{"id":"e238e8d4-0ff1-4a17-b879-65182f143588","arxiv_id":"2503.05706","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Intersections with a higher computed percentage of unobstructed view were associated with more recorded accidents, though the analysis has flaws that make the finding unreliable.","lead":"This study computes the share of view not blocked by buildings at road intersections in London and Manchester, then tests whether that share predicts police-reported accident counts. The authors report that more visible intersections had more accidents, a counterintuitive result that is weakened by serious statistical and reporting problems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Poisson regression is fit to a sample that excludes zero-accident intersections, so the reported positive visibility coefficient is not identifiable as an effect on accident counts.","rationale":"The reader's stated weakest assumption was construct validity of the visibility metric. While that is a legitimate concern, I find the more decisive, internally verifiable flaw to be the zero-truncated sample: the paper itself reports eliminating zero-accident intersections and Table 2 shows a minimum accident count of 2, yet the models use a standard Poisson likelihood. This selection on the response directly biases the coefficient, so the central claim fails regardless of whether the visibility metric is meaningful. The reader did list this among several reasons for rejection, but did not identify it as the load-bearing premise, hence 'disagree' on that specific comparison. Additional problems (overdispersion, impossibly low BIC in Table 5, lack of standard errors) support the same rejection but are secondary. I credit the authors for a reproducible data pipeline using open data and for testing on a second city, but those do not rescue the truncated-sample design. The proposed concrete test is a standard sensitivity analysis that would settle the issue: refit with zero counts included or with a zero-truncated model, and compare the visibility coefficient.","tokens_in":13320,"tokens_out":6456,"duration_ms":56843,"concrete_test":"Reconstruct the full intersection sample in the same 3-km London study area from OSM, compute the same features and the same 33.3-m buffer accident counts without discarding zero-accident intersections, and refit the paper's Poisson model with visibility on (i) the original positive-only subset, (ii) a zero-truncated Poisson on that subset, and (iii) a standard Poisson on the full sample including zeros. If the visibility coefficient in (ii) or (iii) is not significantly positive or differs materially from 1.7450, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the estimated coefficient for 'road visible percentage' being a valid estimate of the relationship between visibility and accident counts. The paper states in the Dataset Overview that preprocessing included 'the elimination of intersections with no recorded accidents,' and Table 2 confirms the accident-count range is 2-114. A Poisson likelihood assigns positive probability to zero counts, so deleting all zero-count observations changes the sampling distribution and the estimand. Fitting an ordinary Poisson model to a zero-truncated sample is selection on the dependent variable; it yields inconsistent coefficients, invalid standard errors, and meaningless AIC/BIC comparisons. The reported coefficient of 1.7450 could be an artifact of this truncation, particularly because visibility may be associated with whether an intersection has any accidents at all before it is associated with the count among intersections that have accidents. If the truncation is corrected, the coefficient could change sign or become insignificant, which would destroy the paper's central claim. This is the most load-bearing concern because it attacks the coefficient itself rather than only its interpretation or uncertainty.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a GIS-derived 'road visible percentage' for intersections, computed from OpenStreetMap building footprints around each intersection using an angular-ray method, and uses it as a predictor in Poisson generalized linear models for counts of accidents during 2010-2021 at London and Manchester intersections. Model 2 includes visible_percentage and is reported to yield a coefficient of 1.7450 with p<0.001, together with lower AIC and BIC and higher pseudo R-squared than Model 1 without it. The authors conclude that building-induced visibility restrictions are associated with accident frequency and recommend visibility-oriented urban planning. The main evidence is presented in Tables 2-5, with a replication attempt for Manchester in Table 6.","tokens_in":13537,"tokens_out":4210,"duration_ms":40920,"significance":"If the central claim were sound, the paper would offer a cheap, scalable visibility metric and a surprising, policy-relevant finding: that more visible intersections have more accidents. The spatial methodology is clearly described, uses open data, and could be reproduced by others. However, the statistical analysis as presented does not support the claim. The estimation sample is selected on the dependent variable, the Poisson model is severely overdispersed, and the reported AIC/BIC values are internally inconsistent. These issues attack the coefficient