Credit-Based Insurance Scores:; Impacts on Consumers of Automobile Insurance:; A Report to Congress By the Federal Trade Commission (July 2007)

CREDIT-BASED INSURANCE SCORES:

IMPACTS ON CONSUMERS

OF AUTOMOBILE INSURANCE

A Report to Congress by the

Federal Trade Commission

July 2007

FEDERAL TRADE COMMISSION

Deborah Platt Majoras Chairman

Pamela Jones Harbour Commissioner

Jon Leibowitz Commissioner

William E. Kovacic Commissioner

J. Thomas Rosch Commissioner

Bureau of Economics

Michael R. Baye Director

Paul A. Pautler Deputy Director for Consumer Protection

Jesse B. Leary Assistant Director, Division of Consumer Protection

Bureau of Consumer Protection

Lydia B. Parnes Director

Mary Beth Richards Deputy Director

Peggy Twohig Associate Director, Division of Financial Practices

Thomas B. Pahl Assistant Director, Division of Financial Practices

Analysis Team

Matias Barenstein, Economist, Bureau of Economics, Div. of Consumer Protection

Archan Ruparel, Research Analyst, Bureau of Economics, Div. of Consumer Protection

Raymond K. Thompson, Research Analyst, Bureau of Economics, Div. of Consumer Protection

Other Contributors

Erik W

. Durbin, Dept. Assistant Director, Bureau of Economics, Div. of Consumer Protection

Christopher R. Kelley, Research Analyst, Bureau of Economics, Div. of Consumer Protection

Kenneth H. Kelly, Economist, Bureau of Economics, Div. of Consumer Protection

Michael J. Pickford, Research Analyst, Bureau of Economics, Div. of Consumer Protection

W. Russell Porter, Economist, Bureau of Economics, Div. of Consumer Protection

TABLE OF CONTENTS i

LIST OF TABLES iii

LIST OF FIGURES iv

I. EXECUTIVE SUMMARY 1

II. INTRODUCTION 5

III. DEVELOPMENT AND USE OF CREDIT-BASED INSURANCE SCORES 7

A. Background and Historical Experience 7

B. Development of Credit-Based Insurance Scores 12

C. Use of Credit-Based Insurance Scores 15

D. State Restrictions on Scores 17

IV. THE RELATIONSHIP BETWEEN CREDIT HISTORY AND RISK 20

A. Correlation Between Credit History and Risk 20

1. Prior Research 20

2. Commission Research 23

a. FTC Database 23

b. Other Data Sources 28

B. Potential Causal Link between Scores and Risk 30

V. EFFECT OF CREDIT-BASED INSURANCE SCORES ON PRICE

AND AVAILABILITY 34

A. Credit-Based Insurance Scores and Cross-Subsidization 35

1. Possible Impact on Car Ownership 39

2. Possible Impact on Uninsured Driving 40

3. Adverse Selection 42

B. Other Possible Effects of Credit-Based Insurance Scores 46

C. Effects on Residual Markets for Automobile Insurance 49

VI. EFFECTS OF SCORES ON PROTECTED CLASSES OF CONSUMERS 50

A. Credit- Based Insurance Scores and Racial, Ethnic, and Income Groups 51

1. Difference in Scores Across Groups 51

2. Possible Reasons for Differences in Scores Across Groups 56

3. Impact of Differences in Scores on Premiums Paid 58

a. Effect on Those for Whom Scores Were Available 58

b. Effect on Those for Whom Scores Were Not Available 59

B. Scores as a Proxy for Race and Ethnicity 61

1. Do Scores Act Solely as a Proxy for Race, Ethnicity, or Income? 62

2. Differences in Average Risk by Race, Ethnicity, and Income 64

3. Controlling for Race, Ethnicity, and Income to Test for a Proxy Effect 67

a. Existence of a Proxy Effect 67

b. Magnitude of a Proxy Effect 69

VII. ALTERNATE SCORING MODELS 73

A. The FTC Baseline Model 74

B. Alternative Scoring Models 78

1. “Race Neutral” Scoring Models 78

2. Model Discounting Variables with Large Differences by Race and

Ethnicity 80

VIII. CONCLUSION 82

TABLES

FIGURES

APPENDIX A. Text of Section 215 of the FACT ACT

APPENDIX B. Requests for Public Comment

APPENDIX C. The Automobile Policy Database

APPENDIX D. Modeling and Analysis Details

APPENDIX E. The Score Building Procedure

APPENDIX F. Robustness Checks and Limitations of the Analysis

iii

TABLES

TABLE 1. Typical Information Used in Credit-Based Insurance Scoring Models

TABLE 2. Claim Frequency, Claim Severity, and Average Total Amount Paid on

Claims

TABLE 3. Median Income and Age, and Gender Make-Up, by Race and Ethnicity

TABLE 4. Change in Predicted Amount Paid on Claims from Using Credit-Based

Insurance Scores, by Race and Ethnicity

TABLE 5. Estimated Relative Amount Paid on Claims, by Race, Ethnicity, and

Neighborhood Income

TABLE 6. Estimated Relative Amount Paid on Claims, by Score Decile, Race,

Ethnicity, and Neighborhood Income

TABLE 7. Change in Predicted Amount Paid on Claims from Using Credit-Based

Insurance Scores Without and With Controls for Race, Ethnicity, and

Income, by Race and Ethnicity

TABLE 8. Change in Predicted Amount Paid on Claims from Using Other Risk

Variables, Without and With Controls for Race, Ethnicity, and Income, by

Race and Ethnicity

TABLE 9. Baseline Credit-Based Insurance Scoring Model Developed by the FTC

TABLE 10. Credit-Based Insurance Scoring Model Developed by the FTC by

Including Controls for Race, Ethnicity, and Neighborhood Income in the

Score-Building Process

TABLE 11. Credit-Based Insurance Scoring Model Developed by the FTC Using a

Sample of Only Non-Hispanic White Insurance Customers

TABLE 12. Credit-Based Insurance Scoring Model Developed by the FTC by

Discounting Variables with Large Differences Across Racial and Ethnic

Groups

FIGURES

FIGURE 1. Estimated Average Amount Paid Out on Claims, Relative to Highest

Score Decile

FIGURE 2. Frequency and Average Size (Severity) of Claims, Relative to Highest

Score Decile

FIGURE 3. "CLUE" Claims Data: Average Amount Paid Out on Claims, Relative to

Highest Score Decile

FIGURE 4. By Model Year of Car: Estimated Average Amount Paid Out on Claims,

Relative to Highest Score Decile (Property Damage Liability Coverage)

FIGURE 5. Change in Predicted Amount Paid on Claims from Using Scores

FIGURE 6. The Ratio of Uninsured Motorist Claims to Liability Coverage Claims

(1996-2003)

FIGURE 7. Share of Cars Insured through States' "Residual Market" Insurance

Programs (1996-2003)

FIGURE 8. Distribution of Scores, by Race and Ethnicity

FIGURE 9. Distribution of Race and Ethnicity, by Score Decile

FIGURE 10. Distribution of Scores, by Neighborhood Income

FIGURE 11. Distribution of Neighborhood Income, by Score Decile

FIGURE 12. Distribution of Scores by Race and Ethnicity, After Controlling for Age,

Gender, and Neighborhood Income

FIGURE 13. By Race and Ethnicity: Change in Predicted Amount Paid on Claims from

Using Scores, by Race and Ethnicity

FIGURE 14. By Race and Ethnicity: Estimated Average Amount Paid Out on Claims,

Relative to Non-Hispanic Whites in Highest Score Decile

FIGURE 15. By Neighborhood Income: Estimated Average Amount Paid Out on

Claims, Relative to People in Highest Score Decile in High Income Areas

FIGURE 16. Estimated Average Amount Paid Out on Claims, Relative to Highest

Score Decile, with and without Controls for Race, Ethnicity, and

Neighborhood Income

FIGURE 17. FTC Baseline Model - Estimated Average Amount Paid Out on Claims,

Relative to Highest Score Decile

FIGURE 18. Distribution of FTC Baseline Model Credit-Based Insurance Scores, by

Race and Ethnicity

FIGURE 19. FTC Score Models with Controls for Race, Ethnicity, and Neighborhood

Income: Estimated Average Amount Paid Out on Claims, Relative to

Highest Score Decile

FIGURE 20. Distribution of FTC Credit-Based Insurance Scores, by Race and Ethnicity

FIGURE 21. An Additional FTC Credit-Based Insurance Scoring Model: The

"Discounted Predictiveness" Model Estimated Average Amount Paid Out

on Claims, Relative to Highest Score Decile

FIGURE 22. Distribution of FTC Credit-Based Insurance Scores, by Race and Ethnicity

I. EXECUTIVE SUMMARY

Section 215 of the FACT Act (FACTA)

requires the Federal Trade Commission

(FTC or the Commission) and the Federal Reserve Board (FRB), in consultation with the

Department of Housing and Urban Development, to study whether credit scores and

credit-based insurance scores affect the availability and affordability of consumer credit,

as well as automobile and homeowners insurance. FACTA also directs the agencies to

assess and report on how these scores are calculated and used; their effects on consumers,

specifically their impact on certain groups of consumers, such as low-income consumers,

racial and ethnic minority consumers, etc.; and whether alternative scoring models could

be developed that would predict risk in a manner comparable to current models but have

smaller differences in scores between different groups of consumers. The Commission

issues this report to address credit-based insurance scores

primarily in the context of

automobile insurance.

Credit-based insurance scores, like credit scores, are numerical summaries of

consumers’ credit histories. Credit-based insurance scores typically are calculated using

information about past delinquencies or information on the public record (e.g.,

bankruptcies); debt ratios (i.e., how close a consumer is to his or her credit limit);

evidence of seeking new credit (e.g., inquiries and new accounts); the length and age of

credit history; and the use of certain types of credit (e.g., automobile loans). Insurance

15 U.S.C. § 1681 note (2006). Appendix A contains the complete text of Section 215 of the FACT Act.

The FRB will submit a report addressing issues related to the use of credit scores and consumer credit

decisions.

The Commission will conduct an empirical analysis of the effects of credit-based insurance scores on

issues relating to homeowners insurance; the FTC anticipates that it will submit a report to Congress

describing the results of this analysis in early 2008.

companies do not use credit-based insurance scores to predict payment behavior, such as

whether premiums will be paid. Rather, they use scores as a factor when estimating the

number or total cost of insurance claims that prospective customers (or customers

renewing their policies) are likely to file.

