ABSTRACT
Poor diets, including those high in fast food, are a leading cause of morbidity and mortality. Exposure to low-quality food environments, such as ‘food swamps’ saturated with fast food outlets (FFO), is hypothesized to negatively impact diet and related disease. However, research linking such exposure to diet and health outcomes has generated mixed findings and led to unsuccessful policy interventions. A major research limitation has been a predominant focus on static food environments around the home, such as food deserts and swamps, and sparse availability of information on mobile food environments people are exposed to and food outlets they visit as they move throughout the day. In this work, we leverage population-scale mobility data to examine peoples’ visits to food outlets and FFO in and beyond their home neighborhoods and to evaluate how food choice is influenced by features of food environments people are exposed to in their daily routines vs. individual preference. Using a semi-causal framework and various natural experiments, we find that 10% more FFO in an area increases the odds of people visiting a FFO by approximately 20%. This strong influence of the food environment happens similarly during weekends and weekdays, is largely independent of individual income. Using our results, we investigate multiple intervention strategies to food environments to promote reduced FFO visits. We find that optimal locations for intervention are a combination of where i) the prevalence of FFO is the highest, ii) most decisions about food outlet visits are made, and most importantly, iii) visitors’ food decisions are most susceptible to the environment. Multi-level interventions at the individual behavior- and food environment-level that target areas combining these features could have 1.7x to 4x larger effects than traditional interventions that alter food swamps or food deserts.
Poor diets, including the over-consumption of foods that are energy-dense but nutrient-poor, that have excess sugar and/or sodium, and that are ultra-processed, are a major cause of diet-related disease and mortality1–3. Poor diets led to 11 million deaths globally in 2017 (more than tobacco)4, largely due to their causal role in major chronic diseases, including obesity, type 2 diabetes, some cancers, and heart disease5. Exposure to, or spending time in, certain built food environments is hypothesized to impact diet and related diseases6,7. Low-quality built food environments are generally categorized into two types. “Food deserts” are defined as areas with low access to healthy foods (e.g., neighborhoods where a majority of residents live more than 0.5 or 1 mile from a supermarket, a key source of affordable, healthy food)8. “Food swamps” are areas saturated with less healthy food outlets, often defined as neighborhoods that have a higher number of fast-food outlets (FFO) and convenience stores, or a high ratio of these outlets relative to healthier food outlets9,10. Both of these types of low-quality food environments are frequently concentrated among low-income communities and communities of color and contribute to inequities in nutritional health11,12. It is hypothesized that exposure to food swamps can nudge people to consume unhealthy food (e.g., fast food) due to a cost decrease via lower food prices or less time needed for transactions13, or through structural or social cues to behavior14. In contrast, food deserts are hypothesized to create barriers to accessing affordable healthy foods, which can lead people to make less healthy food choices that they would otherwise avoid.
To date, research into the relationship between food swamps or deserts and food choice has predominantly focused on predefined local and static food environments15,16, largely of the neighborhood around the home, with schools and workplaces to a lesser extent. While exposure to both types of food environment has been associated with increases in unhealthy eating and diet-related disease, overall, findings are mixed and predominantly null17–20. Furthermore, most of these studies have been cross-sectional and fail to establish a causal relationship between neighborhood food environments and unhealthy diet. Despite this limited evidence, there has been considerable interest from federal and local policymakers and private funders in supporting policy interventions to neighborhood food environments, including investments by the U.S. Healthy Food Financing Initiative of $270 million plus $1 billion in leveraged financing to support healthy food retail in underserved neighborhoods since 201021, and ‘fast food bans’ implemented in select Los Angeles neighborhoods by the city council using zoning regulation to restrict opening of new FFO.22. Across numerous evaluations, these interventions have demonstrated no meaningful impact on diet quality or diet-related disease outcomes22–27. Better understanding of the relationship between food environment exposure, diet, and diet-related disease will be critical to designing more effective interventions to food environments.
The limited focus on residential and static food environments may be one explanation for these mixed results, given that a growing proportion of food acquisition and consumption occurs miles from our homes. For instance, in9 the finding that food swamps predict higher rates of obesity at the neighborhood level is weaker in neighborhoods where residents are more mobile (i.e., more residents who travel to work by car or public transport). Among Americans, “food away from home” (vs. foods prepared at home)–the vast majority coming from fast food and full-service restaurants–constitutes one-third of total energy intake and one-half of food budgets 28. Thus, a major source of exposure to and use of food environments is unlikely to be captured by existing research foci and methods. Additionally, these studies often test whether exposure to food swamps or deserts predicts nutritional health, without incorporating information on the food outlets that individuals actually visit. Given the well-documented biases of survey data to capture detailed human movement and dietary intake29, small studies (often < 100 people) have begun to use tracking technologies to map how people move through their environment to acquire food over brief periods of time (e.g., 1 week)30–34. However, this has not been studied at scales large enough to capture habitual patterns of food environment exposure over extended time intervals, or statistically significant effects of those food environments on peoples’ behavior. Overall, a major gap in the literature is detailed evidence of the food environments people are exposed to as they move around, both at and beyond where they sleep and work (i.e., mobile food environments), the food outlets they actually visit in these contexts, and causal designs capable of investigating how mobile food environments influence diets and diet-related disease.
