Claim at v1.0.0

graph-confounders-trees-surftemp

adb0107331d54da172200c21b1d8709890ad64ce13071902c927569a272865a0

Data

{
  "has_outcome": true,
  "kind": "confirmation",
  "lineage": {
    "audit": "docs/lineage_audit.md; data/derived/lineage_audit.json",
    "class": "none",
    "class_meaning": "the audit found no city-data source fetched before registration",
    "classes": [],
    "early_sources": []
  },
  "outcome": "graph-confounders-trees-surftemp",
  "prereg_branch": {
    "name": "prereg/graph-confounders-trees-surftemp",
    "sha": "48e0642f4d2f3b208445bc29be27cfc610d8c2f3"
  },
  "prior": 0.4,
  "registered_at": "2026-11-12T23:43:38-0500",
  "registrant": "Claude (round 2, graph-derived adjustment sets for heat), 12 November 2026",
  "registration": {
    "already_seen": "Seen before registration by this registrant: data/derived/heat/bridge_cd.csv and bridge_nta.csv (every value of street-tree density, surface temperature, vegetative cover, heat ED rate, heat hospitalization rate, poverty, percent Black and AC, by community district and by NTA), bridge_results.json, and the outcomes of heat-trees-surftemp, heat-surftemp-ed, heat-trees-ed-chain and heat-trees-canopy-within-nta. The registrant recomputed the two original estimates from bridge_cd.csv before registering. PLUTO (64uk-42ks, version 26v2) queries already in data/cache/socrata are citywide sums by land use and by year built, lot area by 2020 tract, land use and owner type (for nbhd-vacant-shootings), and single lots; none is grouped by community district or 2010 tract and none holds building area by area. The EHDP file for indicator 2325 (race and total population) is on disk because the commuting re-analysis read its UHF42 rows; its CD and NTA rows have not been read. The file for indicator 2112 (traffic density) is on disk for the concordance matrix at UHF42; its CD rows have not been read by this registrant. Indicator 2146 (older adults living alone) has not been downloaded. Only EHDP metadata (measure names, geographies, periods) and the PLUTO column list were read to choose the covariates. The registrant's general knowledge of New York (Manhattan is the densest borough, has many one-person households, and the original outcome notes say that hot-surface districts are central) informed the priors. Seen for this claim: unadjusted slope +0.50 (-0.04 to +1.03), adjusted +0.50 (-0.03 to +1.04), with borough terms +0.61, vegetative cover -0.62 (-1.17 to -0.07), NTA adjusted +0.44 (+0.15 to +0.73); Spearman correlations of tree density with vegetative cover -0.23 and with poverty +0.10.",
    "alternatives": [
      "Mechanical coupling: street trees per km2 of land is street trees per km2 of street multiplied by the road share, so adjusting for road share changes the exposure into trees per unit of street. A fall in the slope then reflects a better exposure measure more than removed confounding. The road-only secondary model shows how much of the change this accounts for.",
      "Over-adjustment: built-up share and vegetative cover are nearly complements, and built-up surface is part of what a tree-poor district is. The footprint share may absorb the effect of having few trees.",
      "Any three district covariates move a slope across 55 units. The poverty and percent Black adjustment moved it by 0.01.",
      "PLUTO is eight years later than the Landsat scene."
    ],
    "graph_evidence": [
      {
        "association_id": "CQuKL8Vr5G",
        "concepts": [
          "population density",
          "surface temperature"
        ],
        "direction": "+",
        "n_findings": 1,
        "n_studies": 1,
        "pubmed_ids": [],
        "relationship_id": "WLNXo1tGFr|sSuplgYnDQ",
        "source": "System API v1.0, read from data/cache/sysgraph and data/derived/graph/confounders/links_*.json; Dataverse copy not on this machine, so in_csv unknown",
        "summary": "Population density and land surface temperature, Pearson r = 0.8206, Prizren, Kosovo, Landsat 8 (doi 10.22630/srees.2324). No PubMed id; the Crossref abstract gives r = 0.8206, so the stored value matches. The study compares pixels of one municipality from countryside to town centre."