Credit-based insurance scores evolved from traditional credit scores, and

insurance companies began to use insurance scores in the mid-1990s. Since that time,

their use has grown very rapidly. Today, all major automobile insurance companies use

credit-based insurance scores in some capacity. Insurers use these scores to assign

consumers to risk pools and to determine the premiums that they pay.

Insurance companies argue that credit-based insurance scores assist them in

evaluating insurance risk more accurately, thereby helping them charge individual

consumers premiums that conform more closely to the insurance risk they actually pose.

Others criticize credit-based insurance scores on the grounds that there is no persuasive

reason that a consumer’s credit history should help predict insurance risk. Moreover,

others contend that the use of these scores results in low-income consumers and members

of minority groups paying higher premiums than other consumers.

Pursuant to FACTA, the FTC evaluated: (1) how credit-based insurance scores are

developed and used; and, in the context of automobile insurance (2) the relationship

between scores and risk; (3) possible causes of this relationship; (4) the effect of scores

on the price and availability of insurance; (5) the impact of scores on racial and ethnic

minority groups and on low-income groups; and (6) whether alternative scoring models

are available that predict risk as well as current models and narrow the differences in

scores among racial, ethnic, and other particular groups of consumers. In conducting this

evaluation, the Commission considered prior research, nearly 200 comments submitted in

response to requests for the public’s views, information presented in meetings with a

variety of interested parties, and its own original empirical research using a database of

automobile insurance policies. Based on a careful and comprehensive consideration of

this information, the FTC has reached the following findings and conclusions:

● Insurance companies increasingly are using credit-based insurance scores

in deciding whether and at what price to offer coverage to consumers.

● Credit-based insurance scores are effective predictors of risk under

automobile policies. They are predictive of the number of claims

consumers file and the total cost of those claims. The use of scores is

therefore likely to make the price of insurance better match the risk of loss

posed by the consumer. Thus, on average, higher-risk consumers will pay

higher premiums and lower-risk consumers will pay lower premiums.

● Several alternative explanations for the source of the correlation between

credit-based insurance scores and risk have been suggested. At this time,

there is not sufficient evidence to judge which of these explanations, if

any, is correct.

● Use of credit-based insurance scores may result in benefits for consumers.

For example, scores permit insurance companies to evaluate risk with

greater accuracy, which may make them more willing to offer insurance to

higher-risk consumers for whom they would otherwise not be able to

determine an appropriate premium. Scores also may make the process of

granting and pricing insurance quicker and cheaper, cost savings that may

be passed on to consumers in the form of lower premiums. However, little

hard data was submitted or available to quantify the magnitude of these

benefits to consumers.

● Credit-based insurance scores are distributed differently among racial and

ethnic groups, and this difference is likely to have an effect on the

insurance premiums that these groups pay, on average.

▪ Non-Hispanic whites and Asians are distributed relatively evenly

over the range of scores, while African Americans and Hispanics

are substantially overrepresented among consumers with the

lowest scores (the scores associated with the highest predicted risk)

and substantially underrepresented among those with the highest

scores.

▪ With the use of scores for consumers whose information was

included in the FTC’s database, the average predicted risk (as

measured by the total cost of claims filed) for African Americans

and Hispanics increased by 10% and 4.2%, respectively, while the

average predicted risk for non-Hispanic whites and Asians

decreased by 1.6% and 4.9%, respectively.

● Credit-based insurance scores appear to have little effect as a “proxy” for

membership in racial and ethnic groups in decisions related to insurance.

▪ 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 models with credit-based insurance scores but without controls

for race or ethnicity, African Americans and Hispanics are

predicted to have average predicted risk 10% and 4.2% higher,

respectively, than if scores were not used. In models with scores

and with controls for race, ethnicity, and income, these groups

have average predicted risk 8.9% and 3.5% higher, respectively

than if scores were not used. The difference between these two

predictions for African Americans and Hispanics (1.1% and 0.7%,

respectively) is a measure of the effect of scores on these groups

that is attributable to scores serving as a statistical proxy for race

and ethnicity.

▪ Several other variables in the FTC’s database (e.g., the time period

that a consumer has been a customer of a particular firm) have a

proportional proxy effect that is similar in magnitude to the small

proxy effect associated with credit-based insurance scores.

▪ 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.

● After trying a variety of approaches, the FTC was not able to develop an

alternative credit-based insurance scoring model that would continue to

predict risk effectively, yet decrease the differences in scores on average

among racial and ethnic groups. This does not mean that a model could

not be constructed that meets both of these objectives. It does strongly

suggest, however, that there is no readily available scoring model that

would do so.

II. INTRODUCTION

Over the past decade, insurance companies increasingly have used information

about credit history in the form of credit-based insurance scores to make decisions

whether to offer insurance to consumers, and, if so, at what price. Because of the

importance of insurance in the daily lives of consumers, the widespread use of these

scores raises questions about their impact on consumers. In particular, some have

expressed concerns about the effect of scores on the availability and affordability of

insurance to members of certain demographic groups, especially racial and ethnic

minorities.

In 2003, Congress enacted the Fair and Accurate Credit Transactions Act

(FACTA) to make comprehensive changes to the nation’s system of handling consumer

credit information. In response to concerns that had been raised about credit-based

insurance scores, in Section 215 of FACTA Congress directed certain federal agencies,

including the FTC, to conduct a broad and rigorous inquiry into the effects of these scores

and submit a report to Congress with findings and conclusions. The report is intended to

provide policymakers with critical information to enable them to make informed

decisions with regard to credit-based insurance scores.

Section 215 of FACTA sets forth specific requirements for studying the effects of

credit-based insurance scores in the context of automobile and homeowners insurance. It

directs the agencies to include a description of how these scores are created and used, as

well as an assessment of the impact of scores on the availability and affordability of

automobile and homeowners insurance products. Section 215 also requires a rigorous

and empirically sound statistical analysis of the relationship between scores and

membership in racial, ethnic, and other protected classes. The mandated study further

must evaluate whether scores act as a proxy for membership in racial, ethnic, and other

protected classes. Finally, Section 215 requires an analysis of whether scoring models

could be constructed that both are effective predictors of risk and result in narrower

differences in scores among racial, ethnic, and other protected classes.

Section 215 of FACTA also specifies the process to be used in conducting the

study, and the contents of the report to be submitted. The Act directed the agencies to

seek input from federal and state regulators and consumer and civil rights organizations,

and members of the public concerning methodology and research design. The Act

requires the report to include “findings and conclusions of the Commission,

recommendations to address specific areas of concerns addressed in the study, and

recommendations for legislative or administrative action that the Commission may

determine to be necessary to ensure that . . . credit-based insurance scores are used

appropriately and fairly to avoid negative effects.”

The Commission has conducted a study addressing credit-based insurance scores

in the context of automobile insurance. Pursuant to statutory directive, the FTC

published two Federal Register Notices

soliciting comments from the public concerning

methodology and research design. The Commission supplemented this information with

numerous discussions between its staff and representatives of other government agencies,

private companies, and community, civil rights, consumer, and housing groups. The

public comments and information obtained in meetings with the various interested parties

15 U.S.C. § 1681 note (2006).

Public Comment on Data, Studies, or Other Evidence Related to the Effects of Credit Scores and Credit-

Based Insurance Scores on the Availability and Affordability of Financial Products, 70 Fed. Reg. 9652

(Feb. 28, 2005); Public Comment on Methodology and Research Design for Conducting a Study of the

Effects of Credit Scores and Credit-Based Insurance Scores on Availability and Affordability of Financial

Products, 69 Fed. Reg. 34167 (June 18, 2004).

provided essential information that allowed the Commission to complete this report. In

addition, feedback from state regulators, industry participants, and the consumer, civil

rights, and housing groups had a substantial impact on the methodology and scope of the

analysis.

This report discusses the information that the FTC considered, its analysis of that

information, and its findings and conclusions. Parts I and II above present an Executive

Summary and Introduction, respectively. Part III is an overview of the development and

use of credit-based insurance scores, and Part IV discusses the relationship between

credit history and risk. Part V addresses the effect of credit-based insurance scores on the

price and availability of insurance. Part VI explores the impact of credit-based insurance

scores on racial, ethnic, and other groups. Part VII describes the FTC’s efforts to develop

a model that reduces differences for protected classes of consumers while continuing to

effectively predict risk. Part VIII is a brief conclusion.

III. DEVELOPMENT AND USE OF CREDIT-BASED INSURANCE SCORES

A. Background and Historical Experience

Consumers purchase insurance to protect themselves against the risk of suffering

losses. They tend to be “risk averse,” that is, consumers would prefer the certainty of

paying the expected value of a loss to the possibility of bearing the full amount of the

loss.

For example, assume that a driver faces a 1% risk of being in an automobile

accident that would cause him or her to suffer a $10,000 loss, which means that the

expected value of his or her loss is $100 (1% of $10,000). If the driver is risk averse, he

or she would be willing to pay $100 or more to avoid the possible loss of $10,000.

What makes insurance markets possible is that insurance companies do not

simply take on the risk of their customers, they actually reduce risk. This does not mean

that they reduce the total losses from car accidents or house fires, for example, but rather

that they reduce the uncertainty that individuals face without themselves facing nearly the

same amount of uncertainty. This is possible because the average loss on a large number

of policies can be predicted much more accurately than the losses of a single driver or

homeowner. For instance, while it is extremely difficult to predict who among a group of

100,000 drivers will have an accident, it may be possible to predict the total number of

accidents for these 100,000 drivers with a low margin of error.

By selling many policies

that cover the possible losses for many consumers, an insurance company faces much

lower uncertainty as to total losses than would each consumer if they did not purchase

insurance.

Insurance companies have a strong economic incentive to try to predict risk as

accurately as possible. In a competitive market for insurance in which all firms have

access to the same information about risk, competition for customers will force insurance

companies to offer the lowest rates that cover the expected cost of each policy sold. If an

insurance company is able to predict risk better than its competitors, it can identify

consumers who currently are paying more than they should based on the risk they pose,

and target these consumers by offering them a slightly lower price. Thus, developing and

using better risk prediction methods is an important form of competition among insurance

companies.

This risk reduction is due to the “law of large numbers.” Uncertainty is reduced as long as there is a

sufficient degree of independence among the risk that individual consumers face. For example, selling

flood insurance to those who live in a single flood plain reduces risks less than selling the policies to those

who live in a broader geographic area.

