Effect of Mobile Food Environments on Fast Food Visits

The study investigates how exposure to mobile (dynamic) food environments—encountered as people move through cities—affects visits to fast food outlets (FFO). Prior research has largely examined static, residential food environments (e.g., food deserts and food swamps) and produced mixed or null associations with diet quality and health, often using cross-sectional designs. Yet a substantial share of food acquisition and consumption occurs away from home, suggesting that mobile exposure may be a more relevant driver of behavior. The authors aim to quantify the relationship between mobile food environment exposure and fast food visits using large-scale mobility data, to assess heterogeneity across sociodemographic groups, to explore semi-causal evidence via natural experiments, and to simulate policy interventions that optimally reduce FFO visits by targeting where and when people are most influenced by their environments.

The paper reviews evidence that static neighborhood food environment measures (food deserts/swamps) have inconsistent links to unhealthy eating and obesity, with many null findings and limited causal inference. Policy interventions focused on neighborhood environments (e.g., Healthy Food Financing Initiative investments and the Los Angeles fast-food zoning restrictions) have generally shown little impact on diet or health outcomes. Emerging small-scale studies use GPS and tracking to map exposure and food acquisition but lack the scale and duration to capture habitual patterns or robust effects. Literature also shows that mobility and time constraints shape eating behavior, and that fast food intake often varies more by education and race/ethnicity than by income. These gaps motivate analyzing mobile food environments and actual visitation behavior at population scale with designs that approximate causal inference.

Data: Anonymized mobility data from Spectus for 1.86 million users in 11 US metropolitan areas over 6 months (Oct 2016–Mar 2017). Points of interest (POI) from Foursquare identified food outlets (FO) and fast food outlets (FFO) based on taxonomy and chain-name lists. Demographics from ACS 5-year estimates; health outcomes from CDC PLACES; food deserts from USDA Food Environment Atlas. Definitions: Food environment at location x is φ(x), the ratio of FFO to FO within a 1 km radius (robustness checks include alternative definitions such as nearest 25 FO). Mobile exposure φ^m is the time-weighted average of φ(x_t) across all user stops (>5 minutes) over the study period. Home environment φ^h is φ around the home location. The action y_it=1 if user i selects FFO at time t during a food visit; otherwise y_it=0. Analyses often focus on lunch (11:30–14:00), the peak period for FO/FFO visits. Models: (1) Logistic regression with user and day fixed effects to estimate effect of contextual exposure on FFO choice during lunch: Pr(y_it=1)=logit^{-1}[β0 + α_i + δ_t + β φ(c_t)], where c_t is the last observed morning location (context) before lunch, requiring the context area to contain both FFO and non-FFO. Errors clustered by user and day. Heterogeneity assessed across time, weekdays/weekends, distance, income, and metro area. Semi-causal analyses: (a) Change-point analysis identified users (≈0.46%, ~8.5k) who changed their habitual pre-lunch context from low to high or high to low FFO exposure; Bayesian structural time-series models compared these to counterfactual groups who changed context but maintained similar exposure (Low→Low, High→High). (b) DMV natural experiment: logistic regression analogous to Eq.(1) restricting contexts to Department of Motor Vehicles locations visited, arguing DMV choice is constrained by appointments and distance rather than food environments; individual preference α_i proxied by user’s overall FFO visit fraction μ_i. Policy simulations: Extended the logistic model beyond lunch to all times, treating each stay as a potential context and assessing food decisions within two hours. Derived the marginal effect of changing φ in area Ω on FFO visits: Δ^φ(Ω)=Σ_i [e^{X_i}/(1+e^{X_i})^2] (δφ/δl), where X_i=β0+δ_i+α_i+β φ(c). Compared four strategies under a fixed budget to change 100 outlets: (1) Food Swamp (areas with highest φ), (2) Low Food Access (highest φ within USDA food deserts), (3) Food Hotspots (areas with most FO decisions), (4) Behavior-Environment (areas maximizing Δ^φ(Ω), accounting for exposure, decision volume, and susceptibility). Topic modeling (LDA) characterized area POI profiles associated with selected interventions.

