Impact of Different Policies on Unhealthy Dietary Behaviors in an Urban Adult Population: An Agent-Based Simulation Model

Abstract

Objectives. Unhealthy eating is a complex-system problem. We used agent-based modeling to examine the effects of different policies on unhealthy eating behaviors.

Methods. We developed an agent-based simulation model to represent a synthetic population of adults in Pasadena, CA, and how they make dietary decisions. Data from the 2007 Food Attitudes and Behaviors Survey and other empirical studies were used to calibrate the parameters of the model. Simulations were performed to contrast the potential effects of various policies on the evolution of dietary decisions.

Results. Our model showed that a 20% increase in taxes on fast foods would lower the probability of fast-food consumption by 3 percentage points, whereas improving the visibility of positive social norms by 10%, either through community-based or mass-media campaigns, could improve the consumption of fruits and vegetables by 7 percentage points and lower fast-food consumption by 6 percentage points. Zoning policies had no significant impact.

Conclusions. Interventions emphasizing healthy eating norms may be more effective than directly targeting food prices or regulating local food outlets. Agent-based modeling may be a useful tool for testing the population-level effects of various policies within complex systems.

The literature on policy interventions to address obesogenic dietary behaviors can be divided into 3 distinct categories. The first is the use of economic measures to alter food consumption, such as taxing unhealthy ingredients and subsidizing healthy foods,^1–4 in light of studies that have shown the price of a calorie obtained from unhealthful foods is lower than the price of a calorie from more healthful foods.^5,6 The second is targeting the food environment through zoning polices, including increasing the number of healthy food vendors in “food desert” communities⁷ and restricting the opening of new fast-food restaurants.⁸

The third and final category is related to combating unhealthy eating norms, given research showing the power of food marketing to change dietary behaviors and proposing restrictions on the time, place, and manner in which obesogenic foods are marketed.^9,10 Conversely, pro-nutritional marketing focuses on education as a means of increasing consumers’ awareness of dietary health (e.g., nutrition disclosure on menus and the issuing of dietary guidelines).^11,12

The research to date focusing on the effects of these various approaches has incorporated theoretical and empirical techniques that rely on the stable unit treatment value assumption, according to which there are no interactions among people who experience an intervention that would alter the effectiveness of the intervention. However, this assumption is known to be violated in the case of obesity-related behaviors.^13,14 What is therefore not clear from existing regression-based and experimental empirical work is the potential magnitude of the population-level impact of these policies if they were implemented in the real world. We performed simulations to contrast the potential of different approaches aiming at tackling unhealthy dietary behaviors in a population of urban US adults.

Simulations involving systems dynamics or agent-based modeling (ABM) are increasingly being used in public health,^15,16 particularly in addressing food system issues.^17,18 Such techniques can provide virtual laboratories for testing policies¹⁹ and can be useful in studying complex systems, such as dietary behaviors, wherein decisions are influenced by the interplay of socioeconomic, psychological, and physiological forces.^20–22 We used ABM, a powerful technique when there is heterogeneity among members of the population and when behavior changes as a result of interactions between different individuals or between individuals and their environment (i.e., adaptation).^23,24

We developed a computer model explicitly representing how individuals make decisions to examine the impact of policies on their food consumption. Our model was based on a multilevel theory of population health that emphasizes the role of cognitive habits in human behaviors.²⁵ According to this theory, individual beliefs are influenced by incentives—as in rational-choice theory—but also shaped by cognitive habits that are reinforced within a population by social norms and culture. Thus, our model examined how individual beliefs are influenced by interventions either in the social network or in the food environment.

METHODS

Our agent-based model contained 2 kinds of agents: individuals and food outlets. Individuals make dietary choices, and food outlets adapt to those choices. In our model, a single time period was defined as 1 day, and each simulation was run for more than 3 years (1100 days).

In our simulations, individuals were assigned demographic characteristics (age, gender, educational attainment) matching the demographic distribution of downtown Pasadena, CA. The model accounted for spatial as well as social aspects; individuals were assigned a home location and a set of friends that were both constant throughout the 3-year period.