In this study, we use a large, privacy-preserving, population-scale mobility dataset spanning a 6-month period during 2016-2017 to examine peoples’ visits to food outlets (FO) and FFO in and beyond their home neighborhood, and to investigate how these FFO visits are linked to features of the mobile food environments they are exposed to throughout their daily routines. Mobility data allow us to observe an unprecedentedly diverse and heterogeneous population body35,36. They allow us to observe when and where FFO visits happen among this large, diverse population, and thus to understand the individual- and environment-level variables that condition that decision over other food choice alternatives. Moreover, they allow us to find structural, randomized shocks in people’s routines (e.g., moving, going to a government office), which we can leverage to investigate causal effects of food environments on food outlet decisions. Our analyses focus on visits to FFOs as the key outcome because (i) greater intake of fast food, which is typically ultra-processed, low in nutrients, and energy dense, is a well-established risk factor for poor diets, obesity, and cardiometabolic disease2,37; and (ii) recent work has shown, using the same mobility dataset we utilize in this study, that visits to FFO are strongly associated with self-reported fast-food intake, obesity, and type 2 diabetes, thereby establishing the link between FFO visits observed in mobility data and nutritional health38.
Results
Characterization of mobile food environments
Individuals in large urban areas travel or commute considerable distances40, indicating that for many people, the food environments they are exposed to throughout the day are not near their homes. In our dataset, we find that the median distance from home h to any place visited x is 7.5km, see Fig. 1. The median distance to any type of FO visited is 6.77km but this varies by outlet type: the distance to grocery stores/supermarkets is much smaller (0.44km) while FFO are 6.61km away (median). In fact, only 6.8% of the visits to FFO occur within a user’s home census tract. Thus, most fast food visits occur in food environments outside of a user’s home neighborhood.
A: in their daily life, users navigate the city from home h to different places, until they arrive at a context c where they decide to have food in f. B: Distribution of the distance from home to all visits in the city (black), all retail food outlets (green), and fast food outlets (orange). The distance traveled to food and fast food outlets is much larger than to supermarkets (gray) or the typical size of a census tract (dotted vertical line). C: Heatmap of the average of ratio of fast food outlets ϕ in the Los Angeles metro area. The ratio is calculated within each hexagon of size ∼1 km2. Icons designed by bqlqn/flaticon.com and maps were produced in R using the TIGER shapefiles from the U.S. Census Bureau39.
To characterize a user’s food environment at any given location x, we measure the ratio of FFO to FO within a 1km radius, ϕ(x) (See Methods and Supplementary Section 4 for other definitions). As shown in Fig. 1, most zones in the metro areas have small (average) values of ϕ(x). Because users move around the city, they are exposed to many different food environments: overall mobile exposure to food environments, the time-weighted ratio of FFO to FO that a user is exposed to in our 6-months entire period (
, see Methods) has a median of 14.1%. We also find that users’ mobile food environments are different from the food environments around their homes. Although home environments have a relatively low FFO to FO ratio [median of 
for all users is 9.75%], we find that the correlation between mobile and home environments is small 
across users. In other words, the food environments that users are exposed to throughout the day are different from the ones around their homes.
The ratio of FFO to FO in users’ food environments is associated with various socio-demographic characteristics inferred from users’ home census block group. Using regression models for 
(see Fig. 2 and Methods), we find that users exposed to mobile food environments with a higher proportion of FFO (larger average 
) reside in areas with a higher proportion of residents who are Black, who have long commutes, and who have higher skill jobs, and a lower proportion of residents with more educational attainment and who depend on public transportation. Neighborhood-level household income is not significantly associated with any mobile food environment features. These relationships differ with users’ home food environments (see Fig. 2), where a greater proportion of FFO, larger 
, occurs in neighborhoods with lower income, higher levels of educational attainment, and shorter commuting patterns, similar to findings in41. Despite that, we find a stronger relationship between the socio-demographic characteristic of users and their mobile food environments (R2 = 0.21 for 
) than for their home food environments (R2 = 0.038 for 
). This indicates that socio-demographic inequalities propagate stronger to the mobile food environments than to food spatial accessibility at home.