estimate itself rather than only its interpretation, so the central result cannot currently be regarded as an empirical finding about visibility and accident counts.","major_comments":[{"comment":"The paper states that preprocessing included 'the elimination of intersections with no recorded accidents' and Table 2 reports accident counts with a range of 2-114. A standard Poisson model assigns positive probability to zero counts, so fitting it to a zero-truncated sample changes the likelihood, the estimand, and the validity of standard errors and information criteria. The reported coefficient for visible_percentage (1.7450) and its p-value are therefore not identifiable as an effect on accident counts at all intersections. The authors should either model the full sample, including zero-count intersections, or use a zero-truncated Poisson/negative binomial model that accounts for the sampling scheme.","section":"Dataset Overview, Table 2"},{"comment":"The reported AIC and BIC values are arithmetically impossible. For n=967, BIC = AIC + k(ln n - 2), so BIC must be larger than AIC by roughly 5k units, yet Table 5 gives Model 1 AIC = 15406.98 and BIC = 5035.25, a difference of more than 10,000 in the opposite direction. The same problem appears for Model 2. This means the claimed reduction in AIC and BIC when visibility is added cannot be evaluated from the reported numbers, and the statement that Model 2 is preferred by both criteria is unsupported.","section":"Table 5"},{"comment":"The model exhibits severe overdispersion that is not addressed. For Model 1, the deviance is 11655 on 963 residual degrees of freedom and Pearson chi-squared is 1.75e+04, implying dispersion ratios around 12-18, far above the Poisson value of 1. The p-values in Table 4, including the p-value for visible_percentage, rely on the Poisson variance assumption and are therefore invalid. A negative binomial, quasi-Poisson, or otherwise overdispersion-robust model is required before any coefficient can be interpreted.","section":"Tables 3 and 4"},{"comment":"The sample size is inconsistent: the Dataset Overview says the London dataset consists of 935 records, while Table 3 reports 967 observations for both models. Additionally, the Manchester replication in Table 6 reports only aggregate fit statistics and gives no coefficient for visible_percentage or any other predictor, although the text claims the findings 'align closely' with those from London. Without reconciled sample definitions and a Manchester coefficient table, the claimed cross-city generalization is not verifiable.","section":"Dataset Overview and Table 3; Testing on Manchester City, Table 6"},{"comment":"The view percentage is computed by aggregating the angular intervals subtended by building footprints as seen from the intersection (Algorithm 1, lines 4-7), while the surrounding text also describes an 80-ray, 80-degree sector method with interpolation points along road segments, and Algorithm 1 uses total_angle = 2π. The measure ignores building height, distance-dependent obstruction depth, driver eye height, roadside objects, and road geometry. If building footprints do not correspond to what drivers can actually see, the predictor is measurement error and the regression coefficient has no physical interpretation. The authors should reconcile the algorithmic description and validate the measure against a plausible visibility model or field data.","section":"Methodology, View Percentage Calculation and Algorithm 1"}],"minor_comments":[{"comment":"The paper describes a 'real-time method for estimating intersection risk,' but no real-time data source, latency analysis, or deployment mechanism is described anywhere in the manuscript.","section":"Introduction"},{"comment":"Table 4 reports coefficients and p-values but no standard errors, which makes it impossible to assess the precision of the estimates or to check them against the reported test statistics.","section":"Table 4"},{"comment":"The traffic variable has a range of 0-5,269,093,361 with a standard deviation of 174,662,408, which is implausible for an annual average daily flow measure and suggests a unit or joining error; this should be clarified because the traffic coefficient is essentially zero but still part of the model.","section":"Table 2"},{"comment":"The relationship between the stated 80-degree field of view, the 80 rays from -40 to 40 degrees, and the full-circle total_angle in Algorithm 1 is not reconciled; the pseudocode should match the implemented procedure.","section":"Methodology, View Percentage Calculation"},{"comment":"Table 6 provides only model-level fit metrics for Manchester; without coefficient estimates for visible_percentage and the controls, the claimed replication is not supported by the presented evidence.","section":"Testing on Manchester City and Table 6"},{"comment":"Reference 1 contains the placeholder text 'accessed on Month DD, YYYY', and several other references lack complete access details or page ranges; the reference list should be checked before resubmission.","section":"References"}],"recommendation":"reject","confidential_remarks":"The topic is within the journal's scope and the spatial visibility measure is an interesting idea, but the central statistical evidence is invalid for multiple independent reasons: zero-truncation of the outcome, severe overdispersion, and impossible AIC/BIC values. Any of these alone would require a major rewrite; together they mean the reported positive coefficient for visible_percentage cannot be taken as an empirical result. I do not see this as a fixable presentation issue; the analysis would need to be redone with a different estimation framework and a validated visibility measure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: the paper has a genuinely new and cheap visibility metric, but the main regression result is not identifiable and the reported fit statistics are internally inconsistent. The idea is worth a proper analysis; this draft isn't there yet.