Claim at v1.2.0
heat-extremes-downturn
aa04b5c8e377d01d998de405ca17408b73143bd9cd5fa05dd855a3d690595347
Data
{
"has_outcome": true,
"kind": "confirmation",
"lineage": {
"audit": "docs/lineage_audit.md; data/derived/lineage_audit.json",
"class": "d",
"class_meaning": "the outcome data were read for this claim before registration",
"classes": [
"d",
"a",
"b"
],
"early_sources": [
{
"class": "d",
"dataset": "ckan-csv",
"evidence": "Phoenix aggravated assaults, one of the four series that decide the pooled test, fetched for this claim at 07:38-07:44, before registration (07:58:34). Plan: \"Data, built by analyses/heat_extremes/build.py before this registration\". already_seen: yearly totals and mean daily counts were read; \"No count was tabulated or modelled against temperature\". Referee: \"The repository cannot show that no model was fitted\"; the power run's assumed dispersion (0.05-0.15) differs from the fitted 0.006-0.017, which supports the account.",
"fetched": "2026-11-13T07:38:35-05:00"
},
{
"class": "a",
"dataset": "GHCN-Daily v3 (NCEI), doi:10.7289/V5D21VHZ",
"evidence": "Daily temperature (GHCN-Daily) is the exposure, not the outcome. It is weather, fetched for earlier heat claims and reused.",
"fetched": "2026-11-13T07:39:56-05:00"
},
{
"class": "d",
"dataset": "arcgis-featureserver",
"evidence": "Tucson aggravated assaults, one of the four series that decide the pooled test, fetched for this claim at 07:38-07:44, before registration (07:58:34). Plan: \"Data, built by analyses/heat_extremes/build.py before this registration\". already_seen: yearly totals and mean daily counts were read; \"No count was tabulated or modelled against temperature\". Referee: \"The repository cannot show that no model was fitted\"; the power run's assumed dispersion (0.05-0.15) differs from the fitted 0.006-0.017, which supports the account.",
"fetched": "2026-11-13T07:43:24-05:00"
},
{
"class": "a",
"dataset": "GHCN-Daily v3 (NCEI), doi:10.7289/V5D21VHZ",
"evidence": "Daily temperature (GHCN-Daily) is the exposure, not the outcome. It is weather, fetched for earlier heat claims and reused.",
"fetched": "2026-11-13T07:41:30-05:00"
},
{
"class": "d",
"dataset": "qv6i-rri7",
"evidence": "Dallas aggravated assaults, one of the four series that decide the pooled test, fetched for this claim at 07:38-07:44, before registration (07:58:34). Plan: \"Data, built by analyses/heat_extremes/build.py before this registration\". already_seen: yearly totals and mean daily counts were read; \"No count was tabulated or modelled against temperature\". Referee: \"The repository cannot show that no model was fitted\"; the power run's assumed dispersion (0.05-0.15) differs from the fitted 0.006-0.017, which supports the account.",
"fetched": "2026-11-13T07:43:33-05:00",
"name": "Police Incidents"
},
{
"class": "a",
"dataset": "GHCN-Daily v3 (NCEI), doi:10.7289/V5D21VHZ",
"evidence": "Daily temperature (GHCN-Daily) is the exposure, not the outcome. It is weather, fetched for earlier heat claims and reused.",
"fetched": "2026-11-13T07:43:33-05:00"
},
{
"class": "d",
"dataset": "fdj4-gpfu",
"evidence": "Austin aggravated assaults, one of the four series that decide the pooled test, fetched for this claim at 07:38-07:44, before registration (07:58:34). Plan: \"Data, built by analyses/heat_extremes/build.py before this registration\". already_seen: yearly totals and mean daily counts were read; \"No count was tabulated or modelled against temperature\". Referee: \"The repository cannot show that no model was fitted\"; the power run's assumed dispersion (0.05-0.15) differs from the fitted 0.006-0.017, which supports the account.",
"fetched": "2026-11-13T07:43:39-05:00",
"name": "Crime Reports"
},
{
"class": "a",
"dataset": "GHCN-Daily v3 (NCEI), doi:10.7289/V5D21VHZ",
"evidence": "Daily temperature (GHCN-Daily) is the exposure, not the outcome. It is weather, fetched for earlier heat claims and reused.",
"fetched": "2026-11-13T07:43:40-05:00"
},
{
"class": "b",
"dataset": "63jg-8b9z",
"evidence": "Los Angeles assaults 2010-2019: descriptive curves only. The plan names Phoenix, Tucson, Dallas and Austin as the cities in the pooled test; this city has too few days at 38 C or more to enter it.",
"fetched": "2026-11-13T07:43:45-05:00",
"name": "Crime Data from 2010 to 2019"
},
{
"class": "b",
"dataset": "2nrs-mtv8",
"evidence": "Los Angeles assaults 2020-2023: descriptive curves only. The plan names Phoenix, Tucson, Dallas and Austin as the cities in the pooled test; this city has too few days at 38 C or more to enter it.",
"fetched": "2026-11-13T07:43:49-05:00",
"name": "Crime Data from 2020 to 2024"
},
{
"class": "a",