• Mobility and distance: Median distance from home to any place visited is 7.83 km (IQR 2.47–18.63); to any FO 6.94 km (IQR 2.30–17.23); to FFO 6.74 km (IQR 2.50–16.62); to supermarkets 3.1 km (IQR 1.35–8.22). Only 6.8% of FFO visits occur within the user’s home census tract.
• Exposure metrics: Median mobile exposure φ^m=14.0% (IQR 9.7–19.0); median home φ^h=8.2% (IQR 0–17.5). Correlation between mobile and home exposures is low, ρ(φ^m, φ^h)=0.213±0.001 (≤0.29 across demographic strata), indicating distinct mobile vs. home environments.
• Sociodemographic associations: Higher φ^m is associated with living in areas with higher proportions of Black residents, long commutes, and low-skill jobs, and lower proportions with higher education and public transport use; neighborhood income not significantly associated. Predictive power higher for φ^m (R^2=0.213) than φ^h (R^2=0.038).
• Fast food visiting behavior: Median individual fraction of FFO among FO visits μ=0.133 (IQR 0.025–0.273); 22.9% of users never visited an FFO in six months. Demographic associations with μ mirror exposure patterns but have low explanatory power (R^2=0.052), suggesting many groups visit FFO similarly. FFO visits peak at lunchtime daily.
• Exposure–behavior association: Correlation ρ(φ^m, μ)=0.268±0.001 vs. weaker correlation with home environment (ρ≈0.068), underscoring mobile exposure’s relevance.
• Logistic model effect size: β=1.87±0.033 for φ(c_t) at lunch implies a 10% increase in contextual FFO exposure increases odds of choosing FFO by about 20%. Effects are consistent across weekdays/weekends, times of day, income groups, and metro areas, and attenuate with greater distance between context and lunch choice but remain positive and significant for most observations.
• Semi-causal change-of-context: Users shifting from Low→High exposure increased their FFO visit fraction relative to counterfactual (sustained effect up to 50 days); High→Low decreased accordingly. Cumulatively, Low→High yielded about 4 more FFO visits over 50 days; High→Low about 4 fewer, compared to users with similar-exposure changes.
• DMV natural experiment: Significant contextual effect with β=1.09±0.20, corroborating that exogenous contexts influence FFO choices.
• Policy simulations: Behavior-Environment strategy is 1.93× to 3.85× more efficient at reducing FFO visits than Food Swamp, Low Food Access, or Food Hotspots strategies. Changing 100 outlets (~0.22% of total) under Behavior-Environment could avert ~0.56% of FFO visits (~719k over six months) vs. ≤0.29% (~442k) for other strategies. Effects are broadly independent of income, health conditions, or specific city. Targeted areas under the efficient strategy are enriched for POI topics such as Malls, Industry/Factory, Airports, and Offices.

Findings demonstrate that mobile food environments, rather than static home neighborhoods, are stronger determinants of fast food visitation. Individuals frequently make food decisions far from home, and exposure to higher FFO availability in the immediate pre-meal context significantly increases the likelihood of choosing fast food, with consistent effects across demographic groups and cities. Semi-causal analyses—context changes and the DMV natural experiment—support a causal influence of contextual exposure on choice beyond stable individual preferences. Policy simulations show that incorporating behavioral susceptibility and decision volumes into targeting yields substantially greater reductions in FFO visits than traditional geography-only strategies focusing on home neighborhoods or high-FFO areas. These results suggest a paradigm shift for interventions: prioritize mobile contexts where decisions cluster and where people are most influenced or constrained by available options (e.g., work, travel, and shopping areas). The approach can guide ‘smart city’ interventions that integrate mobility-informed exposure and behavior to improve dietary health.

This work introduces a population-scale, behaviorally informed framework linking mobile food environment exposure to fast food visitation. Using large-scale mobility data, the study shows that a 10% increase in contextual FFO exposure raises FFO visitation odds by ~20%, that changes in habitual contexts shift FFO behavior in expected directions, and that targeting interventions using both environmental context and behavioral susceptibility can be 2–4 times more effective than traditional home-based strategies. Future research should: (1) evaluate the nutritional quality and menu healthfulness within FFO and non-FFO options encountered in mobile contexts; (2) conduct prospective experiments and interventions to strengthen causal evidence; (3) extend analyses to other outlet types (full-service restaurants, grocery stores) and post-pandemic mobility/delivery landscapes; and (4) integrate equity considerations to ensure fair distribution of benefits across communities disproportionately exposed to low-quality environments.

• Outcome proxy: Visits >5 minutes to FFO are used as a proxy for purchases and intake; brief visits (e.g., drive-thrus) may be missed. Attribution to the nearest POI may be imperfect in dense/multi-story venues (e.g., malls).
• Nutritional detail: The study does not assess the healthfulness of menus within FFO visited; FFO can offer heterogeneous nutritional options.
• Causal inference: Although semi-causal designs (context change, DMV) reduce confounding, unobserved factors may remain; randomized interventions would strengthen causal claims.
• Temporal context: Data are from 2016–2017 and may not capture post-pandemic shifts, increased home time, or expansion of delivery platforms, which could alter exposures and choices.
• Generalizability and measurement: Mobility sample and POI definitions, stay detection thresholds, and environment radius were tested for robustness, but residual biases in representativeness and measurement may persist.