publicdata/j40-cejst-2
1
0
Fork 0
mirror of https://github.com/DOI-DO/j40-cejst-2.git synced 2025-07-28 14:31:16 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

Imputing income using geographic neighbors (#1559)

Browse source 
Imputes income field with a light refactor. Needs more refactor and more tests (I spotchecked). Next ticket will check and address but a lot of "narwhal" architecture is here.
This commit is contained in:
Emma Nechamkin 2022-04-27 15:59:10 -04:00 • committed by Emma Nechamkin
commit f047ca9d83
16 changed files with 1245 additions and 81 deletions
Download patch file Download diff file Expand all files Collapse all files

162
data/data-pipeline/data_pipeline/etl/sources/census_acs/etl.py
View file

@ -1,14 +1,26 @@
from collections import namedtuple
import os
import pandas as pd
import geopandas as gpd
from data_pipeline.config import settings
from data_pipeline.etl.base import ExtractTransformLoad
from data_pipeline.etl.sources.census_acs.etl_utils import (
retrieve_census_acs_data,
)
from data_pipeline.utils import get_module_logger
from data_pipeline.etl.sources.census_acs.etl_imputations import (
calculate_income_measures,
)
from data_pipeline.utils import get_module_logger, unzip_file_from_url
from data_pipeline.score import field_names
logger = get_module_logger(__name__)
# because now there is a requirement for the us.json, this will port from
# AWS when a local copy does not exist.
CENSUS_DATA_S3_URL = settings.AWS_JUSTICE40_DATASOURCES_URL + "/census.zip"
class CensusACSETL(ExtractTransformLoad):
def __init__(self):
@ -59,6 +71,23 @@ class CensusACSETL(ExtractTransformLoad):
self.POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME = (
"Percent of individuals < 200% Federal Poverty Line"
)
self.IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME = (
"Percent of individuals < 200% Federal Poverty Line, imputed"
)
self.ADJUSTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME = (
"Adjusted percent of individuals < 200% Federal Poverty Line"
)
self.ADJUSTED_AND_IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME_PRELIMINARY = (
"Preliminary adjusted percent of individuals < 200% Federal Poverty Line,"
+ " imputed"
)
self.ADJUSTED_AND_IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME = (
"Adjusted percent of individuals < 200% Federal Poverty Line,"
+ " imputed"
)
self.MEDIAN_HOUSE_VALUE_FIELD = "B25077_001E"
self.MEDIAN_HOUSE_VALUE_FIELD_NAME = (
@ -136,6 +165,10 @@ class CensusACSETL(ExtractTransformLoad):
"Percent enrollment in college or graduate school"
)
self.IMPUTED_COLLEGE_ATTENDANCE_FIELD = (
"Percent enrollment in college or graduate school, imputed"
)
self.COLLEGE_NON_ATTENDANCE_FIELD = "Percent of population not currently enrolled in college or graduate school"
self.RE_FIELDS = [
@ -188,18 +221,50 @@ class CensusACSETL(ExtractTransformLoad):
self.MEDIAN_INCOME_FIELD_NAME,
self.POVERTY_LESS_THAN_100_PERCENT_FPL_FIELD_NAME,
self.POVERTY_LESS_THAN_150_PERCENT_FPL_FIELD_NAME,
self.POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME,
self.IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME,
self.MEDIAN_HOUSE_VALUE_FIELD_NAME,
self.HIGH_SCHOOL_ED_FIELD,
self.COLLEGE_ATTENDANCE_FIELD,
self.COLLEGE_NON_ATTENDANCE_FIELD,
self.IMPUTED_COLLEGE_ATTENDANCE_FIELD,
]
+ self.RE_OUTPUT_FIELDS
+ [self.PERCENT_PREFIX + field for field in self.RE_OUTPUT_FIELDS]
+ [
field_names.POVERTY_LESS_THAN_200_FPL_FIELD,
field_names.POVERTY_LESS_THAN_200_FPL_IMPUTED_FIELD,
]
)
self.df: pd.DataFrame
def _merge_geojson(
self,
df: pd.DataFrame,
usa_geo_df: gpd.GeoDataFrame,
geoid_field: str = "GEOID10",
geometry_field: str = "geometry",