      },
      {
        "association_id": "w_1tAHxbUU",
        "concepts": [
          "population density",
          "urban heat island"
        ],
        "direction": "+",
        "n_findings": 1,
        "n_studies": 1,
        "pubmed_ids": [],
        "relationship_id": "WLNXo1tGFr|UbF0-c8zHb",
        "source": "System API v1.0, read from data/cache/sysgraph and data/derived/graph/confounders/links_*.json; Dataverse copy not on this machine, so in_csv unknown",
        "summary": "Population density and urban heat island across districts of Ho Chi Minh City, R2 = 0.64 (doi 10.46326/jmes.2021.62(5).06). No PubMed id; the Crossref abstract gives R2 = 0.64. The graph stores no direction; the abstract describes an increase."
      },
      {
        "association_id": "9xCRp0g69S, sVI20gcmIM, hKUE8RwXOV",
        "concepts": [
          "urban built-up area",
          "road area",
          "urban heat island"
        ],
        "direction": "+",
        "n_findings": 3,
        "n_studies": 1,
        "pubmed_ids": [],
        "source": "System API v1.0, read from data/cache/sysgraph and data/derived/graph/confounders/links_*.json; Dataverse copy not on this machine, so in_csv unknown",
        "summary": "One study of Xiamen 2001-2007 (conference paper, doi 10.1109/rsete.2011.5965025): coefficients on urban heat island of +2.333 for urban built-up area, +1.291 for road area and -5.602 for urban greenland. No PubMed id and no abstract on Crossref, so the numbers and their units could not be checked. The three variables are grounded to the concepts 'urban population' or to no concept; they were found by reading variable names under 'urban heat island'."
      },
      {
        "concepts": [
          "poverty",
          "percent Black",
          "surface temperature",
          "green cover"
        ],
        "direction": "0",
        "n_findings": 0,
        "n_studies": 0,
        "pubmed_ids": [],
        "source": "System API v1.0, read from data/cache/sysgraph and data/derived/graph/confounders/links_*.json; Dataverse copy not on this machine, so in_csv unknown",
        "summary": "The analyst's two covariates are not shared neighbours of green cover and temperature in the graph (docs/graph_experiments.md, confounders section)."
      }
    ],
    "id": "graph-confounders-trees-surftemp",
    "kind": "confirmation",
    "literature": [
      {
        "citation": "Ziter CD et al. PNAS 2019;116:7575-7580",
        "doi": "10.1073/pnas.1817561116",
        "finding": "Canopy cools and impervious cover warms daytime air; the two interact. Impervious cover is the common cause the original registration named and did not adjust for."
      }
    ],
    "plan": "Unit: community districts with the original covariates (55 expected). Outcome, exposure, poverty and percent Black are read unchanged from data/derived/heat/bridge_cd.csv. New covariates. Population density: EHDP indicator 2325 measure 992 (total population, ACS 2015-19, CD) divided by the district's land area from EHDP CD.geojson as computed in analyses/heat/bridge.py, in logs. Built-up area: building footprint share of land, the sum over tax lots of bldgarea/numfloors (lots with numfloors > 0 and bldgarea > 0) from PLUTO 64uk-42ks grouped by cd, divided by land area, in percent. Road area: the share of land outside tax lots, 100*(1 - sum of lotarea over lots that are not submerged (lottype not 9) / land area); streets and sidewalks are the land that is in no tax lot. Values are used as computed; districts with a share below 0 or above 60 are listed. Primary model: OLS of surface temperature on z(log tree density), poverty, percent Black, log population density, built-up share and road share, HC3 standard errors; the original model is refitted on the same rows; the change and its paired bootstrap interval (seed 20261113) are computed by analyses/round2_heat_confounders/reanalysis.py, written before this registration and not yet run. VIF is computed for every regressor of the graph-set model with an intercept. Secondary, reported and not deciding: each covariate alone; log built floor area per km2 in place of footprint share; EHDP traffic density (indicator 2112 measure 643, 2019, CD, log) in place of road share; vegetative cover (EHDP 2143) as the exposure; and the same primary model at NTA2010 (n about 190; original +0.44, 95% CI +0.15 to +0.73), with PLUTO grouped by borocode and ct2010 and assigned to NTAs through the 2010 tract-to-NTA equivalency table (8ius-dhrr) and population from indicator 2325 at NTA2010. Descriptive: the share of the tract-level variance of footprint share and floor-area ratio that lies between NTAs, to say what the neighbourhood fixed effects of heat-trees-canopy-within-nta remove. Timing: trees are from 2015, the Landsat scene from 17 July 2018, population from 2015-19 and PLUTO from version 26v2 (2026); buildings completed after 2018 are counted, which is a limit.",
