Background

Prior research has shown that households and communities with higher insurance coverage fare better after disasters.1 2. However, case studies and legal analysis has suggested that lower-income, black, and hispanic policyholders do not experience equal insurance claims processing (how long claims take to resolve, and what is ultimately covered).3 4. Assessing the systemic differences in insurance claims processing is difficult, as most insurer data is private, and consequently this pattern has not been documented in a systemic way across the insurance market. In this analysis, I use on public source of data on property and casualty insurance claims (the National Flood Insurance Program) to explore these potential patterns.

Research Questions

  • How do claims processing times vary by income, wealth and racial demographics neighborhoods?
  • How do claims payouts vary by income, wealth and racial demographics neighborhoods?
  • How does the probability of a claim being closed without payment vary by income, wealth and racial demographics neighborhoods?

Data

Claims data comes from FEMA OpenFEMA NFIP Redacted Claims (v2) and appended by data received by a FOIA request for specific variables not currently published in the public data (claims timing and claims status).

Geography of claims data is 2010 Census Tract boundaries (until 2020). Census data comes from Longitudinal Tract Database (Brown University), which provides interpolated data between historical census results and 2010 census tract boundaries, and ACS data (2008 - 2022).

Current filters:

  • Valid Census Tract IDs
  • Loss year: 1990 - 2022
  • Occupancy type: Residential, Primary Residences
  • Claim Status: Closed

Total claims records after filtering: 1211814
Note: Claims records will be further reduced due to missing data for control variables selected; see missingness below

General Model

Base Model:

\(Y\) is the outcome variable:

  • \(days\): claims processing time (number of days between claim submitted and claim closed)
  • \(clmratio\): the ratio between the claims payment (accounting for deductible) / assessed damage
  • \(CWOP\): binary indicator for a claim closed without payment

Because the outcome variables vary in data type (i.e. count, continuous, etc.) with a different distribution, the model family will vary and depend on the the outcome being analyzed. For details on model selection, see Claims Analysis: Model Selection.

\(X\) is independent variable of interest (i.e. measures of income, wealth, race).

Covariates are observable claims-level (\(i\)) characteristics that account for flood event severity and damage:
* \(damage\): the ratio between assessed damage / property value (omitted when either component value is included)
* \(waterdepth\): the recorded depth of floodwater * \(duration\): the length of the flooding event * \(cat\): a binary indicator if the claim is associated with a catastrophe * \(reopen\): a binary indicator if the claim was reopened during processing * \(RCV\): a binary indicator for whether claim was paid based on replacement cost value (compared to assessed value)
* State Fixed Effects (\(s\))
* Year Fixed Effects (\(t\))

\[Y = \beta_{0} + \beta_{1}X_{it} + \beta_{2}(damage)_{it} + \beta_{3}(waterdepth)_{it} + \beta_{4}(duration)_{it} + \beta_{5}(cat)_{it} + \beta_{6}(reopen)_{it} + \beta_{7}(RCV)_{it} + \gamma_{t} + \theta_{s} + u_{it}\]

Full Model:

Base Model, plus additional covariates not at claim-level, such as building-level characteristics from NFIP claims data:

  • \(SFHA\): binary indicator if home was located in the special flood hazard zone
  • \(elev\): binary indicator if home was elevated
  • \(bldgage\): age of building
  • \(pop\): Census tract population

\[Y = \beta_{0} + \beta_{1}X_{it} + \beta_{2}(damage)_{it} + \beta_{3}(waterdepth)_{it} + \beta_{4}(duration)_{it} + \beta_{5}(cat)_{it} + \beta_{6}(reopen)_{it} + \beta_{7}(repcost)_{it} + \beta_{8}(SFHA)_{it} + \beta_{9}(elev)_{it} + \beta_{10}(bldgage)_{it} + \beta_{11}(pop)_{it} + \gamma_{t} + \theta_{s} + u_{it}\]

Variation: With state and year fixed effects, we are comparing claims treatment variation between census tracts within each state, relative to national trends that may be affecting claims processing. This model structure therefore looks at how within-state sociodemographic characteristics of tracts correlate with claims.

All independent variables are standardized for easier comparison.

