An agent-based model of child sugar-sweetened beverage consumption: implications for policies and practices

ABSTRACT

Background

A strong body of evidence links young children's intake of sugar-sweetened beverages (SSBs) with myriad negative outcomes.

Objectives

Our research provides insight into whether and to what extent potential intervention strategies can reduce young children's consumption of SSBs.

Methods

We built an agent-based model (ABM) of SSB consumption representing participants in the Project Viva longitudinal study between ages 2 and 7 y. In addition to extensive data from Project Viva, our model used nationally representative data as well as recent, high-quality literature. We tested the explanatory power of the model through comparison to consumption patterns observed in the Project Viva cohort. Then, we applied the model to simulate the potential impact of interventions that would reduce SSB availability in 1 or more settings or affect how families receive and respond to pediatrician advice.

Results

Our model produced age-stratified trends in beverage consumption that closely match those observed in Project Viva cohort data. Among the potential interventions we simulated, reducing availability in the home—where young children spend the greatest amount of time—resulted in the largest consumption decrease. Removing access to all SSBs in the home resulted in them consuming 1.23 (95% CI: 1.21, 1.24) fewer servings of SSBs per week on average between the ages of 2 and 7 y, a reduction of ∼60%. By comparison, removing all SSB availability outside of the home (i.e., in schools and childcare) had a smaller impact (0.77; CI: 0.75, 0.78), a reduction of ∼40%.

Conclusions

These results suggest that interventions reducing SSB availability in the home would have the strongest effects on SSB consumption.

Introduction

Consumption of sugar-sweetened beverages (SSBs) remains a major public health challenge (1–3). SSBs, which include carbonated and noncarbonated soda, fruit drinks, and sports drinks that contain added caloric sweeteners such as high-fructose corn syrup, sucrose, or fruit juice concentrates, are the largest source of added sugar in the US diet (4–6) and constitute ∼7% of children's daily caloric intake (7). There is a strong body of evidence linking children's intake of SSBs with adverse health outcomes, including overweight and obesity (1, 3, 8–23). To address these concerns, the CDC (24); the National Academies of Sciences, Engineering, and Medicine (25); and other organizations have called for population-level interventions to reduce childhood SSB consumption for all children (26, 27). However, because of the overlapping influences that shape childhood consumption patterns, it remains largely unknown what types of interventions or combinations of interventions—including those that have yet to be fielded at scale or satisfactorily studied—might have the greatest potential for impact (28).

Our effort to understand the system of factors that shape childhood SSB consumption requires us to grapple with 3 features that make traditional statistical analyses difficult: interdependence, adaptivity, and heterogeneity (29). The factors that affect children's time spent in various settings are intertwined with the factors that affect consumption in those settings, creating substantial interdependence. For example, the amount of time that a child spends in childcare is potentially related to the availability of SSBs in the childcare that they attend. In addition, the behaviors and outcomes of each child can affect other children, for example, through social influence between families. Adaptation arises because the strength of different pathways changes over time as children age (e.g., access to SSBs in home compared with school settings), whereas families alter their behaviors around SSBs to adapt to changing circumstances such as social norms or the food environment. Heterogeneity is introduced by causal mechanisms such as where individual children spend their time, SSB availability across those settings, and access to pediatric care that operate differently across individuals, for example, based on socioeconomic status. These challenges can be addressed using methods that employ a complex adaptive systems perspective (29–31).

We employ one such method, agent-based modeling (ABM), in this study. ABM is a computational simulation approach that has increasingly been applied to a wide variety of topics in public health to guide policy and practice despite challenges imposed by substantial interdependence, adaptivity, and heterogeneity (32–40). We develop, test, and use an ABM of early childhood SSB consumption that explicitly represents individual children, their environments, and their behavior over time. This model is grounded in high-quality, longitudinal cohort data from children between 2 and 7 y of age. After establishing that the model has sufficient explanatory power through comparison to analogous, real-world consumption patterns in this cohort, we use it as a “virtual laboratory” to provide insight into which potential intervention strategies might most effectively reduce children's SSB consumption.

Methods

Agent-based model

We turned to a large body of extant qualitative and quantitative descriptive literature to design an ABM that meaningfully represents the dynamic system of factors that drive young children's SSB consumption. Figure 1 is a conceptual figure that depicts the model design that emerged from this exercise; we summarize the basis for model elements below.

