population density, suggesting that other factors might be at play (Zhao, Kaestner, and Xu 2014).²

Most relevantly, our own prior work, using the same natural experiment as this study, assesses the effects of place as a whole as well as those of specific environmental features on obesity. Military installation assignments generate plausibly exogenous variation in exposure to places, and its associated obesogenic environments, that we used to identify place effects on adolescents’ and parents’ obesity. With the adult obesity rate in the county of the servicemember’s assigned installation as a “realized” or summary measure of all obesogenic environments in the community, we found that adolescents ages 12-13 years and their parents assigned to installations in counties with higher obesity rates were more likely to be obese (Datar and Nicosia 2018). Since time at installation is also plausibly exogenous because the military determines the timing and length of stay at an installation, we also examined whether this relationship varied by time at installation and found that it was stronger among families who had been at the assigned installation longer and, as should be expected, was absent for families who had recently arrived at the assigned installation. These results support the idea that the obesogenic environment, as a whole, has a causal effect on obesity, although the estimated effect size was small. We also explored what specific environmental features contributed to obesity in that study and in a series of additional papers published using the same data (Datar and Nicosia 2018; Richardson et al. 2020; Shier, Nicosia, and Datar 2016; Datar et al. 2015; Datar and Nicosia 2017). The food and physical activity related environments in the installation county, residence neighborhood, and schools (e.g., types of food outlets, fitness and recreation facilities, policies regulating access to unhealthy foods) had either no association, or at best a small association, with BMI and obesity outcomes. However, we did find support for the role of social influence on adolescent obesity (Datar and Nicosia 2018; Nicosia and Datar 2020; Datar, Mahler, and Nicosia 2020) – adolescents exposed to communities with higher obesity reported a higher ideal body type, suggesting social norms related to body size as a potential mechanism.

Understanding whether these findings suggest a limited role for environments or whether they mask important heterogeneity in the environments’ effects is critical for determining the nature of obesity prevention efforts. Heterogeneity can lead to inconclusive or misleading findings regarding place effects (Brand and Thomas 2013). Despite its importance, it remains poorly understood in this literature. Heterogeneity in place effects has mainly been assessed with respect to sociodemographic characteristics such as gender, race-ethnicity, and socioeconomic status. These studies have tended to find stronger associations in females than males but few consistent patterns have emerged with respect to other socio-demographic characteristics (Tcymbal et al. 2020; Duncan et al. 2012; Kranjac et al. 2021; Kranjac et al. 2019; Daniels et al. 2021; Galvez, Pearl, and Yen 2010; Kim, Cubbin, and Oh 2019; Jia et al. 2019). This may be because there is considerable variation in the specific environmental feature being studied (e.g. food environment, walkability, safety, socioeconomic context), the geographic coverage of the study samples (e.g. national versus regional, urban versus rural), and methodologies, which complicates the ability to draw strong conclusions from this literature. But perhaps even more concerning, these studies are based on observational studies and focus primarily on sociodemographic rather than behavioral characteristics. Identifying the critical sources of heterogeneity in a quasi-experimental setting can offer important insights into developing tailored interventions for obesity prevention instead of a one-size-fits-all approach.

Literature on Time Preferences and Health Behaviors

Preferences, particularly time-preferences, are emerging as likely contributors to heterogeneity given the well-documented link between patience and health behaviors and outcomes. In a series of studies, time preferences have been elicited using laboratory methods and then linked to real-world outcomes including health behaviors. Among adults, time preferences predict health, smoking, drinking and drug abuse behaviors (Kirby, Petry, and Bickel 1999; Harrison et al. 2005; Khwaja, Sloan, and Salm 2006; Chabris et al. 2008; Weller et al. 2008; Bradford et al. 2017), and demand for medical screening tests or vaccines (Chapman and Coups 1999; Picone, Sloan, and Taylor 2004). Importantly, a higher level of impatience in childhood and/or adolescence has been linked to greater expenditures on alcohol and cigarettes (Sutter et al. 2013), a greater number of disciplinary referrals at school (Castillo et al. 2011), lower high school completion rates (Castillo, Jordan, and Petrie 2019) and adverse labor market outcomes in adulthood.

Directly relevant to our work is evidence from recent studies that have found that higher impatience in childhood and adolescence is associated with higher BMI (Seeyave et al. 2009; Sutter et al. 2013; Caleza et al. 2016; Samek et al. 2021). Recent studies also show that individuals (including adolescents) with low future-orientation (i.e. low self-control) make poor diet and exercise choices (Conell-Price and Jamison 2015; de Ridder et al. 2012; Wills et al. 2007; Sutter et al. 2013) and have higher BMI (Borghans and Golsteyn 2006). And most recently, time preferences also explained why food purchase decisions made for immediate consumption were less healthy than those made for future consumption (Sadoff, Samek, and Sprenger 2020).