For decades, insurance companies have divided consumers into groups based on

common characteristics which correlate with risk of loss. Automobile insurance

companies divide consumers into groups based on factors such as age, gender, marital

status, place of residence, and driving history, among others. Once insurance companies

have separated consumers into groups based on these characteristics, they use the average

risk of each of these groups in helping to determine the price to charge members of the

group.

Insurance companies report that during the last decade they have begun to use

credit-based insurance scores to assist them in separating consumers into groups based on

risk. Insurers have long used some credit history information when evaluating insurance

applications, for example, considering bankruptcy in connection with offering

homeowners insurance. In the early 1980s, insurance companies and others began

assessing the utility of using additional information about credit history in assessing risk,

leading to a more formal use of such information in a fairly simple manner by the early

1990s.

In the early 1990s, Fair Isaac Corporation (Fair Isaac), drawing on its experience

developing credit scores, led the initial research to develop credit-based insurance scores.

The company developed the first “modern” credit-based insurance score and made it

available to insurance companies in 1993.

This score was developed to predict the

likelihood of claims being submitted for homeowners policies. Fair Isaac introduced a

credit-based insurance score for automobile policies in 1995, and ChoicePoint introduced

Meeting between FTC staff and State Farm (July 13, 2004); Meeting between FTC staff and MetLife

Home and Auto (July 12, 2004); Meeting between FTC staff and Allstate (June 23, 2004).

E-mail from Karlene Bowen, Fair Isaac, to Jesse Leary, Assistant Director, Division of Consumer

Protection, Bureau of Economics (Jan. 30, 2006) (on file with FTC).

a competing score at about the same time.

These scores were developed to predict the

loss ratios – claims paid out divided by premiums received – of automobile policies.

Following the introduction of these third-party scores, some insurance companies began

developing and using their own proprietary scores.

Since the mid-1990s, the use of credit-based insurance scores has grown

dramatically. According to industry sources, some of this growth is attributable to

changes in technology and industry practices that have made it easier for companies to

develop

and use these scores.

For example, during the 1990s insurance company

actuaries began using advanced statistical techniques that made it easier to control for

many predictive variables at the same time.

This made it easier for them to develop

proprietary scores and perhaps made them more receptive to using third-party scores.

Insurers also explained that at this time they began combining more and more data from

throughout their companies into integrated databases, and this “data warehousing” made

it much easier for actuaries and others to engage in the research needed to develop

scores.

More fundamentally, however, insurance companies increasingly used credit-

based insurance scores because their experience revealed that they were effective

Id.; E-mail from John Wilson, ChoicePoint, to Jesse Leary, Assistant Director, Division of Consumer

Protection, Bureau of Economics (June 13, 2005) (on file with FTC).

Developing scores is a fairly expensive process, requiring significant information technology resources

and technical expertise. It also requires a large amount of data on loss experience. Many smaller firms,

and even some larger firms, therefore do not develop their own scores. See, e.g., Lamont Boyd, Fair Isaac

Corporation, Remarks at the Fair Isaac Consumer Empowerment Forum (Sept. 2006) (noting only six firms

use a proprietary scoring model).

Industry participants estimate that of the firms that use credit-based risk scores, one-half (as measured by

market share) use a proprietary score and one-half use a score that others developed. Among insurers who

use a non-proprietary score, about two-thirds use a ChoicePoint score, and one-third use a Fair Isaac score.

These techniques are known as Generalized Linear Models (GLMs). GLMs make it easier to control for

many predictive variables at once, and can be used to develop credit-based scoring models. GLMs play a

central role in the analysis presented in this report, and are discussed in more detail in Appendix D.

Meeting between FTC staff and The Hartford (July 14, 2004).

predictors of risk. For example, according to a published case study, in the early 1990s,

Progressive entered the lower-risk portion of the automobile insurance market.

Progressive used sophisticated risk prediction techniques that it had developed in its other

lines of business to identify consumers who other insurers were overcharging relative to

the risk they posed. Progressive offered these consumers the same coverage at a lower

price, thereby persuading some of them to switch to Progressive.

The success of

Progressive’s strategy provided a powerful incentive for incumbent firms to improve

their own risk prediction techniques to compete more effectively.

Many of them

responded to this incentive by increasing their development and use of credit-based

insurance risk scores.

Insurance companies now widely use credit-based insurance scores. Today, the

fifteen largest automobile insurers (with a combined market share of 72% in 2005) all

utilize these scores.

Many smaller automobile insurers also use credit-based insurance

scores.

The development and increased use of credit-based insurance scores has been

accompanied by concerns and criticisms about the validity of the underlying relationship

between scores and risk and the fundamental fairness of using credit history information

to make decisions about insurance. According to critics, credit-based insurance scores: 1)

See, e.g., F. Frei, Innovation at Progressive (A): Pay as You Go Insurance, Harv. Bus. Sch. Case Study

9-602-175 (Apr. 29, 2004).

Incumbent firms had an incentive to use the new risk prediction technology in any case. The vigorous

competition of Progressive, however, likely spurred incumbent firms to move more aggressively to use this

technology than they otherwise would have.

See id.

National Association of Insurance Commissioners, “Auto Insurance Database Report 2003/2004” (2006)

(on file with the FTC); FTC staff reviews of websites and discussions with industry representatives. No

market share data more recent than 2005 was available.

Fair Isaac Corporation states that it sells credit-based insurance scores to roughly 350 firms. Comment

from Fair Isaac Corp. to FTC at 14 (Apr. 25, 2005), [hereinafter Fair Isaac Comment], available at

http://www.ftc.gov/os/comments/FACTA-implementscorestudy/514719-00090.pdf

unfairly penalize consumers who have suffered from medical or economic crises, or who

have made perfectly legitimate financial decisions that are penalized by scoring models;

2) affect consumers in arbitrary ways, because credit history information may contain

errors; and, 3) have a negative impact on minority and low-income consumers.

B. Development of Credit-Based Insurance Scores

According to score developers and insurance companies, credit-based insurance

scores are developed in the same manner as credit scores generally. To construct a

model, score developers obtain a sample of insurance policies for which losses are

known. The period of time during which losses occurred or could have occurred is called

the “exposure period.” Score developers start with the credit information available about

customers at the beginning of the exposure period and the known losses for them during

the period. Score developers then use various statistical and other techniques to develop

a model that predicts losses based on the credit information that was available at the start

of the exposure period. If the relationship between the credit information and loss is

sufficiently stable over time, the model can be applied to the credit histories of other

consumers to predict the risk of loss they pose.

The details of the credit information used in particular models that produce credit-

based insurance scores generally are not available. As emphasized above, insurance

companies assert that risk prediction techniques are an important form of competition, so

Hearing Before the New York State Assembly Comm. on. Ins. (Oct. 22, 2003) (statement of Birny

Birnbaum, Executive Director, Center for Economic Justice).

firms generally do not want to reveal the credit-based insurance scoring models they

use.

Some states require by law that insurance companies make their models public.

Insurance companies, however, explained that most insurance companies develop and use

different scoring models in these states than they use in other states to minimize the

competitive disadvantage elsewhere as a result of such mandated disclosures. An

important exception is ChoicePoint, which has made its Attract Auto Scoring and other

models available to the public.

Based on the information the agency reviewed, a general picture of what data are

used in credit-based insurance scoring model emerges.

Table 1 presents examples of

the types of information that often are used in models to predict credit-based insurance

scores. Firms, however, vary significantly in the particular information they use in their

models. For example, some insurance companies consider the type of credit granted,

while others do not. Moreover, within a category of information, firms may consider

different variables in calculating credit-based insurance scores. For instance, an

insurance company may use the age of the oldest account in a credit report or may

consider the average age of all accounts in the report.

Insurance companies explained that they use credit-based insurance scoring

models to predict the amount they will pay out in claims, i.e., claims risk. Some models

simply predict the likelihood that a customer will file a claim. These models are most

See Comment from National Association of Mutual Insurance Cos. to FTC at 2 (Apr. 25, 2005)

[hereinafter NAMIC Comment], available at http://www.ftc.gov/os/comments/FACTA-

implementscorestudy/514719- 00088.pdf.

Although credit-based insurance scoring models are developed to predict insurance claims, instead of

credit behavior, many of the same types of information are used. A discussion of the factors that Fair Isaac

Corporation uses in calculating its credit scores of consumers (“FICO scores”) is available at:

http://www.myfico.com/CreditEducation/CreditInquiries

useful in those situations in which credit information is predictive of claim frequency, but

not particularly predictive of the size of claims.

More commonly, however, models are used to predict the “loss ratio,”

which is

the amount that an insurance company pays out on claims divided by the amount that the

customers pay in premiums. This has the advantage of controlling for the effects of non-

credit factors on risk, such as age or driving history, as premiums are determined by those

other factors. For any particular customer, the loss ratio usually will be either zero (i.e.,

no claims paid), or a number greater than one (i.e., claims paid in an amount that exceeds

premiums received). In contrast, for a group of customers, the loss ratio typically will be

a positive number less than one (i.e., some claims paid but in an amount that is less than

total premiums received).

If there is a strong relationship between customers with a

particular credit-related attribute and historic loss ratios, this information can be used to

predict the risk of loss associated with a prospective customer who shares that attribute.

Other models are used to predict “pure premiums.” Pure premiums are the total

amount that an insurance company pays on claims to consumers, not the amount that

From a technical perspective, modeling frequency is relatively straight-forward. There are a number of

standard multivariate techniques that can be used to estimate either the likelihood of a claim occurring,

such as logistic regression, or the number of claims that would be expected during a period of time, such as

Poisson regression.

Loss ratios can be modeled in a variety of ways. Because loss ratios of individuals have such an oddly-

shaped distribution B many zeros and some positive numbers that extend over a wide range B the modeling

is not trivial, but it can be handled by GLMs. Loss ratios can also be modeled by decomposing the ratio

and modeling the two components B claims paid and premiums B separately. For example, some

ChoicePoint models use this technique.

See e-mail from John Wilson to Jesse Leary, supra note 9.

Indeed, for an insurance company to be profitable, the amount that it pays out in claims must be less

than the premiums it receives plus its return on investing those premiums.

MetLife has developed a rules-based system under which credit history information is used to sort

potential customers based on their predicted loss ratio. MetLife’s “Personal Financial Management” uses

combinations of various characteristics in an applicant’s credit report to assign the applicant to one of

several risk categories without ever calculating a numerical score. This type of system essentially is a

sophisticated analog to the simple rules-based approach sometimes used prior to the development of credit-

based scores, under which, for example, some companies would not write homeowners policies to

applicants with recent bankruptcies.

customers pay in to the company. To build a credit-based insurance scoring model based

on pure premiums, it is necessary to control for other risk variables and this can be done

in one of two ways. One approach is to scale each consumer’s losses by an index of how

risky they appear, based on other non-credit risk factors (e.g., age or driving history).