A: Relationship between FF environments and socio-demographic traits. Bar shows the coefficient estimates for OLS regression models of overall mobile and home FFO environments 
and fraction of visits (μi) by user for proportion of workers in low-skill jobs, proportion of people with long (>45min.) commute, proportion of people taking public transportation for commuting, proportion of people employed, proportion of people with higher education level, median household income and proportion of Black people in their home census block group. B: Daily patterns of the number of FO (top) and FFO (bottom) visits in our urban areas. The shaded area corresponds to the lunch observation period taken to determine the action yit in our model (1).
Who, when, and how much people visit FFO
While 
describes the home and mobile exposure to fast food options at any given moment, we encode the actions of users at time t with the variable yit. If individual i chooses to visit a FFO among the FO options, then we set yit = 1. If they select a non-fast food option, then we set yit = 0. The overall averaged fraction of FFO choices to FO options 
(see Methods) in all environments over our observation period is heterogeneous across users with an average of 0.187 ± 0.202. That is, on average, 18.7% of visits to FO are to a FFO. Our results also show that a significant proportion of users never visit FFOs (22.9%) during the 6-month period of observation. Using similar regression models as before, we find some statistically significant differences in the ratio of visits to FFO (μi) across demographic groups. Similar to the results described above, individuals visit FFOs more often if they live in areas with less use of public transportation, with a higher proportion of Black residents, with longer commutes, or with less educational attainment. Similarly, income has a smaller association with the ratio of FFO visits when compared to the other demographic traits. Our results align well with the contradicting evidence of little variability in fast food intake across income levels42–44 despite consistent differences based on educational levels and race and ethnicity. Additionally, we find that traditional demographic traits, specifically, race and ethnicity, type of job, income, and educational level, have a weaker association with fast-food visits than characteristics related to mobility and time constraints; specifically the use of public transportation and long commuting. However, it is important to note that the explanatory power of this association between μi and socio-demographic variables is low, R2 = 0.052. Thus, even though we find statistically significant differences, our results suggest that overall FFO visits do not meaningfully differ across different socio-demographic groups. Many types of people visit FFO in urban areas.
We also found that most food outings happen between midday (lunchtime) and the evening (dinner time), both during weekdays and weekends, see Fig. 2. FFO visits have the same temporal pattern, with a peak of visits to fast-food happening around lunchtime from Monday to Sunday.
Relationship between mobile food environments and fast food visits
To understand the effect of food environments on fast food visits, we first study the relationship between total average exposure to fast food 
and the ratio of FF visits μi. Many studies with small datasets have found null or contradicting results regarding the association between total exposure to FFO and fast food intake32, 45. We find a strong positive relationship between a user’s average daily exposure to FFOs within their mobile food environment, 
, and overall observed visits to FFO, μi. Specifically, the correlation between these two variables is 
. However, the correlation between FFOs within a user’s home food environment and overall visits to FFO is weaker, with 
. These results suggest that an individual’s exposure to FFO across the day, rather than within their home environment, is a more important driver of the decision to get fast food.
To better understand what drives the association between FFO exposure and visits, we need to go beyond total exposure. Visits to FFO might be due to individual preferences, to structural or social cues received from frequent exposure to environments high in FFO14, but also might be a direct response to the food environment where those decisions are made. A person may choose to visit different food outlets in food environments with different features. At the same time, different people exposed to the same food environment may make different decisions about which food outlet to visit. To identify the association between food environment features and FO decisions, we first design an individual analysis of each visit to a FFO. We restrict the data to FO visits during lunch hours (from noon to 13h30 local time) because this time window has the highest FO and FFO visits (see Figure 2). Most importantly, time constraints at this time of day could make the food environment options experienced before going to get lunch even more relevant. Indeed, we find that the relationship between action yit and the food environment features is stronger around noon than at other times of the day (see Supplementary Information Section 7).
For decisions about food outlet visits yit at time t, we take the last place in the morning (until 11:30) where a user was observed (context cit), and characterize the food environment in that context by ϕ(cit). To ensure that the context offered a choice between FFO and non-FFO, we focus exclusively on observations where the context cit contains both types of outlets.
We have extensively checked that our results do not depend on the precise definition of the food environment in the context (see Methods and Supplementary Information Section 10). We use a logistic regression model to estimate the impact of user i’s context environment on their decision to visit a FFO vs. a non-FFO at time t:
where logit−1(x) = ex/(1 + ex), αi is a fixed effect accounting for individual average preferences to visit FFO, δt is a fixed effect accounting for daily variation, and ϕ(ct) is the ratio of FFO in the food environment a user was exposed to prior to the visit. By imposing a fixed factor by user, we separate the effect of individual preferences on visits to FFO from effects of the food environment. For our regression, we only consider users that went to both fast and non-fast food outlets at lunch at least once during the 6 month observation period, and where the food environment they were exposed to before lunch included both FFO and non-FFOs.