While there is emerging evidence linking time preferences to BMI and related behaviors, whether they influence place effects on obesity and related behaviors remains largely unexplored. An exception is Courtemanche and colleagues’ analysis of panel data from the National Longitudinal Study of Youth (NLSY), which finds that impatient adults exhibit the largest weight gain when food prices fall (Courtemanche, Heutel, and McAlvanah 2015). Thus, preliminary evidence suggests that time preferences can be important moderators of the impact of the environment on obesity-related behaviors, albeit in observational data. This finding is consistent with an emerging literature, which shows that time preferences are important moderators with respect to effects other outcomes involving intertemporal tradeoffs, such as educational achievement.³

3. Data

The Military Teenagers Environment Exercise and Nutrition Study (M-TEENS) was designed to leverage the natural experiment generated by the periodic relocation of military families, to assess how place-based environments affect obesity in children. Military servicemembers are periodically reassigned to different installations, typically every 3-4 years, based on the needs of the military. According to Army Regulation 614-200, “the primary goal of the enlisted personnel assignment system is to satisfy the personnel requirements of the Army”. Thus, the primary consideration when assigning a servicemember is their “current qualification and ability to fill a valid requirement”. In exceptional circumstances (e.g. specialized medical needs of a family member, military couples), a servicemember may be assigned among a subset of installations that meets their special needs. As a result, the assignment of servicemembers to installations at a given point in time is arguably exogenous to the outcome of interest, obesity.

Between Spring 2013 and Summer 2014, M-TEENS recruited just over 1,100 children ages 12-13 years from Army enlisted families located primarily at 10 large divisional posts and 2 medium-sized installations spread across all Census regions. Families were eligible to participate if: 1) the service member did not intend to leave the military within the coming year, 2) the child resided with the enlisted parent at least half-time, and 3) the child was enrolled in public or Department of Defense Education Activity schools. Families were recruited between Spring 2013 and Summer 2014 via the military parent’s email and mailing addresses obtained from the Army’s personnel records. The focal child and one parent were invited to complete surveys online at baseline and in three follow-up waves.

Our analysis sample uses data from the 2017-2018 wave, which collected data on adolescents’ time preferences for the first time. In this wave, 491 adolescents completed the survey. The analysis sample was further restricted to adolescents whose school and residence counties were within 100 miles of the installation and lived with their parents at least some of the time. This restriction was necessary to ensure that adolescents’ would be potentially exposed to the parent’s assigned installation. This yielded a final analysis sample of 400 adolescents with complete data.⁴ Table 1 reports summary statistics for this sample (mean age =17 years, 39% Non-Hispanic white, 21% Non-Hispanic Black, 25% Hispanic, 47% have at least one parent with a 4-year college degree).

Table 1:

Descriptive Statistics

Variables	Mean	SD	Min	Max
Dependent Variable
Obese (self-report)	0.13	(0.33)	0	1
Obese (corrected)	0.13	(0.34)	0	1
Preferences
Time Preference Score	3.28	(0.61)	1.33	4.67
Risk Preference Score	5.89	(2.27)	1	10
Obesogenic Environment
Installation County Obesity Rate (%)	30.03	(4.11)	18.3	37.1
Mean County Obesity Rate (%)	29.97	(3.65)	20.27	36.53
Weighted Mean County Obesity Rate (%)	29.96	(3.71)	19.1	37.58
Max County Obesity Rate (%)	30.97	(4.23)	20.6	40.7
Any County Obesity Rate above 30	0.70	(0.47)	0	1
Covariates
Female	0.45	(0.50)	0	1
Male	0.56	(0.50)	0	1
Non-Hispanic White	0.39	(0.49)	0	1
Non-Hispanic Black	0.21	(0.41)	0	1
Hispanic	0.25	(0.43)	0	1
Other	0.15	(0.36)	0	1
Live on installation	0.25	(0.43)	0	1
Time at installation >=2 years	0.35	(0.48)	0	1
Parent rank>=E7	0.47	(0.50)	0	1
Military parent active duty	0.59	(0.49)	0	1
Household income>$70k	0.35	(0.48)	0	1

Notes: N=400. Weighted Mean County Obesity Rate: Weighted mean of the obesity rates for the installation, residence, and school counties. Max County Obesity Rate: maximum obesity rate of the installation, residence, and school counties. Any County Obesity Rate above 30: Indicator for whether the obesity rate for installation, residence, or school county is above the median rate of 30%.

The study was approved by the Institutional Review Boards at RAND, University of Southern California, and the Army’s Human Research Protection Office. Parent consent and child assent were obtained online prior to participation.