FIGURE 1.

Conceptual figure of model design. Ovals represent static model elements and rectangles those that are (potentially) dynamic; a black border indicates that the model element explicitly changes as a function of child age. Arrows denote relations between model elements, with a dashed arrow indicating that there is interaction between individuals (i.e., families adjust home SSB availability based on their observation of others). Our primary model output, SSB consumption, shown in a gray rectangle, is a dynamic function of availability of SSBs in settings and time children spend in those settings. SSB, sugar-sweetened beverage.

Young children tend to consume most SSBs at home but also in other settings such as childcare and school, with children spending less time at home and consuming more SSBs in all settings as they get older (41, 42). Substantial heterogeneity exists in the amount of time and settings in which children spend outside of the home (43–48). Consumption in the home is related to both parental attitudes and SSB availability at home (49–51). Parental attitudes are shaped by social norms among their peers as well as messaging (e.g., advertising) (52–56). Parents’ behavior can also be affected by the advice of their pediatricians: the parents of young children with a high BMI are more likely to be advised to reduce their children's SSB consumption (57). Because SSB consumption can be causally linked with increases in BMI throughout childhood (16, 17, 58), this pediatrician advice represents a source of “negative feedback” that can limit consumption among some children who consume large quantities of SSBs. This pediatrician advice dynamic is also a source of potential heterogeneity, as some families are less likely than others to have access to regular medical care or to be receptive or adherent to medical advice (59, 60).

Project Viva

We drew on data from the Project Viva cohort to parameterize and test our ABM. Project Viva is an ongoing prebirth, prospective observational cohort study. We have previously published details on the full study cohort, recruitment, and follow-up (61). Briefly, we recruited mothers between 1999 and 2002 during their initial prenatal visits at Atrius Harvard Vanguard Medical Associates in eastern Massachusetts. We completed research visits with mothers during pregnancy and with both mothers and children periodically; we also administered annual mailed questionnaires between research visits. Mothers reported on their employment status, household income, and race or ethnicity via interviews and questionnaires administered at enrollment. Specifically, we asked mothers, “Which of the following best describes your race or ethnicity?” Mothers were able to select ≥1 of the following: Hispanic or Latina, White or Caucasian, Black or African American, Asian or Pacific Islander, American Indian or Alaskan Native, and/or Other (please specify); for the participants who chose “other” race/ethnicity, we compared the specified responses to the US census definition for the other 5 races and ethnicities and reclassified them where appropriate. To combine these data with other model inputs (e.g., relevant estimates from literature), we restricted our analyses here to those classified with a single value corresponding to white, Hispanic, or black. When the children were 2 and 3 y of age, we asked about the child's time spent in other care settings, including at “a day care center/preschool,” “day care in someone else's home,” and “day care inside your home, including a nanny or regular babysitter.” We abstracted weight and length/height at 2 y from medical records. We calculated BMI and determined age- and sex-specific BMI z-scores using CDC growth charts (62). On the visit and mailed questionnaires, we asked mothers to report the frequency of their child's SSB consumption using FFQs, including questions on “fruit drinks (Hi-C, Kool-Aid, lemonade)” and “soda (not sugar-free),” annually between ages 2 and 7 y. The Institutional Review Board of Harvard Pilgrim Health Care approved the study protocols, and mothers provided written informed consent. Study protocols and questionnaires are available at https://www.hms.harvard.edu/viva/data-collection-by-domain.html. A relevant participant flowchart is included in our Supplementary Material (Supplementary Materials and Methods, Supplemental Figure 1).

Project Viva and ABM

We built a model that characterized consumption for a simulated population based on Project Viva cohort children aged 2–7 y. Simulated consumption was a dynamic function of effective SSB availability in different settings, including time spent in home, childcare, and grade school. Wherever possible, we parameterized the model elements depicted in Figure 1 using Project Viva data either directly or through calibration. As necessary and appropriate, we supplemented the Project Viva data with additional data (e.g., NHANES) and literature (Table 1). Our model was implemented in Python and uses elements of the Mesa agent-based modeling framework (version 0.8.6) as well as NumPy 1.20.1, pandas 1.2.3, and SciPy 1.6.2 (63–67). We summarize the model below and describe it in full detail in Supplementary Material (Supplementary Materials and Methods).