Our results, illustrated in Figure 3, show a strong association between features of the mobile food environment and visiting a FFO. The model produces a log-odds of β = 1.84 ± 0.072 for all FO visits at lunchtime: when the context includes 10% more FFO, there is an increase in the odds to visit a FFO of (eβ×0.1 − 1) × 100 ≃ 20%. This influence of the food environment one is exposed to before going to lunch was similar during weekdays and weekends, and at different times of the day (not only during lunch hours) (see Figure 3 and SI Section 7). Additionally, the effect is largely independent of an individual’s income (see Figure 3) and other socio-demographic traits (see SI Section 7). Thus, income, socio-demographics, and time of the week do not significantly modulate the effect of the food environment on food outlet decisions, and individuals with different socio-demographic backgrounds appear to respond similarly to mobile food environments at different times.
A: Effect of the mobile food environment on visiting a FFO at different times, locations, or for different income or FFO visitation groups. Values show the coefficient β of the logistic regression in Equation 1 for the food visits (outings) corresponding to the different groups. Bars indicate the standard error of the coefficient. B: Evolution of the fraction of FFO visits (top) and cumulative difference in FFO visits (bottom) for groups of users that change their contexts from Low to High FFO environments (left) and High to Low FFO environments. The dashed line is the predicted counterfactual for groups of users that changed their context but not the FO diversity they were exposed to.
Despite finding an effect of mobile food environments on visits to FFO, it could be that the lack of non-fast food options predominantly affects individuals when they are in a new place. This may be because they are less equipped to navigate the new food environment and identify food options they prefer. It is also plausible that a lack of options for different FO types constantly affects users who would otherwise visit FFO with a lower frequency given their individual preferences. To address these questions, we propose a semi-causal framework using a natural experiment to investigate the relationship between habitual FFO context and FO decisions. In this experiment, we observe people who changed their quotidian context during the study. Using the time series of the different contexts before lunch, we were able to detect a significant proportion of users (0.43%) that changed their habitual context before lunch within our observation period (see Methods and SI Section 5 for details about this detection). Those users were split into four groups depending on whether they changed to a context with similar or different low (ϕ < 0.13) or high (ϕ > 0.13) exposure to fast food. We found that around 34% of the users who changed their context remained exposed to very similar food environment features before and after (High → High or Low → Low). Around 16% of the users changed their context such the FFO in their typical pre-lunch food environment went from Low → High exposure to FFO, and another 16% from High → Low FFO exposure. We construct the time series of the fraction of times that users visit a FFO relative to the day when they change contexts. We study the impact of the change on FFO visits using Bayesian structural time-series models (see Methods) by comparing the group that changed their FFO contexts (Low → High and High → Low) with the counterfactual of those that, despite they changed their context, they were exposed to similar FFO food environments (Low → Low and High → High, respectively, see Methods). Results are presented in Figure 3, which shows that the group that changed from Low → High FFO exposure increased their fraction of FFO visits from ∼ 17% to ∼ 25%. Similarly, users that changed their context from High to Low FFO exposure decreased from ∼ 26% of FFO visits to ∼ 18%. The counterfactual of users that changed contexts but remained exposed to food environments with similar FFO ratios maintained a similar fraction of FFO visits. We also note that this effect is statistically robust and persistent, remaining even 50 days after changing their context. In cumulative numbers, we find that users who changed their context to high (low) exposure visited FFO 4 times more (less) in 50 days than those who remained in food environments with similar FFO ratios. These results suggest that the effect of the food environment is strong even for the same users subject to different habitual contexts, and it is not only driven by visits to new places.
Finally, a third natural experiment study was analyzed to find a variation in the food environment as exogenous and random with respect to food options and individuals as possible. We study visits to the Department of Motor Vehicles (DMVs), because they are commonly visited locations for obtaining a driver’s license, government ID, voting, and other services. Moreover, the food environment around a DMV is unlikely to be a determining factor when choosing a DMV location, compared to other factors such as availability of appointments and distance from home/workplace. While distance to home/workplace can be influenced by latent factors such as income and even food preferences, the set of DMVs in our cities is small enough so that many people exit their home and work neighborhoods to go there. Moreover, the time constraints caused by the scarcity of appointments help make the choice of DMV location less determined by spatial accessibility alone. In our dataset, we detected 47,106 visits to the DMV across a 6-month period. The median distance traveled from home to the DMV was 7.60 km., a distance greater than the median distance to FO. We consider the DMV as the context of a FO visit if such a visit occurs within 2 hours of the DMV visit. To investigate the effect of the DMV food environment on yit, we use a logistic regression model 
similar to Eq. (1) model. Since we typically have one visit to DMV per user, we model individual preference 
as the fraction of visits to FFO of user i and we do not include daily fixed factors. The effect of the DMV food environment is shown in Figure 3. We find a significant effect, although a little smaller in size than the effect of exposure to habitual contexts, with log-odds of β = 1.26 ± 0.25. This third analysis corroborates that features of food environments influence FFO visits.