Standard errors are clustered at the state level, since that is the within variation we are comparing here (though I see the argument for tract-level; we should include both in robustness checks). Some references for SE cluster identification:

Variable Selection

Correlation Check of Variables

A check of the correlation across covariates validates our approach to model structure: there is high collinearity between census variables of income, wealth, and race, such that separate models and interactions will be necessary. There is low collinearity between our claims variables. Monthly claim volume is correlated with damage, and catastrophe indicator, such that we should consider omitting one or the other. Notably, there is little correlation between the claim-level property value, and tract-level wealth and income variables - including median home value.

Missing Data

Not all records have data for variables of interest. Using these variables will limit our analysis sample.

Data Record Completeness
fields record_count record_percent
Damage_over_value_bldg 967873 79.9%
WaterDepth 1061809 87.6%
FloodWaterDuration 1108014 91.4%
flag_cat 1211814 100.0%
flag_reopen 1211814 100.0%
flag_SFHA 1211814 100.0%
flag_elevated 1211814 100.0%
bldg_age 1205530 99.5%
pop_total 1169668 96.5%
med_income_thous 1169406 96.5%
income_buckets 1169406 96.5%
pct_pov 1169573 96.5%
PropertyValue_total 1211814 100.0%
pct_owners 1169573 96.5%
pct_edu_col 1169574 96.5%
pct_pop_white 1169574 96.5%
pct_pop_black 1169574 96.5%
pct_pop_hisp 1169574 96.5%

Claims Processing Time

Data Distribution
  • Data removed:
    • Negative claim time data - data errors
    • Top 5% of claim times - extreme outliers; possible data errors (5% of data)
    • Records with undefined damage
    • Claims closed without payment
      – 22% removed –
    • records w/o census data match (see Claims Data Summary)
      – 4% removed –

Percent Total claims records after filtering: 72%

Model Selection

Negative Binomial

Income & Wealth

Using several measures of income (median HH income, poverty, and income buckets), we can see a statistically positive relationship between income of the census tract and claims processing time, meaning that claims from higher-income tracts tend to take longer to resolve. A one standard deviation increase in tract-level income is associated with a 4% increase in claims processing time. This relationship is more pronounced for lower-income census tracts, with median household income < $75,000. Compared to tracts with income with a median income between $100 - $150k, Tracts with median income less than $50,000 dollars experience 10 - 13% shorter claims processing time.

Because there may be a difference between the impact of income and wealth, we look separately at factors that indicate household wealth. While we do not have perfect census measures of wealth, property ownership, educational attainment, and property value (as assessed by insurer at the property level, and in Census at tract level) may be a sufficient proxy for wealth and class separate from income. We see statistically significant, positive correlations between indicators of wealth and claims processing time, to varying degrees. High educational attainment (measured by the percent of population with a college degree), property value, and high percentage of homeownership, has a statistically significant positive relationship with claims processing time.

Race & Income interaction

Because we see a strong correlation between median income and race variables in census tracts, we look primarily at the interaction between race and income. We estimate three separate models looking a the relationship between the percent of population in three broad racial categories (white, black, hispanic), interacted with median income, to understand how the relationship between income and claims processing time differs for tracts with different racial demographics.

We see that for tracts with average income, tracts with higher percentage white population experience longer claims processing time; tracts with higher percentage black and hispanic population experience shorter claims processing time. The interaction between the income and race therefore suggests that while income is positively associated with claims processing time, that positive relationship is stronger across tracts with higher percentage white population, and weaker across tracts with higher black and hispanic population.

Control Variables

Assessed Damage: Buildings

Data Distribution
  • Data removed:
    • Records with undefined damage
    • Claims closed without payment
    • Top 5% of extreme damage ratios (assuming some data entry errors)
    • Damage amount is less than or equal to insurance coverage total (to make data comparable to sample assessed in claim payment analysis)

Percent Total claims records after filtering: 66%

Model Selection

OLS, asinh transformed

Income & Wealth

Race & Income interaction

Claims Payment: Buildings

First, we look at differential treatment in building claims payments. Claims payment is measured as the proportion of assessed (building) damage actually paid by insurer: the ratio of total building payment (including deductible) / assessed building damage. Claims records with assessed damage that exceeds the insurance coverage limit are excluded, such that the payment of full damage is not limited by the insurance coverage amount. NOTE: This is a crude way of ensuring the claims examined here all fall within the coverage limit; we could consider an alternative, where anything with damage > limit and payment = limit means 100% payment.