Outlets were categorized as selling either fresh fruits and vegetables (FV) or fast food (FF). These 2 options represent healthy and unhealthy dietary choices for the purposes of modeling. Each food outlet had a constant location and sold only one type of food in each period. However, food outlets could change the type of food they sold from one quarter (i.e., 90 days) to the next.

Process Overview

Formation of social networks.

Individuals belong to groups of friends, and some also have friendship ties to members of other groups. The presence of densely connected groups and the fact that any individual is only a few social ties away from any other is known as the small-world effect, which is commonly accounted for in the modeling of social networks.²⁴ Although our baseline scenario relied on small-world networks, we also ran our model under an alternative case with weaker groups, in which social groups were modeled as a random network. (Appendix A, available as a supplement to the online version of this article at http://www.ajph.org, summarizes the models used to generate both network types, together with their calibration.)

Individual decision-making.

Individuals chose to consume FV or FF each period on the basis of taste preferences, health beliefs, a food-price index, price sensitivity, food accessibility, and demographic factors (age, gender, and education). The weight assigned to each factor was based on data derived from the Food Attitudes and Behaviors (FAB) Survey²⁶ and other empirical studies. The sample for the 2007 version of the FAB Survey (from which our data were taken) comprised 3397 adults; estimates from the 2000 US census were used to provide sampling weights so that the sample was representative of the US population. The survey aimed to evaluate various factors, such as attitudes and beliefs, related to eating behaviors among adults.

During each simulation period, individuals’ taste preferences (the stronger their preferences, the more they like the taste of sugary or salty foods) and health beliefs (the stronger their health beliefs, the more they care about the healthiness of food) were updated according to their habits during the preceding period, the influence of their friends, and the food marketing strategies used in the broader environment (the updating function is presented in Appendix B, available as a supplement to the online version of this article at http://www.ajph.org). The models focusing on individuals also evaluated the amount of accessible food within walking distance, defined as 0.25 miles to 1 mile,²⁷ and food prices.

The models then computed 2 probabilities measured in the FAB data set: the weekly probability of consuming more than 1 meal from an FF restaurant and the daily probability of consuming more than 2 servings of FV. Finally, according to model specifications, individuals consumed FV or FF in proportion to their calculated probabilities. (Equations and parameters are provided in Appendix B.)

Food-outlet adaptation.

In our simulations, food outlets evaluated the food consumption of consumers in the neighborhood and decided whether to sell FV or FF. This evaluation took place quarterly (i.e., once every 90 time points). The model tracked consumption of the type of food sold by each outlet. The probability of switching types of food was proportional to the decline between the current quarter’s consumption and that of the preceding quarter (e.g., if average FF consumption decreased by 75%, an FF outlet had a 75% likelihood of switching to sales of FV). In the simulations, outlets did not switch types of food if their sales remained constant or increased.

Model Calibration and Validation

The setting of the simulations was downtown Pasadena, California, an urban district of 1.3 square miles.²⁸ Pasadena was chosen because its moderate population density, racial diversity, and income levels make it fairly typical of Los Angeles County and similar to the nation as a whole (albeit more diverse). The district was represented by a grid of 260 × 260 cells. In 2010, the district included 113 food outlets that could be categorized as FF venues and 16 outlets that could be categorized as FV venues (e.g., farmers’ markets and supermarkets).²⁹ Each outlet was randomly assigned to one cell, and a sensitivity analysis was performed to assess whether uneven distribution of food outlets would affect model predictions (Appendix C, available as a supplement to the online version of this article at http://www.ajph.org).

Data from the 2010 census showed that the district comprised 17 698 residents aged 18 years or older,²⁸ and those residents were assigned to unique locations chosen at random. Also, each resident was assigned an age, a gender, and a level of educational attainment. On the basis of these characteristics, residents were then assigned taste preferences, health beliefs, and price sensitivities (Table 1).