TABLE 1.

Summary of model parameterization¹

Model element	Data source
Initial agent attributes (race, SES, BMI, maternal employment)	Project Viva Cohort
Initial home SSB availability	Project Viva Cohort, NHANES (68)
Agent childcare association
Childcare type	Project Viva Cohort
Childcare SSB availability	Literature (48)
Childcare time in setting	Project Viva Cohort
Agent school association
School attended (assortative based on race and SES)	National Center for Education Statistics (69)
School SSB availability	NHANES
School time in setting	6 h/d (constant for school-age children during school year)
Model dynamics
Change in SSB availability with age	Project Viva Cohort, NHANES
Social influence	Calibration
BMI update from SSB consumption	Literature (16)
Pediatrician access	Literature (60)
Pediatrician advice effect	Calibration

¹SSB, sugar-sweetened beverage; SES, socioeconomic status.

Model initialization

Agents in our model were individual children who collectively were representative of the Project Viva cohort. For the sake of model parsimony, we restricted agents to 3 racial/ethnic groups: white (78% of simulated cohort), black (14%), and Hispanic (8%). We begin the simulation at age 2 y. We probabilistically assigned agent attributes and setting associations based on parameters summarized in Table 1 and detailed in Supplementary Material (Supplementary Materials and Methods).

Model dynamics

Throughout each simulation, we simulated each child aging from 2 to 7 y. SSB availability in home and childcare settings increased as a function of child age (41, 42), and children's time spent in these settings also changed (e.g., as children begin spending time in school, they spend less time elsewhere) (43–48). In addition, we incorporated the following dynamic agent changes during simulations through the following pathways:

1. Social influence. We characterized sources of external social influence on parents’ behavior related to SSB availability in the household (e.g., advertising, social messaging, and behavior of other parents) as a “follow-the-average” mechanism, with parents incrementally moving their offering toward the current population mean for the simulated cohort.

2. BMI update. BMI was allowed to change as a function of recent SSB consumption.

3. Pediatrician guidance. Each year, children's parents probabilistically obtained effective advice (i.e., they consulted a pediatrician and responded to any advice) based on their pediatrician access values. This allowed us to model pediatricians’ advice to parents that can result in a subsequent decrease in home availability.

Specific characterization of these dynamics was based on data sources summarized in Table 1 and detailed in Supplementary Material (Supplementary Materials and Methods).

Model usage

We identified a “baseline” (i.e., no intervention) condition with parameter values taken from data, literature, and calibration (Table 1); parameter selection processes and specific parameter values are detailed in Supplementary Material (Supplementary Materials and Methods, Supplemental Tables 1–7). We compared simulated baseline consumption trends to those observed in the Project Viva cohort to assess whether and to what extent our model reproduced real-world patterns.

Next, we applied the model experimentally, considering a large array of counterfactual conditions that each broadly represent different intervention strategies that might affect how families receive and respond to pediatrician advice or reduce child SSB availability in 1 or more settings. Each condition was defined by which parameters were changed from their values in the baseline condition (e.g., home SSB availability) and the magnitude by which they were altered. We explored the full combination of parameters and magnitudes described in Table 2 (several thousand conditions in total). Formal condition specifications are given in Supplementary Material (Supplementary Materials and Methods, Supplemental Table 8). To reflect variability in the stochastic parts of our model, we conducted 50 simulation runs under each of these conditions. To aid in interpretation, we calculated intervention impact in terms of distributions of average, weekly consumption of SSBs across repeated simulation runs. That is, for each simulation run, we output the average total per-child consumption and divided that by the number of simulated weeks (i.e., 260). We then conducted Welch 2-sample t-test comparisons of distributions of average weekly SSB consumption across 50 repeated runs under each intervention condition to 50 repeated runs of the baseline condition to derive estimates of intervention impact means and 95% CIs.

TABLE 2.

