{"id":"dcefbd31-5a75-49f6-8612-c5cbbde87828","arxiv_id":"2507.08869","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Work zone crash severity in four Florida counties appears related to crash type, work zone location, shoulder type, lighting, weather, and the presence of workers and police, but the paper does not report the statistics needed to support that claim.","lead":"This paper examines work zone crashes in four Florida counties from 2016 to 2023, describing their frequency, the conditions around them, and what a multinomial logistic model says about crash severity. A general reader might care because the findings claim to point toward where safety spending and enforcement could reduce deaths and serious injuries in construction zones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'significant contributing factors' claim is unsupported because the paper reports only coefficient point estimates with no standard errors, p-values, or confidence intervals, while the fatality reference group has only 36 cases.","rationale":"The paper's strongest assertion is the word 'significant' in the abstract and 'significant insights' in the conclusion. For that assertion to hold, the model's coefficients must be estimated with enough precision to rule out chance. The manuscript gives no standard errors, p-values, confidence intervals, or model-fit statistics, and the appendix containing full results is not present. This is not a stylistic omission: without those quantities, 'significant' has no statistical content. The small number of fatal crashes (36) makes the problem worse; with dozens of dummy predictors and a rare reference category, multinomial logit estimates are known to be unstable and can produce large coefficients when data are sparse or perfectly separated. The internal contradictions—the model description says severity is the predictor while the dependent variables are crash attributes, and equations (ii)–(iv) use coefficients that do not correspond to the printed tables—mean a reader cannot even reconstruct what was fit. The descriptive statistics may be useful, but the model-based inference, which is the load-bearing part of the claim, is unsupported. Therefore the REJECT verdict stands, with no change needed. This is a critique of the argument as presented, not of the authors' intent or effort.","tokens_in":9200,"tokens_out":5997,"duration_ms":70206,"concrete_test":"Obtain the underlying crash data (the paper does not name the source; authors would need to supply it or a data-use agreement), then refit the multinomial logit with severity as the outcome (fatality as reference) and the attributes in Tables 1–4 as predictors. Report the full output: coefficients, standard errors, p-values, 95% confidence intervals, and per-cell counts for fatalities. Then count how many reported coefficients have confidence intervals excluding zero and signs matching the paper. Also check for quasi-complete separation or nonconvergence (e.g., large standard errors or coefficients above 5). If most intervals include zero or the model fails to converge, the 'significant contributing factors' claim is not established; if the data cannot be obtained, the central inference is not independently reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the model reveals 'significant contributing factors' is unverifiable from the manuscript because Tables 1–4 report only coefficient point estimates. No standard errors, p-values, confidence intervals, Wald statistics, or model-fit measures appear anywhere, and the promised appendix with full results is absent. Since 'significant' is the operative word in the abstract and conclusion, the reader cannot determine whether any association is distinguishable from noise. This is aggravated by the reference category: only 36 fatalities are in the data (and 151 serious injuries), while the predictor set includes roughly a dozen crash types, eight light conditions, seven weather categories, work-zone types/locations, and shoulder types—dozens of dummy variables. Maximum-likelihood coefficients from such sparse tables can be extreme (e.g., Fog/Smog/Smoke = 8.358, Dawn = 7.485) precisely when data are sparse or separated, not because an effect is real. Equations (ii)–(iv) also do not match Table 1's coefficients for all three outcomes, and the prose swaps 'predictor variable' and 'dependent variables' when describing the model, further blocking any reconstruction of the actual fitted model. Thus, even if the descriptive counts are accurate, the model-based 'significant factors' claim is not supported by the reported evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a descriptive and multinomial-logit analysis of work-zone crashes from 2016 to 2023 in four Florida counties (Broward, Duval, Hillsborough, Orange). It reports yearly crash counts, severity distributions, and percentages for crash type, work-zone type and location, shoulder type, worker and law-enforcement presence, weather, and lighting. The central claim, stated in the abstract and conclusion, is that the multilogit model identifies 'significant contributing factors' to crash severity, offering actionable safety insights.","tokens_in":9459,"tokens_out":3489,"duration_ms":37720,"significance":"If the statistical claims were supported, the paper would provide useful preliminary evidence on which crash attributes are associated with fatality versus non-fatal severity in Florida work zones, potentially guiding targeted interventions. The authors deserve credit for compiling and describing a multi-county dataset and for being transparent about the omission of human factors such as speed and distraction. However, the paper's central inferential claim is not verifiable from the reported results, and the model presentation contains internal inconsistencies that prevent reconstruction. The descriptive statistics alone, while