This is analogous to the modeling of loss-ratios, with the non-credit-variable risk index

playing the role of the premium, but avoids the complications that arise in loss ratio

models if a credit score affected the premiums of the policies in the development

database.

The other approach involves treating credit history variables just like any other

variable in predicting risk. One benefit of this approach is that it allows for certain credit

history variables to have different effects on predicted risk for different groups of drivers.

For example, the age of a consumer’s oldest account might be less predictive for young

drivers than older drivers. Other credit characteristics might be very informative about

drivers without prior claims or violations, but provide limited insight for drivers with

poor driving records. Note that this approach may result in a model that does not produce

a numerical score based solely on credit history information.

C. Use of Credit-Based Insurance Scores

All insurance companies who use credit-based insurance scores explained that

they do so in making decisions concerning potential customers. Insurance companies,

however, also indicated that their use of scores in policy renewals for existing customers

is much more varied and complicated. Some states limit the ability of insurance

companies to use scores when customers renew policies. Even where not precluded by

state law, some insurance companies decide not to use scores when customers renew

policies to avoid damaging their relationship with these customers. Other states mandate

that firms must use, or must use if the customer requests,

updated credit-based insurance

scores to modify premium rates. Even where not mandated by state law, some insurance

companies use scores to modify premium rates for existing customers on request. In

sum, insurance companies use credit-related insurance scores to assess premiums for

potential customers and sometimes in determining premiums for existing customers who

are renewing their policies.

Insurance companies report that they use credit-based insurance scores in a

variety of ways as part of the process of determining whether to offer insurance to

prospective customers, and, if so, at what price. Making these determinations usually

consists of two steps, referred to as “underwriting” and “rating.” In “underwriting,”

insurance companies use certain characteristics of a consumer to assign him or her to a

pool based on the consumer’s apparent risk of loss. The pool into which the consumer is

placed sets the base premium rate for a policy, with the riskier pools having higher base

premium rates. In “rating,” the second step, the insurance company uses other risk

characteristics to adjust the base premium rate up or down to determine the actual amount

the consumer would be charged.

Some insurance companies said that they use credit characteristics in the

underwriting step. For example, a firm might assign a potential customer to a risk pool

based on the number of claims an applicant has filed in the past several years and the

See, e.g., R. I. Ins. Regulation 25 § 11 (although requiring firms to recalculate a consumer’s score upon

request every two years, firms generally can use a change in score only to lower premium rates), available

at http://www.dbr.state.ri.us/documents/rules/insurance/InsuranceRegulation25.pdf

There has recently been some movement towards what can be called Acontinuous rating,” in which the

risk for each applicant is evaluated and priced without first being assigned to a risk pool, but the two-step

process is still standard.

applicant’s credit-based insurance score. Using credit-based insurance scores in

underwriting thus may affect the premiums that a potential customer would have to pay

to obtain coverage, as the risk pool in which the consumer is placed determines his or her

base premium rate.

Other insurance companies report that they use scores in the rating step.

simple way to include scores is to determine a consumer’s base premium using non-credit

factors, such as age or driving history, and then adjust that rate up or down in light of his

or her score. A more complex method of using scores is to include credit as a rating

factor when developing the entire rating scheme. Such an approach allows credit

characteristics to be used interactively with other rating factors. Because how a credit-

based insurance score predicts risk may vary with other rating variables, incorporating

credit more fully into the rating step may assist in determining premiums that more

accurately reflect risk.

D. State Restrictions on Scores

As of June 2006, forty-eight states have taken some form of legislative or

regulatory action addressing the use of consumer credit information in insurance

underwriting and rating; Pennsylvania and Vermont are the only states that have not

regulated insurance scoring.

Most of these laws and regulations are based on the

While we are not aware that any insurance companies consider credit-based insurance scores at both the

underwriting and rating stage, they could do so.

An approach that is intermediate between having credit as an add-on or treating credit like any other

rating factor is to make the size of a credit score discount or mark-up depend on other rating variables. For

example, the good-credit discount for young single male drivers could be larger or smaller than the good-

credit discount for middle-aged married drivers.

The information in this section pertaining to state legislative and regulatory action addressing insurance

scoring is from the National Association of Mutual Insurance Companies’ (NAMIC) 2004 survey of state

laws governing insurance scoring practices. The report is available at:

(continued)

National Conference of Insurance Legislators’ (NCOIL) “Model Act Regarding Use of

Credit Information in Personal Insurance,” which was released in 2002.

The NCOIL Model Act prohibits insurers from using credit information as the

sole basis for increasing rates or denying, canceling, or not renewing an insurance policy.

The model also prohibits consumer reporting agencies from providing or selling

information to others that was submitted to the agency pursuant to an insurance

company’s inquiry about a consumer’s credit information, credit report, or insurance

score. Further, the NCOIL model requires insurers to comply with five conditions:

insurance companies must (1) notify an applicant for insurance if credit information will

be used in underwriting or rating; (2) notify the applicant in the event of an adverse

action based on credit information and explain its reasoning for the adverse action; (3) re-

write and re-rate a policyholder whose credit report was corrected; (4) indemnify

insurance agents and brokers who obtained credit information or insurance scores

according to an insurance company’s procedures and according to applicable laws and

regulations; and (5) file its scoring models with the applicable state department of

insurance.

Twenty-seven states have adopted laws or regulations that adopt verbatim

the language of the NCOIL model or incorporate restrictions that are very similar in

scope and nature to those in the NCOIL model.

http://www.namic.org/reports/credithistory/credithistory.asp. The information in NAMIC’s survey has

been updated to reflect newly enacted legislation and regulation through June 2006. Information on this

new legislation and regulation is from NAMIC’s annual surveys of new state insurance laws and NAMIC’s

2007 state law bulletins. The 2005 survey is available at:

http://www.namic.org/reports/2005NewLaws/default.asp

, the 2006 survey is available at:

http://www.namic.org/reports/2006NewLaws/default.asp

, and the 2007 state law bulletins are available at:

http://www.namic.org/stateLaws/2007stateLawBulletins.asp.

A copy of the text of the NCOIL model is available at: http://www.assureusa.org/docs/NCOIL.doc.

In 2003, the National Association of Insurance Commissions described the NCOIL model in testimony

before the U.S. House of Representative, Committee on Financial Services, Subcommittee on Financial

Institutions and Consumer Credit. This testimony is available at:

http://www.ins.state.ny.us/speeches/pdf/ty030610.pdf

In addition, twenty-one states have adopted some of the same types of restrictions

included in the NCOIL model. Fifteen states prohibit certain uses of credit history

information or ban the use of certain negative credit factors in the calculation of an

insurance score. Eight states have adopted dispute resolution measures governing an

insurance company’s responsibility to re-write and re-rate a policyholder whose credit

report was corrected. Seven states require insurance companies to notify consumers that

their credit information will be used in underwriting or rating. Twelve states require

insurers to notify and explain to consumers any adverse action based on credit

information. Seven states further require insurers to file their insurance scoring

methodologies.

There are several other types of restrictions that have been placed on the use of

scores. Three states (Georgia, Illinois, and Utah) prohibit using credit history

information as the sole basis in making underwriting or rating decisions. Oregon

prohibits the use of credit history information to cancel or not renew existing customers

or increase their rates, and Maryland bans the use of credit history when underwriting or

rating existing customers.

Finally, four states either have or had effective bans on the use of credit history

information in underwriting or rating automobile insurance. Hawaii by statute

specifically bans the use of credit information. California and Massachusetts effectively

ban the use of scores through their rate regulation processes. Formerly, New Jersey had

an effective ban in place, but the use of credit-based insurance scores is now allowed.

IV. THE RELATIONSHIP BETWEEN CREDIT HISTORY AND RISK

Some prior researchers have studied the existence and nature of the relationship

between credit history and insurance risk. To explore this relationship, the Commission

conducted an analysis of a database of automobile insurance policies that the agency

compiled for this study.

A consistent finding of prior research and the FTC’s analysis is

that credit information, specifically credit-based insurance scores, is predictive of the

claims made under automobile policies.

However, it is not clear what causes scores to be

effective predictors of risk.

A. Correlation between Credit History and Risk

1. Prior Research

As discussed above, risk prediction is an important method of competition among

insurance firms. Research that insurance companies have conducted about the

relationship between credit history and insurance risk therefore typically is proprietary

and non-public. Nevertheless, several studies have been made public during the past

decade that show a relationship between credit history and insurance risk.

In 2000, James E. Monaghan, an actuary from MetLife Home and Auto,

published a study analyzing the relationship between credit history variables and claims

on automobile and homeowners insurance policies.

He separately assessed a number of

credit history variables, including delinquencies, inquiries, and debt utilization rates.

Monaghan found that customers with the worst values for these variables posed a greater

See section IV.A.2 and Appendix C for a description of the database.

James N. Monaghan, The Impact of Personal Credit History on Loss Performance in Personal Lines,

Casualty Actuarial Society Ratemaking Discussion Paper (2000) (presented at the Winter 2000 CAS

forum), available at http://www.casact.org/pubs/forum/00wforum/00wf079.pdf

risk (as measured by loss ratios) than customers with the best values - often roughly 50%

more for automobile policies and over 90% more for homeowners policies.

He found

the same pattern of increased risks when he conducted his analysis controlling for other

non-credit risk factors one-by-one.

After this research, several insurance industry trade associations hired EPIC

Actuaries (EPIC) to construct a database of automobile policies with information from a

number of different insurers.

EPIC analyzed the link between credit history and risk,

and described its results in a report issued in 2003.

EPIC reported the relationship

between credit scores and different measures of risk. The study showed a strong

relationship between credit-based insurance scores and the frequency with which claims

were made, as well as between scores and the total dollar amount insurance companies

paid on these claims.

It also showed: (1) no correlation between scores and the size of

liability coverage claims; (2) a weak correlation between scores and the size of collision

coverage claims; and (3) a strong correlation between scores and the size of

comprehensive coverage claims.

In 2003, researchers at the Bureau of Business Research (BBR) at McCombs

School of Business at the University of Texas used data from five automobile insurance

companies in Texas to study the relationship between credit-based insurance scores and

As discussed in the section on the development of credit scores, the loss ratio can be used to control for

the effects of the variables used to determine premiums. However, this relies on the assumption that the

premiums accurately reflect the risks associated with those variables.