state_code_field: str = "STATEFP10",
county_code_field: str = "COUNTYFP10",
) -> gpd.GeoDataFrame:
usa_geo_df[geoid_field] = (
usa_geo_df[geoid_field].astype(str).str.zfill(11)
)
return gpd.GeoDataFrame(
df.merge(
usa_geo_df[
[
geoid_field,
geometry_field,
state_code_field,
county_code_field,
]
],
left_on=[self.GEOID_TRACT_FIELD_NAME],
right_on=[geoid_field],
)
)
def extract(self) -> None:
# Define the variables to retrieve
variables = (
@ -227,6 +292,27 @@ class CensusACSETL(ExtractTransformLoad):
df = self.df
# Here we join the geometry of the US to the dataframe so that we can impute
# The income of neighbors. first this looks locally; if there's no local
# geojson file for all of the US, this will read it off of S3
logger.info("Reading in geojson for the country")
if not os.path.exists(
self.DATA_PATH / "census" / "geojson" / "us.json"
):
logger.info("Fetching Census data from AWS S3")
unzip_file_from_url(
CENSUS_DATA_S3_URL,
self.DATA_PATH / "tmp",
self.DATA_PATH,
)
geo_df = gpd.read_file(
self.DATA_PATH / "census" / "geojson" / "us.json"
)
df = self._merge_geojson(
df=df,
usa_geo_df=geo_df,
)
# Rename two fields.
df = df.rename(
columns={
@ -349,7 +435,7 @@ class CensusACSETL(ExtractTransformLoad):
df["B03003_003E"] / df["B03003_001E"]
)
# Calculate college attendance:
# Calculate college attendance and adjust low income
df[self.COLLEGE_ATTENDANCE_FIELD] = (
df[self.COLLEGE_ATTENDANCE_MALE_ENROLLED_PUBLIC]
+ df[self.COLLEGE_ATTENDANCE_MALE_ENROLLED_PRIVATE]
@ -361,22 +447,64 @@ class CensusACSETL(ExtractTransformLoad):
1 - df[self.COLLEGE_ATTENDANCE_FIELD]
)
# strip columns
df = df[self.COLUMNS_TO_KEEP]
# Save results to self.
self.df = df
# rename columns to be used in score
rename_fields = {
"Percent of individuals < 200% Federal Poverty Line": field_names.POVERTY_LESS_THAN_200_FPL_FIELD,
}
self.df.rename(
columns=rename_fields,
inplace=True,
errors="raise",
# we impute income for both income measures
## TODO: Convert to pydantic for clarity
logger.info("Imputing income information")
ImputeVariables = namedtuple(
"ImputeVariables", ["raw_field_name", "imputed_field_name"]
)
df = calculate_income_measures(
impute_var_named_tup_list=[
ImputeVariables(
raw_field_name=self.POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME,
imputed_field_name=self.IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME,
),
ImputeVariables(
raw_field_name=self.COLLEGE_ATTENDANCE_FIELD,
imputed_field_name=self.IMPUTED_COLLEGE_ATTENDANCE_FIELD,
),
],
geo_df=df,
geoid_field=self.GEOID_TRACT_FIELD_NAME,
)
logger.info("Calculating with imputed values")
df[
self.ADJUSTED_AND_IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME
] = (
df[self.POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME].fillna(
df[self.IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME]
)
- df[self.COLLEGE_ATTENDANCE_FIELD].fillna(
df[self.IMPUTED_COLLEGE_ATTENDANCE_FIELD]
)
).clip(
lower=0
)
# All values should have a value at this point
assert (
df[
self.ADJUSTED_AND_IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME
]
.isna()
.sum()
== 0
), "Error: not all values were filled..."
logger.info("Renaming columns...")
df = df.rename(
columns={
self.ADJUSTED_AND_IMPUTED_POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME: field_names.POVERTY_LESS_THAN_200_FPL_IMPUTED_FIELD,
self.POVERTY_LESS_THAN_200_PERCENT_FPL_FIELD_NAME: field_names.POVERTY_LESS_THAN_200_FPL_FIELD,
}
)
# Strip columns and save results to self.
self.df = df[self.COLUMNS_TO_KEEP]
def load(self) -> None:
logger.info("Saving Census ACS Data")

127
data/data-pipeline/data_pipeline/etl/sources/census_acs/etl_imputations.py Normal file
View file