TABLE 1—

Summary of Model Inputs and Calibrated Statistics at Baseline

Parameter	Definition and Values
Individual attributes
Agei	Age group, y
	18–34: 57.00%
	35–54: 31.27%
	≥ 55: 11.73%
Genderi	Gender
	Male: 48.07%
	Female: 51.93%
Educationi	Educational attainment
	< high school: 10.85%
	High school: 16.27%
	Some college: 11.64%
	College: 61.24%
Ti,ta	Taste preference: person prefers sweet and salty foods; age is negatively associated with taste preference
	Taste ∈ (0.25–1) if age = 18–54 y
	Taste ∈ (0–0.75) if age ≥ 55 y
Hi,ta	Health beliefs: person has a strong value for eating healthfully; age, female gender, and educational attainment are positively associated with health beliefs
	Health ∈ (0–0.75) if age = 18–54 y and gender = male
	Health ∈ (0–0.75) if age ≥ 55 y and gender = male and education < high school
	Health ∈ (0–0.75) if age = 18–54 y and gender = female and education < high school
	Health ∈ (0.25–1) if all other groups
PSia	Price sensitivity: person has a strong preference for cheaper food; age and educational level are negatively associated, and female gender is positively associated, with price sensitivity
	PS ∈ (0–0.25) if age ≥ 55 and gender = male and education = college
	PS ∈ (0–1) if age = 18–54 y and gender = female and education < high school
	PS ∈ (0–0.5) if all other groups
AFV i,t and AFF i,t	Accessibility of fruit and vegetable outlets and accessibility of fast-food restaurants: person is disinclined to purchase food sold outside of walking distance (0.25–1 mile)
	Accessibility = 1 if within walking distance
	Accessibility = 0 if outside of walking distance
α	The extent to which an individual is influenced by friends’ social norms
	α ∈ (0, 0.3), uniform distribution
Environmental attributes
PIfvb	FV price index: weighted FV price index in 2007; set to 0.72 at baseline
PIffc	FF price index: weighted FF price index in 2007; set to 2.71 at baseline
γ	Visibility weights on social norm of friends: γ ∈ (−1, 1), uniform distribution; set to 0 at baseline
FF/FVd	Ratio of stores selling fast food versus fruits and vegetables; set to 7 at baseline
Empirical statistics
Pr(FV)e	Probability of eating at least 2 servings of FV per day among the residents of downtown Pasadena; Pr(FV) = 0.81
Pr(FF)e	Probability of eating at least 1 meal from a fast-food restaurant per wk among the residents of downtown Pasadena; Pr(FF) = 0.47

Note. FF = fast food; FV = fruits and vegetables. The age, gender, and education distributions are assumed to be the same as in downtown Pasadena in 2010 (based on the 2010 US census).

^aPredicted with data from the Food Attitudes and Behaviors Survey.

^bDerived from Powell et al.,³⁰ who estimated the price index with data from the American Chamber of Commerce Researchers Association (ACCRA). The FV price index was based on price data collected in the following food categories: potatoes, bananas, lettuce, sweet peas, tomatoes, peaches, and frozen corn. Each price index was weighted on the basis of expenditure shares provided by ACCRA, derived from the Bureau of Labor Statistics Consumer Expenditure Survey.

^cDerived from Powell et al.³⁰ The fast-food price index is based on the following 3 items included in the ACCRA data: a McDonald’s Quarter Pounder with Cheese, a thin-crust regular cheese pizza at Pizza Hut or Pizza Inn, and fried chicken (thigh and drumstick) at Kentucky Fried Chicken or Church’s Fried Chicken.

^dCalculated with data from the California Department of Public Health.²⁹

^eEstimated with data from the 2007 Los Angeles County Health Survey; statistics for subpopulations of downtown Pasadena were adjusted via population weights.

To validate our equations, we compared baseline model outcomes (i.e., without interventions) with FF and FV consumption data among the adult population in downtown Pasadena estimated from the 2007 Los Angeles County Health Survey³¹ (Table 1).

Policy Experiments

On the basis of the calibrated model and assuming constant demographic characteristics, we conducted policy experiments by manipulating the model’s 4 environmental attributes. The following policies were tested.

• • Tax on unhealthy food: The FF price index in the model was increased 1% up to 20% to reflect the effects of taxes on fat or sugar.