The automobile policy data that form the core of the database that we used to conduct our analysis for

this report are a subset of the data collected for use in the EPIC report. That database is discussed in more

detail below, and in Appendix C.

Michael J. Miller and Richard A. Smith, The Relationship of Credit-Based Insurance Scores to Private

Passenger Automobile Insurance Loss Propensity: An Actuarial Study by EPIC Actuaries, LLC (June

2003) [hereinafter EPIC Study], available at http://www.progressive.com/shop/EPIC_CreditScores.pdf

EPIC also conducted a multivariate analysis that included controls for most non-credit risk variables

used to underwrite and rate automobile polices. While the relationship between scores and the total amount

paid out on claims was not as large once controls were included, it remained quite strong.

losses. The BBR researchers found that customers with lower scores were more likely to

file claims under their automobile insurance policies than customers with higher

insurance scores. In addition, the researchers reported that customers with lower scores

filed claims for larger dollar amounts than customers with higher scores.

To control for

the effects of non-credit risk factors, the BBR researchers used an analysis of loss ratios,

and found that loss ratios were higher for customers with lower scores than for customers

with higher scores.

In 2004, the Texas legislature directed the Texas Department of Insurance (TDI)

to conduct a study and issue a report addressing the relationship between credit-based

insurance scores and risk for automobile and homeowner policies. In reports issued in

late 2004 and early 2005,

TDI analyzed data from six large insurance firms operating in

Texas, using each company’s credit scoring model.

For automobile policies, it found

that scores were negatively correlated with total dollars of claims, i.e., as the scores of

customers increased, the total amount that the insurance companies paid out in claims

decreased. Insurance companies paid out less on automobile policies for customers with

higher scores because they filed fewer claims than customers with lower scores.

For

homeowners insurance, TDI found similar results. TDI found that scores were negatively

Bureau of Business Research, McCombs School of Business, The University of Texas at Austin, “A

Statistical Analysis of the Relationship Between Credit History and Insurance Loss” (Mar. 2003). The

report does not make clear which particular types of automobile coverage were studied.

Id.

Texas Department of Insurance, “Use of Credit Information by Insurers in Texas: The Multivariate

Analysis” (Jan. 31, 2005) (supplemental report) [hereinafter 2005 Texas Report]; Texas Department of

Insurance, “Use of Credit Information by Insurers in Texas” (Dec. 30, 2004) [hereinafter 2004 Texas

Report].

All six insurance companies provided TDI with data on automobile policies, and three of them provided

data on homeowners policies.

TDI’s findings with regard to automobile policies were consistent regardless of whether it controlled for

other risk factors in its analysis.

correlated with both total dollars of claims and loss ratios, i.e., as the scores of customers

increased, the total amount that insurance companies paid out on their policies decreased.

2. Commission Research

a. FTC Database

The FTC undertook an analysis to determine the relationship between credit

history and risk of loss. Five of the firms that provided automobile insurance policy data

for the EPIC study described above provided the same information for the Commission’s

study.

This information included policy and driver characteristics, claims, and a

ChoicePoint Attract Standard Auto credit-based insurance score for the customer who is

named first on the policy. The information submitted to the Commission related to

automobile insurance policies in place at any time between July 1, 2000, and June 30,

2001.

The FTC combined this information from insurance companies with data from a

number of other sources to create its database. The agency included additional

information in the database to broaden the range of credit history variables analyzed; to

improve the set of other risk controls in the analysis; to provide an independent measure

of claims; and to analyze issues relating to race, ethnicity, income, and national origin.

One important feature of the FTC database was that we created weights to make it

The five firms together represented 27% of the automobile insurance market in 2000. The data were

drawn in a way that ensured a nationwide representation of policies. More information about the

companies and the database are provided in Appendix C. A discussion of the limitations of the database

and of our analysis is presented in Appendix F.

We obtained Fair Isaac credit-based insurance scores for a sub-sample of the people in the database. All

of the results presented in the body of the report are for the ChoicePoint Attract score. All of the analysis

was also conducted using the Fair Isaac score. The results were qualitatively similar regardless of whether

the ChoicePoint or the Fair Isaac score was used. Descriptions of all “robustness checks” and other

variations of the analysis are presented in Appendix F.

representative of car owners, by neighborhood income and race and ethnicity, throughout

the United States.

A more detailed description of the construction and contents of the

FTC database is provided in Appendix C.

In assessing the relationship between credit history and risk, the FTC focused its

analysis on four major types of coverage included in automobile policies: property

damage liability coverage, bodily injury liability coverage, collision coverage, and

comprehensive coverage.

4748

Property damage liability coverage insures the customer

against liability for damage he or she causes to the cars and other property of others.

Bodily injury liability coverage protects the customer from liability for bodily injuries he

or she causes to others. Collision coverage insures the customer against damage to his or

her own car from collision or rollover. Comprehensive coverage protects the customer

against losses from theft of his or her own car and for damage to the car other than from

collision or rollover (e.g., vandalism, fire, hail, etc.).

The FTC first analyzed the simple relationship between credit-based insurance

scores and claims for these four coverages. Table 2 shows, for each coverage and for

each score decile, the average number of claims per year of coverage (per hundred cars,

to show detailed differences across deciles), the average size of claims, and the average

total amount paid out on claims per year of coverage (which is the product of the number

of claims and the average size of claims).

The weighting also makes the data representative by geographic area. See Appendix D for a discussion

of the development of the weights.

The FTC database also contains information on two first-party medical coverages, usually referred to as

MedPay and personal injury protection, or “PIP.” Claims on these policies are relatively infrequent, and

the coverages vary from state to state. For these reasons, we do not focus our analysis on these coverages.

These definitions come from the Insurance Information Institute, and are available in more detail at:

http://www.iii.org/individuals/auto/a/basic/.

Figure 1 presents graphs of the relationship between scores and the average total

amount paid out on claims. In Figure 1, the horizontal axis shows automobile drivers

grouped into ten equal groups (“deciles”) based on their credit-based insurance score,

with drivers in the decile with lowest scores located at the far left and drivers in the decile

with the highest scores at the far right. The vertical axis measures the average dollars

paid out on claims per year. This measure of risk is calculated relative to drivers with the

highest credit-based insurance scores, which means that the value of the highest-score

group (i.e., those in the tenth decile) has been defined as one.

Figure 1 shows that there is a relationship between credit-based insurance scores

and risk for all four types of coverage analyzed. Specifically, the downward slopes of the

darker (higher) lines in Figure 1 show that as scores increase, the risk of loss consistently

decreases. (These lines were produced simply by graphing the average total paid on

claims – column (c) – from Table 2, relative to the highest score decile.) They show, for

example, that insurance companies paid out nearly twice as much on the property damage

liability policies of customers in the group with the lowest scores (i.e., those in the first

decile) as they did for the group with the highest scores (i.e., those in the tenth decile).

Credit-based insurance scores thus are predictive of the amount that insurance companies

pay in claims to consumers.

The FTC then constructed statistical models of insurance claims. These models

produce estimates of the relationship between scores and claims, and allow us to control

for the effects of other risk variables.

Score is measured by deciles because the units of scores are arbitrary, so there is no reason to believe that

the relationship between changes in score and changes in risk is constant across the score distribution. For

example, going from a score of 600 to 620 may have a different effect on predicted risk than going from

800 to 820.

The lighter (lower) lines in Figure 1 show the relationship between credit-based

insurance scores and the amount paid out after controlling for other standard risk factors,

such as age and driving history.

The slope of each line demonstrates that the

relationship between scores and risk persists when controls for other risk variables are

included, although the relationship is less strong. Once controls are included, for

instance, the amount that insurance companies paid out on property damage liability

claims to customers with the lowest credit-based insurance scores was 1.7 times the

amount they paid to customers with the highest credit-based insurance scores, down from

paying nearly twice as much if no controls are included. Because the relationship is less

strong when other variables are included, customers who appear more risky based on

non-credit variables are also more likely to have lower credit scores. Nevertheless, even

when non-credit variables are included in the analysis, credit-based insurance scores

continue to predict the amount that insurance companies are likely to pay out in claims to

consumers.

Figure 1 therefore shows that there is a relationship between credit-based

insurance scores and the total dollar amount of claims that insurance companies paid. To

refine this analysis, the FTC assessed whether customers with the lowest scores were

likely to cause insurance companies to pay out more because the customers file more

claims, file claims for higher amounts, or both. As shown by the darker (higher) lines in

Figure 2, customers with lower scores filed substantially more claims than those with

These other factors are controlled for by estimating a Tweedie GLM model of total dollars of claims

using score deciles and all of the other risk factors. Modeling details and the other variables included in the

models are discussed in Appendix C. Race, ethnicity, and income are not included at this stage of the

analysis.

higher scores.

For instance, customers with the lowest credit-based insurance scores

were about 1.7 times more likely to file a property damage liability claim as customers

with the highest credit-based insurance scores. On the other hand, as shown in the lighter

(lower) lines in Figure 2, the average size of the claims paid was nearly constant

regardless of credit-based insurance score. The one exception is comprehensive

coverage, which does show a relationship between claim size and score. The different

result for comprehensive coverage may be attributable to a correlation between having a

lower score and a higher probability of being a victim of automobile theft, because theft

claims are larger than claims resulting from most other events that this type of insurance

covers.

The underlying claims data presented in Table 2 (which are simple averages

without controls for other risk factors) show the same patterns as those in Figures 1 and

2, and provide additional information on the absolute size of claims risk for different

coverages and different score deciles. One important point that comes out in Table 2 is

the difficulty of predicting the claims of individual customers. While the average number

of claims per year in the lowest score decile of collision coverage, for example, was more

than twice that in the highest decile, there were still only 12 claims per hundred cars per

year of coverage for the lowest score decile. So, the vast majority of customers in even

the riskiest decile would not file a claim in a given year. As with other risk variables,

credit-based insurance scores are able to separate consumers into groups with different

average risk, but cannot predict the claims of individual consumers.

The results for the frequency and severity of claims come from models that include controls for other risk

variables. Modeling details and the other variables included in the models are discussed in Appendix C.

b. Other Data Sources

In addition to this analysis of the information in the FTC database, the

Commission evaluated alternative and independent information to assess the relationship

between credit-based insurance scores and risk. ChoicePoint Inc. collects data on claims

from most major automobile insurance firms in the United States. The data allow

insurance companies to learn whether a potential new customer has filed a claim under a

previous policy with another firm, and then use that information in underwriting and

rating. ChoicePoint refers to this data set as the Comprehensive Loss Underwriting

Exchange (“CLUE”).