plausible, are not sufficient to carry the paper's substantive conclusions.","major_comments":[{"comment":"The central claim of 'significant contributing factors' is unverifiable because the manuscript reports only coefficient point estimates. No standard errors, p-values, confidence intervals, Wald statistics, or model-fit measures appear anywhere, and the promised appendix with full results is absent. Since the abstract and conclusion explicitly use the word 'significant,' the reader cannot determine whether any of the reported associations are distinguishable from noise.","section":"Multilogit Model; Tables 1–4; Abstract; Conclusion"},{"comment":"The model description reverses the roles of outcome and predictors: it states that 'severity of the crash (serious injury, injury, and non-injury) [is] the predictor variable, with fatality as the reference value,' while 'the dependent variables included crash type, light condition, weather condition, type of shoulder, crash in work zone, type of work zone, workers present, and law enforcement present.' This is inconsistent with the equations (ii)–(iv), which treat severity as the outcome (log odds of injury, fatal-injury, and non-injury relative to fatality). This internal contradiction blocks any reconstruction of the actual fitted model.","section":"Multilogit Model section"},{"comment":"The coefficients in the reported log-odds equations do not match Table 1. For example, equation (ii) uses 2.894 for 'Animal' while Table 1 reports 2.895; equation (iii) uses -5.544 for 'Animal,' a value that appears nowhere in Table 1. Because Table 1 reports only one column of coefficients (labeled 'Injury') rather than separate columns for serious injury, injury, and non-injury, the three outcome equations cannot be verified against the table.","section":"Equations (ii)–(iv); Table 1"},{"comment":"The inferential basis is undermined by sparse events and extreme coefficients. The analysis period contains only 36 fatalities and 151 serious injuries, while the predictor set includes roughly a dozen crash types, eight light conditions, seven weather categories, work-zone types, location categories, and shoulder types. Coefficients such as Fog/Smog/Smoke = 8.358 and Dawn = 7.485 in Table 4 are the classic signature of sparse-data separation. Without standard errors, penalized estimation, or a reduced predictor set, these extreme values cannot be interpreted as evidence of association.","section":"Yearly Distribution of Crashes by Severity; Multilogit Model; Tables 1–4"},{"comment":"Missing-data handling is not described. The paper reports that 25% of crashes lack a work-zone location, over 28% lack worker-presence information, and 8–19% of severe-crash records have missing covariate information. The multinomial model section does not state whether the analysis used complete cases, imputation, or a missing-data category. If missingness is not completely at random, any reported coefficient is potentially biased, which further undermines the claimed significance.","section":"Descriptive Statistics: Type of Work Zone...; Presence of Workers...; Weather and Light Conditions"}],"minor_comments":[{"comment":"The sentence 'The results in the appendix show the full results of the model' refers to an appendix that is not present in the manuscript.","section":"Multilogit Model section"},{"comment":"Equation (i) and the surrounding notation contain OCR artifacts (e.g., 'AAAAAA1', 'BBBBBl', 'HHHH') and undefined placeholders, making the equations unreadable as printed.","section":"Equations (i)–(iv)"},{"comment":"References 10, 11, and 17 appear to be duplicate citations of the same Khattak, Khattak, and Council work, with inconsistent years (2000, 2002, 2000) and identical titles, which is confusing for readers.","section":"References [10], [11], [17]"},{"comment":"The list of '4Is' differs between the abstract ('Information Intelligence, Innovation, Insight into communities, Investment, and Policies') and the introduction ('Information Intelligence, Innovation, Insight into communities, and Investment and Policies'), and the count is not consistently four items.","section":"Introduction; Abstract"},{"comment":"The table headers 'Crash Predictor Outcome / Crash Attribute Injury' are ambiguous; the manuscript should clearly label the outcome categories (serious injury, injury, non-injury) and the reference category (fatality) for each coefficient column.","section":"Tables 1–4"},{"comment":"The paper does not state the source database (e.g., FDOT crash records, Signal Four Analytics) or the inclusion criteria beyond county and year, which limits reproducibility.","section":"Descriptive Statistics"}],"recommendation":"reject","confidential_remarks":"The paper addresses a relevant safety topic and compiles a useful descriptive dataset, but the statistical analysis as presented cannot support the central 'significant factors' claim. The model description is internally inconsistent, the reported coefficients lack all inferential statistics, and the equations do not match the tables. These are load-bearing issues that would require a full re-analysis of the data with proper model specification, standard errors, and missing-data handling, rather than a routine revision. If the authors redo the analysis and report it transparently, a future version could be suitable for resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper has a usable descriptive picture of work zone crashes in four Florida counties, but its central inferential claim—that the multinomial logit reveals 'significant contributing factors'—is unsupported by anything reported in the manuscript. As it stands, reject.\n\nWhat's actually new: the county-level descriptive statistics for Broward, Duval, Hillsborough, and Orange over 2016–2023, including breakdowns of severity by crash type, work zone type, shoulder, light, weather, and presence of workers or law enforcement. Those counts are easy to read and could be a starting point for local safety planning. The literature review is adequate, and the study is honestly framed as preliminary.