Mapping Housing Instability Across the United States

Estimated Displacement Risk (EDR): The “Soft” Displacement

• Concept: EDR focuses on the displacement pressures that arise from the net out-migration of lower-income households from a census tract over a five-year period. It seeks to capture the less visible forms of displacement where residents may leave due to rising costs or changing neighborhood character, even without a formal eviction. It specifically models migration patterns for households below 80% of the Area Median Income (AMI), broken down into three groups (Low: 50-80% AMI, Very Low: 30-50% AMI, Extremely Low: <30% AMI), using data from Data Axle.
• Methodology: EDR is calculated based on a modeled net migration rate (dis_value) derived from BART (Bayesian Additive Regression Trees) models. Because migration patterns vary significantly across the US, separate models were developed for distinct geographic regions. Negative net migration values (more lower-income households leaving than arriving) indicate displacement pressure.
• Risk Categories: Tracts are categorized into four risk levels based on the calculated dis_value, a net migration rate per 1,000 lower-income renter households summed over the five-year period:
  • Extreme Risk: Net migration rate below -300. (dis_value < -300)
  • High Risk: Net migration rate between -300 and -100. (-300 <= dis_value < -100)
  • Elevated Risk: Net migration rate between -100 and 0. (-100 <= dis_value < 0)
  • At Risk/Early Stage: Net migration rate of 0 or more. (dis_value >= 0)

Estimated Eviction Risk (EER): The “Hard” Displacement

• Concept: EER focuses on the risk of formal eviction proceedings within a census tract. It aims to capture the more direct and immediate forms of displacement resulting from legal eviction filings.
• Methodology: EER is based on a modeled eviction filing rate ratio (final_ev) derived from BART models, incorporating data from the Legal Services Corporation and other sources. Unlike the regional EDR models, the EER is a single national model. It was trained on court records from 15 states with reliable filing data (Arizona, Connecticut, Delaware, Florida, Georgia, Illinois, Indiana, Minnesota, Missouri, Ohio, Oregon, Pennsylvania, South Carolina, Tennessee, and Texas) and then applied to every tract in the lower 48 states, so a tract outside those states receives a model-predicted ratio rather than a measured one.
• Risk Categories: Tracts are categorized into four risk levels based on the calculated eviction filing rate ratio (final_ev):
  • Extreme Risk: Filing rate ratio of 2.0 or higher. (final_ev >= 2.0)
  • High Risk: Filing rate ratio between 1.5 and 2.0. (1.5 <= final_ev < 2.0)
  • Elevated Risk: Filing rate ratio between 1.0 and 1.5. (1.0 <= final_ev < 1.5)
  • At Risk: Filing rate ratio between 0.8 and 1.0. (0.8 <= final_ev < 1.0) (Note: Tracts with a ratio below 0.8 are considered to have lower or minimal eviction risk relative to the baseline and are not assigned a risk category in the final HPRM score calculation).

Census Tract Level Data

• HPRM Scores: Combined housing precarity risk scores (0-8)
• EDR Metrics: Displacement risk category and underlying dis_value metrics
• EER Metrics: Eviction risk category and underlying final_ev metrics
• Demographic Data: Key demographic indicators including renter count, rent burden percentages, median rent and income
• Geographic Identifiers: GEOID, county, and state for each tract

File Formats

• CSV format for easy use in spreadsheet applications and data analysis tools
• GeoJSON format with tract boundaries for use in GIS software

These data files are generated from the 2019 and 2022 American Community Survey (ACS) 5-year datasets, using the methodology described in this document. Approved requests receive files that enable researchers, policymakers, and community organizations to conduct their own analyses of housing precarity risks in their communities.