    "power": "The original slope has HC3 SE 0.272 (n = 55). The SE of the change is about sqrt(SE_graph^2 - SE_original^2). If the graph covariates explain half of the variance of tree density (VIF 2) and none of the residual variance of temperature, SE_graph is 0.39 and the SE of the change 0.27, so the smallest change detectable with 80% power at the 5% level is about 0.76 F per SD; with VIF 1.5 it is 0.54. The lower part of the predicted band (changes of -0.3 to -0.5) would therefore probably end as unsettled at district level. At NTA2010 the original SE is 0.148 and the same calculation gives a detectable change of 0.41 (VIF 2), which is why the NTA model is registered as a secondary. The test can fail in three ways: the change can be small, it can be positive (adding borough terms to the original raised the slope to +0.61), or the covariates can be too collinear.",
    "precedence": "Verdict by concord.claims.PRECEDENCE rules 1 to 4.",
    "prediction": {
      "direction": "-",
      "magnitude": "The change in the slope of surface temperature on z(log live street trees per km2) when log population density, built-up share and road share are added is between -0.3 and -1.2 F per SD, so the slope under the graph set lies between -0.7 and +0.2 F per SD (original +0.50, 95% CI -0.03 to +1.04, n = 55). Derivation: on a standardised scale the one-confounder formula is b_adj = (r_xy - r_xz*r_zy)/(1 - r_xz^2). r_xy = 0.28 is seen (tree density and surface temperature across the 55 districts, SD of temperature 2.03 F). The graph gives r_zy = 0.8 for population density and surface temperature (r = 0.82; R2 = 0.64). r_xz, the correlation of street-tree density with population density, has not been queried; for r_xz between 0.3 and 0.6 the formula gives b_adj between +0.04 and -0.31 SD, that is +0.1 to -0.6 F, a change of -0.4 to -1.1 F. The band is widened at both ends because the graph's two correlations come from study areas that run from farmland to city centre, a wider range of density than New York's districts, so the correlation here is expected to be lower than 0.8, and because built-up and road area are added as well.",
      "otherwise": "Unsettled. no_test if the largest VIF exceeds 10, or if fewer than 50 districts have all variables. Whether the slope under the graph set is negative with an interval excluding zero (which would reverse the reading of heat-trees-surftemp) is reported and does not decide this verdict.",
      "refute_if": "The change is positive and the bootstrap interval of the change excludes zero (the graph's covariates move the slope further from the research direction), with the largest VIF 10 or less.",
      "support_if": "All three hold in the primary model: (a) the change in the point estimate (graph set minus original, same districts) is -0.30 F per SD or more negative; (b) the 95% percentile interval of the change from a paired bootstrap over districts (5000 resamples, both models refitted on each) excludes zero; (c) the largest variance inflation factor among the regressors of the graph-set model is 10 or less."
    },
    "prior": 0.4,
    "registered_at": "2026-11-12T23:43:38-0500",
    "registrant": "Claude (round 2, graph-derived adjustment sets for heat), 12 November 2026",
    "relates_to": "heat-trees-surftemp",
    "sha256": "e7c71bc91bc4eae44eab1e33c2fea2f1d1f198f2e8d38b485a74d021ea9a66b4",
    "statement": "Across New York City community districts, the positive adjusted association between street-tree density and daytime surface temperature (+0.50 F per SD, the opposite of the research direction) becomes smaller by at least 0.3 F per SD when population density, built-up area and road area, the common causes named by findings in the System Public Graph, are added to the analyst's adjustment set of poverty and percent Black."