• • Subsidies for healthy food: The FV price index was lowered by 1% up to 20% to reflect the effects of increased agricultural subsidies.

• • Promotion of healthy norms: Food advertising is associated with a perception that others (peers, friends, and family) like the advertised foods.³² Therefore, we examined the impact of advertising healthy foods by increasing the visibility (parameter γ in Table 1) assigned to healthy peers from 1% to 10%.

• • Regulation of local food environment: We modeled zoning regulations that would alter the makeup of the district with respect to the ratio of stores selling fast food versus fruits and vegetables (FF/FV in Table 1). We lowered the ratio from its baseline value of 7 down to 4 and accounted for the mediating role of accessibility.

These policy experiments and their magnitudes were chosen as realistic policy options. Each run of the simulation took place over 1100 days and included a sample of 5000 participants. For each set of variables, we conducted the simulation 5 times to account for random variation.

RESULTS

Results showed that price-based strategies (i.e., taxing unhealthy food, subsidizing healthy food) could lower FF consumption and increase FV consumption (Table 2). However, improvements were small given the tax increases considered (Figure 1a). Promotion of healthy norms exhibited more of an effect: a 10% improvement in the visibility of positive social norms yielded an increase of 7 percentage points in FV consumption and a decrease of 6 percentage points in FF consumption, whereas the impact of a 20% increase in taxes on FF was only 3 percentage points.

TABLE 2—

Comparison of the Results of Different Policy Interventions on Food Consumption Decisions

		Increasing Taxes on FFa		Increasing Subsidies for FVb		Zoning Policiesc		Promoting Healthy Normsd
	Baseline	10%	20%	10%	20%	6	4	1%	10%
Probability of consuming ≥ 2 servings of fruits and vegetables/d: Pr(FV) × 100	75.8 (1.8)	76.0 (1.0)	76.0 (1.8)	76.4 (1.6)	77.7 (1.6)	75.5 (1.4)	76.5 (2.3)	76.7 (1.0)	83.1* (1.6)
Probability of consuming ≥ 1 fast-food meals/wk: Pr(FF) × 100	52.8 (0.8)	51.2* (0.7)	49.9* (1.1)	52.4 (0.9)	52.8 (1.4)	52.3 (0.7)	52.0 (1.7)	52.2 (1.1)	47.3* (1.5)

Note. FF = fast food; FV = fruits and vegetables. Average probabilities and standard deviations are provided.

^aIncreasing the FF price index in the model by 10% and 20%.

^bDecreasing the FV price index in the model by 10% and 20%.

^cAdding FV outlets in the region so that the ratio of FF to FV outlets falls from 7 to 6 (first column) or from 7 to 4 (second column).

^dIncreasing the visibility (parameter γ) given to healthy peers from 1% to 10%.

*P < .05 (significant difference from baseline).

FIGURE 1—

Impact of (a) taxes on the consumption of fast foods and (b) zoning policies on the consumption of fruit and vegetables.

Note. Impact of zoning policies was defined as altering the ratio of stores selling fast food vs fruit and vegetables in the district. We evaluated zoning policies by assuming 3 different food accessibility buffers: 1 mile (squares), 0.5 miles (triangles), and 0.25 miles (circles).

The results also showed that zoning policies aiming at improving the density of stores selling FV would have no significant impact if individuals’ walking distance was set to either 1 mile or 0.5 miles (Table 2 and Figure 1). Even if individuals were to walk only 0.25 miles (Figure 1), the likelihood of finding healthy stores would be improved by only 2 percentage points through extensive rezoning (and locating such stores in that small area would be more difficult). However, one interesting result illustrated in Figure 1b is that expanding the buffer from 0.25 miles to 1 mile greatly increases FV consumption, which may indicate that improving transportation is a more effective strategy than rezoning.

Furthermore, promotion of healthy norms continued to have strong effects in both small-world networks and random networks (Figure 2). The impact on FV consumption remained largely unchanged for each decrease of 10% in the importance of peers’ influence (Figure 2b), and a similar effect was witnessed with respect to the impact on FF consumption (Figure 2a).