Policy implications
The observed relationship between food environments with high ratios of FFOs and increased visits to FFO, specifically for mobile food environments, implies that more targeted interventions to reduce visits to FFO can be designed. Many intervention approaches have focused on improving food environment quality around the home neighborhood or in geographic regions with poor food environments, without accounting for where people more frequently visit food outlets. Notable examples include the over one billion dollars leveraged by the U.S. Healthy Food Financing Initiative to finance healthy food retail in under-served local neighborhoods21, and the ‘fast food ban’ implemented in 2008 in neighborhoods in South Los Angeles with high prevalence of FFO22. Our findings highlight that FFO visits often take place well beyond the home neighborhood, and suggest that strategies that solely focus on geography and spatial access to food outlets (ignoring human behavior) are likely to lead to sub-optimal intervention effects. Indeed, evaluations of major policies and interventions to improve the quality of neighborhood food environments have demonstrated they have little impact on diet or diet-related diseases25,46. Here, we use the results of our observational study to identify the optimal locations to intervene in food environments to have the greatest impact on decreasing FFO visits. Specifically, these are contexts demonstrating the highest ratios of FFO, highest frequencies of user exposure and FFO visits, and the largest observed impact of food environment features on a population’s FFO decisions. We investigate the likely effects of intervention strategies that change the ratio of FFO to FO in these optimal impact locations vs. locations such as neighborhood food deserts and food swamps, the traditional choice locations for intervention. These targeted food environment interventions could be implemented through similar strategies tried in the past, such as ‘fast-food ban’ zoning or subsidies for opening new healthier food outlets, but with potentially much greater benefits to population diet and health.
Assuming that our intervention I changes the context in an area Ω by δϕ/δ I and that users are still making their decisions according to the model in Eq. (1), the change in the number of FFO visits made immediately after being exposed to the food environment of Ω can be obtained as (see Methods)
where Xit = β0 + δt + αi + βϕ (cit). This expression shows that the effect of an intervention in an area Ω depends on three factors: i) the susceptibility of FFO visits with respect to the availability of FO options in the area Ω, expressed through 
and ultimately by the balance between individual preferences αi and the context ϕ(cit), ii) the number of decisions made in area Ω, expressed by the sum, and iii) the effect of the intervention on the context ϕ, expressed by δϕ/δ I. For example, for the same intervention δϕ/δ I, we can have an area where many people go, but they have a large individual preference for FFO (αi ≫ 0) causing the effect of the intervention to be small because they are not influenced by the food environment features. On the contrary, we can have an area that is not visited by many people, but the people who do visit are highly influenced by that contexts’ food environment and the number of FFO around (αi ≃ 0). An intervention in this latter area can meaningfully change a number of food outlet decisions.
To illustrate this, we consider a simple intervention in which we convert a FFO into a non-FFO in a particular area Ω. In this case, ϕ(cit) changes by approximately δϕ/δ I ≃ −1/nΩ where nΩ is the number of FO in the area. We have also extended the model to all times of the day and week to describe the full effect of the strategy (see Methods). Assuming that we have limited resources to change hundred food outlets, where are the areas in which our intervention maximizes its impact? Here, we compare four different strategies.