  },
  "registration_commit": {
    "commit": "e0719f9b6f917f1dbcf2341b7790287aa8d56e21",
    "committed_at": "2026-11-12T23:43:38-05:00"
  },
  "registration_file_sha256": "7e74be08e3b4fb1c4e5582c6b01c122f8cc09b6b6c4d60bca0aa654bf4e3c3e2",
  "registration_path": "claims/registered/graph-confounders-trees-surftemp.json",
  "registration_sha256": "e7c71bc91bc4eae44eab1e33c2fea2f1d1f198f2e8d38b485a74d021ea9a66b4",
  "registration_text": "{\n \"id\": \"graph-confounders-trees-surftemp\",\n \"kind\": \"confirmation\",\n \"relates_to\": \"heat-trees-surftemp\",\n \"statement\": \"Across New York City community districts, the positive adjusted association between street-tree density and daytime surface temperature (+0.50 F per SD, the opposite of the research direction) becomes smaller by at least 0.3 F per SD when population density, built-up area and road area, the common causes named by findings in the System Public Graph, are added to the analyst's adjustment set of poverty and percent Black.\",\n \"graph_evidence\": [\n  {\n   \"association_id\": \"CQuKL8Vr5G\",\n   \"relationship_id\": \"WLNXo1tGFr|sSuplgYnDQ\",\n   \"concepts\": [\n    \"population density\",\n    \"surface temperature\"\n   ],\n   \"n_findings\": 1,\n   \"n_studies\": 1,\n   \"direction\": \"+\",\n   \"summary\": \"Population density and land surface temperature, Pearson r = 0.8206, Prizren, Kosovo, Landsat 8 (doi 10.22630/srees.2324). No PubMed id; the Crossref abstract gives r = 0.8206, so the stored value matches. The study compares pixels of one municipality from countryside to town centre.\",\n   \"pubmed_ids\": [],\n   \"source\": \"System API v1.0, read from data/cache/sysgraph and data/derived/graph/confounders/links_*.json; Dataverse copy not on this machine, so in_csv unknown\"\n  },\n  {\n   \"association_id\": \"w_1tAHxbUU\",\n   \"relationship_id\": \"WLNXo1tGFr|UbF0-c8zHb\",\n   \"concepts\": [\n    \"population density\",\n    \"urban heat island\"\n   ],\n   \"n_findings\": 1,\n   \"n_studies\": 1,\n   \"direction\": \"+\",\n   \"summary\": \"Population density and urban heat island across districts of Ho Chi Minh City, R2 = 0.64 (doi 10.46326/jmes.2021.62(5).06). No PubMed id; the Crossref abstract gives R2 = 0.64. The graph stores no direction; the abstract describes an increase.\",\n   \"pubmed_ids\": [],\n   \"source\": \"System API v1.0, read from data/cache/sysgraph and data/derived/graph/confounders/links_*.json; Dataverse copy not on this machine, so in_csv unknown\"\n  },\n  {\n   \"association_id\": \"9xCRp0g69S, sVI20gcmIM, hKUE8RwXOV\",\n   \"concepts\": [\n    \"urban built-up area\",\n    \"road area\",\n    \"urban heat island\"\n   ],\n   \"n_findings\": 3,\n   \"n_studies\": 1,\n   \"direction\": \"+\",\n   \"summary\": \"One study of Xiamen 2001-2007 (conference paper, doi 10.1109/rsete.2011.5965025): coefficients on urban heat island of +2.333 for urban built-up area, +1.291 for road area and -5.602 for urban greenland. No PubMed id and no abstract on Crossref, so the numbers and their units could not be checked. The three variables are grounded to the concepts 'urban population' or to no concept; they were found by reading variable names under 'urban heat island'.\",\n   \"pubmed_ids\": [],\n   \"source\": \"System API v1.0, read from data/cache/sysgraph and data/derived/graph/confounders/links_*.json; Dataverse copy not on this machine, so in_csv unknown\"\n  },\n  {\n   \"concepts\": [\n    \"poverty\",\n    \"percent Black\",\n    \"surface temperature\",\n    \"green cover\"\n   ],\n   \"n_findings\": 0,\n   \"n_studies\": 0,\n   \"direction\": \"0\",\n   \"summary\": \"The analyst's two covariates are not shared neighbours of green cover and temperature in the graph (docs/graph_experiments.md, confounders section).\",\n   \"pubmed_ids\": [],\n   \"source\": \"System API v1.0, read from data/cache/sysgraph and data/derived/graph/confounders/links_*.json; Dataverse copy not on this machine, so in_csv unknown\"\n  }\n ],\n \"literature\": [\n  {\n   \"citation\": \"Ziter CD et al. PNAS 2019;116:7575-7580\",\n   \"doi\": \"10.1073/pnas.1817561116\",\n   \"finding\": \"Canopy cools and impervious cover warms daytime air; the two interact. Impervious cover is the common cause the original registration named and did not adjust for.