Do Streetscapes Still Matter for Customer Ratings of Eating and Drinking Establishments in Car-Dependent Cities?

Chaeyeon Han; Seung Jae Lieu; Subhrajit Guhathakurta; Uijeong Hwang

REVIEW 4 major objections 3 minor 7 references

Do Streetscapes Still Matter for Customer Ratings of Eating and Drinking Establishments in Car-Dependent Cities?

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that the positive relationship between perceived streetscape safety and Yelp ratings for eating and drinking establishments weakens as the neighborhood around the establishment becomes more car-dependent, and at high car…

desk verdict A useful moderation hypothesis in an otherwise solid empirical paper, but the safety odds ratio is misreported and the car-dependent reversal is left undiscussed. read the letter →

arxiv 2508.06513 v1 pith:624O7NNH submitted 2025-07-29 physics.soc-ph cs.CYcs.LG

Chaeyeon Han , Seung Jae Lieu , Uijeong Hwang , Subhrajit Guhathakurta This is my paper · ORCID

classification physics.soc-phcs.CYcs.LG

keywords WalkabilityPerceivedSafetyCarDependencyServicescapePoint-of-InterestsComputerVisionYelpRatingsOrdinalLogisticRegression

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tests whether the visual quality of the street around a restaurant, bar, or café still shapes customer ratings when people mostly drive there. Using 744 Yelp-listed eating and drinking establishments in Washington, DC, it pairs user review photos with Google Street View imagery, scores both with computer vision models, and models Yelp rating categories with ordinal logistic regression. It finds that indoor aesthetics and perceived street safety are both positively linked to ratings, but the safety effect shrinks as a neighborhood-level Car Dependency Index rises; at high car dependency the effect turns negative. This matters because it suggests that the value of streetscape investment depends on the dominant travel mode of the area, not on the streetscape alone.

What carries the argument

The load-bearing object is the Car Dependency Index (CDI), a neighborhood score for each establishment computed as 0.5 times the normalized commuting car modal share plus 0.5 times the inverse of the normalized sum of population and employment density across block groups within a 1-km buffer. The decisive term is the interaction Perceived Safety × CDI in an ordinal logistic regression of the five-category Yelp rating; this term carries the paper's claim that the safety effect depends on how car-oriented the surrounding area is. Two computer vision pipelines supply the perceptual inputs: one scores interior visual appeal from Yelp review photos using a model trained on aesthetic ratings, and one scores perceived safety from Google Street View images using a model trained on crowdsourced safety judgments.

What would settle it

Observe actual arrival modes at the sampled establishments, for example from mobile-location traces or parking utilization, and re-run the ordinal logistic model with that real car-access measure replacing the neighborhood Car Dependency Index; the central moderation claim would be falsified if the safety interaction is not negative when actual car access is used.

Watch

Extended reading notes

Core claim

The central discovery is a statistical moderation: the positive association between the perceived safety of the surrounding streetscape and an eating establishment's rating score weakens as the car dependency of its neighborhood increases. In the main ordinal logistic model, the Perceived Safety coefficient is 4.468 (p<0.001) and the interaction with the Car Dependency Index is -0.068 (p<0.001), so the safety advantage declines by about 0.068 log-odds per unit of car dependency. The paper also reports positive associations for interior perceived aesthetics and for WalkScore, showing that both indoor and outdoor visual qualities predict higher rating categories, with the outdoor effect conditional on car dependency.

Load-bearing premise

The Car Dependency Index, a neighborhood score built from commuting mode share, population density, and employment density, is assumed to reflect whether customers actually reach the establishment by car, even though it ignores the visitor's origin, trip distance, and actual mode of arrival.