We obtained the CLUE reports for each person in the FTC database for the period

July 1995 – June 2003. This encompasses three time periods: (1) the five years prior to

the period of the firm-submitted data; (2) the period of the firm-submitted data (July 2000

– June 2001); and (3) the two-year period following the period of the firm-submitted

data. The data on claims prior to the firm-submitted data (i.e., prior to July 2000) were

used to construct controls in the risk models that the FTC ran.

The CLUE data also give

us an alternative and independent source of data on claims to use to measure the

relationship between credit-based insurance scores and claims.

Figure 3 shows the average dollars paid out for each decile on policies for each of

the four main coverages studied.

Each panel includes average claims for three data

We used three years of prior claims data to construct the risk variables used in the risk models. The use

of information on prior claims is an improvement over previously published analyses of credit-based

insurance scores, which have not included controls for prior claims filed on policies with consumers’ prior

insurers.

The results in Figures 1 and 2 are for a stratified sub-sample of the database. The stratification was based

on which policies had claims in the company-provided data. The sub-sample is discussed in Appendix C.

The results in Figure 3 are for the entire sample of 1.4 million policies. We use the full sample because the

stratified sub-sample does not have sufficient information to reliably measure claims in the CLUE data for

the six-month period starting July 1, 2001. The results shown on these graphs are not controlled for other

(continued)

sources and samples: (1) claims in the data set we received from the firms; (2) claims in

CLUE for the year over-lapping with the company data set (July 2000 B June 2001); and,

(3) claims in CLUE for the six-month period following the company data set (July 2001 B

December 2001).

These results show a consistent pattern of average total dollars paid out on claims

being higher for individuals with lower credit-based insurance scores. The relationship is

generally similar across the data sources for the year of overlap, with the exception that it

is somewhat weaker for bodily injury liability coverage.

For the six months starting

July 1, 2001, the results vary for different types of automobile insurance coverage.

Comprehensive coverage results look very similar in the two time periods. The overall

slope is similar for bodily injury but the relationship is less stable. The relationship

becomes much flatter in the later time period for collision coverage, and somewhat flatter

for property damage liability. This may be evidence that credit-based insurance scores

become less predictive of claims for these coverages as more time passes from when the

scores were calculated.

non-credit risk variables, because we do not have reliable information about those variables outside of the

time period covered by the company data and because CLUE does not contain information at the car level.

For the same reasons, we use the sum of the earned car years for each coverage on each policy when

analyzing the CLUE data.

We used a six-month period because we were concerned that information on the number of insured

vehicles and coverage choices would become less reliable the further in time the data were from the data

that the companies provided. We also measured claims for the six-month period starting July 1, 2001, for a

sample of drivers limited to those who did not have any claims during the period covered by the company-

provided data. This gave results for that time period that were very similar to the results for the full sample

for that same time period.

Given the time it can take for the full cost of bodily injury liability claims to be determined, this may

affect how claims for bodily injury coverage are reported to the CLUE database.

B. Potential Causal Link between Scores and Risk

Thus, two different data sets, and previously published research, show that credit-

based insurance scores are correlated with the total amount that insurance companies pay

out on claims under automobile insurance policies.

The question that naturally arises is

why a customer’s credit history makes it more or less likely that he or she will suffer a

loss and file an insurance claim. The FTC considered various proposed explanations of

such a link and the data available bearing on those explanations. The information

available, however, does not allow the agency to draw any broad or definitive

explanations why there is a relationship between credit-based insurance scores and risk.

We emphasize that assessing the relationship between credit history and

insurance risk necessarily involves addressing the attributes and circumstances on

average of consumers with particular levels of credit-based insurance scores. Of course,

these attributes and circumstances do not necessarily apply to each consumer with a

particular level of score. People may have negative information on their credit histories

for reasons that would seem to be totally unrelated to insurance risk. The starkest

example is when the information is simply incorrect. Consumers also may wind up in

financial distress for all sorts of reasons that have no bearing on how risky they are as

drivers.

In addition, consumers may have credit histories that lead to low scores

because of a lack of an extensive credit history. This may reflect societal effects like a

lack of mainstream credit offerings where a consumer lives, or a lack of sophistication

Section VII of this report contains the results of the FTC’s successful efforts to build scoring models that

are predictive of risk. The FTC’s scoring model predicts risk in the company-provided claims data, and in

the CLUE data for an entirely different set of people and a different time period. These results provide

additional evidence that credit history information can be used to predict automobile insurance claims.

Hearing Before the New York State Assembly of Comm. on Ins. (Oct. 22, 2003) (statement of Birny

Birnbaum, Executive Director, Center for Economic Justice).

about mainstream credit markets. Again, it is not apparent that these types of

circumstances should lead to higher insurance risk.

A strong credit history, however, might indicate that a consumer has taken care in

managing his or her financial affairs B avoiding loans that might be difficult to repay,

avoiding high balances on credit cards, making sure that bills are not misplaced and are

paid on time, etc. A consumer who is prudent in financial matters may also be cautious

in other matters related to insurance, such as being more likely to put time, effort, and

money into things like car and home maintenance, cautious driving habits, etc. An

overall inclination to be prudent may lead a consumer both to have a strong credit history

and file fewer insurance claims.

There is ongoing research reflected in the behavioral economics literature that

tends to show that people who engage in risky behavior in an area of their lives are often

willing to take on more risk in other areas, as well. Researchers have studied attitudes

toward risk, as well as behavior, in financial settings and driving, as well as a range of

other areas including smoking, occupational choice, and migration.

One recent article

argues that existing research shows that physiological and psychological factors affect

how much risk individuals are willing to take in their financial, driving, and other

behavior. Many of the psychological studies surveyed in that article analyze the

relationship between psychological factors and risk-taking in a single aspect of life. The

authors connect these results between financial behavior and driving from studies on

separate groups of people, and posit the theory that credit-based insurance scoring works

See, e.g., Thomas Dohmen, et al., Individual Risk Attitudes: New Evidence from a Large, Representative,

Experimentally-Validated Survey (Sept. 2005), available at http://ftp.iza.org/dp1730.pdf

because scores reflect the psychological makeup of the individual in ways that affect

insurance risk.

Others have suggested that credit history provides information about a consumer’s

circumstances and those circumstances affect the likelihood or size of claims. One

example is that a driver with a low credit-based insurance score may be in a distressed

financial situation. This may cause stress that makes the consumer a less attentive

driver.

Being in a distressed financial situation also might give the driver a greater

incentive to try to obtain payment under an insurance policy. For example, he or she may

be more likely to file a claim for a small amount of damage to an automobile rather than

paying for those expenses out of pocket.

Another circumstance that could explain a correlation between credit-based

insurance scores and risk of loss under automobile insurance policies is differences in the

number of miles driven. The number of miles that a car is driven is directly related to

automobile insurance risk, but companies find it difficult to capture information on

“miles driven” with a great deal of accuracy. Consumers with lower scores may put more

miles on their cars than consumers with higher scores. For example, consumers with

lower scores may put more miles on their cars because they have more drivers per car in

their household, they share cars with others, etc. If there is a link between credit-based

insurance scores and number of miles driven, this could lead to a correlation between

credit-based insurance scores and risk.

Patrick L. Brackett and Linda L. Golden, Biological and Psychobehavioral Correlates of Risk Scores and

Automobile Insurance Losses: Toward an Explication of Why Credit Scoring Works, 74 J.

OF RISK AND

INS. 23 (2007).

Id.

See, e.g., Patrick Butler, Driver Negligence vs. Odometer Miles: Rival Theories to Explain 12

Predictors of Auto Insurance Claims (Aug. 9, 2006) (presented at the American Risk & Insurance

(continued)

As discussed above, a circumstance that could explain the relationship between

credit-based insurance scores and risk under automobile insurance policies is differences

in the resources that consumers put into maintaining their cars. Consumers with lower

scores may not be willing or able to spend as much money to maintain their cars. This

may, in turn, make the cars more dangerous to operate and lead to more or larger claims.

If this were an important part of the explanation for the relationship between scores and

risk, one would expect the relationship to be weaker for newer cars, which presumably

would not have had the chance to develop maintenance-related safety problems.

The FTC used its database to test this hypothesis. We divided cars in our

database into three groups: model years 1992 and older, model years 1993 – 1996, and

model years 1997 and later. Using policy information from 2000 to 2001, we estimated

the relationship between credit-based insurance scores and property damage liability risk

separately for these three groups.

Figure 4 shows that credit-based insurance scores are

strongly correlated with risk for each group, that is, the slope of the lines reveal that

within each of the three model-year categories, consumers with lower scores pose a

greater risk of loss than consumers with higher scores.

The relationship between credit-based insurance scores and risk was slightly

stronger for the oldest cars. For the oldest cars, consumers with the lowest scores are

1.81 times riskier than consumers with the highest scores. By contrast, for the newest

cars, consumers with the lowest scores are 1.68 times riskier than consumers with the

highest scores, and for middle-aged cars, consumers with the lowest scores are 1.64 times

Association Annual Meeting), available at http://www.aria.org/meetings/2006papers/butler.pdf.

We used property damage liability because (unlike collision or comprehensive coverage) the size of

claims does not depend on the value of the car covered by the policy. Car values will vary with model

year, so using coverages where the size of claims varies with the value of the car would complicate the

analysis.

riskier than consumers with the highest scores. Our results are weakly consistent with the

hypothesis that some of the relationship is attributable to consumers with lower scores

spending less to maintain their vehicles, but also show that difference in maintenance is

not the primary cause of the relationship.

In short, many explanations have been offered as to why the characteristics or

circumstances of consumers might account for the relationship between scores and risk.

Little empirical data testing these possible explanations are available. The FTC tested

one possible explanation for the relationship between scores and risk under automobile

policies, and the results were weakly consistent with the hypothesis that some of the

relationship could be attributable to the lower amount that consumers with lower scores

may spend on maintenance. Although this result provides some insight, the information

available does not allow the agency to draw any broad or definitive conclusions as to the

reason that there is a relationship between scores and risk.

V. EFFECT OF CREDIT-BASED INSURANCE SCORES ON PRICE AND

AVAILABILITY

Credit-based insurance scores are predictive of risk for automobile policies.