\n\nThe soft spots are not minor. The multilogit section reverses predictor and outcome variables in the prose, making the model impossible to interpret as written. Equations (ii)–(iv) use coefficients that don't match Table 1, and the promised appendix with full results is absent. None of the coefficient tables report standard errors, p-values, or confidence intervals, so 'significant' in the abstract is not backed by any inferential statistic. The reference category is only 36 fatalities, and the model includes dozens of dummy variables, which can produce extreme coefficients (e.g., fog at 8.358, dawn at 7.485) from sparse or separated data—not evidence of real effects. The paper never names the crash database or describes missing-value handling, even though a quarter of records lack crash location and 28% lack worker presence. If missingness is not random, every coefficient is biased.\n\nThe descriptive counts may be salvageable. But the load-bearing claim about significant contributing factors cannot be checked from the manuscript, and the internal contradictions block any reconstruction of the actual model. This is not a case of a solid paper with one weak section; the statistical core is not in a usable state.\n\nWho is this for: someone wanting a quick look at Florida work zone crash distributions by county. It is not a paper whose model results should be relied on yet. The authors could fix this with a complete reanalysis, proper inference, and a clear model definition. I would not send this out for peer review in its current form; I'd desk reject and invite a resubmission if the statistical work is redone properly.","headline":"The descriptive county data are useful, but the logistic-regression results are unverifiable as reported, so the paper's central claim of 'significant contributing factors' does not hold.","tokens_in":10005,"tokens_out":2510,"would_cite":false,"duration_ms":26412,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Crash type, work zone location, shoulder type, lighting, weather, and the presence of workers or law enforcement separate fatal from nonfatal crashes in four Florida urban counties, a multinomial logistic analysis of 2016–2023 data finds.","keywords":["Construction work zones","crash severity","road safety","Florida","multinomial logistic regression","work zone crashes","law enforcement presence"],"falsifier":"Take the same four-county 2016–2023 data, add indicators for missing work-zone location and worker-presence fields, and re-estimate the multilogit model with imputed or complete-case records; if the coefficients for law enforcement presence, paved shoulder, or dawn/fog conditions lose significance or flip sign, the paper's claim that these are significant contributing factors would be falsified.","tokens_in":8989,"feed_emoji":"🚧","tokens_out":7602,"duration_ms":74180,"temperature":0.7,"pith_summary":"This paper seeks to establish which observable crash and work-zone attributes separate fatalities and serious injuries from less severe outcomes in construction zones in four urban Florida counties: Broward, Duval, Hillsborough, and Orange. A multinomial logistic regression with fatality as the reference category is fitted to 2016–2023 crash data, and the paper reports systematic associations for crash type, location within the work zone, shoulder type, lighting, weather, and the presence of workers or law enforcement. If these associations are real, agencies could use them to target signage, lighting, and enforcement at the configurations most likely to kill or seriously injure people, and to train predictive tools that alert drivers and construction managers.","feed_headline":"Work zone crash severity hinges on shoulders, lighting, and police","feed_subtitle":"Four-county Florida study: crash type, work zone location, and worker or officer presence separate fatal from nonfatal outcomes.","key_machinery":"The carrying mechanism is the multinomial logistic (multilogit) regression model, which estimates the log odds of each severity outcome—serious injury, injury, and non-injury—relative to the reference category of fatality. The log-odds equation is $\\log\\left(\\frac{P(\\text{category})}{P(\\text{fatality})}\\right) = \\beta_{0,\\text{category}} + \\beta_{\\text{attribute 1, category}} A_1 + \\beta_{\\text{attribute 2, category}} A_2 + \\cdots$, and the fitted coefficients for crash type, work zone type and location, shoulder type, weather, lighting, and the presence of workers and law enforcement are what carry the argument: positive coefficients indicate a higher likelihood of that outcome relative to fatality, while negative coefficients indicate a lower likelihood.","core_discovery":"The central claim is that identifiable features of a work-zone crash predict whether it ends in death rather than injury or property damage. In the fitted multilogit model, positive coefficients for bicycle, rear-end, and sideswipe crashes indicate a higher likelihood of serious injury relative to fatality, while animal, head-on, and pedestrian crashes show the opposite. The paper reports that 81% of fatalities and 62% of serious injuries occurred in the activity area; 56% of fatalities occurred on paved shoulders; 53% of fatalities occurred in dark-lighted conditions and 83% in clear weather; and 83% of fatalities and 72% of serious injuries occurred when no law enforcement was present. Positive coefficients for the presence of workers and law enforcement across severity categories are read as a lower likelihood of fatalities when they are present. The paper concludes that these associations provide significant insights into the factors contributing to crash severity in