FIGURE 2—

Impact of promoting healthy norms on (a) consumption of fast foods and (b) consumption of fruit and vegetables.

Note. We evaluated 4 scenarios for each outcome. The first 3 scenarios involved a small-world network and varied the strength of the peer influence value (α; see Table 1) from 0.3 (squares) to 0.2 (triangles) and 0.1 (circles). The fourth scenario involved a random network with a peer influence value of up to 0.3.

DISCUSSION

A typical unhealthy diet in the United States is composed of too much high-calorie food and an insufficient amount of fruits and vegetables. Our agent-based simulation model suggests that social networks have significant potential to influence dietary decisions. Although this general point about social networks has been made before,¹³ the value of our simulation is to put this insight into a comparative perspective: not only do social networks affect dietary behavior, but they may be the most significant modifiable contributor to such behavior, far stronger than other proposed interventions (e.g., strategies designed to affect the relative prices and availabilities of healthy and unhealthy foods).

Targeting food prices by introducing a tax on fat or sugar or by subsidizing healthy foods may not change dietary behaviors to a great extent. Attracting more FV stores may not exert a powerful influence on behaviors in moderately dense urban areas such as downtown Pasadena, given that people have easy access to both healthy and unhealthy foods nearby. However, improving transportation networks may be an alternative to enhance the accessibility of healthy food in instances in which routes are less walkable.

The policies tested in this study are comparable, and all have been proposed as politically feasible interventions to curb the obesity epidemic.^2,33 Moreover, the magnitudes of the effects derived from our model simulations are similar to what might be expected from other results in the literature. For instance, a few states have imposed small taxes on carbonated beverages, and their experience has suggested that there is no evidence that a tax on these products would have a positive impact on obesity rates.³³ However, one study predicted a slightly larger impact of taxes on food consumption than predicted here.² There has been increasing adoption of social marketing in public health, and these efforts have achieved important successes.³⁴

The use of ABM to understand this public health problem has certain important strengths. First, it provides a measure of the population-level impact of various policies. Although previous research has carefully tested the significance of effects of various determinants of dietary behavior, the next step is to combine these insights into a single, population-based model. ABM may be a fruitful way forward in assessing the overall determinants of population-health outcomes.

Second, unlike statistical models, ABM allows individuals to be connected and influenced by each other, which has considerable effects, as shown in this and other studies.^13,24 The assumption of a stable unit-value treatment effect, on which nearly all statistical analyses depend, makes the statistical testing of social norms difficult³⁵; however, this difficulty should not be taken to imply that social norms are not important in determining behavior.

Moreover, ABM allows interactions between individuals and the external food environment, which itself can adapt to individuals’ consumption. By relaxing this assumption, it is possible to observe the dynamic evolution of obesogenic eating behaviors beyond simply explaining correlations between variables, as statistical models typically do. Unlike most other agent-based models that have studied food systems,^24,36 our model is based on existing theories and an empirical database. A baseline validation was conducted to increase the model’s reliability.

Notwithstanding these advantages, our model also has important limitations. First, data are lacking regarding the extent to which people are influenced by their friends. We used a conservative value¹³ for this variable. Even a small change in the value led to substantial changes in outcomes, suggesting the necessity of carefully parameterizing this factor in food-choice models. Second, there is little empirical information regarding how much adults’ perceptions of social norms can be biased by food advertising, particularly healthy food advertising. The mechanisms relating to how advertising information is disseminated and how it modifies perceptions remain unclear. Third, throughout the simulation, we assumed that population characteristics (age, gender, and education) remained constant at 2010 values and that agents did not change locations. These assumptions may not hold in the presence of in-migration or out-migration, some of which might be endogenous as the food environment evolves.

In summary, our model simulation suggests that norm-based interventions may be more effective in changing unhealthy eating behaviors than tax-based policies or zoning policies. It is beneficial to reinforce and reproduce healthy norms even if individuals do not form communities, suggesting that this strategy does not need to be community based but can also be implemented, for example, via mass media. ABM is useful in developing and testing potential interventions targeting complex health problems.