Causal spillover effects of electric vehicle charging station placement on local businesses: a staggered adoption study

Abstract

Understanding the economic impacts of the placement of electric vehicle charging stations (EVCSs) is crucial for planning infrastructure systems that benefit the broader community. Theoretical models have been used to predict human behavior during charging events, however, these models have often neglected the complexity of trip patterns, and have underestimated the real-world impacts of such infrastructure on the local economy. In this paper, we design a quasi-experiment using mobile phone GPS location and EVCS deployment history data to analyze the causal impact of EVCS placement on visitation patterns to businesses. More specifically, we leverage the staggered placement of EVCSs in New York City and California Bay Area to match treated and control businesses that share similar characteristics including the business sector, location, and pre-treatment visitation count. By comparing three alternative matching strategies, we show that staggered adoption avoids selecting controls from non-treated clusters, and yields greater spatial overlap in dense urban areas. We find that EVCS installations significantly increase customer traffic, with effects concentrated in recreational venues in New York City and routine destinations such as groceries, pharmacies, and cafes in California Bay Area. Our results suggest that the economic spillovers of EVCSs vary across urban contexts and highlight the effectiveness of leveraging the staggered nature of adoption timings for evaluating infrastructure impacts in heterogeneous urban environments.

Figures (4)

• Figure 1: Staggered EV charging station deployment enables POI matching for robust causal inference. a) Selection criteria for control and treatment POIs, where POIs i and j are situated within the same borough and business category, with similar pre-visitation count. These POIs form a structured comparison set, each impacted by an EVCS at different times. The staggered adoption strategy ensures temporal alignment in treatment, with minimum of three months between successive EVCS installations. The green line denotes POI j, acting as a control for treated POI i, subsequently impacted by an EVCS in a later date. b) Covariate balance between treatment and control groups, derived using the staggered adoption approach and non-staggered adoption approach in Manhattan New York City (NYC)(left) and San Francisco (right). In the non-staggered approach, covariates are imbalanced when zooming into the downtown areas, with imbalanced covariates highlighted by black borders. This imbalance is not observed in staggered matching pairs, where covariates are balanced across both urban settings. Spatial and structural properties of matched POI pairs, categorized by treatment type and distance proximity, obtained using the staggered adoption and non-staggered matching approaches, are shown for two different urban settings: Manhattan, NYC (c), and San Francisco (d). The yellow points represent treated POIs, the green points indicate control POIs, and the blue points (in the staggered approach) denote POIs initially treated but later used as controls in subsequent time periods. The Moran’s I statistics confirm greater spatial mixing between treated and control POIs under the staggered approach (Manhattan, NYC: 0.183; San Francisco: 0.156) compared to the non-staggered approach (Manhattan, NYC: 0.713; San Francisco: 0.719; all p-values < 0.001). Results suggest that staggered POI matching effectively enhances spatial overlap between treated and control groups, especially in high-density EVCS settings where non-staggered matching often produces spatial segregation. Maps were produced in Python using the TIGER shapefiles from the U.S. Census BureauUnitedStatesCensus2022.
• Figure 2: Temporal variations in treatment effects and placebo tests for the impact of EV charger installations. Time variation of the treatment effects on customer counts at surrounding POIs analyzed through event studies in NYC (a) and California Bay Area (b), with the center of shaded error bands representing point estimates and error bands representing the 95% confidence intervals. The red vertical line represents the baseline period, which is one month before the treatment. Effects were statistically insignificant prior to EVCS installation across both urban settings, supporting the parallel trend assumption of the DID strategy, but turned significantly positive afterward, highlighting EVCS installations increase visits to neighboring businesses. c) The placebo test for the staggered adoption approach is performed by randomly shuffling the opening dates of EVCSs and assigning a hypothetical placebo opening time point for each EVCS. d) Difference-in-Differences (DID) estimates for the impact of installing