{"id":"131126c3-78db-43ca-a63b-c142a22da4c3","arxiv_id":"2411.11390","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"School runs reduce the probability of smooth traffic around Beijing schools by 8.34 percentage points, and congestion is higher near school clusters, bus stops, and business-oriented streetscapes.","lead":"This study uses Beijing traffic data and street-view images to show that school drop-off and pick-up trips measurably increase traffic congestion around schools, and that congestion is worse near clusters of schools, bus stops, and commercial areas. It gives city planners a data-driven checklist for siting new schools and redesigning school neighborhoods to reduce rush-hour gridlock.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 8.34pp causal claim is unidentified: it rests on a single Gaokao-week Thursday as the no-school counterfactual, with no control roads to absorb day-of-week or exam-week effects; the marginal-effect contrast may also be misread.","rationale":"The reader's weakest assumption identifies exactly the load-bearing problem: Thursday morning of Gaokao week is the sole no-school counterfactual, and the design lacks control roads to absorb non-school-specific temporal shocks. My reading of the paper confirms this is the central identification weakness for the 8.34pp claim. I also note a secondary interpretational ambiguity in Fig. 5(b) that could mean the causal effect is even smaller than claimed, but the counterfactual problem is sufficient to undermine the headline result as stated. The paper's descriptive findings — the spatial patterns and built-environment correlations — are not invalidated by this concern, and the authors could reasonably address it with a control-road DID or multiple no-school weeks. Therefore, the existing CONDITIONAL verdict remains appropriate: acceptance should require the robustness checks the reader already requested. I do not see a reason to move the verdict, since the paper's broader descriptive contribution and the plausibility of school-run congestion are supported by other evidence in the paper, even if the precise 8.34% causal magnitude is not yet credible.","tokens_in":18649,"tokens_out":5749,"duration_ms":58113,"concrete_test":"Re-estimate the school-run effect with a difference-in-differences design: use roads located more than 1000 m from any school as controls, and compare the change in congestion from Thursday morning (Gaokao, no school) to Monday/Tuesday mornings (school open) between school-adjacent and control roads. If the treatment-control interaction is small, insignificant, or materially different from -8.34pp, the headline effect is confounded by day-of-week or exam-week effects. A secondary check: repeat the analysis using an additional non-school workday outside Gaokao week (e.g., a regular school holiday or another June 2023 Thursday) to test whether the same 8.34pp estimate is reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that school-escorted trips reduce the probability of free-flow by 8.34 percentage points is identified entirely from a within-week comparison: roads within 500 m of schools on Monday/Tuesday mornings (schools open) versus the same roads on Thursday morning of June 8, 2023, during Gaokao week (schools closed). No control roads outside school neighborhoods are used, so any citywide difference between Monday/Tuesday and Thursday — including Gaokao-related traffic management, road closures, exam-center trips, day-of-week travel patterns, or transient weather/events — is attributed to school runs. Because schools are closed all week and Gaokao itself generates atypical traffic, Thursday morning is not a clean 'no-school workday' counterfactual. The paper's identification assumption is thus untested and likely violated. Additionally, Fig. 5(b) is ambiguous: the text says the Monday/Tuesday marginal effect 'explicitly measures the impact of adding school-escorted trips into Thursday morning commute hours,' yet the reported -8.34% is the level for the school+commute scenario, not the -4.73pp difference from the -3.61% no-school commute baseline. Either the magnitude is overstated or the displayed quantities are mislabeled. Without control roads, multiple no-school days, or a correctly defined contrast, the 8.34% point estimate does not support the strength of the causal claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the relationship between school-run trips and traffic congestion around 846 schools in Beijing, using one week (June 5–11, 2023) of hourly congestion indices from Baidu Map, combined with multi-source built-environment data including road topology, POIs, building footprints, mobile signaling, and street-view imagery. The authors estimate a generalized ordered logit model (Model 1) with workday, school-run, and exam-period dummies to quantify the impact of school-escorted trips on the probability of each congestion state, and a linear