Businesses in high-income zip codes often saw sharper visit reductions during the COVID-19 pandemic

The study investigates how COVID-19 and related non-pharmaceutical interventions altered visits to businesses located in neighborhoods differing by median income. Motivated by dramatic, policy-induced changes to urban mobility beginning March 2020 in Minnesota, the authors ask: did businesses in high-income zip codes experience sharper declines in visits relative to similar businesses in low-income zip codes, and did these patterns differ across lockdown and reopening periods? Using pre-pandemic behavior as a benchmark, the paper examines heterogeneous visitation responses by business type and neighborhood affluence. Conceptually, changes in visitation reflect a mix of factors including income shocks, risk aversion to infection, ability to work from home, and policy mandates. The authors posit that if affluent areas saw larger visit reductions for certain categories while incomes and preferences likely remained relatively stable, higher infection-risk aversion and substitution to alternatives (e.g., delivery, outdoor dining) may be driving the differences. The work aims to document these disparities and inform targeted relief and policy design.

Prior research documents aggregate declines in mobility and visits after COVID-19 onset (e.g., Gao et al., Lee et al., Sevtsuk et al., Weill et al.). Many studies benchmarked early 2020, potentially confounding seasonal effects; this paper instead uses same-period 2019 visits as a counterfactual to account for seasonality. Literature shows higher-income areas increased days at home more than lower-income areas (Jay et al.; Ruiz-Euler et al.), partly due to greater work-from-home capacity (Dingel & Neiman) and differing risk perceptions (Bundorf et al.; Erchick et al.). Prior work on policy impacts suggests shutting bars/restaurants and closing non-essential businesses strongly affects disease spread and mobility (Wellenius et al.). Studies of broader human mobility patterns show strong regularity and links to socioeconomic characteristics (González et al.; Lu et al.). Research gaps remain in disaggregating impacts by business type within rich vs. poor areas, over extended time, and controlling for seasonality. This paper contributes a novel pairwise-comparison approach holding business type constant while contrasting high- vs. low-income zip codes across multiple pandemic phases using 2019 parallel periods as baseline.

Setting and periods: Minnesota, USA. Five periods aligned with state policy phases: (1) Pre-pandemic baseline: 02/03/2020–03/16/2020; (2) Lockdown 1: 03/16/2020–06/01/2020; (3) Reopening/Interim 1: 06/02/2020–11/23/2020; (4) Lockdown 2: 11/24/2020–01/11/2021; (5) Reopening/Interim 2: 01/11/2021–05/31/2021. Each period is compared to the corresponding weeks in 2019 to account for seasonality. Data: SafeGraph COVID-19 Data Consortium. Weekly Patterns provides weekly visit counts to each POI; Core Places supplies POI attributes (NAICS code, location, zip, etc.). Visits are counted when a device GPS location falls within a POI polygon. Data cover about 45 million anonymized devices and ~3.6 million POIs in the U.S. POI records were merged by safegraph_id. Zip code-level demographics (median income, population) are from the 2015–2019 ACS 5-year estimates. Income classification: Zip codes ranked by median household income; top one-third labeled high-income (H), bottom one-third low-income (L). Sensitivity checks with one-fourth and two-fifths thresholds yielded similar results. Constructing measures: For business category i in zip code j during period t, total weekly visits across all businesses in category i are summed, V_ijt. The counterfactual is the same weeks in 2019, V_ijtpre. Visits are divided by the zip code population to make changes comparable across zip codes. The identifying assumption is that, absent the pandemic, V_ijt would resemble V_ijtpre for the same seasonal window. Pairwise comparison approach: With j = 1,...,574 zip codes (287 low-income; the remainder high-income), all possible H×L pairings are compared, yielding 82,369 pairwise differences per category-period. For each pair (h ∈ H, l ∈ L), compute Δ_ih = (V_iht − V_ihtpre) and Δ_il = (V_ilt − V_iltpre), then take (Δ_ih − Δ_il). A negative value indicates a sharper reduction in high-income zip codes relative to low-income for that business category and period. Distributions of these differences are summarized via kernel density estimates (KDE) and by reporting the share of comparisons with (Δ_ih − Δ_il) < 0 (Table 2). This simulation-based approach avoids strong parallel-trends assumptions required for difference-in-differences, which may be violated due to differing risk aversion, occupation mix, and policy responses across income groups. Business categories: Twenty-one NAICS-based categories, including full- and limited-service restaurants, religious organizations, movie theaters, hospitals and clinics, pharmacies, groceries, gasoline stations, salons, gyms, libraries, etc. Additional notes: The analysis focuses on visits as a proxy for demand; spending, mode of travel, and individual-level demographics are not observed. KDE plots illustrate distributional shifts across periods for selected categories (e.g., full-service restaurants and groceries).