Measures

BMI and Obesity

Adolescents’ height and weight information was collected primarily via self-reports because it was the most cost-effective way of collecting this data from a geographically-dispersed sample. However, to address concerns about potential measurement error in self reports, a subsample (N=216) of adolescents were also measured by trained staff using guided videoconference sessions using measurement equipment shipped to respondents, an approach that yielded high accuracy in a pilot study (Ghosh-Dastidar, Nicosia, and Datar 2020)⁵. This subsample was used as a validation sample to correct for bias in self-reports of adolescents’ height and weight for the full sample using regression calibration (Ghosh-Dastidar et al. 2016). The self-reported and “corrected” height and weight were used to construct age- and gender-specific BMI z-scores and BMI percentiles based on the 2000 BMI-for-age and gender growth charts issued by the Centers for Disease Control and Prevention. A child was classified as obese if the BMI percentile was greater than or equal to 95. In our sample, 13% of adolescents were classified as obese based on the age- and sex-specific growth charts (Table 1).

Installation County Obesity Rate

Exposure to obesogenic environments was measured using the adult obesity rate in the county where the family’s current installation was located. The installation county obesity rate (InstaCOR) is a useful summary (or realized) measure of all potential obesogenic influences in the county. Installation county, instead of residence county, was used to construct our primary exposure measure (akin to an intent-to-treat analysis) because, as explained later, residential choice at a given installation may be less exogenous. Moreover, military families regularly access the installation for work, health care, shopping (e.g. Commissary, Post-Exchange), recreation (e.g. Youth programs), or education and so are exposed to the installation county regardless of where they live.

The “assigned” installations of the study sample were spread across 30 states, 57 counties, in all Census regions. County obesity rates for each of these installation counties were obtained from the Robert Wood Johnson Foundation’s County Health Rankings data⁶ and were linked to the M-TEENS sample by installation. The county obesity rates are from the 2018 release of the County Health Rankings data, which are based on estimates computed by the US Centers for Disease Control and Prevention by pooling the 2013-2015 Behavioral Risk Factor Surveillance System dataset. The lagged county obesity data were preferred because of the typical length of time at installation for our sample.

Descriptive statistics for the installation and composite COR measures are provided in Table 1. The mean (SD) of installation COR in our sample was 30% (4.2%) and ranged from 18.3% (Arlington county, VA) to 37.1% (Christian county, KY).

Time Preferences

Our time preference measure is a validated 12-item survey called the Consideration of Future Consequences (CFC) scale (Strathman et al. 1994). In the survey, adolescents rated how characteristic each statement was of them on a 5-point Likert scale, including statements such as “Often I engage in a particular behavior in order to achieve outcomes that may not result for many years” and “I only act to satisfy immediate concerns, figuring the future will take care of itself.” The responses were averaged over the 12 questions (with higher numbers generally indicating more patience, but reverse-coding statements like “I only act to satisfy immediate concerns”). The time preference scores ranged from 1.8 to 4.7 with a mean (SD) of 3.28 (0.61) (Table 1).

We chose to use the CFC scale rather than alternative measures of time preferences for a few reasons. First, the CFC Scale has been widely used in the psychology literature to study self-regulating behaviors in health and finance. Second, the CFC scale has been shown to be correlated with personality traits associated with self-control (Joireman et al. 2008; Joireman, Anderson, and Strathman 2003; Joireman, Strathman, and Balliet 2006). A recent meta-analysis also showed that the CFC scale correlates well with a host of health-related behaviors including diet and exercise (Murphy and Dockray 2018; Samek et al. 2021). Third, in a separate study aimed at understanding the effectiveness of different ways to elicit time preferences in our sample of adolescents, we showed that time preferences elicited using the CFC were associated with BMI and health-related behaviors (Samek et al. 2021).⁷ In contrast, time preferences elicited using monetary trade-off tasks commonly used in the experimental economics literature did not correlate with BMI in our sample (Samek et al., 2021). One benefit of the CFC scale over questions about monetary trade-offs is that it is easier to explain and therefore easier to implement remotely. The CFC is similar to the “preference survey module” – a survey with questions on risk, time and social preferences that has been proposed by Falk et al. (Falk et al. 2016) and is gaining widespread use in economics (Falk et al. 2018).

