technical_summary.md

Technical Methodology Appendix

This analysis evaluates whether flood mitigation funding in Vermont aligns with underlying need. “Need” is treated as a normative construct (what funding should target), while observed funding patterns reflect real-world allocation dynamics. Models are used as validation tools, not as ends in themselves.

Data Sources

This project integrates multiple federal and spatial datasets at the town level:

• FEMA National Risk Index (NRI) – expected annual loss (EAL), risk, vulnerability, and resilience indicators
• FEMA Hazard Mitigation Assistance (HMA) – project-level mitigation funding data
• National Flood Insurance Program (NFIP) – claims and policy data
• U.S. Census Bureau American Community Survey (ACS) – demographic and socioeconomic variables
• FEMA National Flood Hazard Layer (NFHL) and Vermont river corridors – spatial flood exposure proxies
• Vermont town boundaries – base geography for aggregation

Data Processing and Integration

Spatial Data

• All spatial datasets were standardized to an equal-area coordinate system (EPSG:5070) for accurate area calculations, then converted to WGS84 (EPSG:4326) for web mapping.
• Flood exposure was estimated by calculating the percent of each town’s land area within:
  • FEMA high-risk flood zones (NFHL)
  • State-defined river corridors
• NFHL coverage varies across the state; a coverage flag was included to account for incomplete mapping.

NRI Aggregation

• FEMA NRI data is available at the census tract level and was spatially intersected with town boundaries.
• Tract-level values were allocated to towns using area-weighted proportions.
• Sliver overlaps (<1%) were removed and weights renormalized.
• Aggregated metrics include:
  • Expected annual loss (EAL)
  • Population, building value, and agricultural value
  • Risk, vulnerability (SOVI), and resilience indices

ACS Socioeconomic Data

• Multiple ACS tables were cleaned and merged using GEOID identifiers.
• Key indicators were calculated following Census guidance, including:
  • Percent below poverty line
  • Percent elderly
  • Percent without vehicle access
• Margins of error (MOE) were retained and propagated where appropriate.

FEMA HMA Funding Data

• Project-level funding was:
  • Inflation-adjusted to 2025 dollars using CPI data (FRED)
  • Filtered to flood-related mitigation projects
• Projects were assigned to towns using:
  • Regex parsing
  • Manual mapping
  • Fuzzy matching for ambiguous cases
• Funding was aggregated to the town level, including:
  • Total funding
  • Funding per capita and per housing unit (log-transformed)
  • Funding by time period (pre-2011, 2011–2022, 2023+)

NFIP Claims and Policies

• Claims and policies were cleaned and matched to towns.
• Claims were inflation-adjusted and aggregated by town and period.
• Insurance penetration rates were calculated using active policy counts.

Final Dataset

• All datasets were merged at the town level using GEOID and town name.
• Consistency checks ensured alignment across population and housing variables.
• The final dataset includes all Vermont towns, including those with zero funding.

Construction of the Need Index

A composite need index was developed to estimate relative flood mitigation need across towns.

Components

• Risk (Exposure):
  • Primary measure: Expected Annual Loss (EAL) from FEMA NRI
• Vulnerability (Socioeconomic):
  • Percent below poverty
  • Percent elderly
  • Percent without vehicle access

Method

• Variables were normalized using both:
  • Z-score standardization
  • Rank-based scaling
• Risk and vulnerability components were equally weighted (50/50) and summed.

Rationale

• EAL was selected as the primary risk variable due to its ability to capture multi-dimensional loss (buildings, agriculture, population).
• A parsimonious set of vulnerability variables was chosen to balance interpretability and signal strength.

Model Evaluation and Validation

Robustness Checks

• Need indices were highly consistent across normalization methods
  • Spearman correlation ≈ 0.91–1.00
• Rankings were stable, with only modest variation at the extremes.

Sensitivity Analysis

Two approaches were used:

• Leave-One-Variable-Out (LOVO):
  Tested how removing each variable affected rankings and underfunded town identification.
• Weight Variation:
  Tested alternative weightings (e.g., 70/30, 30/70 risk vs. vulnerability)

Key result:

• The model is robust to weighting choices
• The choice of risk variable (EAL vs. exposure) is the most consequential decision

Model Comparison

Multiple specifications were tested, including:

• Exposure-based models (river corridors, NFHL)
• EAL-based models
• Per capita risk models
• FEMA NRI composite indices

Finding:

• EAL-based models are more stable and policy-relevant than simple exposure measures
• Adding more variables provides limited additional explanatory power (parsimony preferred)

Funding Alignment Analysis

Gap Index

To evaluate allocation effectiveness:

• Gap = Need − Funding

Interpretation:

• Positive values → underfunded relative to need
• Negative values → overfunded relative to need

The gap index shows:

• Moderate positive correlation with need (~+0.32 to +0.49), well below +1.0, indicating it captures allocation mismatch rather than simply re-labeling need.

Correlation Analysis

• Need vs. funding: weak correlation (~0.10–0.30)
• Claims vs. funding: stronger correlation (~0.55)

Interpretation:

• Funding aligns more with past damage (claims) than with forward-looking need.

Regression Analysis

• Logistic regression (funding access):
  • AUC ≈ 0.6–0.8 → moderate ability to predict which towns receive funding
• OLS regression (funding amount):
  • R² ≈ 0 → near-zero explanatory power

Interpretation:

• Structural variables weakly explain who gets funding
• They do not explain how much funding is allocated

Spatial Analysis

• Moran’s I indicates significant spatial clustering of need
• Choropleth maps confirm geographic consistency of risk patterns
• Funding does not exhibit the same spatial alignment

Quadrant Analysis

Towns were categorized based on need and funding:

• Zero funding
• Underfunded (high need, low funding)
• Aligned
• Overfunded
• Low priority (low need, low funding)

Key observations:

• ~52% of towns received zero funding
• High-need towns are disproportionately represented among underfunded and zero-funded groups
• Distribution across quadrants is consistent across model specifications, though specific towns vary

Key Limitations

• NFHL flood mapping is incomplete across Vermont
• NFIP participation is extremely low (fewer than 2% of Vermont housing units), limiting claims as a comprehensive risk proxy
• Town-level aggregation may obscure within-town variation
• Funding data reflects approved projects, not unmet demand or unsuccessful applications

Summary

• The need index is methodologically robust and stable across specifications
• Funding is weakly aligned with structural need, but more strongly aligned with past damages
• Results suggest a reactive funding system, rather than a proactive, risk-based allocation model
• The framework reliably identifies systemic patterns, though precise rankings of individual towns should be interpreted with caution