Updating Reduced Match Criteria for the Forest Restoration & Wildfire Risk Mitigation Grant
Introduction
In 2020, our research team at Colorado State University developed the Wildfire Social Vulnerability Index (WFSVI) in order to set a criteria for reduced match for the Forest Restoration & Wildfire Risk Mitigation (FRWRM) Grant. This report describes updates to the WFSVI including updated data from the US Census, alignment with the CO HB21-1266 EnviroScreen tool, and further analysis of the WFSVI and reduced match criteria. The number of census block groups qualifying for reduced match has increased due to updated data and the inclusion of disproportionately impacted communities information from CO EnviroScreen.
American Community Survey Update
The original WFSVI was built using 5-year (2016 - 2020) American Community Survey (ACS) estimates. We have updated the WFSVI using the most current 5-year (2017-2021) ACS data. In some cases, no data is reported for a census block group because of poor response to the ACS or insufficient sample size to meet data disclosure requirements. We use two methods to impute missing values. First, if the data is reported at the census tract1, we use the census tract data to fill any missing records. While these data may be less spatially accurate, they represent the next best estimate of the block group value. Second, we use a machine learning imputation model to complete missing records when tract level information is unavailable. The machine learning method is known as random forest and is implemented via missForest package in R. The random forest method uses nonmissing data for one variable to train a model based on other nonmissing variables from other block groups around the state and uses that model to estimate missing values. It applies this procedure to each variable and iteratively updates until the model converges on its best estimate (see the package documentation for more information).
The original WFSVI used ACS data but sourced the data from Safegraph. While the data was accurate, it was less accessible and relied on analysts at SafeGraph to update the data when the Census released new data. We have rebuilt the code to access the Census Application Programming Interface (API) via the tidycensus package. The code now downloads data for the Census Block Groups (CBG) and Census Tracts, then calculates the index component for both geographies using tract information to fill in missing block group information. The user can now define the year from which to query census data, making the layer easy to update in the future.
Colorado EnviroScreen Data
Colorado’s Environmental Justice Act (HB21-1266) prompted the development of the EnviroScreen mapping tool to identify “disproportionately impacted communities”. We consider any disproportionately impacted community to automatically qualify for the FRWRM reduced match requirement. According to the EnviroScreen documentation, a disproportionately impacted community is defined as
… census block groups where more than 40% of the population are low-income (meaning that median household income is at or below 200% of the federal poverty line), 50% of the households are housing cost-burdened (meaning that a household spends more than 30% of its income on housing costs like rent or a mortgage), 40% of the population are people of color (including all people who do not identify as non-Hispanic white), or 20% of households are linguistically isolated (meaning that all members of a household that are 14 years old or older have difficulty with speaking English). Also included in this definition are mobile home communities, the Ute Mountain Ute and Southern Ute Indian Reservations, and all areas that qualify as disadvantaged in the federal Climate and Economic Justice Screening Tool. The definition also includes census block groups that experience higher rates of cumulative impacts, which is represented by an EnviroScreen Score (Percentile) above 80.
The CO EnviroScreen tool data was built using 5-year ACS data from 2015-2019 along with many other data sources. The challenge is that this updated uses data from the 5-year ACS 2017-2021, which are based on census block groups redefined in 2020. Consequently, the CO EnviroScreen tool data cannot be directly merged with the WFSVI data. We use a geographic crosswalk designed to harmonize data across time (Manson et al., 2021). The crosswalk develops a set of weights based on the estimated fraction of the population or households in the 2010-2020 block group definition to the 2020 block group definition. We use the population weights to translate the following data between block group definitions: the low-income, people of color, the cumulative impact score, and the Justice 40 criteria, and the household weights to map the housing burden and linguistic isolation criteria. We then apply the qualifying thresholds defined above to determine which of the 2020 block groups qualifies as a disproportionately impacted community.
The integration of the CO EnviroScreen tool aligns the WFSVI with HB21-1266 in Colorado and the federal Justice 40 initiative.
Analysis
Data from the ACS is sampled with uncertainty. We develop a method to investigate how the uncertainty affects the qualifying classification of a particular block group. A margin of error is reported for each variable (component of the WFSVI) in each block group. We use the margin of error to estimate the standard deviation of the reported estimate. We define a normal sampling distribution with the estimate as the mean and its standard deviation for each census variable for each block group. We then conduct a Monte Carlo simulation by sampling values from each distribution and recomputing the WFSVI. We run 1000 simulations and calculate the fraction of simulations that a block group qualifies based on its WFSVI above the 75th percentile cutoff. For example, a block group may qualify for reduced match in every one of the 1000 simulations indicating strong confidence that that block group indeed should qualify. In contrast, a block group that only qualifies 50% of the time suggests less confidence that it should qualify. The simulation is conducted in the script 07_monte_carlo_simulation.R.