Data Sources

The model integrates multiple administrative and survey datasets to capture the multifaceted nature of housing precarity:

1. American Community Survey (ACS): 5-year estimates (2015-2019 for the 2019 model, 2018-2022 for the 2022 model) provide demographic, economic, and housing characteristics at the census tract level.
2. Data Axle (formerly InfoGroup): Proprietary household-level migration data tracking residential moves over time, enabling calculation of net migration rates by income group.
3. Court eviction records: Filing records for 2018-2019 from 15 states, used to build the eviction filing rates the EER model is trained on. Most were assembled by the Legal Services Corporation; the Eviction Research Network collected the Cook County, Illinois records, and Portland State University collected the Oregon records.
4. EPA Smart Location Database: Accessibility and urban form metrics including job access, walkability, and transit availability.
5. Housing Choice Voucher (HCV) Data: HUD administrative data on subsidized housing participation rates.
6. County-Level Voting Data: 2020 presidential election results aggregated to county level.
7. Fair Market Rent (FMR) Data: HUD’s rent benchmarks by metropolitan area.
8. Time-to-evict data: State-level minimum statutory timelines for the notice, grace, court, and post-judgment phases of a nonpayment eviction, compiled from state landlord-tenant statutes. These rank among the strongest EER predictors.
9. Additional sources: HUD public housing inventory, LEHD/LODES employment counts, and a divergence index computed from ACS race/ethnicity distributions.

Geographic Framework

Due to significant regional variations in housing markets and migration patterns, the EDR model was developed separately for eight distinct regions:

1. Midwest Great Lakes: Illinois, Ohio, Michigan, Indiana, Wisconsin, Minnesota
2. Northeast New England: Massachusetts, Connecticut, Rhode Island, New Hampshire, Maine, Vermont, New York
3. Northeast Penn-Washington: Pennsylvania, New Jersey, Maryland, Delaware, District of Columbia, Virginia
4. South Atlantic: Florida, Georgia, North Carolina, South Carolina
5. South Central: Texas, Oklahoma
6. South Deep: Tennessee, Alabama, Louisiana, Kentucky, Mississippi, West Virginia, Arkansas, Missouri
7. West Mountain Midwest: Iowa, Kansas, Nebraska, South Dakota, North Dakota, Colorado, Utah, Idaho, Montana, Wyoming, Arizona, Nevada, New Mexico
8. West Pacific: California, Oregon, Washington

These regions are population-balanced groupings of 33 to 53 million people each, informed by a composite state-similarity index and constrained by contiguity of housing-market character, so that each model captures region-specific dynamics. Alaska and Hawaii are not covered by any model.

Income Group Definitions

The EDR model focuses on households below 80% of Area Median Income (AMI), stratified into three groups (the EER filing rate is computed over all renter households):

• Extremely Low Income (EL): < 30% AMI
• Very Low Income (VL): 30-50% AMI
• Low Income (L): 50-80% AMI

These thresholds are calculated each year at the county level from ACS 1-year median household income, and each household in the migration panel is assigned to a tier by its income.

Analytical Workflow

The modeling process follows a systematic workflow across all regions:

BART Modeling Approach

Bayesian Additive Regression Trees (BART) was selected for its ability to handle complex data relationships while providing robust predictions.

• Handle complex non-linear relationships
• Automatically detect interactions between variables
• Provide uncertainty quantification through posterior distributions
• Avoid overfitting through built-in regularization

Let’s examine how each of these capabilities works in practice:

# From code/m2_bart_model_edr_v4.R (condensed; comments added here)
bart <- bartMachine(
    Xy = m_df,
    num_trees = nt,  # number of trees, chosen by cross-validation
    k = k,           # shrinkage on the tree leaf values
    nu = nu,         # degrees of freedom of the error-variance prior
    q = q            # quantile of the error-variance prior
)

# From code/archive/xnortheastatlantic/m2_bart_batch_v0_el.r.Rout
# The model achieves strong performance with 480 total features
# without manually specifying interactions:
"bartMachine after preprocess... 480 total features..."