In the first strategy (Food Swamp intervention), we select the areas with the largest (average) values of ϕ, i.e., the areas where FFO predominate. For comparison and to resemble prior food desert interventions around home neighborhoods, in our second strategy (Low Food Access intervention) we select the areas that have the largest values of ϕand are classified by the USDA as food deserts (both low-income and low-supermarket-access)8. The third strategy (Food Hotspots intervention) is implemented by selecting the areas where most FO visit decisions are made. However, these strategies do not incorporate individual preferences or susceptibility to food environments. Thus, in our fourth strategy (Behavior-Environment intervention), we select areas Ω as the top areas ranked by ΔFFO(Ω) in Eq. (2), which includes not only the context but also the individual preferences of people deciding in those contexts. Figure 4A shows the relationship between the change in FFO visits ΔFFO(Ω) and the average context ϕin the different areas (census tracts) in our cities. The figure illustrates the dependence between these two variables, but there is still significant variability. For the same average of ϕ, we have areas with changes in ΔFFO(Ω) that span two orders of magnitude. In this representation, our Food Swamp and Behavior-Environment interventions are very easy to interpret. They consist of choosing the rightmost (greatest change in context) or topmost (greatest change in FFO visits) areas, respectively. In the case of Food Swamps and Low Food Access interventions, we can see those strategies choose areas Ω in which ΔFFO(Ω) is small because not many decisions are made (Low Food Access) or because users are less affected by the FFO environment in those areas (Food Swamps). The Food Hotspots intervention chooses areas where the most FO decisions are made, but without considering whether users are affected by the FFO environment. As a result, the total effect of the four strategies is very different, see Figure 4B. Overall, the Behavior-Environment intervention would be 1.66x to 4.07x times more efficient in decreasing FFO visits than interventions that used only the FFO context where decisions are made or around the home neighborhood. In relative terms, by changing one hundred FFO (0.22% of the total), our Behavior-Environment intervention could avert around 0.56% of the visits to FFO, while other strategies only could affect 0.3% of those visits at most. If we scale these numbers to the total population, our Behavior-Environment could avert around 622k visits to FFO in 6 months, compared with only 320k at most in the other interventions.Furthermore, the impact of the interventions is predominantly independent of the income (see Figure 4B) and health risks of people (see Supplementary Material Section 9), and thus its effect does not concentrate on particular groups. For example, we found that our Behavior-Environment intervention is still 2.2x to 2.5x more effective than the rest of the interventions at targeting decisions made by high obesity or diabetes prevalence groups (see Supplementary Material Section 9).
A: Change in the number of visits to FFO in an area Ω after deploying an intervention I as a function of the ratio of FFO to FO, ϕ, in that area. Dashed lines marked the thresholds for the strategies to choose the top 100 areas by FFO ratio (vertical line, red points, Food Swamp intervention) or by the total change in the number of visits (horizontal line, green points, Behavior-Environment intervention). Orange points are those 100 areas chosen as the top home areas according to the FFO environment (Low Food Access intervention), and blue points correspond to those 100 areas chosen as the top areas containing more FO visits (Food Hotspots intervention). B: Total effect of each intervention strategy in the different urban areas. Shades correspond to the number of actions changed by different income quantiles. C: Relative frequency (to all areas) to find different groups of POIs (topics) in the areas selected in each intervention strategy.
Finally, to understand what types of areas Ω are targeted in our Behavior-Environment intervention, we use latent topic analysis to determine the groups of points of interest (POIs) that appear more frequently in those areas (see Supplementary Material Section 8). As shown in Figure 4, we find that the groups of POIs related to “Malls”, “Industry / Factory”, “Airport” or “Office” are more likely to appear in our targeted areas than in the rest of the areas in the city and the rest of interventions. Other groups of POIs like “Restaurant” or “Entertainment” are less likely to be areas selected in our Behavior-Environment intervention, but more likely to be selected in the “Food Hotspots” intervention. However, this intervention is less efficient, meaning that despite a lot of food visits happening around “Restaurant” and “Entertainment” areas, people making food decisions there are less affected by the environment, and thus the intervention is less successful. In summary, these results suggest that more efficient interventions could be deployed to specific types of food environments where FO decisions are most influenced (and likely constrained) by the environment, i.e., closer to work, travel, or shopping areas.
Discussion
The effect of food environment exposure on diet and related diseases has been studied extensively, however, in many cases, small, limited datasets have hampered the ability to understand this complex process in urban areas18–20,22,23,25,27,30–33,45. A major research limitation has been a focus on cross-sectional observations of static food environments around the home, and limited study of the mobile food environments people are exposed to and food outlets they visit as they navigate their day-to-day environments20,30–32,45. Our dataset and semi-causal study designs have allowed us to analyze the effect of exposure to food environments on food choice at an unprecedented granularity and across diverse populations. This longitudinal, individual-level behavior data has also allowed us to analyze how food choice is motivated by features of food environments people are exposed to in their daily routines vs. by individual preference. We find that most visits to fast food outlets occur relatively far (a median of 6.61 km) from home and that exposure to low-quality food environments beyond the home is strongly and significantly linked to increased visits to FFO, across diverse sociodemographic groups.
Previous policy interventions to food environments have been targeted to low-quality neighborhood food environments, such ‘food swamps’ characterized by an abundance of fast food options. Interventions focused on static features of neighborhood environments do not reflect the complex intertwined process of human mobility, food environment exposure, and eating decisions in urban areas. This mischaracterization may partially explain the unsuccessful neighborhood interventions that aim to regulate the fast-food environments near where people live22,23,34. To reduce visits to FFO and improve overall dietary quality, our results suggest that we may need to intervene in the mobile food environments that are not only characterized as low-quality, but also where most FFO decisions happen, and importantly, where people have been demonstrated to be most influenced or constrained by the options available in that area — observations possible with this mobility data. The most efficient interventions may be further from people’s homes, in areas where food environments are more determinant to food decisions, like work, school, travel, or shopping areas.