"dataset": "GHCN-Daily v3 (NCEI), doi:10.7289/V5D21VHZ",
"evidence": "Daily temperature (GHCN-Daily) is the exposure, not the outcome. It is weather, fetched for earlier heat claims and reused.",
"fetched": "2026-11-13T07:43:50-05:00"
},
{
"class": "b",
"dataset": "carto-sql",
"evidence": "Philadelphia assaults or shootings: descriptive curves only. The plan names Phoenix, Tucson, Dallas and Austin as the cities in the pooled test; this city has too few days at 38 C or more to enter it.",
"fetched": "2026-11-13T07:43:55-05:00"
},
{
"class": "b",
"dataset": "carto-sql",
"evidence": "Philadelphia assaults or shootings: descriptive curves only. The plan names Phoenix, Tucson, Dallas and Austin as the cities in the pooled test; this city has too few days at 38 C or more to enter it.",
"fetched": "2026-11-13T07:43:59-05:00"
},
{
"class": "a",
"dataset": "GHCN-Daily v3 (NCEI), doi:10.7289/V5D21VHZ",
"evidence": "Daily temperature (GHCN-Daily) is the exposure, not the outcome. It is weather, fetched for earlier heat claims and reused.",
"fetched": "2026-11-13T07:44:00-05:00"
},
{
"class": "b",
"dataset": "ijzp-q8t2",
"evidence": "Chicago assaults: descriptive curves only. The plan names Phoenix, Tucson, Dallas and Austin as the cities in the pooled test; this city has too few days at 38 C or more to enter it.",
"fetched": "2026-11-13T07:48:29-05:00",
"name": "Crimes - 2001 to Present"
},
{
"class": "b",
"dataset": "gumc-mgzr",
"evidence": "Chicago gunshot victims: descriptive curves only. The plan names Phoenix, Tucson, Dallas and Austin as the cities in the pooled test; this city has too few days at 38 C or more to enter it.",
"fetched": "2026-11-13T07:48:46-05:00",
"name": "Violence Reduction - Victims of Homicides and Non-Fatal Shootings"
},
{
"class": "a",
"dataset": "GHCN-Daily v3 (NCEI), doi:10.7289/V5D21VHZ",
"evidence": "Daily temperature (GHCN-Daily) is the exposure, not the outcome. It is weather, fetched for earlier heat claims and reused.",
"fetched": "2026-11-12T22:33:29-05:00"
},
{
"class": "b",
"dataset": "qgea-i56i",
"evidence": "New York felony assaults: descriptive curves only. The plan names Phoenix, Tucson, Dallas and Austin as the cities in the pooled test; this city has too few days at 38 C or more to enter it.",
"fetched": "2026-11-13T07:48:50-05:00",
"name": "NYPD Complaint Data Historic"
},
{
"class": "b",
"dataset": "5ucz-vwe8",
"evidence": "New York shootings: descriptive curves only. The plan names Phoenix, Tucson, Dallas and Austin as the cities in the pooled test; this city has too few days at 38 C or more to enter it.",
"fetched": "2026-11-13T07:48:57-05:00",
"name": "Shootings (2006-Present)"
},
{
"class": "a",
"dataset": "GHCN-Daily v3 (NCEI), doi:10.7289/V5D21VHZ",
"evidence": "Daily temperature (GHCN-Daily) is the exposure, not the outcome. It is weather, fetched for earlier heat claims and reused.",
"fetched": "2026-11-12T20:23:55-05:00"
}
]
},
"outcome": "heat-extremes-downturn",
"prereg_branch": {
"name": "prereg/heat-extremes-downturn",
"sha": "96174c4f69c06edbe6b0d073867492a3ee81612e"
},
"prior": 0.3,
"registered_at": "2026-11-13T07:58:34-0500",
"registrant": "Claude (agent, heat-extremes study, 13 Nov 2026)",
"registration": {
"already_seen": "Before registering I read, for each city, yearly totals and the overall mean daily count (printed by build.py) and the count of days by temperature. I saw seven Chicago daily counts from January 2006 while timing a query. No count was tabulated or modelled against temperature for any of the eight cities' aggravated assaults. Known from earlier work in this project, for shootings and not assaults: the New York shootings curve keeps rising to 35 C (docs/heat.md), and unadjusted Chicago gunshot cases by 3 C bin (8.9 a day at 27-30 C, 9.5 at 30-33, 8.9 at 33-36). Neither city is in the pooled test. test.py was run once on simulated counts to check that it executes.",
"alternatives": [
"Plateau: the rate stops rising above about 35 C and neither falls nor rises (verdict unsettled).",
"Linear rise: the rate at 38 C and above exceeds the rate at 32 to 35 C by the amount the slope below 35 C predicts (verdict contested).",
"Seasonal confounding: days at 38 C and above fall in June to August and reference days in spring and autumn; month and year terms, and the year-by-month sensitivity, address this.",
"Recording: on very hot days fewer assaults may be reported or recorded on the day; the counts are by date of occurrence, which limits this, and it cannot be excluded.",
"Adaptation: the turning point may lie at a percentile of the local climate and not at a fixed temperature, in which case Phoenix and Austin would differ at 38 C."