• Businesses in affluent zip codes, especially those encouraging extended indoor stays, experienced sharper visit reductions relative to low-income areas outside strict lockdowns. Categories include full-service restaurants, religious organizations, and movie theaters.
• In low-income zip codes, the drop in visits during lockdowns was more pronounced across many business types; blue cells in Table 2 concentrate in lockdown periods, indicating larger reductions in poorer areas during those phases.
• Essential services (e.g., supermarkets/grocery stores, gasoline stations, and many medical services) showed broadly similar reductions between rich and poor zip codes; KDE for groceries remained centered near zero across periods, indicating comparable changes.
• Dispersion in visit changes increased during the pandemic within both income groups, suggesting heterogeneous responses beyond median income (e.g., differences in remote-work capacity, age composition, local perceptions, or inequality).
• Selected Table 2 statistics (share of pairwise comparisons where high-income zip codes saw larger reductions than low-income, i.e., (Δ_ih−Δ_il)<0): • Full-service restaurants: 74.5% (Lockdown 1), 66.7% (Interim 1), 68.3% (Lockdown 2), 58.9% (Interim 2). • Religious organizations: 63.2% (Lockdown 1), 62.3% (Interim 1), 60.1% (Lockdown 2), 57.3% (Interim 2). • Movie theaters: 58.8% (Lockdown 1), 66.0% (Interim 1), 56.1% (Lockdown 2), 64.9% (Interim 2). • Supermarkets & grocery stores: distributions centered near zero overall; shares hover near parity pre-lockdown and modestly above 50% in later periods (e.g., 60.4% Interim 1, 58.5% Lockdown 2). • Nursing care & child day care: 58.8% (Interim 1) and 63.8% (Interim 2).
• KDE for full-service restaurants shifts left of zero during lockdowns and remains left through reopening phases, indicating persistently larger visit reductions in affluent areas for extended indoor dining.

The findings directly address the research question: relative to pre-pandemic baselines, affluent areas reduced visits more to businesses involving prolonged indoor exposure, while poorer areas curtailed visits more during lockdowns across many categories. These patterns align with mechanisms such as higher infection-risk aversion and greater capacity to substitute (e.g., delivery, outdoor dining, working from home) in affluent areas, and stronger constraints on low-income populations during mandated lockdowns. Essential services saw broadly similar changes across income groups, pointing to universal necessity overriding income-driven behavioral differences. Increased variance in visit changes within income groups highlights additional determinants (remote-work share, demographics, local perceptions, inequality) influencing mobility beyond median income. The results are relevant for designing targeted relief and nuanced non-pharmaceutical interventions that balance disease control with minimizing economic and social disruption.

This study contributes a seasonality-adjusted, pairwise-comparison methodology to quantify how business visits changed across high- and low-income zip codes through multiple pandemic phases in Minnesota. It documents that businesses in affluent areas—particularly those requiring longer indoor stays—sustained larger, persistent visit declines outside strict lockdowns, whereas poorer areas exhibited greater visit reductions during lockdowns across many categories. Policy implications include targeting relief to affected business types and neighborhoods (e.g., extended indoor venues in affluent areas) and considering income-specific vulnerabilities during lockdowns. Future research should link visits to spending and revenue to assess economic impact; incorporate job composition, remote-work capacity, transportation modes, and demographic structure; examine individual-level mobility disparities; and extend the analysis to other states with different policy regimes.

• Visits, not expenditures: the study lacks granular spending or revenue data; visit-sales correlations may have changed due to shifts to online ordering and delivery.
• No mode-of-transport data; individual-level identities are unobserved; analyses are at the zip code level.
• Essential worker classifications and other externally defined job categories may confound observed patterns.
• Delivery and pickup may distort visit counts (e.g., one driver picking up multiple orders in a single visit); brief-duration visits do not fully distinguish delivery vs. pickup.
• Causal mechanisms (income shock vs. risk preferences) cannot be cleanly disentangled; DID was avoided due to likely violations of parallel trends from differential risk aversion and overlapping policies.
• Income groups defined by pre-pandemic median income rank; assumes relative rankings remained stable during the pandemic.