While food environment interventions based on these factors were demonstrated to impact all groups equally, including lower income groups and those with higher rates of diet-related disease, these data also allow us to observe vast inequities in exposure to food environments of different quality for historically marginalized communities. Based off of these observations and innumerable studies demonstrating the structural inequities between socioeconomic groups11,12,14,47, future targeted interventions should be designed to account for the additional and complex dimensions of fairness and equity48,49, while accounting for individual preferences and projected decisions.
In addition to identifying locations for intervention, our methodology could also be used to inform individual-level interventions promoting or encourage visits to food outlets located in food environments that have more diverse, healthy food options, for example using a mobile app. Design of such individual-level interventions would require combination with other data about food intake (e.g., delivery, nutrients50, degree of food processing51), food preference and sentiment52,53, and price sensitivity46,54,55.
Population-scale mobility data provide useful, dynamic behavioral indicators of FF visits and consumption35,38, re-defining static notions of “food deserts” or “food swamps” to mobile food environments determined not only by the diversity of FO available in those environments but also by their frequency of use and peoples’ susceptibility to what they offer. We hope our results and our complex-systems methodology using large-scale mobility data can inform more efficient policies and interventions on food environments complementing and extending those around home neighborhoods50,56–58 or efforts to increase the healthfulness of food items being sold at FFO59.
Our study has several limitations. Although it is well established that eating at FFOs is linked to poorer diet quality28, and there is a strong association between observed visits to FFO and FF intake38, foods of diverse nutritional quality are sold across FFO60. We have currently not examined the extent to which healthier options are offered at the FFO visited by our sample, nor how these options may impact purchase and consumption behavior. Our results, therefore, serve as a proxy and bound for the potential FF intake. Also, since visits are attributed to the closest POI, there are limitations to the detection of visits to certain food outlets, such as those in multi-story or multi-purpose buildings (e.g., malls) where FFO are frequently found. Additionally, because we only detect visits greater than five minutes in duration, we may miss very brief FF outlet visits (e.g., drive-thrus). Finally, although our semi-casual framework provides robust evidence about the impact of mobile food environments on people’s FFO visits, we believe our results may be further tested through carefully designed experiments and interventions. These interventions should also explore potentially different effects of food environments on other types of food outlet visits (e.g., visits to restaurants or grocery stores). Finally, our mobility data sample from 2017 may not reflect changes in exposure to and impact of food environment on FO visit behavior following the pandemic and related changes that have occurred in the intervening years, including increased time spent in home neighborhoods and the great expansion of food delivery apps and their coverage. However, in settings where individuals are constrained by their environment, our findings and population-scale mobility framework likely still apply.
Methods
Mobility data
We use individual-level anonymized mobility data of 1.86 million anonymized users in 11 US metropolitan areas over a period of 6 months, from October 2016 to March 2017. The mobility data were collected with the informed consent of the users, who opted-in to anonymized data sharing for research purposes under a GDPR and CCPA-compliant framework. Our mobility data were obtained from Cuebiq, a location intelligence, and measurement company. To identify visits to FO and FFO, we extracted from the mobility data the stays (stops) of people around a large collection of points of interest (POI) obtained from Foursquare. FFOs are quick-service restaurants where patrons typically pay before eating and were defined using Foursquare’s taxonomy and a name search using a list of chain FFO; FOs, which represent all retail food outlets including grocery stores, supermarkets, big box stores, convenience stores, restaurants, were based on Foursquare’s existing taxonomy (see Supplementary Information Section 3). We have comprehensibly checked that our results do not depend on the choices made on the definition of stays, the categorization of the FFO, the POI database, or the definition of the environment and the population representativity of our data. A full description of those definitions and robustness checks is provided in the Supplementary Information Sections 1, 2, 3, 4, and 10.
Definition of the home and mobile food environment
To characterize the food environment users are exposed around a given place x we measure the ratio of FFO to any FO within a 1km radius, ϕ(x). We have extensively checked that our results do not depend on other definitions of the food environment. For example, in Supplementary Information Section 10 we show that similar results are obtained when we take ϕ(x) as the ratio of FFO to FO of the closest 25 FO to x, a definition that accounts for the different density of FO around the city. Home food environments are described by the value of ϕaround where people live 
. While food environment exposure around place x is given by ϕ(x) we also computed the total exposure a user gets by moving around as 
for all times t the user stops for more than five minutes (irrespectively of whether a FO is visited) and where τ(xi,t) is the duration of the stop of individual i at xi,t. Finally, the overall averaged fraction of visits to fast-food of individual i is computed as 
for all times t the user visits a FO and where Ni is the total number of FO visits of individual i.
