Red Cross Fire Risk Map V2

A project to support Red Cross smoke detector installation campaigns.

See the source code on GitHub.

TODO

Help us improve the documentation! Follow the links for the items below to see where we need input. Then submit a pull request on GitHub.

Todo

Document the NFIRS raw data.

  • Identify the years we use.
  • Describe the cleaning and geocoding steps.
  • Describe how the data will grow.
  • Link to relevant code in docs.

original entry

Todo

Document the ACS raw data.

  • Describe the types of variables we’re interested in.
  • Describe the geographic granularity & extent.
  • Identify the years we use.
  • Describe how the data will grow.
  • Link to relevant code in docs.

original entry

Todo

Document the Red Cross raw data.

  • Describe how to acquire the data.
  • Identify the years we use.
  • Describe how the data will grow.
  • Link to relevant code in docs.

original entry

Todo

Describe the fire severity model.

  • Describe the general purpose.
  • Describe the input data and any transformations.
  • Describe the outcome variable.
  • Describe the features.
  • Describe the modeling approach.
  • Link to any relevant code.

original entry

Todo

  • Objective: Predict which Census Geographies will be in >90th percentile of total # of home fires the following year

  • Inputs: NFIRS (national fire incidents reporting) and ACS (census demographics)
    1. ‘NFIRS Fire Incident Data.csv’ (2009-2016)

    System Message: WARNING/2 (/home/docs/checkouts/readthedocs.org/user_builds/red-cross-fire-risk-map-v2/checkouts/latest/docs/data-pipeline.rst, line 121)

    Enumerated list ends without a blank line; unexpected unindent.

    • Hundreds of thousands of records with each record representing a fire incident in a US location

    System Message: WARNING/2 (/home/docs/checkouts/readthedocs.org/user_builds/red-cross-fire-risk-map-v2/checkouts/latest/docs/data-pipeline.rst, line 122)

    Block quote ends without a blank line; unexpected unindent.

    1. ‘ACS 5YR Block Group Data.csv’ OR ‘ACS 5YR Tract Data.csv’ (2013-2017)

    System Message: INFO/1 (/home/docs/checkouts/readthedocs.org/user_builds/red-cross-fire-risk-map-v2/checkouts/latest/docs/data-pipeline.rst, line 2)

    Enumerated list start value not ordinal-1: “2” (ordinal 2)

    System Message: WARNING/2 (/home/docs/checkouts/readthedocs.org/user_builds/red-cross-fire-risk-map-v2/checkouts/latest/docs/data-pipeline.rst, line 123)

    Enumerated list ends without a blank line; unexpected unindent.

    • Hundreds of thousands of records with each record representing the demographics of a census geoid

    System Message: WARNING/2 (/home/docs/checkouts/readthedocs.org/user_builds/red-cross-fire-risk-map-v2/checkouts/latest/docs/data-pipeline.rst, line 124)

    Block quote ends without a blank line; unexpected unindent.

    • New NFIRS & ACS data is released each year
  • Data Transformation:
    • NFIRS:
      1. Aggregated individual fire incidents into tabular format (records # of fires in each year in each geoid)
      2. Normalized number of fires by total population in the respective geoid
    • ACS:
      1. Removed non-numeric variables
      2. Removed 9 variables that are highly correlated with other variables in the dataset
  • Outputs: Label predictions (0: not Top 10%, 1: Top 10%) & prediction probabilities for each census geography
    1. ‘PropensityBlockModel_2014_2016.csv’ - provides predictions for years 2014-2016, but can be updated each year
  • Model: Balanced Random forest Model using Python imb-learn package
    • Script: ‘NFIRS_Block_level.ipynb’
    • Training Method: leverage all available data prior to target prediction year
      • Training Method Example: to predict 2016 top 10%, train on NFIRS 2009-2015 & ACS 2013-2015
      • Each record represents the feature values for a single geoid
    • Features: the below features were selected from a larger group of ~100 b/c model importance score is > 0.01:
      1. Sum of fires over selected years in the given geoid
      2. Mean number of fires over selected years in the given geoid
      3. Max number of fires over selected years in the given geoid
      4. % of houses in geoid built before 1939
      5. % of houses in geoid that are occupied
      6. % of houses in geoid whose heating source is kerosene
      7. % of houses in geoid with just a single occupant
      8. % of houses in geoid inhabited by married couple
      9. % of people in geoid with a bachelors degree
      10. % of people in geoid that have worked within the past year
      11. % of people in geoid that are black
      12. % of houses in geoid occupied by the owner
      13. % of people in geoid that have attended college but not graduated
      14. % of houses in geoid with value between $175-200K
      15. % of houses in geoid with value between $200-250K
      16. % of houses in geoid with a home equity loan
    • Preliminary Results: - 65-75% weighted Avg Recall at the Census block level - 75-85% Recall at the Census tract level

original entry

Todo

Describe the smoke alarm presence model.

  • Describe the general purpose.
  • Describe the input data and any transformations.
  • Describe the outcome variable.
  • Describe the features.
  • Describe the modeling approach.
  • Link to any relevant code.

original entry

Todo

Describe the combined model.

  • Describe the general purpose.
  • Describe the input data and any transformations.
  • Describe the outcome variable.
  • Describe the features.
  • Describe the modeling approach.
  • Link to any relevant code.

original entry