Editorial extensions

If this is right

In walkable parts of a city, a one-point higher WalkScore is associated with a 3% higher odds of being in a better rating category, so pedestrian-friendly location itself appears to carry a satisfaction premium.
Perceived indoor aesthetics consistently predict higher ratings, so interior design quality matters across all car-dependency contexts.
Perceived street safety raises the odds of a higher rating category by a large factor at low car dependency, but this advantage shrinks as CDI rises, meaning the same streetscape improvement will not produce the same rating gain in a walkable district and a car-oriented one.
In highly car-dependent areas, the total association between perceived safety and ratings becomes negative, implying that safe streets alone will not raise customer satisfaction there.
For planning practice, the paper implies that street-level interventions should be calibrated to the local transportation context rather than applied uniformly across a city.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A consequence the authors do not state is the crossover point: solving $4.468 - 0.068 \times \text{CDI} = 0$ gives CDI $\approx 65.7$, so perceived safety's total effect on ratings turns negative above roughly that index value.
The same logic would predict that other destination types, such as shops, services, and offices, also show a diminished streetscape premium in car-dependent neighborhoods; this is testable with the same Yelp and street-view pipeline applied to other point-of-interest categories.
The moderation could reflect selection rather than perception: drivers sort into car-oriented destinations, so their satisfaction is less tied to the walking environment; using the foot-traffic origin data to construct actual arrival-mode shares would separate these mechanisms.
In practice, this suggests that street investments in car-oriented districts may need to be paired with parking or transit-access changes before they move customer ratings, while in walkable districts the streetscape itself is the lever.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The useful thing here is the interaction between car dependency and perceived street safety, which I don't think is in the cited literature. The authors test it with an ordinal logistic on 744 Yelp-scraped EDEs in DC, using computer vision for street safety and indoor aesthetics. That's a legitimate extension, and the data work looks honest: they report a Brant test for proportional odds, VIFs, and the standard errors don't look fishy. If the interaction holds up, it gives planners context-sensitive guidance about when streetscape investment matters for local business ratings.

The problems are mainly in the reporting and interpretation. The Highlights say a 0.1-point increase in safety raises the likelihood of a higher rating by 8.7 times, but the Table 3 coefficient is 4.468, and exp(4.468*0.1) is about 1.56, not 8.7. They also describe safety as ranging from 0 to 1 while Table 2 shows mean 6.0, SD 0.3, range 4.7–6.6. That's not a harmless typo; it's the headline effect.

More substantively, the model itself implies that above a Car Dependency Index of roughly 66 (from 4.468/0.068), the total effect of safety on ratings is negative. The paper never acknowledges this; instead the Discussion says streetscape quality remains important in car-dependent areas, which is hard to reconcile with a negative effect at high CDI. This needs a direct address, because it could mean the interaction is too strong or the model is picking up something else.

Third, the CDI is a neighborhood proxy built from car modal share and density, and the authors concede in the limitations that it may not capture whether visitors actually drive. That's load-bearing for the moderation claim. I agree with the reader that this is the weakest point, and it deserves more than a sentence in limitations—ideally a robustness check with trip-level or origin-destination data, or at least an acknowledgment that the result is about neighborhood context, not individual travel behavior.

Credit where earned: they test model assumptions, use a validated aesthetic model, control for visitor income and footfall, and are candid about computer vision bias. Self-citation is not a real problem here; the safety model is their own validated tool, and using it is appropriate.

This paper will interest urban planners and design researchers working on streetscapes and business vibrancy. It deserves a serious referee, but I'd send it back for major revision before publication. I would not cite it in its current form, mainly because of the odds-ratio error and the undiscussed sign reversal.

Referee Report

4 major / 3 minor

Summary. The paper asks whether the association between perceived streetscape safety (measured by a computer vision model applied to Google Street View imagery) and Yelp ratings of eating and drinking establishments (EDEs) in Washington, DC, is moderated by neighborhood car dependency. The authors build a Car Dependency Index (CDI) from adjacent census block groups' car modal share and population/employment density, then estimate an ordinal logistic regression with a perceived-safety × CDI interaction. They report that higher car dependency weakens the positive safety–rating association, with the interaction coefficient -0.068 (p < 0.001). The paper also includes indoor aesthetics from Yelp photos, WalkScore, foot traffic, and several controls.

Significance. If the moderation result holds, it would be a useful contribution to servicescape and walkability research by showing that the value of streetscape improvements depends on the transport context, with implications for context-sensitive planning. Strengths include the integration of multiple novel data sources (Yelp review photos, Street View imagery, mobile-device foot traffic), the explicit testing of the proportional odds assumption and multicollinearity, and a candid limitations section. However, the central empirical claim is currently undermined by an arithmetic/scale error in the reported safety effect, an unresolved construct-validity issue with the CDI, and the paper's silence on the sign reversal implied by its own interaction model. These issues are load-bearing and require revision.

major comments (4)

