Credit-Based Insurance Scoring And Measurement Error In The FTC Race Proxy Finding

Following is the testimony presented by Lawrence S. Powell, PhD before the United States House of Representatives Financial Services Committee Oversight & Investigations Subcommittee on May 21, 2008, by:

Lawrence S. Powell, Ph.D.
Research Fellow - The Independent Institute (http://www.independent.org), and
Whitbeck-Beyer Chair of Insurance & Financial Services
University of Arkansas-Little Rock
***************************************

Chairman Watt, Ranking Member Miller, and Members of the Subcommittee, it was my pleasure to share information with you about insurance scoring during the hearing on May 21, 2008. In response to Representative Waters' request, I offer the following explanation of my disagreement with the FTC finding that insurance scores display a proxy effect for race.

The FTC reports, with reservation, statistically significant race and ethnicity "proxy effects" within the predictive ability of insurance credit scores. I absolutely believe this finding is mathematically incorrect.

The FTC study also includes other findings contrary to insurance scores being proxies for race and ethnicity. They state "the relationship between scores and claims risk remains strong when controls for race, ethnicity, and neighborhood income are included in statistical models of risk." In addition, they find "tests also showed that scores predict insurance risk within racial and ethnic minority groups (e.g., Hispanics with lower scores have higher estimated risk than Hispanics with higher scores). This within-group effect of scores is inconsistent with the theory that scores are solely a proxy for race and ethnicity." Collectively, the lack of objective confidence in the existence of race or ethnicity proxy effects, and the evidence inconsistent with a proxy effect, demonstrate that public policy should not be altered to address this weak finding.

The Race Proxy "Finding"

The FTC report claims to find statistically significant evidence that insurance scores include a "proxy effect" for race. To understand what they find, it is important to understand - at least a little bit - about what the analytical models they use actually test. The "Tweedie GLM" model used in the FTC report is a modified regression model. A regression model measures how much of the variation in a dependent variable (predicted loss) can be explained by variation in an independent variable (race, ethnicity, income, or credit score) while controlling for other independent or control variables (geographic location, age, driving record, etc.).

The authors define a race proxy effect as a change in expected losses due to using insurance scores in the model that cannot be explained by a factor other than race. To test for the effect, they estimate predicted losses for individuals in the sample with and without insurance scores and explicit race and ethnicity controls in the model. Next, they compare predicted losses of each group (African Americans, Hispanics, Asians, and Non-Hispanic Whites) with and without insurance scores. This is reported in Column (a) of Table 7 from the FTC report (copied below). Column (b) shows the same percentage differences when race, ethnicity, and income are explicitly controlled for in the model. The results for African Americans in Column (a) is 10%, and in Column (b) it is 8.9%, a difference of 1.1%. Thus, if all other aspects of the model were reliable, one might conclude that, of the 10% difference in expected losses from using insurance scores 1.1% is attributable to race.

Careful objective review of the FTC analysis leads me to conclude without reservation that flaws in the model render the race and ethnicity proxy findings invalid. The technical term for the flaw in the model is omitted variable bias. It is a form of measurement error in which one variable is not measured accurately and, as a result, its effect is attributed to another variable. Results from the FTC model suggest strongly that territorial risk is not measured adequately, and the incorrect finding of a "proxy effect" is actually attributable to this measurement error.

The territorial risk variable used in the FTC study is not the same as territorial risk controls used by insurance companies. The FTC created a national territorial risk variable, whereas insurers make rates within each state. Calculation of the variable is described in the report as follows:

Territorial Risk Variable
The five firms also submitted to EPIC data on earned car years and claims on property damage liability policies by ZIP code for a three-year period from 2000 to 2002, for their full book of business. EPIC combined the data from the five firms to calculate ZIP-code level average property damage liability pure premiums (i.e., average dollars paid out per year of coverage per car)*. This is an improvement over the original Census-based population density measure that EPIC used in its report. The new ZIP code risk variable was included in the policy database EPIC forwarded to the FTC.
(*For ZIP codes with fewer than 3,000 property damage liability claims, data from surrounding ZIP codes were also used to calculate average pure premiums.)

The zip codes were then ranked by quintile of property damage liability claims. The variables that enter the final model are a series of indicators of the zip code quintile.