Covariates

Covariates included adolescents’ gender and race-ethnicity, parents’ marital status, military parent’s rank, indicator for military parent’s active duty status⁸, annual household income, time at installation, and on-installation residence.⁹ Besides adolescent’s gender and race-ethnicity, which was reported by the respondent, all other covariates were parent-reported. Covariates also included the adolescent’s risk preferences, as measured by a single question where respondents were asked to rate their willingness, in general, to take or avoid risk on a scale of 1 (very unwilling to take risks) to 10 (very willing to take risks). Time preferences are closely linked to risk preferences, since choosing to wait for a reward also assumes some level of risk taking. As such, in our preferred regression specifications, we also control for risk preferences. Descriptive statistics for these covariates are reported in Table 1.

4. Empirical Approach

Our estimation model is shown by the linear regression in Eq (1), where Obese_ic is an obesity indicator for adolescent i in county c; InstaCORc is the obesity rate for the installation county; TimePref_i is the CFC scale score for adolescent i, the vector X includes individual and family level covariates described earlier, and ε_ic is the error term. The variables InstaCOR and TimePref are de-meaned for easier interpretation of the estimates. The coefficient β₁ captures the effect of InstaCOR on obesity risk for an adolescent with time preferences at the mean. We expect β₁ to be greater than zero i.e. exposure to more obesogenic counties should increase the adolescent’s risk for obesity. The coefficient β₂ captures the effect of time preferences on obesity risk for adolescents in counties with obesity rates at the mean. We expect β₂ to be less than zero, as adolescents with higher time preference scores (i.e. more patient) would be less likely to be obese. The coefficient of interest, β₃, captures whether the effect of InstaCOR on adolescent’s obesity risk varies by time preference. We hypothesize that β₃ will be less than zero as patient adolescents would be less adversely affected by exposure to obesogenic environments.

$$Obes{e}_{ic}={\beta }_{1}InstaCO{R}_c+{\beta }_{2}TimePre{f}_i+{\beta }_{3}InstaCO{R}_c\ast TimePre{f}_i+\gamma X_i+{\epsilon }_{ic}$$ (1)

Identification in our model comes from the fact that assignment to installations is based on the Army’s needs.¹⁰ This assignment creates plausibly exogenous variation in exposure to obese communities, proxied by the obesity rate in the county where the assigned installation is located. We provide support for this claim in Table 2, which compares the sociodemographic characteristics of adolescents assigned to installation counties with above-median (>=30%) versus below-median (<30%) InstaCOR.¹¹ We also show estimates from univariate regressions of continuous InstaCOR on each of the covariates individually. We find that the composition of adolescents does not vary systematically by InstaCOR.¹² We also estimate models with and without covariates given that stability in the estimates for the effect of InstaCOR across adjusted and unadjusted models would lend further support to the exogeneity of InstaCOR.

Table 2:

Balance table

Variables	Installation County Obesity Rate (InstaCOR)			Coeff (Std Error) from univariate linear regressions of InstaCOR on this variable
	Below Mediana (n=176)	At or Above Median (n=224)	Difference
Dependent Variable
Obese (self-report)	0.091	0.152	−0.061	0.784**(0.391)
Obese (corrected)	0.091	0.165	−0.074**	0.829**(0.387)
Covariates
Time preference score	3.295	3.272	0.023	−0.259 (0.411)
Risk preference score	5.716	6.018	−0.302	−0.026 (0.090)
Female	0.466	0.429	0.037	Reference
Male	0.534	0.571	−0.037	0.109 (0.322)
Non-Hispanic White	0.409	0.379	0.030	Reference
Non-Hispanic Black	0.182	0.228	−0.046	0.749 (0.543)
Hispanic	0.284	0.219	0.065	−0.636 (0.562)
Other race-ethnicity	0.125	0.174	−0.049	0.698 (0.681)
Live on the installation	0.284	0.219	0.065	−0.834 (0.673)
Time at installation >= 2 years	0.358	0.339	0.019	−0.626 (0.709)
Parent Rank E7 or higher	0.426	0.509	−0.083	0.765 (0.448)
Parent Active Duty	0.540	0.625	−0.085	0.595 (0.346)
Household annual income <=$70,000	0.653	0.643	0.011	Reference
Household annual income >$70,000	0.347	0.357	−0.011	0.175 (0.605)

^**p<0.05

^aMedian value of the Installation County Obesity Rate is 30%.

An additional identification assumption in our empirical model is that adolescents’ ex ante obesity risk is uncorrelated with InstaCOR, conditional on time preferences. To provide support for this assumption, we predict adolescents’ ex ante obesity risk using all covariates used in our models (except InstaCOR) and then regress this prediction on InstaCOR and its interaction with time preference. The coefficients on both of these variables in this regression are statistically insignificant and close to zero. These results are reported in Appendix Table 4.