],
"can_it_fail": {
"D": {
"D_downturn_vs_flat": 5.3,
"D_downturn_vs_linear": 24,
"D_linear_vs_flat": 6.8
},
"days_by_city": {
"austin": {
"d32_35": 1188,
"d35_38": 1270,
"d38p": 710,
"d41p": 91,
"d44p": 1,
"days": 8644,
"mean_32_35": 33.34882154882155,
"mean_38p": 39.44352112676056,
"p95": 38.9,
"p99": 41.1
},
"chicago": {
"d32_35": 353,
"d35_38": 81,
"d38p": 7,
"d41p": 0,
"d44p": 0,
"days": 7546,
"mean_32_35": 33.07365439093484,
"mean_38p": 38.628571428571426,
"p95": 32.2,
"p99": 35
},
"dallas": {
"d32_35": 481,
"d35_38": 481,
"d38p": 187,
"d41p": 35,
"d44p": 1,
"days": 3530,
"mean_32_35": 33.39293139293139,
"mean_38p": 39.60909090909091,
"p95": 38.3,
"p99": 40.6
},
"los_angeles": {
"d32_35": 249,
"d35_38": 87,
"d38p": 18,
"d41p": 3,
"d44p": 1,
"days": 5113,
"mean_32_35": 33.12369477911646,
"mean_38p": 39.98888888888889,
"p95": 32.8,
"p99": 36.1
},
"new_york": {
"d32_35": 283,
"d35_38": 53,
"d38p": 2,
"d41p": 0,
"d44p": 0,
"days": 7305,
"mean_32_35": 33.022614840989405,
"mean_38p": 39.7,
"p95": 31.7,
"p99": 34.4
},
"philadelphia": {
"d32_35": 517,
"d35_38": 156,
"d38p": 9,
"d41p": 0,
"d44p": 0,
"days": 7548,
"mean_32_35": 33.12630560928432,
"mean_38p": 38.8,
"p95": 33.3,
"p99": 35.6
},
"phoenix": {
"d32_35": 280,
"d35_38": 395,
"d38p": 1132,
"d41p": 712,
"d44p": 212,
"days": 3683,
"mean_32_35": 33.35,
"mean_38p": 41.96369257950529,
"p95": 44.4,
"p99": 46.7
},
"tucson": {
"d32_35": 328,
"d35_38": 469,
"d38p": 667,
"d41p": 247,
"d44p": 18,
"days": 3165,
"mean_32_35": 33.30823170731708,
"mean_38p": 40.428485757121436,
"p95": 41.7,
"p99": 43.9
}
},
"reading": "With the larger dispersion tried (alpha 0.15) the pooled standard error is 0.022 on the log scale. The squared standardised distance between a downturn of 0.95 and a linear rise of 1.06 is 24, between 0.95 and 1.00 it is 5.3, and between 1.06 and 1.00 it is 6.8; all exceed the rule 8 threshold of 4. In simulation a true ratio of 1.00 gives 'unsettled' in 97% of series, a true 0.95 gives 'confirmed' in 40 to 68%, and a true 1.06 gives 'contested' in 63 to 83%. The test can therefore return any of the three verdicts.",
"simulated": {
"alpha0.05_downturn_0.95": {
"pooled_random_se": 0.018294165559280203,
"se_by_city": {
"austin": 0.02599596920591679,
"dallas": 0.04278212143784217,
"phoenix": 0.02826698518473979,
"tucson": 0.036402732061468636
},
"verdict_share": {
"confirmed": 0.6833333333333333,
"contested": 0,
"unsettled": 0.31666666666666665
}
},
"alpha0.05_flat_1.00": {
"pooled_random_se": 0.01738303041217024,
"se_by_city": {
"austin": 0.025846301098494607,
"dallas": 0.04204049580945773,
"phoenix": 0.0283502249920985,
"tucson": 0.03614899408663677
},
"verdict_share": {
"confirmed": 0,
"contested": 0.03333333333333333,
"unsettled": 0.9666666666666667
}
},
"alpha0.05_linear_1.06": {
"pooled_random_se": 0.017837761487406054,
"se_by_city": {
"austin": 0.02552774919091732,
"dallas": 0.04200090321886709,
"phoenix": 0.028119881148714797,
"tucson": 0.036122519058279255
},
"verdict_share": {
"confirmed": 0,
"contested": 0.8333333333333334,
"unsettled": 0.16666666666666666
}
},
"alpha0.15_downturn_0.95": {
"pooled_random_se": 0.022031771361080612,
"se_by_city": {
"austin": 0.03136341001056193,
"dallas": 0.05303062048028674,
"phoenix": 0.03917445921635349,
"tucson": 0.044537115438925474
},
"verdict_share": {
"confirmed": 0.4,
"contested": 0,
"unsettled": 0.6
}
},
"alpha0.15_flat_1.00": {
"pooled_random_se": 0.02216219292024002,
"se_by_city": {
"austin": 0.031150121851035693,
"dallas": 0.05176777866804278,
"phoenix": 0.03893612289709238,
"tucson": 0.04433292426094333
},
"verdict_share": {
"confirmed": 0,
"contested": 0.03333333333333333,
"unsettled": 0.9666666666666667
}
},
"alpha0.15_linear_1.06": {
"pooled_random_se": 0.022379723399789782,
"se_by_city": {
"austin": 0.030929020562270544,
"dallas": 0.05145589414099309,
"phoenix": 0.038995161250984144,
"tucson": 0.0440613494555562
},
"verdict_share": {
"confirmed": 0,
"contested": 0.6333333333333333,
"unsettled": 0.36666666666666664
}
}
},
"source": "data/derived/heat_extremes/power.json, from analyses/heat_extremes/power.py (seed 20261113, 60 simulated series per scenario on the real calendars and temperatures)."