@ -0,0 +1,127 @@
from typing import List, NamedTuple
import pandas as pd
import geopandas as gpd
import numpy as np
from data_pipeline.utils import get_module_logger
logger = get_module_logger(__name__)
def _get_fips_mask(
geo_df: gpd.GeoDataFrame,
row: gpd.GeoSeries,
fips_digits: int,
geoid_field: str = "GEOID10_TRACT",
) -> pd.Series:
return (
geo_df[geoid_field].str[:fips_digits] == row[geoid_field][:fips_digits]
)
def _get_neighbor_mask(
geo_df: gpd.GeoDataFrame, row: gpd.GeoSeries
) -> pd.Series:
return geo_df["geometry"].touches(row["geometry"])
def _choose_best_mask(
geo_df: gpd.GeoDataFrame,
masks_in_priority_order: List[pd.Series],
column_to_impute: str,
) -> pd.Series:
for mask in masks_in_priority_order:
if any(geo_df[mask][column_to_impute].notna()):
return mask
raise Exception("No mask found")
def _prepare_dataframe_for_imputation(
impute_var_named_tup_list: List[NamedTuple],
geo_df: gpd.GeoDataFrame,
geoid_field: str = "GEOID10_TRACT",
) -> tuple[list, gpd.GeoDataFrame]:
imputing_cols = [
impute_var_pair.raw_field_name
for impute_var_pair in impute_var_named_tup_list
]
# prime column to exist
for impute_var_pair in impute_var_named_tup_list:
geo_df[impute_var_pair.imputed_field_name] = geo_df[
impute_var_pair.raw_field_name
].copy()
# generate a list of tracts for which at least one of the imputation
# columns is null
tract_list = geo_df[geo_df[imputing_cols].isna().any(axis=1)][
geoid_field
].unique()
# Check that imputation is a valid choice for this set of fields
logger.info(f"Imputing values for {len(tract_list)} unique tracts.")
assert len(tract_list) > 0, "Error: No missing values to impute"
return tract_list, geo_df
def calculate_income_measures(
impute_var_named_tup_list: list,
geo_df: gpd.GeoDataFrame,
geoid_field: str,
) -> pd.DataFrame:
"""Impute values based on geographic neighbors
We only want to check neighbors a single time, so all variables
that we impute get imputed here.
Takes in:
required:
impute_var_named_tup_list: list of named tuples (imputed field, raw field)
geo_df: geo dataframe that already has the census shapefiles merged
geoid field: tract level ID
Returns: non-geometry pd.DataFrame
"""
# Determine where to impute variables and fill a column with nulls
tract_list, geo_df = _prepare_dataframe_for_imputation(
impute_var_named_tup_list=impute_var_named_tup_list,
geo_df=geo_df,
geoid_field=geoid_field,
)
# Iterate through the dataframe to impute in place
for index, row in geo_df.iterrows():
if row[geoid_field] in tract_list:
neighbor_mask = _get_neighbor_mask(geo_df, row)
county_mask = _get_fips_mask(
geo_df=geo_df, row=row, fips_digits=5, geoid_field=geoid_field
)
state_mask = _get_fips_mask(
geo_df=geo_df, row=row, fips_digits=2, geoid_field=geoid_field
)
# Impute fields for every row missing at least one value using the best possible set of neighbors
# Note that later, we will pull raw.fillna(imputed), so the mechanics of this step aren't critical
for impute_var_pair in impute_var_named_tup_list:
mask_to_use = _choose_best_mask(
geo_df=geo_df,
masks_in_priority_order=[
neighbor_mask,
county_mask,
state_mask,
],
column_to_impute=impute_var_pair.raw_field_name,
)
geo_df.loc[index, impute_var_pair.imputed_field_name] = geo_df[
mask_to_use
][impute_var_pair.raw_field_name].mean()
logger.info("Casting geodataframe as a typical dataframe")
# get rid of the geometry column and cast as a typical df
df = pd.DataFrame(
geo_df[[col for col in geo_df.columns if col != "geometry"]]
)
# finally, return the df
return df

1
data/data-pipeline/data_pipeline/etl/sources/census_acs/etl_utils.py
View file

@ -4,6 +4,7 @@ from typing import List
import censusdata
import pandas as pd
from data_pipeline.etl.sources.census.etl_utils import get_state_fips_codes
from data_pipeline.utils import get_module_logger