While most military families are exposed to the installation county, the majority of families live in surrounding communities off-base (Buddin et al. 1999; Bissell, Crosslin, and Hathaway Feb 2010; MilitaryOneSource 2020) (75% of our sample), some of whom choose to live outside the installation county, exposing them to those counties as well.¹³ Likewise, adolescents who attend schools in counties that are different from the installation and/or the residence county, are exposed to those counties as well (Department of Defense 2015).¹⁴ To account for exposure to obesogenic environments across up to three different counties (installation, residence, and school), we construct a composite measure of the obesity rates in the three counties as an alternate measure of the obesogenic environment. The composite COR is constructed in several different ways. First, we take a simple average of the COR in the three counties (MeanCOR). Second, we compute a weighted average of the COR in the three counties, where the weights are based on the proportion of waking hours spent in the three locations (WgtmeanCOR)¹⁵. Since adolescents spend a sizeable fraction of their waking time at school, we assign 50% of the weight to the school county and divide the rest equally between installation and residence.¹⁶ Third, we use the maximum obesity rate of the three counties as the composite measure (MaxCOR). And finally, we use an indicator for whether the COR for any of the three counties was greater than 30% (AnyCORabove30), which is the mean and median in the sample.

Table 3 compares the adult obesity rates for the installation, residence, and school counties. Concordance between school and residence county obesity rates is highest – 90% of the sample goes to school in the same county as their residence. In comparison, the concordance is 58% between school and installation county obesity rates, and is 55% between residence and installation county obesity rates.¹⁷

Table 3:

Comparison of Adult Obesity Rates for Installation, School, and Residence Counties

	% of Sample	Difference in COR (percentage points)
		Minimum	Median	Maximum
School COR < Installation COR	29%	−10.7	−2.0	−0.2
School COR = Installation COR	58%
School COR > Installation COR	13%	0.1	5.3	12.5
Residence COR < Installation COR	29%	−10.7	−2.0	−0.2
Residence COR = Installation COR	55%
Residence COR > Installation COR	15%	0.1	5.1	12.5
Residence COR < School COR	4%	−10.9	−2.7	−0.4
Residence COR = School COR	90%
Residence COR > School COR	6%	0.4	2.6	10.7

COR: County Obesity Rate. Correlation between school and installation COR is 0.65, between residence and installation COR is 0.62, and between residence and school COR is 0.92.

While the composite measures encompass more of the environment that adolescents are actually exposed to, they are likely to be endogenous because even though installation county is exogenously assigned families can choose where to live and attend school around the assigned installation. To address this concern, we also estimate two-stage least squares (2SLS) models that use the assigned InstaCOR as an instrument for the composite COR.¹⁸ Our identification of the interaction effect of COR and time preferences relies similarly on the assumption that time preferences are not affected by assignment to installation. This assumption is consistent with the long-standing tradition in economics of assuming that preferences are predetermined and stable, at least in adults (Meier and Sprenger 2015).¹⁹

Because time preferences were collected after assignment to installation, we need to assess the plausibility of our assumption. To do so, we examine whether time preferences are associated with COR. Specifically, we estimate models that regress time preferences on InstaCOR or composite measures of COR. Since both are measured contemporaneously, a significant association between time preferences and COR may indicate that preferences are shaped by environments. In contrast, an insignificant association would be consistent with the idea that preferences are likely stable and not affected by environments, which would reduce concerns about endogeneity of the TimePref variable in our models. Furthermore, an insignificant association would provide further evidence that the natural experiment does, in fact, balance the sample with respect to preferences, and that there is no sorting of adolescents into places based on preferences.

5. Results

Effect of County Obesity Rate on Adolescent Obesity

Table 4 reports results from estimating Equation (1) using linear probability models for obesity based on the child’s self-reported height and weight.²⁰ Corresponding results using the obesity indicator based on height and weight data “corrected” for measurement error are reported in Appendix Table 6. For ease of interpretation, InstaCOR and TimePref variables were centered on their respective means.

Table 4:

Effect of Installation County Obesity Rate and Time Preferences on Adolescent Obesity

Explanatory Variables	Dependent Variable = Obese
	(1)	(2)	(3)	(4)	(5)
Installation County Obesity Rate	0.005** (0.002)	0.006* (0.003)	0.005* (0.002)	0.005* (0.002)	0.005* (0.003)
Time Preference Score			−0.061*** (0.020)	−0.061*** (0.019)	−0.058*** (0.020)
Installation County Obesity Rate *
Time Preference Score			−0.005* (0.003)	−0.005* (0.003)	−0.006** (0.003)
Controls
Covariatesa	No	Yes	No	No	Yes
Risk Preference Score	No	No	No	Yes	Yes
R-squared	0.004	0.034	0.018	0.018	0.047
Observations	408	406	404	402	400

Estimates in each column are from Ordinary Least Squares Regressions. Standard errors are in parentheses and are clustered at the installation county level. The Installation County Obesity Rate and Time Preference Score are centered on their respective means. Higher values of time preference score indicate greater patience.