},
"graph_evidence": [
{
"concepts": [
"temperature",
"homicide"
],
"direction": "+",
"n_findings": 4,
"n_studies": 2,
"pubmed_ids": [
32738768,
36858215
],
"relationship_id": [
"4FaXcuNJRt|YZku29igbl"
],
"summary": "Xu et al. 2020 (nine US cities) report a linear temperature-homicide association, significant in Chicago and New York; Rahman et al. 2023 (California) report a monotonic association, +4.9% homicide per 5 C of same-day maximum. Both abstracts were read in PubMed on 13 Nov 2026. The graph stores three of the four values wrongly (percent stored as odds ratios), as docs/heat.md records; the shape words are from the abstracts."
},
{
"concepts": [
"temperature",
"firearm"
],
"direction": "+",
"n_findings": 1,
"n_studies": 1,
"pubmed_ids": [
36525273
],
"relationship_id": [
"temperature|firearm"
],
"summary": "Lyons et al. 2022 (100 US cities, shootings 2015 to 2020): risk rises almost monotonically with daily maximum temperature, with a local peak at the 84th percentile of each city's temperature range (relative risk 1.17, 1.12 to 1.21, against the median). Abstract read 13 Nov 2026. This is the only graph finding that describes a peak, and it places it at a percentile of the local climate, not at a fixed temperature."
}
],
"graph_evidence_note": "Taken from the concept-level search cached in data/derived/graph/transfer_chicago/temp_violence_studies.json (13 temperature-violence relationships, 23 studies, most on suicide and injury). The graph's findings on temperature and violence therefore predict a monotonic rise (the refuting result here), with one hint of a percentile-based peak. A wider search of study variable names (analyses/heat_extremes/graph_shape.py) was still running when this was registered; its results are reported in docs/heat_extremes.md and did not inform the prediction.",
"id": "heat-extremes-downturn",
"kind": "confirmation",
"literature": [
{
"citation": "Gamble JL, Hess JJ. Temperature and violent crime in Dallas, Texas. West J Emerg Med 2012;13:239-46.",
"doi": "10.5811/westjem.2012.3.11746",
"finding": "Daily aggravated crime in Dallas 1993-1999 rises with daily mean temperature, moderates above 80 F (26.7 C) and turns negative above 90 F (32.2 C) daily mean."
},
{
"citation": "Rotton J, Cohn EG. Violence is a curvilinear function of temperature in Dallas: a replication. J Pers Soc Psychol 2000;78:1074-81.",
"doi": "10.1037/0022-3514.78.6.1074",
"finding": "Assault calls in Dallas 1994-1995 are an inverted-U function of temperature in daylight hours and linear at night."
},
{
"citation": "Bushman BJ, Wang MC, Anderson CA. Is the curve relating temperature to aggression linear or curvilinear? J Pers Soc Psychol 2005;89:62-6.",
"doi": "10.1037/0022-3514.89.1.62",
"finding": "The Minneapolis inverted U (peak at 74.9 F in 3-hour data) disappears when time of day is controlled; the relation is positive and linear in the evening and night."
},
{
"citation": "Stevens HR et al. Hot and bothered? Associations between temperature and crime in Australia. Int J Biometeorol 2019;63:747-62.",
"doi": "10.1007/s00484-019-01689-y",
"finding": "Daily assaults in New South Wales rise with maximum temperature and the rise slows above 30 C; thefts decline above 30 C."
}
],
"plan": "Data, built by analyses/heat_extremes/build.py before this registration. Daily counts of police-recorded aggravated assaults by date of occurrence from each city's own portal: Phoenix (phoenixopendata.com crime CSV, UCR category AGGRAVATED ASSAULT, distinct incident numbers, Nov 2015 to Nov 2025), Tucson (Tucson Police reported-crimes feature layers 2018 to 2026, UCR summary ASSAULT, AGGRAVATED, distinct incident numbers), Dallas (dallasopendata.com qv6i-rri7, nibrs_crime beginning AGG ASSAULT, distinct incident numbers, 2017 on), Austin (data.austintexas.gov fdj4-gpfu, category Aggravated Assault, 2003 on), and for the descriptive curves Los Angeles (data.lacity.org 63jg-8b9z and 2nrs-mtv8, codes 230, 231, 235, 236, 2010 to 2023), Philadelphia (phl.carto.com incidents_part1_part2, the two Aggravated Assault codes, by dispatch date, 2006 on), Chicago (ijzp-q8t2, FBI codes 04A and 04B, 2006 on) and New York (qgea-i56i, felony assault ky_cd 106, 2006 to 2025). Series end 31 Aug 2026 unless stated. Temperature: GHCN-Daily TMAX at Phoenix Sky Harbor USW00023183, Tucson USW00023160, Dallas Love Field USW00013960, Austin Camp Mabry USW00013958, Los Angeles downtown USW00093134, Philadelphia USW00013739, Chicago Midway USW00014819, Central Park USW00094728.