# From code/m2_bart_model.R
imp_permute <- var_selection_by_permute(bart, 
    bottom_margin = 10, 
    num_permute_samples = 10)
imp_permute$important_vars_local_names  # shows locally important interactions

# From code/m5_eer_fits_v13_2022.R
# Calculate prediction intervals on national data
cpi <- calc_prediction_intervals(bm_model, rhs %>% select(names(bm_model$X)))

# Calculate margins of error
margins_of_error <- (cpi$interval[, "pi_upper_bd"] - cpi$interval[, "pi_lower_bd"]) / 2

# The results include uncertainty estimates
results <- tibble(
    fit = fit,
    log_moe = margins_of_error
) %>%
mutate(
    exp_fit = case_when(exp(fit)-1 <= 0 ~ 0, TRUE ~ exp(fit)-1),
    exp_moe = exp(log_moe)-1,
    lower_bound = exp_fit-exp_moe,
    upper_bound = exp_fit+exp_moe,
    p_diff = exp_moe/exp_fit  # proportion of uncertainty
)

# From code/m3_var_importance_eer_v13.R
predictions[q, , ] <- bart_machine_get_posterior(bart_machine, test_data)$y_hat_posterior_samples

# From cross-validation results in code/archive/xnortheastatlantic/m2_bart_batch_v0_el.r.Rout
# Optimal parameters found through CV:
"bartMachine CV win: k: 5 nu, q: 10, 0.75 m: 200"

# From code/m4_bart_diagnostics.R
# Multiple comments about overfitting checks:
"# Check for overfitting with in and out of sample data"
"# If a model performs well on the training set but poorly on unseen data, it may be overfitting."

# Models with too little regularization (k=2) perform worse:
"[16,] 2  3 0.90        50  117.5082         5.60494571"
"[17,] 2  3 0.99       200  118.6784         6.65657258"

# From code/m2_bart_model.R
bartcv <- bartMachineCV(
    Xy = m_df,
    seed = seed,
    serialize = TRUE,
    use_missing_data = FALSE
)
# Extract optimal parameters
k = bartcv$k
nu = bartcv$nu
q = bartcv$q
nt = bartcv$num_trees

Outcome Variable Construction

Model Validation and Diagnostics

Quality assurance includes:

1. Convergence diagnostics: Monitoring BART chain convergence
2. Predictive performance: Evaluating R² and RMSE in cross-validation and on held-out data
3. Held-out validation: For the EER, leaving each training state out in turn and testing the model’s predictions against that state’s observed filing rates
4. Temporal stability: Comparing 2019 and 2022 results for consistency

Final Risk Categorization

Risk categories were determined through analysis of prediction distributions:

EDR Categories (based on net migration):

• At Risk/Early Stage: ≥ 0 households
• Elevated Risk: -100 to 0 households
• High Risk: -300 to -100 households
• Extreme Risk: < -300 households

EER Categories (based on filing rate ratio):

• At Risk: 0.8 to 1.0
• Elevated Risk: 1.0 to 1.5
• High Risk: 1.5 to 2.0
• Extreme Risk: ≥ 2.0

Data Quality Considerations

Tracts are excluded from the map when:

• Population density is below 20 people per square kilometer
• There are no renter households
• A military installation is present

Tracts are excluded from the EDR component (but keep their EER score) when:

• Student populations are high (>30% enrolled)
• Retired populations are high (>40% receiving retirement income)

Tracts with too few sampled lower-income households in the Data Axle panel are left out of model training but still receive model-predicted values.

These exclusions ensure that risk assessments reflect genuine displacement pressures rather than demographic artifacts.

Limitations and Future Directions

While HPRM 2.0 represents a significant advance in measuring housing precarity, several limitations should be noted:

• Data Axle coverage varies by region and may underrepresent certain populations
• Eviction data availability limits EER to specific states
• Models capture associations, not causal relationships
• Five-year ACS estimates may lag current conditions

Future iterations will incorporate additional data sources, expand geographic coverage, and develop real-time updating capabilities.