^aCovariates include adolescents’ gender and race-ethnicity, parents’ marital status, military parent’s rank, indicator for military parent’s active duty status, annual household income, time at installation, and on-installation residence.

^***p<0.01

^**p<0.05

^*p<0.1

In the unadjusted model with only InstaCOR as the covariate, a 10 percentage point increase in InstaCOR increases the likelihood of obesity by 5 percentage points (Column 1). Adding the full set of sociodemographic covariates to the model does not change this estimate (Column 2), as expected if the natural experiment effectively randomizes adolescents to different InstaCORs.

The effect of time preferences, and of their interaction with InstaCOR, on obesity is estimated starting in Column 3, first without covariates (Column 3) and then with the addition of controls for risk preferences (Column 4) and sociodemographic covariates (Column 5). Since InstaCOR and TimePref are de-meaned, each of their coefficients capture the effects on obesity for an adolescent with TimePref and InstaCOR at their respective means. The estimated effects of InstaCOR, TimePref and their interaction remain stable across all columns, indicating that inclusion of covariates and risk preference does not alter the estimates. An increase in the time preference score, which indicates greater patience, is consistently associated with a lower probability of being obese. Estimates in column 5, the fully specified model, indicate that a 1 standard deviation increase in the time preference score (0.6) is associated with a 3.5 percentage point (=−0.058*0.6) decrease in the likelihood of being obese.

Importantly, the interaction effect indicates that the effect of InstaCOR is smaller for adolescents with more patient time preferences. A 10 percentage point increase in InstaCOR increases the likelihood of obesity by 5 percentage points for an adolescent with time preferences at the mean. The likelihood of obesity increases by 7 percentage points for an adolescent with a time preference score at the 25^th percentile, but increases by only 2 percentage points for an adolescent with a time preference score at the 75^th percentile. These results are statistically significant at the 10% level in all models, and statistically significant at the 5% level in the model that incorporates all controls.²¹

Table 5 reports estimates from alternate specifications where InstaCOR is replaced with composite measures of the obesogenic environment that capture environments from the installation county, the school county, and the residence county.²² Results from specifications that use the mean COR, weighted mean COR, and MaxCOR measures are qualitatively similar to the main models. When using AnyCORabove30, for example, adolescents exposed to a county with obesity rate above 30% are 5.3 percentage points more likely to be obese and this association is smaller among more patient adolescents.

Table 5:

Effect of Composite County Obesity Rate Measures on Adolescent Obesity

	Dependent Variable = Obese
	(1)	(2)	(3)	(4)
A. Mean County Obesity Rate	0.005 (0.004)
Time Preference Score	−0.056** (0.021)
Mean County Obesity Rate *	−0.009** (0.004)
Time Preference Score
B. Weighted Mean County Obesity Rate		0.005 (0.004)
Time Preference Score		−0.056** (0.022)
Weighted Mean County Obesity Rate *		−0.009** (0.004)
Time Preference Score
C. Max County Obesity Rate			0.005 (0.003)
Time Preference Score			−0.046** (0.023)
Max County Obesity Rate *			−0.009*** (0.003)
Time Preference Score
D. Any County Obesity Rate above 30				0.042* (0.024)
Time Preference Score				−0.005 (0.027)
Any County Obesity Rate above 30 *				−0.077** (0.033)
Time Preference Score
Observations	400	400	400	400
R-squared	0.047	0.048	0.049	0.048

Estimates in each column are from Ordinary Least Squares Regressions. Standard errors are in parentheses and are clustered at the installation county level. Mean County Obesity Rate (COR) is the average obesity rate of the installation, school, and residence counties. Weighted mean COR is the weighted average of the obesity rate of the three counties. Max COR is the highest obesity rate of the three counties. Any COR above 30 is an indicator for whether any of the three counties (installation, residence, school) have an obesity rate above 30%, which is the mean county obesity rate in our sample and is also the mean and median obesity rate across all counties in the U.S. COR, Mean COR, Weighted mean COR and Time Preference Score are all centered on their respective means. All models include the full set of covariates - adolescents’ risk preference score, gender and race-ethnicity, parents’ marital status, military parent’s rank, indicator for military parent’s active duty status, annual household income, time at installation, and on-installation residence.

^***p<0.01

^**p<0.05

^*p<0.1

Table 6 reports results from 2SLS models, which instrument the different composite measures of obesogenic environment and their interaction with time preferences using InstaCOR and InstaCORxTimePref.²³ The first stage tests show a strong positive association between InstaCOR and composite COR measures; the F-statistic for excluded instruments ranged from 25.8-35.8. The second stage results show that the IV estimate of the effect of weighted mean COR on obesity is 0.008 (SE=0.004). The interaction effect is −0.008 (SE=0.004). Results for the models that instrument for the other composite CORs are similar.