[Section 4 (Street-level effects)] The reported effect of perceived safety is internally inconsistent. The text states that a 0.1-point increase in the perceived safety score (described as ranging from 0 to 1) raises the likelihood of a higher rating category by 8.7 times, but exponentiating the coefficient 4.468 times 0.1 gives about 1.56, not 8.7. In addition, Table 2 reports Perceived Safety with mean 6.0, SD 0.3, and range 4.7–6.6, which contradicts the claimed 0–1 scale and makes a 0.1-point change far smaller than the reported SD. The odds ratio of 87.18 in Table 4 corresponds to a one-unit change on the actual scale (if the coefficient is 4.468 per unit), not a 0.1-point change. Please correct the scale description, the odds-ratio interpretation, and the corresponding statement in the Highlights, and report an easily interpretable effect size (e.g., per SD or per realistic increment).
[Section 3.1.4 and Section 5] The Car Dependency Index is built from neighborhood-level averages (car modal share, population density, employment density of adjacent block groups), and the authors concede that it may not capture whether visitors actually use cars to access the EDE, nor trip origin or trip distance. Since the central claim is that car dependency moderates the safety–rating relationship, this construct-validity limitation is load-bearing. The interaction could be driven by unmeasured neighborhood attributes correlated with CDI, such as parking supply, transit accessibility, or land-use mix, rather than by the actual travel behavior of customers. To support the interpretation, please add a robustness check controlling for plausible neighborhood confounders, or validate CDI against observed visitor travel behavior using the Advan foot-traffic data that the paper already uses for visitor income.
[Section 4 and Section 5] The model's interaction implies a total effect of perceived safety of 4.468 − 0.068 × CDI, which becomes negative when CDI exceeds approximately 65.7. Given that CDI ranges from 1.1 to 99.1 with a mean of 54.6, a substantial share of sampled EDEs fall in the region where the model predicts safety has a negative association with ratings. The paper does not acknowledge or discuss this sign reversal, despite its direct relevance to the policy recommendations about improving streetscape quality in car-dependent areas. Please address this explicitly, for example by plotting marginal effects of safety across the CDI range and discussing whether a negative safety effect is plausible or likely an artifact of extrapolation or confounding.
[Section 3.1.4] The definition of the Car Dependency Index is ambiguous. The text defines PEDi as the sum of the normalized population density and employment density, but the formula CD_i = 0.5 × MS_i + 0.5 × (1 − PED_i), with PED as a sum of two 0–1 variables, can produce negative values and would not map to a 0–100 range after multiplying by 100. The reported sample range of 1.1–99.1 suggests that PED was actually an average of the two densities or that the normalization was different. Because CDI is the key moderating variable, please clarify the exact construction, including how each density was normalized and how the final index was scaled to lie in the reported range.

minor comments (3)

[Section 3.1.2] The variable names are inconsistent between Table 2 and Table 3: Table 2 lists 'Neighborhood Income level (10k)' while Table 3 uses 'Income Level of Neighborhood'; please harmonize the terminology.
[Section 3.1.3] The perceived safety scores from Hwang et al. (2023) are described as TrueSkill scores, but the paper does not explain how these scores were transformed into the reported range of 4.7–6.6. Please provide the transformation or rescaling details.
[Section 5] There is a grammatical error in the final paragraph of the Discussion: 'this study provides suggests targeted interventions' should be 'this study suggests targeted interventions.'

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central interaction is an empirical regression result estimated from independent data; only minor, non-load-bearing self-citation is present.

full rationale

The paper's central claim — that car dependency moderates the perceived-safety effect on Yelp ratings — is an empirical interaction estimated in an ordinal logistic model (Table 3: Perceived Safety × CDI coefficient -0.068, p<0.001). It is not definitionally tied to the outcome. The Car Dependency Index is constructed from ACS car modal share, population density, and employment density in adjacent block groups (Section 3.1.4), the perceived-safety score is produced by a computer vision model trained on the external Place Pulse 2.0 dataset (Section 3.1.3), and ratings are scraped from Yelp. No equation in the paper defines the interaction term as a function of the rating categories or of the fitted parameters, so no prediction reduces by construction to its inputs. The use of Hwang et al. (2023) for safety scoring is a self-citation, but that work supplies a measurement model validated on external data rather than the paper's substantive result; the other self-citations (Koo et al., 2023; Han et al., 2025; Lieu & Guhathakurta, 2025) are contextual and not load-bearing. The Discussion's explicit limitation that the Car Dependency Index 'may not fully capture whether visitors actually use cars to access EDEs' and the apparent inconsistency in the reported safety odds ratio (coefficient 4.468 vs. the '8.7 times per 0.1-point increase' claim) are construct-validity and reporting concerns, not circularity. No uniqueness theorem, ansatz, or renaming step is invoked to force the conclusion.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claim rests on the Car Dependency Index, a hand-constructed composite with no external validation, and on computer vision scores that are treated as proxies for human perception. Several modeling choices (weights, buffers, binning) are free parameters that could affect the results.