Statistical Models
To test for the effect of mobile food environments, we have run a number of statistical models. For the main results in Figure 2 we used a logistic regression to link the binary output yit to the ratio of FFO options around the context ϕ(cit), see Equation (1). We control individual preferences and daily patterns by introducing a fixed effect by user (αi) and day (δt). Regression was only performed for those individuals that have at least one FFO and non-FFO visit. To account for potential heterogeneity in our regression, we also cluster errors by day and user. Similarly, for the visits to DMV, we used a simpler logistic regression. Since we typically have only one observation and day per user, the fixed factor αi was substituted by the actual observed fraction of visits to FFO of each individual and we dropped the daily fixed factor. Finally, for the analysis of the different interventions, we have extended the model (1) to the rest of the day by considering each stay within our dataset as a context cit and we evaluate if there is a food visit yit in the next two hours after that stay. A full description of those models is provided in the Supplementary Information Section 7.
Detecting and analyzing change of context
To identify those users that change their context before lunch, we have used a statistical methodology to detect change points in time series (see Supplementary Material Section 5). Using this method, we detect 7913 users in our dataset that changed context during our observation period. To provide a statistically robust estimation of the impact of that change in FFO visits at lunch, we define four groups of users depending on whether their contexts before/after the change have Low (ϕ < 0.13) or High (ϕ > 0.13) ratio of FFO in their contexts. We investigate the FFO ratio of visits of those groups of users that change from Low to High and from High to Low using those that change from Low to Low and from High to High as counterfactuals, respectively. Note that we did not use as counterfactual those users that stayed in the same geographical context, but only those that changed their geographical context. This was done to reduce the possibility of some endogeneity between changing contexts and the food environment in the previous context. To analyze the difference in response to the change, we use Bayesian Structural Time Series to predict how the response would have evolved after the change to a different context if the change had never happened61. Further details about this methodology can be found in the Supplementary Information 5.
Interventions
To investigate the effect of an intervention strategy in an area Ω we evaluate the change in the probability P(yit = 1) for each action with context cit in that area using the extension of the model in Eq. (1) for the full day with and without the intervention I. The total increase in the number of non-FFO visits can be approximated by the derivative of the model in Eq. (1):
where we are assuming that we only change the food environment ϕ in the intervention I. This expression is evaluated for a fixed amount (100) of areas Ω chosen by different criteria in each intervention strategy. See Supplementary Information Section 11 and 9 for further details about how interventions are defined and how we evaluate their impact on FFO visits.
Topic analysis of the areas
To identify the type of areas where the most efficient interventions happen, we use topic modeling to describe the different groups (topics) of POIs in each category that co-occur in the 18896 census tracts in our urban areas. Using Latent Dirichlet allocation (LDA) we found 20 groups of POIs and analyzed their composition. The topics are easily recognizable (see Supplementary Figure S7) and we manually annotated them as “Airports”, “Malls”, “Office”, etc. Each census tract can be described then by the frequency of each of the 20 topics within it. Further details about this methodology can be found in the Supplementary Information Section 8.
Data Availability
All data produced in the present study is property of Cuebiq (https://www.cuebiq.com), and if required, should be asked to Cuebiq or Spectus (https://spectus.ai)
Data availability
Mobility data are available from Cuebiq, available upon request submitted to https://www.cuebiq.com/about/data-for-good/. Other data used come from the American Community Survey (5y) from the Census62 or the PLACES Local data for Better Health from the CDC63, which is publicly available on their websites.
0.1 Author Contributions
B. G-B., A. L. H., K. de la H., and E. M. designed research; B. G-B. and E. M. performed research and analyzed the results. B. G-B. and E. M. wrote the first draft of the manuscript. All authors discussed results and edited the manuscript. All authors approved the final version.
Acknowledgments
E.M. acknowledges support by Ministerio de Ciencia e Innovación/Agencia Española de Investigación (MCIN/AEI/10.13039/501100011033) through grant PID2019-106811GB-C32. K. de la H. acknowledges support from the University of Southern California Keck School of Medicine Dean’s Pilot Funding Program. B.M.B. is supported by the Yale Cancer Prevention and Control Training Program, funded by the National Cancer Institute (T32 CA250803). A.L.H. acknowledges support from an NIH Ruth L Kirschstein National 591 Research Service Award (NRSA) Institutional Training Grant (T32 5T32CA009492-35). The funders had no role in the study design, data collection, and analysis, decision to publish, or preparation of the manuscript.