Table 6:

Instrumental Variables Regression for the Effect of Composite County Obesity Rate on Adolescent Obesity

	Dependent Variable = Obese
Explanatory Variables	(1)	(2)	(3)
Mean County Obesity Rate	0.007* (0.004)
Time Preference Score	−0.057*** (0.021)
Mean County Obesity Rate *
Time Preference Score	−0.009** (0.004)
Weighted Mean County Obesity Rate		0.007* (0.004)
Time Preference Score		−0.055*** (0.021)
Weighted mean County Obesity Rate *
Time Preference Score		−0.009** (0.004)
Max County Obesity Rate			0.007* (0.004)
Time Preference Score			−0.047** (0.021)
Max County Obesity Rate*
Time Preference Score			−0.008* (0.004)
First Stage Tests
F-stat of excluded instruments	35.78 (p=0.000)	27.78 (p=0.000)	34.57 (p=0.000)
Andersen Rubin Wald test (Chi-2)b	8.04 (p=0.018)	8.04 (p=0.018)	8.04 (p=0.018)
Sanderson and Windmeijer multivariate F-test	71.11 (p=0.000)	55.23 (p=0.000)	65.79 (p=0.000)
Test of exogeneity of instrumentsa:	0.545	0.479	0.332
Robust F-stat (p-value)	(p=0.583)	(p=0.622)	(p=0.719)
Observations	400	400	400
R-squared	0.047	0.047	0.049

Estimates in each column are from Two Stage Least Squares Regressions where the composite County Obesity Rate measure and its interaction with time preference are instrumented with the Installation County Obesity Rate and its interaction with the Time Preference Score. Standard errors are in parentheses and are clustered at the installation county level. All models include the full set of covariates - adolescents’ risk preference score, gender and race-ethnicity, parents’ marital status, military parent’s rank, indicator for military parent’s active duty status, annual household income, time at installation, and on-installation residence.

^aThe F-stat reported here is a robust score test (Wooldridge 1995) equivalent of the Durbin-Wu-Hausman test.

^bThe Andersen Rubin Wald Chi-2 statistic tests the joint significance of endogenous regressors in main equation.

^***p<0.01

^**p<0.05

^*p<0.01

Effect of County Obesity Rate on Time Preferences

Table 7 reports results from models that estimate the relationship between the county obesity rate and adolescents’ time preferences. There is no significant association between InstaCOR and preferences, which may have two implications given that both measures are collected contemporaneously. One implication may be that there is no systematic sorting of adolescents into more versus less obesogenic communities based on their time preferences. This would provide further support to our identification assumption that the assignment to installations is exogenous. In addition, the null findings may also imply that obesogenic environments have no impact on adolescents’ time preferences. This would be consistent with the long-standing tradition in economics of assuming that preferences are predetermined and stable, at least in adults (Meier and Sprenger 2015; Chuang and Schechter 2015). There is emerging evidence that time preferences evolve substantially during childhood (Bettinger and Slonim 2007; Angerer et al. 2015; Kosse et al. 2020; Sutter, Yilmaz, and Oberauer 2015; Andreoni et al. 2019), but there is not much evidence about whether time preferences respond to environmental changes. In their survey of the literature, Chuang and Schecter (2015) find mixed evidence about whether events like economic shocks or natural disasters affect time preferences, potentially due to the difficulties with collecting such data.

Table 7:

Regression of Time Preferences on County Obesity Rate

	Time Preference Score
Explanatory Variables	(1)	(2)	(3)
Installation County Obesity Rate	−0.004 (0.008)
Weighted Mean County Obesity Rate		−0.007 (0.008)
Max County Obesity Rate			−0.003 (0.008)
Observations	400	400	400
R-squared	0.046	0.047	0.049

Estimates in each column are from Ordinary Least Squares Regressions. Standard errors are in parentheses and are clustered at the installation county level. All models include the full set of covariates - adolescents’ risk preference score, gender and race-ethnicity, parents’ marital status, military parent’s rank, indicator for military parent’s active duty status, annual household income, time at installation, and on-installation residence.