Insurance companies therefore are able to use these scores to underwrite and rate policies

in ways that correspond more closely to individual risk, on average. Enhanced accuracy

results in decreased premiums for lower-risk consumers and in increased premiums for

higher-risk consumers, and reduces the extent to which lower-risk consumers subsidize

higher-risk consumers. Enhanced accuracy also may have broader effects in the

marketplace. It may make insurance companies willing to offer policies to consumers

posing a wider range of risk and it may reduce adverse selection among consumers.

A. Credit-Based Insurance Scores and Cross-Subsidization

Every insurance policy written for a consumer can be thought of as posing a true

level of claims risk, that is, the expected cost to the insurance company of claims that the

customer will submit. If the firm knew this true level of risk, it could base premiums on

this risk. Because of practical limitations on the ability of firms to obtain and process

information, they cannot determine the true level of risk that any particular consumer

poses.

Instead, they must use the information available to them to estimate the expected

claims cost for each consumer. Traditionally, insurance companies have divided

customers into groups based on their characteristics and calculated expected average

losses for the group, after which group members are charged premiums based on these

expected losses.

Because the true expected claims costs will vary within any group of customers,

some in the group will be paying premiums that are higher and others will be paying

premiums that are lower than their own individual true expected claims cost. Those in

the group with lower expected claims costs (i.e., the lower-risk customers) subsidize

those with the higher expected claims cost (i.e., the higher-risk customers).

In the

absence of perfect information about individual customer risks, there will always be some

consumers in an insured group who subsidize other consumers in the group.

Because insurers never have complete information about consumers, their estimates of expected claims

costs are, at best, only correct on average; some estimates are over-estimates and others under-estimates.

Such a situation is referred to as “imperfect information” about consumer risk.

This is ex ante cross-subsidization (or, cross-subsidization “in expectation”). It is a distinct concept from

ex post cross-subsidization. Inherent in the concept of insurance is ex post cross-subsidization, that is,

customers who do not experience loss subsidize customers who do.

Note that if information is symmetric between insurers and consumers (i.e., they both have the same

imperfect beliefs about expected claims costs), consumers will not know whether they are beneficiaries or

contributors to the subsidization. Given that consumers do not know whether they are paying more or less

(continued)

Better risk prediction techniques allow insurance companies to more effectively

separate higher-risk consumers from lower-risk consumers. This information assists

insurance companies in charging consumers prices that correspond more closely to the

true risk they pose, on average. This, in turn, decreases the premiums of lower-risk

consumers and increases the premiums of higher-risk consumers, on average. Improved

risk prediction techniques therefore reduce the extent to which lower-risk consumers

subsidize higher-risk consumers.

Even though improved risk prediction techniques will make firms’ estimates of

the riskiness of consumers on average more accurate, the predicted risk of some

individual consumers may become less accurate. For example, there are some consumers

who are very safe drivers but have low credit-based insurance scores. If scores are used,

the predicted risk for these specific individuals will become less accurate. This result is

unavoidable in any scheme used to make predictions about the risk consumers pose.

Therefore, even if risk predictions become more accurate overall as additional predictive

information is considered, there will always be some people who are much safer – or

much riskier – than they appear.

The FTC analyzed the information in its automobile insurance database to

estimate the extent to which the use of credit-based insurance scores (a risk prediction

technique) could reduce cross-subsidization. Many of the premiums for policies included

in the database were calculated without using scores,

and the data do not indicate which

than their true risk, their decisions will be unaffected by the existence of cross-subsidization.

This is true even for customers of firms that do not adopt the more accurate prediction method, because

those firms will wind up with a riskier and more homogenous pool of customers. Because the pool of

customers is more homogenous, there will be less cross-subsidization within that group of consumers.

E-mail from Rick Smith, Towers Perrin, to Jesse Leary, Assistant Director, Division of Consumer

Protection, Bureau of Economics (Apr. 13, 2005) (on file with FTC).

policies these were. The FTC database contains information from 2000-2001, shortly

after the introduction of scores. As discussed above, scores typically are used in

determining the premiums to be charged to prospective customers. Customers who

renewed their policies during 2000-2001 thus were not likely to have had scores used to

determine their premiums. In addition, although by 2000 insurance companies were

using scores to determine premiums in many states, their use was not universal.

Accordingly, many, and probably most, of the premiums charged to consumers during

this period of time were determined without the use of credit-based insurance scores.

Because most of the premiums in the database likely do not reflect the use of

credit-based insurance scores, the FTC used risk, measured in expected total dollars of

claims, as a substitute for premiums in an analysis of the effects of scores. We believe

that this calculation of risk is a reasonable substitute for premiums in this context,

because the premiums that an insurance company charges consumers in a competitive

marketplace should be roughly proportional to the risk they appear to pose.

The FTC used a three step analysis to evaluate how expected risk changes if

insurance companies consider credit-based insurance scores. The first step was to use a

model to calculate a predicted dollar risk for each consumer using all risk factors in the

database, except score.

The second step was to calculate a predicted dollar risk for each

consumer using all risk factors plus a score. Both of these steps to calculate predicted

Some industry participants have stated that homeowners and automobile insurance markets are fiercely

competitive. See, e.g., Comment from State Farm Ins. Co. to FTC at 3-4 (Apr. 25, 2003) [hereinafter State

Farm Comment], available at http://www.ftc.gov/os/comments/FACTA-implementscorestudy/514719-

00100.pdf; Comment from the American Ins. Ass’n to FTC at 14 (Apr. 25, 2005) [hereinafter AIA

Comment], available at http://www.ftc.gov/os/comments/FACTA-implementscorestudy/514719-

00084.pdf.

This was done using a Tweedie GLM model. Modeling details are provided in Appendix C. Race,

ethnicity, and income were not considered at this stage of the analysis.

dollar risk were conducted separately for property damage liability, bodily injury

liability, collision, and comprehensive coverage.

The third and final step was to sum

the predicted dollar risks for all four types of insurance coverage with and without the use

of credit-based insurance scores.

This produced two estimates of total risk for each

insurance policy in the database: an estimate without using a score, and an estimate using

a score.

The FTC’s analysis predicts that the use of credit-based insurance scores

redistributes premium costs from consumers with higher scores to those with lower

scores.

This is a zero-sum calculation: the total increases in premiums predicted if

scores are used must be exactly the same as the total decreases in premiums predicted.

Figure 5 shows the results of the FTC’s analysis of the effect of credit-based

insurance scores on changes in premiums. It shows what share of consumers would be

predicted to have changes of different sizes. Figure 5 also reveals that if credit-based

insurance scores are used, more consumers (59%) would be predicted to have a decrease

in their premiums than an increase (41%).

We also conducted this analysis using a single-equation model of all coverages, instead of separate

models by coverage. As discussed in Appendix F, this yielded similar results.

This approach uses the actual coverage choices of individuals. That is, we predict claims cost for

individuals only for the coverages they had, and measure the change in their total predicted claims for those

coverages. This has the advantage of taking into account the real choices people made when purchasing

insurance, but the disadvantage of not allowing for the possibility that individuals would change their

coverage choices in response to changing premiums. To generate the overall distribution of changes, we

weighted consumers by the earned car years on their property damage liability coverage.

We emphasize that this is not a measure of how firms are actually using scores to price consumers.

Scores are not used to underwrite or rate all customers, especially existing customers. In addition, firms

may not adjust premiums in response to scores as much as our analysis would predict. For these two

reasons, this exercise may overstate the redistributive effects of using scores. The fundamental assumption

of the analysis, that premiums will be proportional to predicted risk, is likely to be violated in the short-

term, especially if existing customers are not fully re-underwritten and re-rated every year. In sum, this

approach probably overstates the redistributive effects to date of using scores, but should be a reasonable

substitute for the long-term effects of using scores.

The increased premiums for consumers whose premiums would rise are larger

than the decreased premiums for those whose premiums would fall. This can be seen in

the longer “tail” on the right-hand side of the graph, which shows larger changes in the

direction of an increase. The median increase for those with an increase in predicted risk

is 16% (i.e., one-half the increases in predicted risk are greater than 16% and one-half are

less than 16%), while the median decrease is 13%.

1. Possible Impact on Car Ownership

If using credit-based insurance scores results in consumers paying premiums that

are closer to the true risk that they pose, this could result in car owners incurring costs

closer to the real costs of owning and operating their cars. Internalizing these costs could

affect consumer decisions whether to own cars, thus resulting in more efficient car

ownership.

If consumers decide how many cars to own based on the benefits and costs of car

ownership, their decisions can be said to be “efficient” in that they will choose to own

cars only when the benefits are at least as great as the costs. If consumers pay premiums

that are lower than the risk they actually pose, they will own more cars than is efficient,

because other people are helping to pay for the cost of their driving.

And, if consumers

pay premiums that are higher than the risk they pose, they will own fewer cars than

would be efficient, because they face costs higher than the true total costs of their driving.

The use of credit-based insurance scores to charge premiums that more accurately reflect

This is a classic “negative externality.” Negative externalities arise any time consumers or businesses

pay costs for a product that are less than the costs to society. Because societal costs are not considered in

the decision of consumers or businesses, their decisions will be inefficient for society.

the true cost of driving, on average, thus could lead to a more efficient level of car

ownership.

The FTC was not able to determine whether and to what extent credit-based

insurance scores have an effect on automobile ownership. We are not aware of

information addressing specifically how much of an increase or decrease in the cost of

driving will cause consumers to decide whether to own a vehicle.

Moreover, even if we

were able to determine the effect of insurance scores on car ownership, this study does

not assess whether such an outcome would be equitable.

2. Possible Impact on Uninsured Driving

Using credit-based risk scores to determine premiums also could have an effect on

the number of drivers who drive without insurance. Although most states have

requirements that drivers carry specified minimum amounts of liability insurance, there

are still significant numbers of drivers who drive without insurance. Raising premiums

of drivers with lower scores could lead to more of them driving without insurance.

Lowering premiums of drivers with higher scores could lead to fewer of them driving

without insurance. Whether the use of scores on balance leads to more or fewer people

driving without insurance depends on which of these two effects is greater.

Even though there is published work on the effects of prices on new car sales, see, e.g., Patrick S.

McCarthy, Market Price and Income Elasticity of New Vehicle Demands, 78 R

EV. OF ECON. AND STATS.

543 (1996), we are not aware of studies that measure the effect of the cost of insurance on the number of

cars that households choose to own.