\n\nWhat is new: computing a 'view percentage' from OSM building footprints around intersections appears to be original. It is reproducible from open data, and the attempt to replicate on Manchester is good practice. The writing is clear and the data integration steps are transparent.\n\nThe soft spots are serious. The dataset drops all intersections with zero recorded accidents (stated in the Dataset Overview). Fitting a Poisson model to that sample conditions on the outcome, so the coefficient on visibility is not an estimate of its effect on accident counts—it's a description of intersections that already have at least one crash. The stress-test note is correct that this is load-bearing.\n\nThe internal numbers also don't add up. Table 5 reports AIC=15406.98 and BIC=5035.25 for Model 1. With log-likelihood -7699.5 and 967 observations, BIC must be around 15,400, not 5,000. That's not a typo; it means the fit statistics were not computed from the reported model. Table 3 says 967 observations while the text says 935. Table 4 reports p-values but no standard errors. The traffic variable's range (0 to 5.3 billion) looks like a unit or data error.\n\nThe view percentage metric itself is a rough proxy—building footprints only, no heights, no driver eye height, no roadside objects. That could be acceptable as a first-cut predictor if the model were correct, but combined with the truncation issue, the positive coefficient is uninterpretable.\n\nWho gets value: traffic-safety researchers interested in GIS-based visibility measures might want to know the metric exists, but they should not rely on these results. The paper deserves a serious referee because the idea is timely and the flaws are fixable, but it needs a major rework. I'd recommend a reject with an invitation to resubmit, or a very demanding major revision.","headline":"Novel visibility metric from building footprints, but the zero-truncated sample and impossible BIC make the headline coefficient unsupported.","tokens_in":14035,"tokens_out":4396,"would_cite":false,"duration_ms":39088,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a road visible percentage computed from OpenStreetMap building footprints is a significant positive predictor of intersection accident counts, with a coefficient of 1.7450 in a Poisson model applied to London and…","keywords":["road safety","intersection accidents","visibility","OpenStreetMap","Poisson regression","building footprints","accident prediction","GIS analysis"],"falsifier":"Recompute the analysis at the same London and Manchester intersections using a 3D line-of-sight visibility measure taken from a driver's eye height, and check whether the map-derived view percentage still predicts accidents once the 3D measure is included; if its coefficient drops to zero or changes sign, the central claim is falsified.","tokens_in":13136,"feed_emoji":"🚦","tokens_out":10596,"duration_ms":91743,"temperature":0.7,"pith_summary":"Road accidents at intersections are common, but building-induced visibility restrictions are rarely quantified in safety models. This paper builds a 'road visible percentage' for each intersection from OpenStreetMap building footprints, measuring the share of the surrounding view not blocked by buildings, and feeds it into a Poisson model of accident counts. In 967 London intersections the visibility coefficient is $1.7450$ and statistically significant, meaning intersections with more open views are associated with more accidents, not fewer. Adding the visibility measure raises the model's pseudo $R^2$ from 0.9046 to 0.9318 and lowers AIC and BIC, and a replication on 243 Manchester intersections shows the same pattern. The authors present this as evidence that architectural layout is a meaningful accident predictor, while leaving the mechanism behind the positive direction open.","feed_headline":"More visible intersections record more crashes","feed_subtitle":"Building-footprint visibility predicts accident counts in London and Manchester, against standard sightline intuition.","key_machinery":"The central object is the road visible percentage, computed by Algorithm 1 (View Percentage Calculation): from each interpolation point on road segments approaching an intersection, the procedure sums the angular intervals covered by building footprints as seen from that point, then sets view percentage $=1-\\frac{\\text{blocked angle}}{2\\pi}$. In a second implementation step the authors cast 80 rays across an 80-degree field of view at one-degree increments, take the nearest ray-building intersection to form the visible polygon, and average over interpolation points. This scalar enters a Poisson GLM