regression model (Model 2) with SHAP explanations to link built-environment features to school-neighborhood congestion frequency. The headline result is that school runs reduce the probability of 'smooth' traffic by 8.34 percentage points; the paper also reports that congestion is worse near multiple schools, bus stops, and certain business/financial scenescapes, and proposes a built-environment score for school siting and upgrading.","tokens_in":18904,"tokens_out":8793,"duration_ms":88368,"significance":"If the 8.34% estimate were credible, it would provide a spatially explicit, policy-relevant magnitude for school-run congestion and would extend existing work by linking congestion to a richer set of built-environment measures, including street-view scenescapes. The paper assembles a large multi-source dataset and its descriptive school-level model (Model 2) identifies plausible correlates with a moderate fit (adjusted R² = 0.45). However, the causal interpretation of Q1 rests on a fragile single-week, no-control-group comparison, and the reported effect size appears to be misread from the displayed marginal effects. These issues are load-bearing for the paper's central claim, although they are addressable in revision, so the current version does not support the strength of the conclusions.","major_comments":[{"comment":"The paper's headline claim that 'school runs reduce the probability of smooth traffic by 8.34%' is not the contrast shown in Fig. 5(b). The figure reports -8.34% for the combined school-work commute state and -3.61% for the no-school work commute state; the quantity described in the text as 'adding school-escorted trips into Thursday morning commute hours' is the difference between these two states, which is -4.73 percentage points, not -8.34. The Abstract and Section 6 repeat the level as if it were the effect, so either the reported magnitude is overstated or the labels/quantities in Fig. 5(b) are mis-specified. This needs to be corrected and consistently reported.","section":"Section 4.1 (Fig. 5(b)) and Abstract"},{"comment":"The Q1 estimate is identified from a single week (June 5–11, 2023) with one non-school weekday, Thursday morning of Gaokao week, serving as the counterfactual for Monday/Tuesday mornings. Schools are closed for the entire exam week, and Gaokao itself induces atypical traffic management, road closures, and exam-center trips; no control roads outside school neighborhoods are used to absorb citywide day-of-week or exam-week shocks. The text calls this a 'DID setting,' but there is no control group, so the estimate conflates school-run effects with any Thursday-specific or exam-week-specific changes. The authors should add control roads and multiple non-school days, or explicitly downgrade the claim to a descriptive within-week comparison.","section":"Section 3.2.1 and Section 4.1"},{"comment":"The coding of the temporal dummies is not precise enough to reproduce the analysis. Table 3 defines 'work' as 'Monday and Friday' (presumably a typo for Monday–Friday); 'school' is defined only as 'pick-up/drop-off hours on Monday and Tuesday' without listing the exact timestamps that enter the model; and 'exam' excludes Thursday morning but does not state how Thursday afternoon or Saturday (also listed as an exam day) are treated. Since the 8.34% result depends on the exact coding of these dummies, the supplement should provide the full set of indicators, the time slots used, and the complete regression output with standard errors.","section":"Section 3.3.1 and Table 3"},{"comment":"The built-environment score in Eq. (6) is a linear combination of the coefficients βi estimated in Model 2, using the same variables and the same p<0.1 threshold that define significance in that model. The reported R²=0.461 in Fig. 8(a) is therefore an in-sample measure of Model 2's own fit, not an independent validation of the score's predictive ability. The statement that the score 'has the ability to distinguish school neighborhoods with a high risk of congestion' is circular; an out-of-sample or cross-validated assessment is needed before the score is presented as a planning tool.","section":"Section 5.1 and Fig. 8(a)"},{"comment":"The SHAP interaction analysis is presented as evidence that the impact of one scenescape on congestion changes sign depending on the presence of another scenescape, and the policy section draws planning recommendations from this 'contradictory effect.' However, Model 2 is a linear regression without interaction terms, so SHAP interaction values can only reflect correlations among features rather than causal moderation. The claim that converting scenescape 5 to scenescape 1 'will cancel out' the congestion-promoting effect is not supported by the estimated model and should be reframed as a hypothesis, not a model-derived conclusion.","section":"Section 5.2 and Appendix A.1"}],"minor_comments":[{"comment":"The manuscript contains numerous