Data Distribution
  • Data removed:
    • Records with undefined damage
    • Claims closed without payment
    • Top 5% of extreme claim ratios (assuming some data entry errors) Note: Excluding these outliers seems to make a big effect on the patterns observed below
    • Damage amount is less than or equal to insurance coverage total

Percent Total claims records after filtering: 65%

Model Selection

OLS

Income & Wealth

We find a statistical relationship between a tract’s income and wealth and the ratio of the building claim paid, across several measures. Median income is positively associated, and poverty negatively associated, with claims payment. The percentage of population who are college educated and homeowners is also positively associated with claims payment ratios. The relationship is strongest in tracts at the lowest income bracket: tracts with a median income of <$35,000 year receive, on average, significantly less of their assessed damage paid via their building claim. Because we have excluded claims where the assessed damage is above the insurance coverage limit, this result is not a product of differential insurance coverage limits.

Race & Income interaction

Generally, we see that race has a relationship with the claims payment ratio: controlling for income, tracts with higher percentage white population experience statistically higher claims payment ratios, and tracts with higher percentage black and hispanic population see statistically lower claims payments. The interaction between race and income is not statistically significant, meaning the overall relationship between income and payment is not dependent on race.

Claims Processing Time + Claims Amount

We have seen strong evidence that income and wealth is positively associated with income and wealth; we now see that it is also positively associated with claims payments as well. This suggests that claims that are paid faster - while critically important for households recovering from flooding or disasters - may not result in equal claims payment. Here, we estimate the association of claims processing time (loss to claim time, and claim to close time).

We find that claims processing time has a small, but statistically significant, relationship with building claims payment. The time between the loss and claim made, which represents how long it takes a household to enter a claim to the insurer, is negatively associated with claims payment ratios. The time it takes for a claim to be processed, however, is associated with higher claims payment ratios.

Control Variables

Claims Payment: Contents

Next, we look at differential treatment in contents claims payments. Claims payment is measured as the proportion of assessed (contents) damage actually paid by insurer: the ratio of total contents payment (including deductible) / assessed contents damage. Claims records with assessed damage that exceeds the insurance coverage limit are excluded, such that the payment of full damage is not limited by the insurance coverage amount. Note: this is a smaller sample of claims, as many records do not have a recorded contents value

Data Distribution
  • Data removed:
    • Records with undefined damage
    • Claims closed without payment
    • Damage amount is less than or equal to insurance coverage total

Percent Total claims records after filtering: 28%

Model Selection

OLS

Income & Wealth

Generally, we do not see a statistically significant relationship between income and wealth, and content claims payment ratios. We may lack statistical power in these models, since the sample of contents claims is much smaller.

Race & Income interaction

Similarly, we do not see a significant difference in how income affects content claims payment by race.

Claims Processing Time + Claims Amount

We do see a statistically positive relationship between claims processing and claims submission time, but the relationship is not practically significant.

Control Variables

Claims Closed without Payment

Many claims in the insurance claims record (~ 23%) are closed without payment. Here, we consider any claims that are either labeled as “CWOP” in claims status, or labeled as “closed” but resulting in $0 claim payment, and the assessed value of damage exceeds the deductible.

Income & Wealth

There is a statistically significant negative relationship between several measures of income and wealth, and claims closed without payment. Tracts with more homeownership, higher property value, and higher income are associated with lower rates of claims closed without payment. Just as in the claims payment and claims timing, the relationship is more significant in tracts within the lowest income bracket.

Race & Income interaction

Controls

Claims from RCV vs ACV

Income & Wealth

Further Analysis

  • Incident-specific dynamics (i.e. Hurricane)
  • County fixed effects
  • Other dependent variables to test:
    • Usage or denials of ICC

  1. You, X. & Kousky, C. Improving household and community disaster recovery: Evidence on the role of insurance. Journal of Risk & Insurance. (2024) doi: 10.1111/jori12466↩︎

  2. Gallagher, J., Hartley, D. & Rohlin, S. Weathering an Unexpected Financial Shock: The Role of Federal Disaster Assistance on Household Finance and Business Survival. Journal of the Association of Environmental and Resource Economists 10, 525-567 (2023)↩︎

  3. Squires, G.D. Racial Profiling, Insurance Style: Insurance Redlining and the Uneven Development of Metropolitan Areas. Journal of Urban Affairs 25, 391-410 (2003)↩︎

  4. Schwarcz, D. Transparently Opaque: Understanding the Lack of Transparency in Insurance Consumer Protection. UCLA Law Review 61, 394 (2014)↩︎