with traffic volume, speed limit, and primary/secondary road dummies, so the visibility measure is the new predictor whose incremental fit is tested by comparing models with and without it.","core_discovery":"On its own terms, the paper's central finding is that the 'road visible percentage'—a scalar obtained by projecting surrounding OpenStreetMap building footprints from points near an intersection and measuring what fraction of the view they fail to block—is a statistically significant predictor of intersection accident counts in a Poisson generalized linear model. In the London model the coefficient is $1.7450$, with $p<0.001$; because the model is log-linked, a one-unit increase in visible percentage corresponds to raising the expected accident count by a factor of $e^{1.7450}\\approx 5.7$. Adding the visibility term improves fit on every reported metric: pseudo $R^2$ rises from 0.9046 to 0.9318, log-likelihood rises from $-7699.5$ to $-7537.1$, AIC falls from 15406.98 to 15084.18, and BIC from 5035.25 to 4717.32. The sign is positive, which the authors call intriguing and not fully understood; they suggest higher visibility could invite more aggressive driving. The same qualitative pattern appears in Manchester, which the paper takes as evidence of robustness and generalizability.","pith_inferences":["Inference: the visibility measure is likely entangled with intersection size and centrality; large open junctions carry more traffic and turning movements, so the positive coefficient may partly reflect exposure that the single traffic-flow control does not fully absorb.","Inference: recomputing visibility with 3D building heights and driver eye height would produce a much different scalar, and the paper's positive result may or may not survive a physically realistic sight-line model.","Inference: a natural causal test is before-after accident counts around building construction or demolition, where the view percentage changes while road geometry stays roughly constant.","Inference: if the positive sign survives exposure controls, the policy implication is not to block views but to pair open intersections with speed-management measures; the paper does not test this."],"forward_implications":["Road safety performance functions should include a visibility term derived from building footprints; omitting it leaves a statistically significant predictor out of the model.","The positive sign undercuts the simple assumption that more open sightlines automatically reduce intersection crashes, and the paper's own explanation is that openness may encourage more aggressive driving.","Because the view percentage is built from OpenStreetMap and public accident and traffic datasets, the same pipeline can be applied to other cities without new field data.","Urban planners should treat building layout as a risk-relevant design variable, not purely an aesthetic or land-use concern."],"supporting_citations":[{"why":"Supplies the OpenStreetMap road network, intersection locations, and building footprints from which the view percentage is computed.","marker":"(34)"},{"why":"Supplies the geocoded police-reported accident records used as the response variable for each intersection.","marker":"(35)"},{"why":"Supplies annual average daily traffic flows assigned to intersections as the exposure covariate.","marker":"(36)"},{"why":"Provides the driver-visibility assessment standard that supports the 80-degree field-of-view assumption.","marker":"(22)"},{"why":"Provides the driver eye-position model used with (22) to define the visibility cone.","marker":"(41)"},{"why":"Provides the established statistical modelling approach for at-grade intersection accidents that the Poisson GLM builds on.","marker":"(23)"}],"fun_headline_variants":["Clearer intersections see more crashes","Visibility paradox: more sight, more crashes","Visible roads predict higher accident counts","Study: Visibility boosts intersection crash risk","Building-free views, more intersection accidents"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a driver's view at an intersection is captured by the angular width of two-dimensional building footprints on a map; if those outlines do not track what a driver can actually see, the view percentage is noise and the positive coefficient loses its physical meaning.","fun_headline_variants_meta":{"raw":{"variants":["Clearer intersections see more crashes","Visibility paradox: more sight, more crashes","Visible roads predict higher accident counts","Study: Visibility boosts intersection crash risk","Building-free views, more intersection accidents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1565,"prompt_tokens":994,"completion_tokens":571,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":511}},"tokens_in":610,"tokens_out":571,"duration_ms":7016,"temperature":1.0,"reasoning_tokens":511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:07:35.595122+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the analysis at the same London and Manchester intersections using a 3D line-of-sight visibility measure taken from a driver's eye height, and check whether the map-derived view percentage still predicts accidents once the 3D measure is included; if its coefficient drops to zero or changes sign, the central claim is falsified.","supporting_citations":[],"review_version":1}