\nQualifying cities. A city enters the pooled test with at least 100 days at or above 38.0 C and 100 days at 32.0 to 34.9 C. From temperatures alone (power.py) these are Phoenix, Tucson, Dallas and Austin. US stations record whole degrees F, so the hot bin is 101 F and above and the reference bin is 90 to 94 F.\nModel (analyses/heat_extremes/model.py and test.py, committed with this registration). Per city, a negative binomial GLM of the daily count on indicators for bins of TMAX (below 8, 8-14, 14-20, 20-23, 23-26, 26-29, 29-32, 32-35 as reference, 35-38, 38 and above) with C(dow) + C(month) + C(year) + hol + dom1 + wet, HC0 standard errors; alpha estimated once per city by maximum likelihood with a linear temperature term. The city estimate is the coefficient of the 38-and-above bin. The four log rate ratios are pooled by DerSimonian-Laird random effects; the pooled 95% interval decides the verdict. No resampling and no seed are involved.\nRegistered sensitivity, reported and not deciding: year-by-month strata in place of month and year; May to September only; without 2020 and 2021; the steps 35-38, 38-41 and 41-and-above each against 32-35; the fixed-effect pool; the linear prediction from the slope below 35 C.\nPlanned description, not deciding (analyses/heat_extremes/describe.py): the curve by 3 C bin for all eight cities and pooled; the curve by percentile of each city's own TMAX distribution; a comparison of stacked models with one curve common to all cities in degrees C against one common in percentiles (the adaptation question); piecewise-linear slopes below and above 35 C. Exploratory and labelled so: New York shootings by position in runs of days at or above 32 C.",
"prediction": {
"direction": "-",
"magnitude": "Downturn hypothesis: pooled rate ratio for days at or above 38 C against days at 32 to 35 C between 0.88 and 0.98. Linear-rise alternative: the slope below 35 C carried to the mean temperature of the days at or above 38 C, which is 6 to 9 C above the mean of the reference days; at the published slopes of 5 to 15% per 5 C this is a rate ratio of 1.06 to 1.25.",
"otherwise": "Unsettled: the pooled interval includes 1.00, which is what a plateau gives. The report then says whether the interval excludes the linear prediction computed in test.py (linear_alternative), and gives each city's estimate.",
"refute_if": "The 95% lower limit of the same pooled rate ratio is above 1.00 (assaults are still more frequent at 38 C and above than at 32 to 35 C; no downturn).",
"support_if": "The 95% upper limit of the DerSimonian-Laird random-effects pooled rate ratio over Phoenix, Tucson, Dallas and Austin is below 1.00."
},
"prior": 0.3,
"registered_at": "2026-11-13T07:58:34-0500",
"registrant": "Claude (agent, heat-extremes study, 13 Nov 2026)",
"rules": "Verdict precedence rules 1, 2, 8, 9 and 10 of concord.claims.PRECEDENCE apply; the conditions are on interval bounds.",
"sha256": "d22d9dabfb9e58b688d59f50dc84e59e196ccd70ccc8a0a4826a42f20beadee6",
"statement": "In US cities with many very hot days (Phoenix, Tucson, Dallas, Austin), police-recorded aggravated assaults are less frequent on days with a maximum temperature of 38 C or more than on days at 32 to 35 C in the same city, year, month and weekday position: violence turns down at extreme heat and does not keep rising in a straight line."
},
"registration_commit": {
"commit": "479e77abb8faa17835f8f6201f39b0305b64220b",
"committed_at": "2026-11-13T08:15:25-05:00"
},
"registration_file_sha256": "27212a1907f271674c55af2f5e1dd5615a0e76b1e3b613de0b3ed27fb48b8926",
"registration_path": "claims/registered/heat-extremes-downturn.json",
"registration_sha256": "d22d9dabfb9e58b688d59f50dc84e59e196ccd70ccc8a0a4826a42f20beadee6",
"registration_text": "{\n \"id\": \"heat-extremes-downturn\",\n \"kind\": \"confirmation\",\n \"statement\": \"In US cities with many very hot days (Phoenix, Tucson, Dallas, Austin), police-recorded aggravated assaults are less frequent on days with a maximum temperature of 38 C or more than on days at 32 to 35 C in the same city, year, month and weekday position: violence turns down at extreme heat and does not keep rising in a straight line.\",\n \"graph_evidence\": [\n {\n \"relationship_id\": [\n \"4FaXcuNJRt|YZku29igbl\"\n ],\n \"concepts\": [\n \"temperature\",\n \"homicide\"\n ],\n \"n_findings\": 4,\n \"n_studies\": 2,\n \"direction\": \"+\",\n \"summary\": \"Xu et al. 2020 (nine US cities) report a linear temperature-homicide association, significant in Chicago and New York; Rahman et al. 2023 (California) report a monotonic association, +4.9% homicide per 5 C of same-day maximum. Both abstracts were read in PubMed on 13 Nov 2026. The graph stores three of the four values wrongly (percent stored as odds ratios), as docs/heat.md records; the shape words are from the abstracts.