Summary of broad interventions applied alone and in combination with one another

Intervention description	Intervention effect	Magnitudes explored
Pediatrician access	Increases probability of families receiving effective consultation in a given year	20%, 50% increase
Pediatrician advice effect	Larger reduction in home SSB availability upon physician guidance to do so	20%, 50% greater reduction
School SSB availability	Reduction in school SSB availability	10%, 20%, 100% reduction in soda only, 100% SSB reduction
Childcare SSB availability	Reduction in all childcare SSB availability (both family centers and childcare centers)	20%, 30%, 100% reduction
Childcare center SSB availability	Reduction in childcare center SSB availability only	20%, 30%, 60%, 100% reduction1
Home SSB availability	Reduction in home SSB availability	10%, 20%, 100% reduction in soda only, 100% SSB reduction

¹We only explore childcare sugar-sweetened beverage (SSB) availability and childcare center SSB availability combinations where the childcare center SSB availability reduction magnitude is equal to or greater than the childcare SSB availability magnitude.

Our selection of intervention strategies to examine using our ABM was informed by a review of policies and practices that have been recently implemented or discussed by policymakers and practitioners. For example, some SSB policies, such as California's Healthy Beverages in Child Care Act (CA AB 2084) (70), target the childcare environment; thus, we modeled the impact of reducing SSB availability in childcare settings. Because they differ in salient respects, we differentiated between childcare that occurs in 2 types of settings: daycare in someone else's home (“family center”) and at a larger facility (“childcare center”) (71–73). Other policies, such as SSB taxes, would raise prices of SSBs regardless of where they are sold and thus are likely to affect all settings where young children consume SSBs (i.e., home, school, and childcare) (74). Finally, some interventions, such as a pediatrician's advice to reduce SSB consumption, would primarily influence SSB availability in the home environment (75, 76). Although many interventions focus on beverages that have added caloric sweeteners, such as sodas and fruit drinks, juice is increasingly recognized as a potential contributor to obesity (77, 78). Thus, we considered scenarios that would reduce availability of all sugary drinks (including juice) as well as those that would only affect beverages with added caloric sweeteners.

Results

Model testing

Our model was able to generate SSB consumption patterns similar to those observed in the Project Viva cohort between the ages of 2 and 7 y (Figure 2). In the baseline condition, the model projected a total, average per-child consumption of ∼521 servings of SSBs (with servings given in 4 oz/120 mL) over the 5 years between ages 2 and 7 y across repeated simulation runs. Linear interpolation of mean consumption among Project Viva participants produced an estimate of ∼505 servings per child consumed during this age range. These results suggest the model had sufficient explanatory power for subsequent experimental analyses of potential intervention impacts to be meaningful.

FIGURE 2.

Comparison of cohort SSB consumption trends in simulated runs under 10 “baseline” conditions (solid lines) to those observed in Project Viva cohort data (dashed black line). SSB, sugar-sweetened beverage.

Intervention experiments

A select set of intervention impact estimates is presented in Table 3; others are available in Supplemental Materials. Overall, we found that interventions that increase only access to a physician or the effectiveness of physician advice on home availability of SSBs had a minimal impact: even the highest magnitude adjustments to these parameters were estimated to reduce consumption by approximately one-hundredth of a serving per week (0.013; 95% CI: −0.007, 0.032). Interventions that focused directly on reducing availability of SSBs in specific settings had greater beneficial effects. Of these, reducing availability in the home—where young children spend the greatest amount of time—had the greatest potential impact. Removing access to all SSBs in the home resulted in young children consuming about one and a quarter (1.227; 95% CI: 1.211, 1.244) fewer serving of SSBs per week on average between the ages of 2 and 7 y, a reduction of ∼60% relative to the baseline (no intervention) condition.

TABLE 3.