typographical errors, including 'trafffic', 'envionment', 'seperation', 'inplausible', 'buidings', 'occurence', 'convinience', and 'meantime'; a thorough copyedit is needed.","section":"Throughout"},{"comment":"The axis label 'Coefficient of marginal effect (%)' is misleading because the two bars per category represent the combined work+school state and the work-only state; please provide a explicit legend and state which base category each bar refers to.","section":"Fig. 5(b)"},{"comment":"The text says the screening period includes 'four national college entrance examination days (June 7–10; Wednesday–Saturday)' but Table 3's exam variable only includes Wednesday and Friday; please clarify which days and hours are treated as exam periods and why Thursday afternoon and Saturday are excluded.","section":"Section 3.2.1 and Table 3"},{"comment":"The K-means cluster counts (4, 3, 3 for ordinary, main, and express roads) are stated without justification or stability checks, and the scenescape labels in Table 2 are assigned by 'expert knowledge' without any validation; please describe the labeling procedure and any reliability assessment.","section":"Section 3.2.3 and Table 2"},{"comment":"The paper reports p<0.01 for all marginal-effect coefficients in Model 1 but does not provide standard errors, confidence intervals, or information about clustering of road-level observations within school neighborhoods; these should be reported.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's contribution is better framed as a descriptive association study; the causal language on Q1 should be substantially softened. No data or code is provided, which makes the single-week identification hard to audit. The reported 8.34% headline appears to be a misread of Fig. 5(b), so the authors need to reconcile the text with the displayed marginal effects. The scope of the journal may favor a revision that strengthens the identification (e.g., control roads, multiple weeks) or clearly separates causal and descriptive claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, know this: the empirical core is a new combination—Gaokao week as a no-school workday comparison, street-view scenescapes as built-environment predictors—and the descriptive work is volume-solid. But the headline claim that school runs lower the probability of free-flow by 8.34 percentage points is not supported as stated. The contrast in Fig 5(b) is between a work+school morning (-8.34) and a no-school commute morning (-3.61). The school-specific marginal effect is the difference, -4.73pp, not the -8.34 reported. And even that difference is identified from a single Thursday in Gaokao week, with no control roads outside school neighborhoods, so day-of-week and exam-week effects are not absorbed. That is a load-bearing flaw.\n\nWhat is genuinely good: the dataset is large—846 schools, 1.4M congestion records—and the built-environment measurement is more comprehensive than most prior work, with the Places365-derived scenescapes a welcome complement to the standard 5Ds. The generalized ordered logit is appropriate for the ordinal congestion index, and the SHAP analysis gives a useful ranking of features. The policy scoring formula is interesting in principle, though it is an in-sample refit and the reported R² of 0.461 is partly circular.\n\nOther soft spots: Model 2 uses arbitrary binarizations (bus stop >5, population >10,000, building height >6 stories), ignores spatial autocorrelation, and the SHAP interaction effects are described without a reproducible model specification. The paper would be materially stronger with multi-week or multi-year data, control roads, confidence intervals on the marginal effects, and an out-of-sample check of the scoring function.\n\nWho this is for: urban planners and transport geographers who want a rich descriptive map of what built-environment features correlate with school-neighborhood congestion in a large Chinese city. The causal claim should be ignored until the identification is fixed.\n\nRecommendation: this deserves a serious referee, not a desk reject. A competent reviewer can push for the robustness checks and a corrected contrast. If the authors can supply a clean no-school counterfactual and honest uncertainty, the underlying idea is publishable.","headline":"The paper's descriptive analysis of school-neighborhood congestion is solid and the street-view scenescape approach is a real contribution, but the headline 8.34% school-run effect is misidentified and likely misreported; the actual contrast is about half that and rests on one Thursday with no control roads.","tokens_in":19490,"tokens_out":3404,"would_cite":false,"duration_ms":31989,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J12","62J05","62P25"],"pacs":[],"model":"deepseek-v4-flash","headline":"School-escorted trips lower the probability of free-flowing traffic around Beijing schools by 8.34 percentage points during pick-up