\",\n \"pubmed_ids\": [\n 32738768,\n 36858215\n ]\n },\n {\n \"relationship_id\": [\n \"temperature|firearm\"\n ],\n \"concepts\": [\n \"temperature\",\n \"firearm\"\n ],\n \"n_findings\": 1,\n \"n_studies\": 1,\n \"direction\": \"+\",\n \"summary\": \"Lyons et al. 2022 (100 US cities, shootings 2015 to 2020): risk rises almost monotonically with daily maximum temperature, with a local peak at the 84th percentile of each city's temperature range (relative risk 1.17, 1.12 to 1.21, against the median). Abstract read 13 Nov 2026. This is the only graph finding that describes a peak, and it places it at a percentile of the local climate, not at a fixed temperature.\",\n \"pubmed_ids\": [\n 36525273\n ]\n }\n ],\n \"graph_evidence_note\": \"Taken from the concept-level search cached in data/derived/graph/transfer_chicago/temp_violence_studies.json (13 temperature-violence relationships, 23 studies, most on suicide and injury). The graph's findings on temperature and violence therefore predict a monotonic rise (the refuting result here), with one hint of a percentile-based peak. A wider search of study variable names (analyses/heat_extremes/graph_shape.py) was still running when this was registered; its results are reported in docs/heat_extremes.md and did not inform the prediction.\",\n \"literature\": [\n {\n \"citation\": \"Gamble JL, Hess JJ. Temperature and violent crime in Dallas, Texas. West J Emerg Med 2012;13:239-46.\",\n \"doi\": \"10.5811/westjem.2012.3.11746\",\n \"finding\": \"Daily aggravated crime in Dallas 1993-1999 rises with daily mean temperature, moderates above 80 F (26.7 C) and turns negative above 90 F (32.2 C) daily mean.\"\n },\n {\n \"citation\": \"Rotton J, Cohn EG. Violence is a curvilinear function of temperature in Dallas: a replication. J Pers Soc Psychol 2000;78:1074-81.\",\n \"doi\": \"10.1037/0022-3514.78.6.1074\",\n \"finding\": \"Assault calls in Dallas 1994-1995 are an inverted-U function of temperature in daylight hours and linear at night.\"\n },\n {\n \"citation\": \"Bushman BJ, Wang MC, Anderson CA. Is the curve relating temperature to aggression linear or curvilinear? J Pers Soc Psychol 2005;89:62-6.\",\n \"doi\": \"10.1037/0022-3514.89.1.62\",\n \"finding\": \"The Minneapolis inverted U (peak at 74.9 F in 3-hour data) disappears when time of day is controlled; the relation is positive and linear in the evening and night.\"\n },\n {\n \"citation\": \"Stevens HR et al. Hot and bothered? Associations between temperature and crime in Australia. Int J Biometeorol 2019;63:747-62.\",\n \"doi\": \"10.1007/s00484-019-01689-y\",\n \"finding\": \"Daily assaults in New South Wales rise with maximum temperature and the rise slows above 30 C; thefts decline above 30 C.\"\n }\n ],\n \"prediction\": {\n \"direction\": \"-\",\n \"magnitude\": \"Downturn hypothesis: pooled rate ratio for days at or above 38 C against days at 32 to 35 C between 0.88 and 0.98. Linear-rise alternative: the slope below 35 C carried to the mean temperature of the days at or above 38 C, which is 6 to 9 C above the mean of the reference days; at the published slopes of 5 to 15% per 5 C this is a rate ratio of 1.06 to 1.25.\",\n \"support_if\": \"The 95% upper limit of the DerSimonian-Laird random-effects pooled rate ratio over Phoenix, Tucson, Dallas and Austin is below 1.00.\",\n \"refute_if\": \"The 95% lower limit of the same pooled rate ratio is above 1.00 (assaults are still more frequent at 38 C and above than at 32 to 35 C; no downturn).\",\n \"otherwise\": \"Unsettled: the pooled interval includes 1.00, which is what a plateau gives. The report then says whether the interval excludes the linear prediction computed in test.py (linear_alternative), and gives each city's estimate.\"\n },\n \"plan\": \"Data, built by analyses/heat_extremes/build.py before this registration. Daily counts of police-recorded aggravated assaults by date of occurrence from each city's own portal: Phoenix (phoenixopendata.com crime CSV, UCR category AGGRAVATED ASSAULT, distinct incident numbers, Nov 2015 to Nov 2025), Tucson (Tucson Police reported-crimes feature layers 2018 to 2026, UCR summary ASSAULT, AGGRAVATED, distinct incident numbers), Dallas (dallasopendata.com qv6i-rri7, nibrs_crime beginning AGG ASSAULT, distinct incident numbers, 2017 on), Austin (data.austintexas.gov fdj4-gpfu, category Aggravated Assault, 2003 on), and for the descriptive curves Los Angeles (data.lacity.org 63jg-8b9z and 2nrs-mtv8, codes 230, 231, 235, 236, 2010 to 2023), Philadelphia (phl.carto.com incidents_part1_part2, the two Aggravated Assault codes, by dispatch date, 2006 on), Chicago (ijzp-q8t2, FBI codes 04A and 04B, 2006 on) and New York (qgea-i56i, felony assault ky_cd 106, 2006 to 2025). Series end 31 Aug 2026 unless stated. Temperature: GHCN-Daily TMAX at Phoenix Sky Harbor USW00023183, Tucson USW00023160, Dallas Love Field USW00013960, Austin Camp Mabry USW00013958, Los Angeles downtown USW00093134, Philadelphia USW00013739, Chicago Midway USW00014819, Central Park USW00094728.