a single EVCS on the increase in customer counts at surrounding POIs in NYC (left) and California Bay Area (right). The significant DID estimates from the staggered adoption approach (purple bar line) highlight its robustness in capturing the causal impacts of EVCS installations, particularly in EVCS dense areas of NYC, where both non-staggered matching (orange bar line) and propensity score matching (green bar line) fail to address biases arising from limited spatial overlap. In California Bay Area, an area with moderate urban density and high EV adoption rates, both staggered and non-staggered approaches yield positive results, supporting the robustness of the staggered adoption method across varying urban densities. The placebo test results (red bar line) demonstrate no significant effects, reinforcing the validity of the parallel trends assumption followed in the staggered matching approach. Error bars represent 95% confidence intervals.
• Figure 3: Heterogeneous causal effects of EVCS placement across POI types. These figures illustrate the causal impacts of EVCS installations on customer counts in nearby businesses, stratified by POI category, in NYC (a) and California Bay Area (b). In NYC (a), treatment effects (purple bar lines) are largest for Shopping Centers $(\beta = 117.83,\ \text{p-value} < 0.001)$, Buildings and Entertainment Centers $(\beta = 85.73, \text{p-value} < 0.001)$, Leisure $(\beta = 67.76,\ \text{p-value} < 0.001)$, Hotels and Casinos $(\beta = 45.81,\ \text{p-value} < 0.05)$ and Dining $(\beta = 14.94,\ \text{p-value} < 0.05)$, suggesting that newly attracted customers are primarily recreational or exploratory visitors rather than regular patrons. Placebo tests (red bar lines) show significant effects for Shopping Centers $(\beta = 179.11, p < 0.05)$ and Hotels and Casinos $(\beta = 51.79,\ \text{p-value} < 0.05)$, indicating potential data imbalances. b) In contrast, California Bay Area (b) exhibits distinct behavioral preferences among EV drivers, with stronger effects for Groceries $(\beta = 9.94,\ \text{p-value} < 0.05)$, Dining $(\beta = 7.70,\ \text{p-value} < 0.001)$, Apparel $(\beta = 7.32,\ \text{p-value} < 0.05)$, Shops and Services $(\beta = 4.90,\ \text{p-value} < 0.001)$, Medical and Health $(\beta = 4.33,\ \text{p-value} < 0.05)$, and Beauty and Spa $(\beta = 1.77,\ \text{p-value} < 0.05)$, reflecting the city’s moderate urban density and different charging use contexts compared to NYC’s dominance of recreational POIs. In both plots, error bars represent 95% confidence intervals.
• Figure 4: Spatial variations in treatment effects and heterogeneous causal effects across clientele. a) Variations in the treatment effects of EVCS installations on customer counts based on distance between the POIs and nearby EVCSs, in NYC (left) and California Bay Area (right). The effects (purple bar lines) estimated using the staggered adoption approach are more pronounced for POIs within 100 meters of an EVCS in NYC, where customer counts increase by $\sim 39$, whereas in California Bay Area, POIs located 100--200 meters from an EVCS attract more customers, with counts increasing by $\sim 10$. Compared to non-staggered matching approaches and the placebo test, distance-varying treatment effects reveal bias in high density EVCS settings in NYC, whereas in California Bay Area, with a moderate urban density, the results are consistent across approaches. Thus, these results provide nuanced insights into how we can generate more reliable treated and control POI pairs using the staggered adoption approach, further emphasizing its robustness across spatial categories. b) The heterogeneous effect of an additional EVCS on customer counts among different income groups at surrounding businesses, in NYC (left) and California Bay Area (right). We find that EVCS installations attract customers from mid high-income households $(\$100-150k, \beta = 42.43, \text{p-value} < 0.001)$ most strongly in NYC, whereas in California Bay Area, businesses serving high income households $(>\$150k, \beta = 10.02, \text{p-value} < 0.001)$ experience the largest positive effects. Results suggest that businesses serving higher-income customers experience greater impacts, aligning with the assumption that EV drivers are more likely to belong to higher-income groups. Insignificant results from the non-staggered matching approaches and placebo test across income segments in NYC further confirm the integrity of the staggered adoption approach, highlighting the risk of bias due to limited spatial overlap in high-density EVCS settings. In contrast, California Bay Area, where urban density is moderate, non-staggered approaches yield consistent and significant effects for higher income segments, illustrating that the reliability of non-staggered methods depends on spatial context. In both plots, error bars represent 95% confidence intervals.