Territorial risk is an important predictor of risk because it describes the area where insured automobiles are garaged and driven. However, territorial risk may differ for several reasons. There could be differences in claim frequency due to traffic density, propensity to litigate, moral hazard and fraud, the population that could be injured, or many other reasons. These differences cannot be measured adequately by simply grouping territories into national zip code quintiles.

Further evidence that measurement error in the territorial risk variable is responsible for the race proxy effect is found by comparing results across coverage categories. Because the territorial risk measure is calculated using property damage liability data, it should be most accurate when applied to property damage liability claims. In the Property Damage Liability Coverage column, there is not a statistically significant valid proxy finding. In the other three columns, the magnitude of the estimated proxy effect grows as the expected measurement error from using a territorial risk calculated with PDL claims grows. However, the FTC study reports the sum of the effect on all four types of coverage to arrive at the 1.1% effect for African Americans and the 0.7% effect for Hispanics.

The differences in accuracy of the territorial risk measure across coverage types merit further explanation. As mentioned above, the territory measure is a function of third party property damage liability (PDL) losses by zip code. It is then applied, with noted reservation in the study, to claims for bodily injury liability (BIL), and first party property damage claims referred to in the study as collision (COL) and comprehensive (COM).

PDL claims pay for damage to property owned by a third party that is damaged as a result of the insured driver's negligence. They provide the most accurate estimate of driving ability because they represent almost all potential BIL claims**, and the majority of potential COL claims. They are also not subject to claiming behavior influences such as the impact of deductibles and moral hazard.
(**One exception is an automobile versus pedestrian collision in which the injured party's property is not damaged. However, such claims represent only a small portion of BIL claims.)

BIL claims pay for bodily injury to a third party resulting from the insured driver's negligence. Nearly all BIL losses also involve a PDL loss, but the amount of damage from a BIL loss is much harder to determine objectively because it may include damages for pain and suffering. This lack of objectivity is found to create large differences in claiming behavior across territories (Hoyt, Mustard and Powell, 2006; Cummins and Tennyson, 1996). Therefore, a measure of territorial risk derived from PDL losses will not be accurate when applied to BIL losses.

A similar problem exists for COL claims. COL claims pay for property damage to one's own vehicle when another party is not liable for the loss. A territorial risk measure derived from PDL claims will not accurately reflect COL claim risk for two reasons. First, while PDL and BIL coverage is mandatory for all drivers, COL is not. Because many drivers do not carry COL coverage, the measure will be biased by differences in this coverage across territories. Second, claiming behavior affects these losses because a deductible applies to each occurrence or claim.

Finally, COM claims pay for damage to an insured automobile from perils other than collision with another vehicle or object. Perils covered by COM include fire, earthquake, windstorm, larceny, and malicious mischief. Thus, as the FTC study implies, there is no reason to assume a territorial risk measure derived from PDL claims would apply to COM claims. Nonetheless, the FTC study includes the race proxy finding estimated from COM claims in its conclusion.

The preceding discussion of differences across coverage type leaves little room for doubt that the estimates of a proxy effect, while still questionable, would be most accurate for PDL losses. The FTC report does not find a proxy effect for African Americans or Hispanics when analyzing PDL claims. However, the proxy effect finding presented in the FTC study represents the sum of the effects measured for PDL (proxy effect =0), BIL, COL, and COM, potentially leading readers to an incorrect conclusion. Table 1 shows that as the degree of expected measurement error from the territorial risk variable increases across coverage types, the estimated proxy effect increases, suggesting the effect is actually the result of measurement error in the territorial risk variable.

Collectively, the lack of objective confidence in the existence of race or ethnicity proxy effects, and the evidence inconsistent with a proxy effect, demonstrate that public policy should not be altered to address this weak finding.

References

• Cummins, J. David, and Sharon Tennyson, 1996. "Moral Hazard in Insurance Claiming: Evidence from Automobile Insurance." Journal of Risk and Uncertainty 12 (1996): 29-50.
• Hoyt, Robert E., David B. Mustard, and Lawrence S. Powell, 2006. "The Effectiveness of State Legislation in Mitigating Moral Hazard: Evidence from Automobile Insurance," Journal of Law and Economics, v49 (October 2006): 427-450.

The views expressed in this article/commentary are solely those of the author and do not necessarily represent the views of MyNewMarkets.com, the Insurance Journal or Wells Publishing.