\\nQualifying cities. A city enters the pooled test with at least 100 days at or above 38.0 C and 100 days at 32.0 to 34.9 C. From temperatures alone (power.py) these are Phoenix, Tucson, Dallas and Austin. US stations record whole degrees F, so the hot bin is 101 F and above and the reference bin is 90 to 94 F.\\nModel (analyses/heat_extremes/model.py and test.py, committed with this registration). Per city, a negative binomial GLM of the daily count on indicators for bins of TMAX (below 8, 8-14, 14-20, 20-23, 23-26, 26-29, 29-32, 32-35 as reference, 35-38, 38 and above) with C(dow) + C(month) + C(year) + hol + dom1 + wet, HC0 standard errors; alpha estimated once per city by maximum likelihood with a linear temperature term. The city estimate is the coefficient of the 38-and-above bin. The four log rate ratios are pooled by DerSimonian-Laird random effects; the pooled 95% interval decides the verdict. No resampling and no seed are involved.\\nRegistered sensitivity, reported and not deciding: year-by-month strata in place of month and year; May to September only; without 2020 and 2021; the steps 35-38, 38-41 and 41-and-above each against 32-35; the fixed-effect pool; the linear prediction from the slope below 35 C.\\nPlanned description, not deciding (analyses/heat_extremes/describe.py): the curve by 3 C bin for all eight cities and pooled; the curve by percentile of each city's own TMAX distribution; a comparison of stacked models with one curve common to all cities in degrees C against one common in percentiles (the adaptation question); piecewise-linear slopes below and above 35 C. Exploratory and labelled so: New York shootings by position in runs of days at or above 32 C.\",\n \"alternatives\": [\n \"Plateau: the rate stops rising above about 35 C and neither falls nor rises (verdict unsettled).\",\n \"Linear rise: the rate at 38 C and above exceeds the rate at 32 to 35 C by the amount the slope below 35 C predicts (verdict contested).\",\n \"Seasonal confounding: days at 38 C and above fall in June to August and reference days in spring and autumn; month and year terms, and the year-by-month sensitivity, address this.\",\n \"Recording: on very hot days fewer assaults may be reported or recorded on the day; the counts are by date of occurrence, which limits this, and it cannot be excluded.\",\n \"Adaptation: the turning point may lie at a percentile of the local climate and not at a fixed temperature, in which case Phoenix and Austin would differ at 38 C.\"\n ],\n \"prior\": 0.3,\n \"registrant\": \"Claude (agent, heat-extremes study, 13 Nov 2026)\",\n \"can_it_fail\": {\n \"source\": \"data/derived/heat_extremes/power.json, from analyses/heat_extremes/power.py (seed 20261113, 60 simulated series per scenario on the real calendars and temperatures).\",\n \"days_by_city\": {\n \"phoenix\": {\n \"days\": 3683,\n \"d32_35\": 280,\n \"d35_38\": 395,\n \"d38p\": 1132,\n \"d41p\": 712,\n \"d44p\": 212,\n \"p95\": 44.4,\n \"p99\": 46.7,\n \"mean_38p\": 41.96369257950529,\n \"mean_32_35\": 33.35\n },\n \"tucson\": {\n \"days\": 3165,\n \"d32_35\": 328,\n \"d35_38\": 469,\n \"d38p\": 667,\n \"d41p\": 247,\n \"d44p\": 18,\n \"p95\": 41.7,\n \"p99\": 43.9,\n \"mean_38p\": 40.428485757121436,\n \"mean_32_35\": 33.30823170731708\n },\n \"dallas\": {\n \"days\": 3530,\n \"d32_35\": 481,\n \"d35_38\": 481,\n \"d38p\": 187,\n \"d41p\": 35,\n \"d44p\": 1,\n \"p95\": 38.3,\n \"p99\": 40.6,\n \"mean_38p\": 39.60909090909091,\n \"mean_32_35\": 33.39293139293139\n },\n \"austin\": {\n \"days\": 8644,\n \"d32_35\": 1188,\n \"d35_38\": 1270,\n \"d38p\": 710,\n \"d41p\": 91,\n \"d44p\": 1,\n \"p95\": 38.9,\n \"p99\": 41.1,\n \"mean_38p\": 39.44352112676056,\n \"mean_32_35\": 33.34882154882155\n },\n \"los_angeles\": {\n \"days\": 5113,\n \"d32_35\": 249,\n \"d35_38\": 87,\n \"d38p\": 18,\n \"d41p\": 3,\n \"d44p\": 1,\n \"p95\": 32.8,\n \"p99\": 36.1,\n \"mean_38p\": 39.98888888888889,\n \"mean_32_35\": 33.12369477911646\n },\n \"philadelphia\": {\n \"days\": 7548,\n \"d32_35\": 517,\n \"d35_38\": 156,\n \"d38p\": 9,\n \"d41p\": 0,\n \"d44p\": 0,\n \"p95\": 33.3,\n \"p99\": 35.6,\n \"mean_38p\": 38.8,\n \"mean_32_35\": 33.12630560928432\n },\n \"chicago\": {\n \"days\": 7546,\n \"d32_35\": 353,\n \"d35_38\": 81,\n \"d38p\": 7,\n \"d41p\": 0,\n \"d44p\": 0,\n \"p95\": 32.2,\n \"p99\": 35.0,\n \"mean_38p\": 38.628571428571426,\n \"mean_32_35\": 33.07365439093484\n },\n \"new_york\": {\n \"days\": 7305,\n \"d32_35\": 283,\n \"d35_38\": 53,\n \"d38p\": 2,\n \"d41p\": 0,\n \"d44p\": 0,\n \"p95\": 31.7,\n \"p99\": 34.4,\n \"mean_38p\": 39.7,\n \"mean_32_35\": 33.022614840989405\n }\n },\n \"simulated\": {\n \"alpha0.05_downturn_0.95\": {\n \"se_by_city\": {\n \"phoenix\": 0.02826698518473979,\n \"tucson\": 0.036402732061468636,\n \"dallas\": 0.04278212143784217,\n \"austin\": 0.02599596920591679\n },\n \"pooled_random_se\": 0.018294165559280203,\n \"verdict_share\": {\n \"confirmed\": 0.6833333333333333,\n \"contested\": 0.0,\n \"unsettled\": 0.31666666666666665\n }\n },\n \"alpha0.05_flat_1.00\": {\n \"se_by_city\": {\n \"phoenix\": 0.0283502249920985,\n \"tucson\": 0.03614899408663677,\n \"dallas\": 0.04204049580945773,\n \"austin\": 0.025846301098494607\n },\n \"pooled_random_se\": 0.01738303041217024,\n \"verdict_share\": {\n \"confirmed\": 0.0,\n \"contested\": 0.03333333333333333,\n \"unsettled\": 0.9666666666666667\n }\n },\n \"alpha0.05_linear_1.06\": {\n \"se_by_city\": {\n \"phoenix\": 0.028119881148714797,\n \"tucson\": 0.036122519058279255,\n \"dallas\": 0.04200090321886709,\n \"austin\": 0.02552774919091732\n },\n \"pooled_random_se\": 0.017837761487406054,\n \"verdict_share\": {\n \"confirmed\": 0.0,\n \"contested\": 0.8333333333333334,\n \"unsettled\": 0.16666666666666666\n }\n },\n \"alpha0.15_downturn_0.95\": {\n \"se_by_city\": {\n \"phoenix\": 0.03917445921635349,\n \"tucson\": 0.044537115438925474,\n \"dallas\": 0.05303062048028674,\n \"austin\": 0.03136341001056193\n },\n \"pooled_random_se\": 0.022031771361080612,\n \"verdict_share\": {\n \"confirmed\": 0.4,\n \"contested\": 0.0,\n \"unsettled\": 0.6\n }\n },\n \"alpha0.15_flat_1.00\": {\n \"se_by_city\": {\n \"phoenix\": 0.03893612289709238,\n \"tucson\": 0.04433292426094333,\n \"dallas\": 0.05176777866804278,\n \"austin\": 0.031150121851035693\n },\n \"pooled_random_se\": 0.02216219292024002,\n \"verdict_share\": {\n \"confirmed\": 0.0,\n \"contested\": 0.03333333333333333,\n \"unsettled\": 0.9666666666666667\n }\n },\n \"alpha0.15_linear_1.06\": {\n \"se_by_city\": {\n \"phoenix\": 0.038995161250984144,\n \"tucson\": 0.0440613494555562,\n \"dallas\": 0.05145589414099309,\n \"austin\": 0.030929020562270544\n },\n \"pooled_random_se\": 0.022379723399789782,\n \"verdict_share\": {\n \"confirmed\": 0.0,\n \"contested\": 0.6333333333333333,\n \"unsettled\": 0.36666666666666664\n }\n }\n },\n \"D\": {\n \"D_downturn_vs_linear\": 24.0,\n \"D_downturn_vs_flat\": 5.3,\n \"D_linear_vs_flat\": 6.8\n },\n \"reading\": \"With the larger dispersion tried (alpha 0.15) the pooled standard error is 0.022 on the log scale. The squared standardised distance between a downturn of 0.95 and a linear rise of 1.06 is 24, between 0.95 and 1.00 it is 5.3, and between 1.06 and 1.00 it is 6.8; all exceed the rule 8 threshold of 4. In simulation a true ratio of 1.00 gives 'unsettled' in 97% of series, a true 0.95 gives 'confirmed' in 40 to 68%, and a true 1.06 gives 'contested' in 63 to 83%. The test can therefore return any of the three verdicts.\"\n },\n \"already_seen\": \"Before registering I read, for each city, yearly totals and the overall mean daily count (printed by build.py) and the count of days by temperature. I saw seven Chicago daily counts from January 2006 while timing a query. No count was tabulated or modelled against temperature for any of the eight cities' aggravated assaults. Known from earlier work in this project, for shootings and not assaults: the New York shootings curve keeps rising to 35 C (docs/heat.md), and unadjusted Chicago gunshot cases by 3 C bin (8.9 a day at 27-30 C, 9.5 at 30-33, 8.9 at 33-36). Neither city is in the pooled test. test.py was run once on simulated counts to check that it executes.\",\n \"rules\": \"Verdict precedence rules 1, 2, 8, 9 and 10 of concord.claims.PRECEDENCE apply; the conditions are on interval bounds.\",\n \"registered_at\": \"2026-11-13T07:58:34-0500\",\n \"sha256\": \"d22d9dabfb9e58b688d59f50dc84e59e196ccd70ccc8a0a4826a42f20beadee6\"\n}",
"statement": "In US cities with many very hot days (Phoenix, Tucson, Dallas, Austin), police-recorded aggravated assaults are less frequent on days with a maximum temperature of 38 C or more than on days at 32 to 35 C in the same city, year, month and weekday position: violence turns down at extreme heat and does not keep rising in a straight line."
}History in this collection
- addedv1.0.0Nov 14, 2026this version