free parameters (3)

Car Dependency Index weights = w=0.5 for each of MS and (1-PED)
The composite CDI is defined with hand-chosen weights, giving car modal share a weight of 0.5 and the summed density measures a combined weight of 0.5. The estimates are sensitive to this weighting; if actual car use were measured directly, the interaction could differ.
Buffer distances = 300m for street view, 1km for CDI, 500m for footfall
These distances are chosen by the authors and determine which street segments and block groups are averaged for each EDE. The paper notes that the 300m buffer may not match the streetscape customers actually experience.
Rating category cut points = 1.5-3.0, 3.0-3.5, 3.5-4.0, 4.0-4.5, 4.5-5.0
The ordinal outcome is binned by hand; different binning could alter the proportional odds test and coefficient estimates.

assumptions (5)

domain assumption Car dependency is associated with high rates of car travel and car-oriented land use.
Invoked in Section 3.1.1 to justify constructing the Car Dependency Index from car modal share and densities.
domain assumption Computer vision safety scores approximate human perceived safety.
The street view model is trained on Place Pulse 2.0 crowd votes; the paper acknowledges it is an approximation of perception rather than an objective measure (Section 3.1.3).
domain assumption AADB aesthetic scores capture the visual appeal that matters for customer ratings.
Interior aesthetics are measured by a pre-trained model on general photo aesthetics; validation on restaurant images is reported, but demographic representativeness is limited.
standard math Proportional odds assumption holds for the ordinal regression.
Tested with the Brant test and graphical checks in Section 3.2; the paper reports no significant violation.
domain assumption Yelp rating categories reflect an ordinal satisfaction scale.
The rating is treated as a 5-level ordinal outcome, with thresholds chosen by the authors.

invented entities (1)

Car Dependency Index (CDI)
purpose: Composite measure of neighborhood car dependency for each EDE, using car modal share, population density, and employment density.
The index is constructed by the authors with hand-chosen weights and has no external validation against actual travel behavior; the paper itself concedes it may not reflect whether visitors drive.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Do Streetscapes Still Matter for Customer Ratings of Eating and Drinking Establishments in Car-Dependent Cities?." pith.science (2026). https://pith.science/paper/624O7NNH

@misc{pith2026250806513,
  author       = {Pith},
  title        = {Pith review of: Do Streetscapes Still Matter for Customer Ratings of Eating and Drinking Establishments in Car-Dependent Cities?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/624O7NNH}},
  note         = {Machine review of arXiv:2508.06513}
}

read the original abstract

This study examines how indoor and outdoor aesthetics, streetscapes, and neighborhood features shape customer satisfaction at eating and dining establishments (EDEs) across different urban contexts, varying in car dependency, in Washington, DC. Using review photos and street view images, computer vision models quantified perceived safety and visual appeal. Ordinal logistic regression analyzed their effects on Yelp ratings. Findings reveal that both indoor and outdoor environments significantly impact EDE ratings, while streetscape quality's influence diminishes in car-dependent areas. The study highlights the need for context-sensitive planning that integrates indoor and outdoor factors to enhance customer experiences in diverse settings.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 6 canonical work pages

[1]

servicescape,

Introduction In recent years, digital place rating platforms like Yelp and Google Maps have transformed how people choose eating and drinking establishments like bars, cafes, and restaurants, henceforth called EDEs, by offering more detailed information than word-of-mouth recommendations (Xiang et al., 2015; Book et al. 2018). These platforms provide not ...

work page 2015
[2]

Literature Review 2.1. Place Characteristics and Customer Satisfaction In the business and marketing literature, the physical environment within EDEs is recognized as a critical factor influencing customer satisfaction and patronage. The physical environment of an EDE includes elements such as furniture/seating arrangement and quality, lighting, color sch...

work page 2023
[3]