and drop-off hours.","keywords":["school runs","traffic congestion","built environment","generalized ordered logit","street-view imagery","Beijing","school siting","Shapley values"],"falsifier":"Find a normal workday in Beijing on which schools close for a reason unrelated to traffic (such as a teacher in-service day), rerun Model 1 comparing that day with ordinary school mornings, and check whether the smooth-traffic marginal effect is near $-8.34$ percentage points; if it is not, the Gaokao-week Thursday, not school runs, is the source of the effect.","tokens_in":18369,"feed_emoji":"🚌","tokens_out":10686,"duration_ms":96981,"temperature":0.7,"pith_summary":"This paper aims to establish that school-escorted pick-up and drop-off trips are a measurable cause of traffic congestion concentrated around schools, and that the surrounding built environment determines which school neighborhoods suffer most. Using roughly 1.4 million road-level congestion records from Beijing's core within the 5th Ring Road, the authors compare roads within 500 m of 846 schools on Monday/Tuesday school mornings with Thursday morning of the national college entrance examination week, when schools were closed; the difference yields an 8.34 percentage-point drop in the probability of free-flowing traffic. A second model links the frequency of congestion to 25 built-environment features and finds higher risk near clusters of schools, bus stops, high-betweenness roads, tall buildings, and business/financial street scenes, with lower risk farther from the city center. If these associations hold, they give planners concrete siting and street-design levers for reducing school-run congestion.","feed_headline":"School runs cut free-flow odds near schools by 8.34 points","feed_subtitle":"A Gaokao-week shutdown isolates school-escorted trips; clusters of schools and bus stops worsen the jam.","key_machinery":"The temporal argument runs through a generalized ordered logit model for the ordinal congestion index (smooth, slow, congested, severely congested), $P(Y>j)=\\exp(\\alpha_j+X\\beta_j)/(1+\\exp(\\alpha_j+X\\beta_j))$, with dummy variables for workday hours, school-run hours, and national-exam hours. Its marginal effects, computed from Equation (3), translate estimated coefficients into percentage-point changes in the probability of each congestion level, which lets the authors subtract a commute-without-school baseline from a commute-with-school baseline. The spatial argument runs through a multiple linear regression of each school neighborhood's congestion frequency on Z-scored built-environment variables, with K-means clustering of street-view scene probabilities producing ten scenescapes, and Shapley values used to rank contributions and interactions.","core_discovery":"The paper's central discovery is that school runs are a distinct congestion shock that can be separated from ordinary commute congestion. In the generalized ordered logit model, the marginal effect of the Monday/Tuesday school-run hours relative to the Thursday morning no-school commute baseline shows that smooth traffic probability falls by 8.34 percentage points, with slow traffic up 3.08, congested up 4.20, and severely congested up 1.05 percentage points. The paper further claims that the spatial distribution of this congestion is explained by the built environment: the multiple linear regression (adjusted $R^2 = 0.4524$) and Shapley explanations identify neighboring schools, bus stops, road betweenness centrality, average building height, and scenescapes tied to financial, business, and educational/highway functions as congestion-increasing, while distance from the city center and certain residential or well-organized urban scenescapes are congestion-reducing.","pith_inferences":["The paper does not test the 8.34-point estimate against control roads or a second school-closure week; an inference is that the effect could be re-estimated as a difference-in-differences with roads beyond 500 m from any school as controls, which would remove citywide traffic shocks that the single exam-week Thursday cannot absorb.","The scenescape interaction effects are correlational, but they suggest testable planning interventions: converting business/highway scenescapes to residential or organized-urban streetscapes should reduce congestion, and a before-after study around a street redesign or bus-stop relocation could verify this.","Because the data cover one week and one city, an out-of-sample test in another Chinese city or in a different season would clarify which built-environment coefficients are stable; the paper's own scoring function is not validated on held-out schools.","A direct extension of the built-environment score is to screen proposed school sites before construction by computing the congestion risk score for candidate plots, though such predictive use would require validation on sites built after the model was fit."],"forward_implications":["School-escorted