^***p<0.01

^**p<0.05

^*p<0.1

Decomposing the Effect of County Obesity Rate

The results presented thus far indicate that time preferences play an important role in understanding the effects of obesogenic environments on adolescent obesity. Here, we explore what features of the environment interact with time preferences to influence adolescent obesity. Because we lack the power to assess the role of specific environmental factors (e.g. fast food restaurants, access to parks), we focus on the role of broad categories of environments via some exploratory analyses. Using a county-level regression of COR on built and social environment measures ( Appendix table 11), we partition COR into three components: a) the portion predicted by the built environment, b) the portion predicted by the socioeconomic environment, and c) the residual, defined as the difference between actual COR and that predicted by the built and social environment measures.²⁴ We then interact these three components of COR with the time preference measure to estimate their interactive effects on adolescent obesity (Appendix Table 12). Similar to our own and other prior studies, we find little support for an average effect of the built environment on obesity (column 1). However, results from the interaction do support an effect of the built environment on impatient adolescents (Column 2). Or put another way, patience offers significant protection against obesogenic built environments, which may explain the fact that many studies find little or no effect, on average. With respect to social environment, we find no evidence of an effect on obesity, on average, or that it varies by adolescent time preferences. Lastly, we find a significant impact of the residual component, which, following Datar and Nicosia (2018) and Datar, Mahler, and Nicosia (2020), may be interpreted as a social contagion effect: for example, seeing more obese people may influence adolescents’ own likelihood of obesity via changes in social norms about body type, behavior mirroring or other social influence mechanism. In contrast to the built environment, social contagion appears to be equally important for all adolescents regardless of time preferences. Overall, these results suggest that understanding the role of time preferences is important for understanding the effects of some features of the environment more so than others.

6. Conclusion

This paper provides the first quasi-experimental evidence on heterogeneity in place effects on adolescent obesity with respect to time preferences. Specifically, we examine whether adolescents with more patient time preferences are less affected by their obesogenic environments compared to their less patient counterparts.

Our results show that exposure to obesogenic environments has a greater effect on obesity among less patient adolescents. For an adolescent with time preferences at the mean, a 10 percentage point increase in the county obesity rate for the assigned installation increases the likelihood of obesity by 5 percentage points. This effect varies considerably across adolescents with varying time preferences. For example, the likelihood of obesity increases by 7 percentage points for an adolescent with a time preference score at the 25^th percentile, but only increases by 2 percentage points for an adolescent with a time preference score at the 75^th percentile. In fact, patience almost completely offsets the adverse effects of obesogenic environments for adolescents near the 90^th percentile of the time preference score distribution. Finally, while exploratory, our analyses also suggest that adolescents’ time preferences appear to be most important for protecting against the adverse effects of built environments on obesity. Such heterogeneity should be explored in more detail in future research.

While our study suggests an important role of time preferences in understanding heterogeneity in environments’ effects on adolescent obesity, it is worth noting that time preferences are not exogenously assigned. They are likely to be correlated with other factors, including family socioeconomic circumstances and risk preferences. We control for these characteristics in our analyses, but other unobserved factors (e.g. optimism (Boehm et al. 2018) about the future, trust (Albanese, de Blasio, and Sestito 2017) or belief that effort today will pay off in the future) may also contribute to this heterogeneity. These other potential mechanisms should be explored in future work.

Nonetheless, our findings have several implications. First, they may explain why obesity prevention interventions, policies, and programs may appear to have small or no effects on average (Tseng et al. 2018; Wang et al. 2015; Bramante et al. 2019). They may also explain why childhood obesity rates have continued to rise despite significant efforts to reverse the trends (Skinner et al. 2018). Efforts to address adolescent obesity may benefit from assessing time preferences during childhood and adolescence and targeting interventions towards those at higher risk. Second, our findings suggest a greater role for interventions that seek to alter adolescents’ time preferences. In our analysis, we treat time preferences as a stable trait, and show that time preferences are not affected by place. However, a relatively new area of research examines whether (and how) children’s time preferences can be modified via targeted interventions. For example, Alan and Ertac (2014) find that a program targeted at helping children imagine their future selves yields more patient time preferences in a separately-elicited laboratory task. And, Luhrmann et al. (2014) find that a financial education program administered with adolescents affects time preferences by making treated subjects less present-biased. Our research implies that such interventions, by making children more patient, could also shield children from the potentially harmful effects of an obesogenic environment. Finally, our findings speak to the literature on socioeconomic disparities in obesity. Low income families are not only more likely to live in obesogenic neighborhoods (Lovasi, Hutson, Guerra, & Neckerman, 2009) but there is also a growing literature suggesting that poverty is linked to impatience (Haushofer and Fehr 2014). Therefore, our findings suggest that the combination of time preferences and obesogenic environments might exacerbate socioeconomic disparities in obesity.

Highlights.

• Natural experiment assessing heterogeneity in place effects on adolescent obesity

• Exposure to counties with higher obesity rates increased adolescents’ obesity risk

• Adverse impacts of obesogenic environments are larger for less patient adolescents