Washington DC

Methods 3.1 Data 3.1.1 Selection of the Study Area This study aims to explore how different levels of car dependency influence the relationship between the built environment and user satisfaction with urban destinations. To achieve this, Washington, DC, was selected as an ideal case to study given its balanced distribution of car modal share. Data on tran...

work page 2016
[4]

Odds ratios, calculated by exponentiating coefficients, indicate the likelihood of higher category rating with changes in the independent variables (Table 4)

Results Table 3 presents the outcomes of the regression model. Odds ratios, calculated by exponentiating coefficients, indicate the likelihood of higher category rating with changes in the independent variables (Table 4). Also, marginal effects (or marginal percentages), which makes the interpretation of ordinal logistic regression results easier, were ca...

work page 1980
[5]

black box

Discussion This study demonstrates the critical role of both indoor and outdoor environments in shaping customer satisfaction with dining destinations. The strong association between interior aesthetics and customer ratings aligns with the servicescape literature, which emphasizes the impact of ambiance on customer experience (Bitner, 1992; Ryu & Jang, 20...

work page 2024
[6]

Bikelash

Conclusion In conclusion, this study contributes to the existing literature by examining how customer satisfaction with EDEs is influenced by both interior and exterior built environment characteristics across different urban contexts, particularly focusing on the role of car dependency. By integrating data from Yelp reviews and Google Street View images,...

work page arXiv 2022
[795]

Montgomery, J

https://doi.org/10.1108/01443571211250077. Montgomery, J. (1998). Making a city: Urbanity, vitality and urban design. Journal of Urban Design, 3(1), 93–116. https://doi.org/10.1080/13574809808724418. Moudon, A. V ., Lee, C., Cheadle, A. D., Garvin, C., Johnson, D., Schmid, T. L., Weathers, R. D., & Lin, L. (2006). Operational definitions of walkable neigh...

arXiv 1998

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

servicescape,

Introduction In recent years, digital place rating platforms like Yelp and Google Maps have transformed how people choose eating and drinking establishments like bars, cafes, and restaurants, henceforth called EDEs, by offering more detailed information than word-of-mouth recommendations (Xiang et al., 2015; Book et al. 2018). These platforms provide not ...

work page 2015

[2] [2]

Literature Review 2.1. Place Characteristics and Customer Satisfaction In the business and marketing literature, the physical environment within EDEs is recognized as a critical factor influencing customer satisfaction and patronage. The physical environment of an EDE includes elements such as furniture/seating arrangement and quality, lighting, color sch...

work page 2023

[3] [3]

Washington DC

Methods 3.1 Data 3.1.1 Selection of the Study Area This study aims to explore how different levels of car dependency influence the relationship between the built environment and user satisfaction with urban destinations. To achieve this, Washington, DC, was selected as an ideal case to study given its balanced distribution of car modal share. Data on tran...

work page 2016

[4] [4]

Odds ratios, calculated by exponentiating coefficients, indicate the likelihood of higher category rating with changes in the independent variables (Table 4)

Results Table 3 presents the outcomes of the regression model. Odds ratios, calculated by exponentiating coefficients, indicate the likelihood of higher category rating with changes in the independent variables (Table 4). Also, marginal effects (or marginal percentages), which makes the interpretation of ordinal logistic regression results easier, were ca...

work page 1980

[5] [5]

black box

Discussion This study demonstrates the critical role of both indoor and outdoor environments in shaping customer satisfaction with dining destinations. The strong association between interior aesthetics and customer ratings aligns with the servicescape literature, which emphasizes the impact of ambiance on customer experience (Bitner, 1992; Ryu & Jang, 20...

work page 2024

[6] [6]

Bikelash

Conclusion In conclusion, this study contributes to the existing literature by examining how customer satisfaction with EDEs is influenced by both interior and exterior built environment characteristics across different urban contexts, particularly focusing on the role of car dependency. By integrating data from Yelp reviews and Google Street View images,...

work page arXiv 2022

[7] [795]

Montgomery, J

https://doi.org/10.1108/01443571211250077. Montgomery, J. (1998). Making a city: Urbanity, vitality and urban design. Journal of Urban Design, 3(1), 93–116. https://doi.org/10.1080/13574809808724418. Moudon, A. V ., Lee, C., Cheadle, A. D., Garvin, C., Johnson, D., Schmid, T. L., Weathers, R. D., & Lin, L. (2006). Operational definitions of walkable neigh...

arXiv 1998