trips add a congestion burden on roads within 500 m of schools that is separable from ordinary commuting: smooth traffic probability drops by 8.34 percentage points during pick-up and drop-off hours.","School-run congestion is structurally tied to local land use and transport supply: schools near other schools, bus stops, roads with high betweenness centrality, and business/financial street scenes face higher congestion risk.","School siting can be treated as a congestion-management decision: new schools should avoid the city center and the northeast sector within the 5th Ring Road, stay outside the 500 m catchment of existing schools, and avoid high-traffic roads and bus-stop-dense blocks.","Existing school neighborhoods can be ranked by a built-environment congestion score that explains roughly 46% of observed variation, giving an upgrade priority list for street and scenescape improvements.","Contradictory scenescape interactions imply that a street type that reduces congestion can amplify another type's congestion-promoting effect, so area-level redesign should consider combinations rather than single features."],"supporting_citations":[{"why":"Establishes the prior result that traffic congestion is 5 to 20 percent lower on non-school holidays, the comparison this paper refines into a school-run-specific DID setting.","marker":"Lu et al. 2017"},{"why":"Provides the earlier school-run congestion evidence from China and school-level attributes that the paper extends to broader built-environment features.","marker":"Sun et al. 2021"},{"why":"Supplies the generalized ordered logit model and its marginal-effect interpretation used to derive the 8.34-point estimate.","marker":"Williams 2016"},{"why":"Supplies the Shapley additive explanations used to rank built-environment features and analyze their interactions.","marker":"Lundberg & Lee 2017"},{"why":"Provides the street-view scene recognition model whose output is clustered into the ten scenescapes used as built-environment variables.","marker":"Zhou et al. 2018"},{"why":"Defines the space-syntax measures of integration and choice used as road-topology variables.","marker":"Van Nes and Yamu (2021)"},{"why":"Introduces the scenescape concept that organizes the street-view classifications into meaningful place types.","marker":"Silver & Clark 2016"}],"fun_headline_variants":["School runs cut free-flow odds by 8.34 points in Beijing","8.34-point free-flow drop from school runs","School-run congestion spikes near school clusters and bus stops","Built environment explains where school runs jam traffic","Gaokao shutdown reveals school-run congestion: 8.34-point loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The counterfactual comparison treats Thursday morning of the national college entrance examination week as a school-free version of an ordinary workday; if exam-week traffic management, school closures at exam sites, or citywide conditions make that Thursday unlike Monday and Tuesday morning, the 8.34-point effect is misattributed to school runs.","fun_headline_variants_meta":{"raw":{"variants":["School runs cut free-flow odds by 8.34 points in Beijing","8.34-point free-flow drop from school runs","School-run congestion spikes near school clusters and bus stops","Built environment explains where school runs jam traffic","Gaokao shutdown reveals school-run congestion: 8.34-point loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00187,"raw_usage":{"total_tokens":7352,"prompt_tokens":968,"completion_tokens":6384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":6301}},"tokens_in":584,"tokens_out":6384,"duration_ms":45170,"temperature":1.0,"reasoning_tokens":6301,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:35:19.392250+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a normal workday in Beijing on which schools close for a reason unrelated to traffic (such as a teacher in-service day), rerun Model 1 comparing that day with ordinary school mornings, and check whether the smooth-traffic marginal effect is near $-8.34$ percentage points; if it is not, the Gaokao-week Thursday, not school runs, is the source of the effect.","supporting_citations":[{"cited_title":"APACrefauthors \\ 2016","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized ordered logit model and its marginal-effect interpretation used to derive the 8.34-point estimate."},{"cited_title":"\\ Lee, S I","cited_arxiv_id":null,"evidence_quote":"Supplies the Shapley additive explanations used to rank built-environment features and analyze their interactions."},{"cited_title":"\\ Yamu, C","cited_arxiv_id":null,"evidence_quote":"Defines the space-syntax measures of integration and choice used as road-topology variables."},{"cited_title":"\\ Clark, T N","cited_arxiv_id":null,"evidence_quote":"Introduces the scenescape concept that organizes the street-view classifications into meaningful place types."}],"review_version":1}