Selected intervention impacts on sugar-sweetened beverage consumption

Intervention condition description						Baseline consumption1
	School beverage	Childcare beverage	Childcare center beverage	Pediatrician	Pediatrician	Mean2.002	SD.050	95% CI1.983, 2.022
Home beverage	availability	availability	availability	access	advice effect	Intervention impact2
availability reduction	reduction, %	reduction, %	reduction,3 %	increase, %	increase, %	Mean	SD	95% CI
—	—	—	—	20	—	0.003	0.050	−0.017, 0.023
—	—	—	—	50	—	0.007	0.050	−0.013, 0.027
—	—	—	—	—	20	0.002	0.050	−0.018, 0.021
—	—	—	—	—	50	0.003	0.050	−0.016, 0.023
—	—	—	—	20	20	0.005	0.050	−0.015, 0.025
—	—	—	—	50	50	0.013	0.049	−0.007, 0.032
—	100	100	—	—	—	0.766	0.039	0.748, 0.783
100	—	100	—	—	—	1.753	0.019	1.737, 1.768
100	100	—	—	—	—	1.477	0.018	1.462, 1.492
—	—	100	—	—	—	0.516	0.045	0.497, 0.535
—	100	—	—	—	—	0.248	0.043	0.230, 0.267
100	—	—	—	—	—	1.227	0.030	1.211, 1.244
—	—	100	—	20	—	0.517	0.045	0.498, 0.536
—	—	100	—	50	—	0.518	0.044	0.500, 0.537
—	—	100	—	—	20	0.516	0.045	0.498, 0.535
—	—	100	—	—	50	0.517	0.045	0.498, 0.536
—	1.8 (10% soda reduction)	100	—	—	—	0.521	0.045	0.502, 0.539
—	3.6 (20% soda reduction)	100	—	—	—	0.525	0.044	0.506, 0.544
—	17.8 (100% soda reduction)	100	—	—	—	0.561	0.043	0.542, 0.579
—	100	—	—	20	—	0.251	0.043	0.232, 0.269
—	100	—	—	50	—	0.255	0.043	0.236, 0.273
—	100	—	—	—	20	0.250	0.043	0.231, 0.268
—	100	—	—	—	50	0.251	0.043	0.233, 0.270
—	100	—	60	—	—	0.331	0.044	0.312, 0.349
—	100	20	—	—	—	0.350	0.042	0.332, 0.369
—	100	30	—	—	—	0.401	0.042	0.383, 0.420
4.4 (10% soda reduction)	—	100	—	—	—	0.577	0.043	0.559, 0.596
4.4 (10% soda reduction)	100	—	—	—	—	0.308	0.042	0.290, 0.327
8.8 (20% soda reduction)	—	100	—	—	—	0.637	0.040	0.619, 0.655
8.8 (20% soda reduction)	100	—	—	—	—	0.368	0.039	0.350, 0.385
20	—	100	—	—	—	0.789	0.038	0.771, 0.807
20	100	—	—	—	—	0.518	0.037	0.501, 0.536
30	—	100	—	—	—	0.923	0.032	0.907, 0.940
30	100	—	—	—	—	0.651	0.031	0.635, 0.668
44.1 (100% soda reduction)	—	100	—	—	—	1.108	0.026	1.093, 1.124
44.1 (100% soda reduction)	100	—	—	—	—	0.835	0.026	0.819, 0.851
100	—	—	—	20	—	1.227	0.030	1.211, 1.244
100	—	—	—	50	—	1.227	0.030	1.211, 1.244
100	—	—	—	—	20	1.227	0.030	1.211, 1.244
100	—	—	—	—	50	1.227	0.030	1.211, 1.244
100	—	—	60	—	—	1.311	0.030	1.294, 1.327
100	—	20	—	—	—	1.332	0.027	1.316, 1.348
100	—	30	—	—	—	1.385	0.026	1.369, 1.401
100	1.8 (10% soda reduction)	—	—	—	—	1.232	0.029	1.216, 1.248
100	3.6 (20% soda reduction)	—	—	—	—	1.236	0.029	1.220, 1.253
100	17.8 (100% soda reduction)	—	—	—	—	1.272	0.027	1.256, 1.288

¹Summary statistics from average weekly sugar-sweetened beverage (SSB) consumption across 50 repeated baseline condition runs.

²Impact estimates of means and 95% confidence intervals taken from Welch 2-sample t-test comparison of distributions of average weekly SSB consumption across 50 repeated runs under each intervention condition to 50 repeated runs of the baseline condition.

³If different (i.e., greater) than overall childcare beverage availability reduction value given in column to the left.

Model estimates suggested that the impact of interventions targeted at the home setting was linear in nature. Figure 3 shows a variety of modeled effects from interventions that reduced (but did not completely remove) SSBs from the home setting, ranging from only removing all soda availability in the home to reducing soda availability in the home by 10%; the former represents a reduction of overall SSB consumption by ∼45% and the latter by ∼4.5%.