Florida construction work zones.","pith_inferences":["Because 25% of records lack work-zone location and over 28% lack worker-presence information, the reported coefficients could shift if missingness is not random; a robustness check with imputation or missing-indicator terms would test this.","The paper's own exclusion of speed, distracted driving, and driver condition means the reported associations may partly reflect these omitted factors; if those factors are correlated with crash type or lighting, the severity gradients could be confounded.","The high fatality share in dark-lighted conditions combined with clear weather points to nighttime visibility as a plausible operative mechanism, a distinction the paper's separate weather and lighting variables cannot fully separate.","The paper's proposed machine-learning alerting would need to convert these retrospective odds into real-time predictions; the current model is a frequency-based snapshot, not a validated predictive system."],"forward_implications":["If the associations hold, increasing law enforcement presence at work zones is a concrete lever: crashes without officers accounted for 83% of fatalities and 72% of serious injuries, and the model's positive coefficients for enforcement presence imply lower fatality odds.","If the associations hold, safety engineering should focus on the activity area, lane shifts/crossovers, and work on shoulders or medians, since these configurations carry the largest shares of fatal and serious outcomes.","If the associations hold, weather- and light-specific countermeasures such as extra lighting for dark-lighted conditions and visibility aids for fog and smoke could reduce the most severe crashes, even though most crashes occur in clear daylight.","If the associations hold, the same variables could feed machine-learning alert systems for drivers and construction managers, as the paper proposes."],"supporting_citations":[{"why":"State-of-the-art review that frames work zone safety analysis and modeling approaches.","marker":"[8]"},{"why":"Severity modeling of work zone crashes using machine learning, providing a baseline for coefficient-based severity factors.","marker":"[12]"},{"why":"Documents unobserved heterogeneity and temporal instability in work zone crash severity models, the modeling challenge this study simplifies.","marker":"[13]"},{"why":"Provides an empirical analysis of driver injury severities in work-zone crashes, supporting the severity-outcome framing.","marker":"[15]"},{"why":"Investigates work zone crash casualty patterns using association rules, contributing to the set of crash attributes examined.","marker":"[16]"},{"why":"Identifies highway work zone risk factors and their impact on crash severity, the attribute set this paper extends.","marker":"[18]"}],"fun_headline_variants":["Florida work zone deaths: paved shoulders, darkness, and no police","Head-on and pedestrian crashes often fatal in Florida work zones","Work zone fatality factors: crash type, light, shoulder, police","Police and worker presence cut fatal work zone crash risk in Florida","Dark, clear, unpatrolled: common factors in Florida work zone fatalities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the crash records and the fitted multinomial logistic model are adequate to support statements about severity: the study excludes speed, driver condition, and distraction, does not name the crash database, and leaves 25% of records unclassified for work-zone location and over 28% unclassified for worker presence, so if missingness or omitted factors correlate with the included attributes, the reported coefficients could be biased.","fun_headline_variants_meta":{"raw":{"variants":["Florida work zone deaths: paved shoulders, darkness, and no police","Head-on and pedestrian crashes often fatal in Florida work zones","Work zone fatality factors: crash type, light, shoulder, police","Police and worker presence cut fatal work zone crash risk in Florida","Dark, clear, unpatrolled: common factors in Florida work zone fatalities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000755,"raw_usage":{"total_tokens":3399,"prompt_tokens":1026,"completion_tokens":2373,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":2282}},"tokens_in":642,"tokens_out":2373,"duration_ms":17825,"temperature":1.0,"reasoning_tokens":2282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:48:03.664215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same four-county 2016–2023 data, add indicators for missing work-zone location and worker-presence fields, and re-estimate the multilogit model with imputed or complete-case records; if the coefficients for law enforcement presence, paved shoulder, or dawn/fog conditions lose significance or flip sign, the paper's claim that these are significant contributing factors would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"State-of-the-art review that frames work zone safety analysis and modeling approaches."},{"cited_title":"S., Kabir, M","cited_arxiv_id":null,"evidence_quote":"Severity modeling of work zone crashes using machine learning, providing a baseline for coefficient-based severity factors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents unobserved heterogeneity and temporal instability in work zone crash severity models, the modeling challenge this study simplifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides an empirical analysis of driver injury severities in work-zone crashes, supporting the severity-outcome framing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Investigates work zone crash casualty patterns using association rules, contributing to the set of crash attributes examined."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Identifies highway work zone risk factors and their impact on crash severity, the attribute set this paper extends."}],"review_version":1}