\"\n  }\n ],\n \"prediction\": {\n  \"direction\": \"-\",\n  \"magnitude\": \"The change in the slope of surface temperature on z(log live street trees per km2) when log population density, built-up share and road share are added is between -0.3 and -1.2 F per SD, so the slope under the graph set lies between -0.7 and +0.2 F per SD (original +0.50, 95% CI -0.03 to +1.04, n = 55). Derivation: on a standardised scale the one-confounder formula is b_adj = (r_xy - r_xz*r_zy)/(1 - r_xz^2). r_xy = 0.28 is seen (tree density and surface temperature across the 55 districts, SD of temperature 2.03 F). The graph gives r_zy = 0.8 for population density and surface temperature (r = 0.82; R2 = 0.64). r_xz, the correlation of street-tree density with population density, has not been queried; for r_xz between 0.3 and 0.6 the formula gives b_adj between +0.04 and -0.31 SD, that is +0.1 to -0.6 F, a change of -0.4 to -1.1 F. The band is widened at both ends because the graph's two correlations come from study areas that run from farmland to city centre, a wider range of density than New York's districts, so the correlation here is expected to be lower than 0.8, and because built-up and road area are added as well.\",\n  \"support_if\": \"All three hold in the primary model: (a) the change in the point estimate (graph set minus original, same districts) is -0.30 F per SD or more negative; (b) the 95% percentile interval of the change from a paired bootstrap over districts (5000 resamples, both models refitted on each) excludes zero; (c) the largest variance inflation factor among the regressors of the graph-set model is 10 or less.\",\n  \"refute_if\": \"The change is positive and the bootstrap interval of the change excludes zero (the graph's covariates move the slope further from the research direction), with the largest VIF 10 or less.\",\n  \"otherwise\": \"Unsettled. no_test if the largest VIF exceeds 10, or if fewer than 50 districts have all variables. Whether the slope under the graph set is negative with an interval excluding zero (which would reverse the reading of heat-trees-surftemp) is reported and does not decide this verdict.\"\n },\n \"plan\": \"Unit: community districts with the original covariates (55 expected). Outcome, exposure, poverty and percent Black are read unchanged from data/derived/heat/bridge_cd.csv. New covariates. Population density: EHDP indicator 2325 measure 992 (total population, ACS 2015-19, CD) divided by the district's land area from EHDP CD.geojson as computed in analyses/heat/bridge.py, in logs. Built-up area: building footprint share of land, the sum over tax lots of bldgarea/numfloors (lots with numfloors > 0 and bldgarea > 0) from PLUTO 64uk-42ks grouped by cd, divided by land area, in percent. Road area: the share of land outside tax lots, 100*(1 - sum of lotarea over lots that are not submerged (lottype not 9) / land area); streets and sidewalks are the land that is in no tax lot. Values are used as computed; districts with a share below 0 or above 60 are listed. Primary model: OLS of surface temperature on z(log tree density), poverty, percent Black, log population density, built-up share and road share, HC3 standard errors; the original model is refitted on the same rows; the change and its paired bootstrap interval (seed 20261113) are computed by analyses/round2_heat_confounders/reanalysis.py, written before this registration and not yet run. VIF is computed for every regressor of the graph-set model with an intercept. Secondary, reported and not deciding: each covariate alone; log built floor area per km2 in place of footprint share; EHDP traffic density (indicator 2112 measure 643, 2019, CD, log) in place of road share; vegetative cover (EHDP 2143) as the exposure; and the same primary model at NTA2010 (n about 190; original +0.44, 95% CI +0.15 to +0.73), with PLUTO grouped by borocode and ct2010 and assigned to NTAs through the 2010 tract-to-NTA equivalency table (8ius-dhrr) and population from indicator 2325 at NTA2010. Descriptive: the share of the tract-level variance of footprint share and floor-area ratio that lies between NTAs, to say what the neighbourhood fixed effects of heat-trees-canopy-within-nta remove. Timing: trees are from 2015, the Landsat scene from 17 July 2018, population from 2015-19 and PLUTO from version 26v2 (2026); buildings completed after 2018 are counted, which is a limit.