FIGURE 3.

Points and lines denote means and standard deviations of SSB consumption (weekly average) across repeated runs under different intervention conditions that reduce home SSB availability. One hundred percent home availability corresponds to the baseline condition. The line shows the trend in intervention impacts on SSB consumption. The nature of this trend indicates a highly linear relation between the magnitude of these interventions and their impact. SSB, sugar-sweetened beverage.

Figure 4 shows the effects of several interventions that reduce SSB availability outside of the home setting. School-based interventions alone had a small to moderate impact on SSB consumption. A complete removal of access to SSBs in childcare (applied to both types of childcare settings) resulted in reductions of ∼0.5 servings per week, about a 25% reduction overall. Similarly, a combination of more modest out-of-home reduction strategies resulted in a reduction of 0.25 servings per week, or about a 12% reduction.

FIGURE 4.

Points and lines denote means and standard deviations of SSB consumption (weekly average) across repeated runs under different intervention conditions that reduce SSB availability in settings outside of the home. The dotted line falls at the baseline mean. SSB, sugar-sweetened beverage.

Additional results, including those from sensitivity analyses, are included in our Supplementary Material (Supplemental Results, Supplemental Tables 9 and 10, Supplemental Figures 2 and 3).

Discussion

We designed a parsimonious agent-based model focusing on how young children divide their waking hours across specific contexts, the availability of SSBs across each of these contexts, and social influence and physician guidance mechanisms. This model drew from cohort data, nationally representative data, and existing literature. Simulation results indicated our model has a high degree of explanatory power based on its ability to produce consumption trends similar to observational data from the Project Viva cohort. This builds confidence that we have captured the underlying mechanisms sufficiently well so that results from counterfactual scenarios may meaningfully represent what consumption patterns might look like under conditions in which interventions successfully effect changes in policy and practice. Thus, we believe that our model is a useful tool for prospectively providing guidance to policymakers and intervention experts.

Our most interesting and potentially policy-relevant finding is that home-based interventions have the greatest potential for reducing consumption. This makes sense, as young children spend the most time in this setting and SSB availability at home is not restricted by any formal rules. It also suggests a tantalizing opportunity for future action. To date, interventions aimed at reducing childhood consumption of SSBs have had greater observable success targeting childcare and school settings than those that attempt to affect the home environment or parent behavior (79). However, this might be due to such interventions being easier to implement or study. Our research indicates that finding ways to effectively intervene in the home setting, optimally supplemented by action that targets other settings, might yield much greater impact. Thus, investment of resources into identifying such strategies is merited.

Limitations and future research

This study applied the model to a single cohort of young children. These results may not be generalizable to populations with different demographics, sociocultural contexts, or geographic environments. Planned future work will apply the model to other longitudinal cohorts, with the added benefit that this “out-of-sample” application provides an opportunity to iteratively add sophistication to model (e.g., in the characterization of social influence). We can also either extend the model to older children or supplement this model with a similar one that captures dynamics that affect these children's consumption behaviors (e.g., peer influences).

In conclusion, the application of ABM to prospectively guide the design and deployment of interventions to reduce childhood consumption of SSBs is highly innovative. It allows us to build upon, but not remain limited by, extant observations of SSB consumption and interventions intended to decrease it. Based on our use of this model, we suggest that researchers, intervention experts, policymakers, and other stakeholders focus resources and efforts into identifying strategies to effectively reduce children's SSB consumption in the home setting, optimally supplemented and supported by action that targets other settings.

Notes

Supported by the NIH (grants 4UH3OD023286-03 and R01 HD 034568). RAH also receives support from NIH grant P30DK092950.

Data described in the manuscript, code book, and analytic code will be made available upon request pending review and approval by study principal investigators (participant consent does not allow for public data sharing).

Supplemental Figures 1–3 and Supplemental Tables 1–10 are available from the “Supplementary data” link in the online posting of the article and from the same link in the online table of contents at https://academic.oup.com/ajcn/.

Abbreviations used: ABM, agent-based model/modeling; SSB, sugar-sweetened beverage.