\",\n \"power\": \"The original slope has HC3 SE 0.272 (n = 55). The SE of the change is about sqrt(SE_graph^2 - SE_original^2). If the graph covariates explain half of the variance of tree density (VIF 2) and none of the residual variance of temperature, SE_graph is 0.39 and the SE of the change 0.27, so the smallest change detectable with 80% power at the 5% level is about 0.76 F per SD; with VIF 1.5 it is 0.54. The lower part of the predicted band (changes of -0.3 to -0.5) would therefore probably end as unsettled at district level. At NTA2010 the original SE is 0.148 and the same calculation gives a detectable change of 0.41 (VIF 2), which is why the NTA model is registered as a secondary. The test can fail in three ways: the change can be small, it can be positive (adding borough terms to the original raised the slope to +0.61), or the covariates can be too collinear.\",\n \"alternatives\": [\n  \"Mechanical coupling: street trees per km2 of land is street trees per km2 of street multiplied by the road share, so adjusting for road share changes the exposure into trees per unit of street. A fall in the slope then reflects a better exposure measure more than removed confounding. The road-only secondary model shows how much of the change this accounts for.\",\n  \"Over-adjustment: built-up share and vegetative cover are nearly complements, and built-up surface is part of what a tree-poor district is. The footprint share may absorb the effect of having few trees.\",\n  \"Any three district covariates move a slope across 55 units. The poverty and percent Black adjustment moved it by 0.01.\",\n  \"PLUTO is eight years later than the Landsat scene.\"\n ],\n \"prior\": 0.4,\n \"registrant\": \"Claude (round 2, graph-derived adjustment sets for heat), 12 November 2026\",\n \"already_seen\": \"Seen before registration by this registrant: data/derived/heat/bridge_cd.csv and bridge_nta.csv (every value of street-tree density, surface temperature, vegetative cover, heat ED rate, heat hospitalization rate, poverty, percent Black and AC, by community district and by NTA), bridge_results.json, and the outcomes of heat-trees-surftemp, heat-surftemp-ed, heat-trees-ed-chain and heat-trees-canopy-within-nta. The registrant recomputed the two original estimates from bridge_cd.csv before registering. PLUTO (64uk-42ks, version 26v2) queries already in data/cache/socrata are citywide sums by land use and by year built, lot area by 2020 tract, land use and owner type (for nbhd-vacant-shootings), and single lots; none is grouped by community district or 2010 tract and none holds building area by area. The EHDP file for indicator 2325 (race and total population) is on disk because the commuting re-analysis read its UHF42 rows; its CD and NTA rows have not been read. The file for indicator 2112 (traffic density) is on disk for the concordance matrix at UHF42; its CD rows have not been read by this registrant. Indicator 2146 (older adults living alone) has not been downloaded. Only EHDP metadata (measure names, geographies, periods) and the PLUTO column list were read to choose the covariates. The registrant's general knowledge of New York (Manhattan is the densest borough, has many one-person households, and the original outcome notes say that hot-surface districts are central) informed the priors. Seen for this claim: unadjusted slope +0.50 (-0.04 to +1.03), adjusted +0.50 (-0.03 to +1.04), with borough terms +0.61, vegetative cover -0.62 (-1.17 to -0.07), NTA adjusted +0.44 (+0.15 to +0.73); Spearman correlations of tree density with vegetative cover -0.23 and with poverty +0.10.\",\n \"precedence\": \"Verdict by concord.claims.PRECEDENCE rules 1 to 4.\",\n \"registered_at\": \"2026-11-12T23:43:38-0500\",\n \"sha256\": \"e7c71bc91bc4eae44eab1e33c2fea2f1d1f198f2e8d38b485a74d021ea9a66b4\"\n}",
  "statement": "Across New York City community districts, the positive adjusted association between street-tree density and daytime surface temperature (+0.50 F per SD, the opposite of the research direction) becomes smaller by at least 0.3 F per SD when population density, built-up area and road area, the common causes named by findings in the System Public Graph, are added to the analyst's adjustment set of poverty and percent Black."
}

History in this collection

  1. addedv1.0.0Nov 14, 2026this version