Although it is not obvious which change would be larger, there is strong intuition that suggests people

with higher scores are relatively less likely to be driving without insurance even when scores are not used

to determine premiums. This would be true if scores were correlated with wealth or if scores were a

measure of caution or responsibility. The value of liability insurance to an individual depends in part on

that individual’s wealth, because people with very little wealth may be nearly “judgment proof,” and

therefore face very little effective risk from liability claims. A company that issues a policy, however, is

liable up to the policy limits. So, liability insurance may be worth less to a low-wealth driver than it costs,

(continued)

The FTC sought to estimate the impact of credit-based insurance scores on the

prevalence of consumers driving without insurance. It is difficult to obtain reliable data

concerning the number or share of drivers who drive without insurance, because this

conduct is illegal in most states. In an effort to derive such an estimate, the FTC

compared the number of uninsured motorist claims relative to other claims filed during

1996 to 2003 (i.e., when credit-based insurance scores were becoming more widely used)

for states in which these scores were used and in states in which they were not.

We assessed how often consumers filed uninsured motorist claims relative to how

often they filed bodily injury claims and property damage liability claims. Figure 6

shows that the number of uninsured motorist claims filed compared to the number of

bodily injury claims filed increased in states where credit-based insurance scores were

allowed, but decreased slightly in states where they were not.

Figure 6 also shows the

number of uninsured motorist claims filed compared to the number of property damage

claims filed was basically unchanged in states where scores were allowed and decreased

somewhat in states where they were not.

These results are consistent with the hypothesis that scores, because they raise the

premiums of some consumers, cause a larger share of consumers to drive without

insurance

and/or more risky consumers to drive without insurance.

These results,

because – if uninsured – the driver would have to pay out less to cover others’ losses from an accident than

would the insurance company if the drivers bought insurance.

The states identified as not allowing the use of scores during the relevant period of time are California,

New Jersey, Massachusetts, and Hawaii. Because of limitations in the data, Texas and South Carolina are

not included in either group.

If reduced cross-subsidization leads to more consumers driving without insurance, this could actually

lead to lower overall losses from accidents. Research shows that the effect on accidents of requiring

drivers to buy liability insurance in order to operate a car legally is unclear. Some drivers may choose not

to purchase insurance and then either not drive or drive less often or more carefully, to avoid detection,

leading to fewer accidents. Other drivers may purchase insurance they otherwise would have foregone, and

then drive more often or more riskily, because they no longer bear the liability risk of causing an accident,

(continued)

however, should be treated with caution. First, the relative change between the groups of

states took place during the period 1997 – 2000. While scores were becoming widely

used during this period, credit-based insurance scoring had probably not yet affected

most consumers’ premiums, given that insurance companies generally do not use scores

when renewing customers. Perhaps more importantly, the FTC’s analysis could be

affected by any state-specific changes in insurance markets. Because the number of

states not allowing the use of credit-based insurance scores for automobile insurance was

small (California, Hawaii, Massachusetts, and New Jersey), any such changes could

render them unreliable as a comparison group. In addition, because the analysis relies on

uninsured motorist claims to indirectly measure the level of driving without insurance,

differences over time in which consumers carried uninsured motorist coverage in states

which allow the use of scores and those that do not could affect the results.

3. Adverse Selection

Credit-based insurance scores also may make insurance markets more efficient if

they decrease the extent to which consumers make insurance purchasing decisions using

better risk information than that available to insurance companies. In a competitive

market, insurers will offer prices to groups of consumers reflecting the average expected

risk of loss for each group. But if a consumer has better information than the insurance

leading to more accidents. One study has reported that the latter effect predominates over the former effect.

Alma Cohen and Rajeev Dehejia, The Effect of Automobile Insurance and Accident Liability Laws on

Traffic Fatalities, 47 J.

OF LAW AND ECON. 357 (Oct. 2004).

If these results do reflect effects of credit-based insurance scores, they could have the indirect effect of

mitigating some of the savings that higher-score drivers get from the use of scores. If more higher-risk

drivers are uninsured, this could increase the expected cost of uninsured motorist claims that insured

drivers submit under their policies. In turn, this could increase the premiums that insurance companies

must charge lower-risk consumers to cover these increased uninsured driver claims. Accordingly, even

these increases in the cost of uninsured motorist coverage may offset somewhat the decrease in premiums

that higher-score consumers receive from the use of scores.

company about his or her own true expected risk of loss, he or she will know whether the

group price is higher or lower than his or her true expected claims cost. The consumer

may use this superior knowledge to determine whether and how much insurance to

purchase, a phenomenon known as “adverse selection.”

Adverse selection may be occurring if higher-risk consumers are more likely to

have insurance or more complete coverage than lower-risk consumers. A higher-risk

consumer who realizes that he or she is being charged a price that is lower than his or her

actual risk of loss cost will have an incentive to purchase more insurance coverage.

lower-risk consumer who realizes that he or she is being charged a price higher than his

or her actual risk of loss will have an incentive to purchase less insurance coverage.

higher-risk consumers purchase more insurance coverage and lower-risk consumers

purchase less insurance coverage, the average risk of the group of consumers who do buy

insurance will be higher. Premiums then would have to increase for insurance companies

to cover the total claims costs, providing a further disincentive for lower-risk consumers

to purchase insurance. If consumers know more about the risk of loss they pose than

The discussion here is of market-wide adverse selection, where consumers know more about their risk

than any firm does. A firm competing in an insurance market can face another form of adverse selection if

one of its competitors is able to do a better job of predicting risk and entices away low-risk customers while

leaving behind high-risk customers. See State Farm Comment, supra note 66, at 8.

It is conventional wisdom in the insurance industry, however, that the riskiest drivers are those who

choose to buy the least amount of coverage possible, and would buy no insurance if it were not legally

required. This conventional wisdom probably reflects, at least in part, that the “riskiest drivers” in question

are riskiest based on characteristics that are used to underwrite and rate policies, like driving history, and

they are charged the highest rates. If these drivers really are very risky, simple theory would predict that

they would still be willing to pay very high rates. Explanations for why these drivers would be unwilling to

buy insurance at rates that reflect their true risks could include: that the drivers have very limited assets and

are therefore “judgment proof,” and therefore face less actual risk than the firm would face; that these

drivers are less risk-averse than other drivers, or even risk-loving, and therefore unwilling to buy insurance

at market rates; that the drivers believe themselves to be less risky than firms judge them to be; or that the

drivers are cash-constrained, and do not buy insurance even though they would rather have the insurance

than face the risk of a large loss.

A consumer may do this by purchasing a policy with large deductibles or low liability limits, by not

purchasing certain types of coverage, or by not purchasing insurance at all.

insurance companies, it therefore can affect insurance purchasing decisions in ways that

cause economic inefficiency.

If scores allow insurance companies to predict risk more accurately, it could

decrease the difference between what consumers and insurance companies know about

the risk that individual consumers pose. Insurance companies therefore would be able to

charge consumers premiums that more accurately reflect the true risk. This would reduce

the incentive of higher-risk consumers to purchase more insurance and lower-risk

consumers to purchase less. Accordingly, scores may reduce the extent of adverse

selection and make insurance markets more efficient.

The FTC considered whether adverse selection exists in automobile insurance

markets in the United States.

It seems unlikely that consumers have better information

about the risk they pose than do insurance companies. Although consumers might have

some sense of how much risk they pose based on their own experience, it seems unlikely

that this sense is more accurate than the assessment insurance companies can make.

Insurers who realize that adverse coverage selection is occurring may attempt to separate the higher-

from the lower-risk consumers by offering different price-coverage combinations. One theoretical analysis

suggests that under some conditions, this approach can reduce the inefficiency caused by adverse selection.

Michael Rothschild and Joseph Stiglitz, Equilibrium in Competitive Insurance Markets: An Essay on the

Economics of Imperfect Information, 90 T

HE Q. J. OF ECON. 629 (Nov. 1976). In fact, firms do offer

different deductible choices, which could be a mechanism for separating high-risk and low-risk customers.

It may also be a pricing response by firms to differing levels of risk-aversion among customers.

The American Insurance Association has stated that an insurance company found that, after introducing

the use of credit-based insurance scores, “there is some evidence that higher limits of liability coverage and

lower physical damage deductibles are being purchased…” AIA Comment, supra note 66, at 14 (emphasis

in original). This would be consistent with a reduction in adverse selection resulting from the use of scores.

Empirical studies have found only limited evidence for adverse selection in automobile markets in other

countries, and even then only in very special circumstances. See Alma Cohen, Asymmetrical Information

and Learning: Evidence from the Automobile Insurance Market, 87 R

EV. OF ECON. AND STATS. 197 (May

2005) (finding evidence that lower risk drivers in Israel purchased less insurance coverage than higher risk

drivers, but only for experienced drivers for a limited period of time after switching policies and in a

country in which insurance companies do not share data on prior claims); Pierre-Andre Chiappori, et al.,

Asymmetrical Information in Insurance: General Testable Implications, (Feb. 24, 2004) (finding no

evidence that lower risk drivers in France bought less insurance than higher risk drivers), available at

http://www.iue.it/FinConsEU/papers2004/salanie.pdf

Specifically, companies make their assessment also using information about the

consumer’s past experience, such as extensive prior claim information included in a

database that insurance companies share and public record information, such as

convictions for driving while intoxicated or speeding.

Moreover, even assuming that consumers have better knowledge than insurance

companies of the risk they pose, there are significant limitations on the extent to which

consumers can use this advantage to alter their insurance purchasing decisions. Most

states mandate minimum liability coverage for cars, and lenders typically require even

greater coverage on cars they finance. Even though consumers retain some ability to

make choices concerning insurance coverage, such as deductibles and limits, these

choices are limited considerably.

The FTC analyzed its automobile insurance database to test whether there was

any indication that adverse selection may be occurring. We found that lower-risk drivers

tend to have policies with higher deductibles than do higher-risk drivers, that is, lower-

risk drivers have less insurance coverage than higher-risk drivers. This is consistent with

(but does not prove)

adverse selection is occurring in automobile insurance markets.

If credit-based insurance score information is considered in the analysis, i.e., the

risk information available to insurance companies relative to consumers is enhanced,

then adverse selection would be expected to decrease. However, when the FTC

considered scores in its analysis, lower-risk drivers were still found to have insurance

It is clear that adverse selection experienced by a single firm can be a powerful force. When different

firms have significantly different risk prediction technology, consumers will see the different prices

charged and will tend to choose the firms with lower premiums. This can lead to a negative-feedback loop

that can even cause a firm to collapse.

One alternative explanation is moral hazard. If people with more complete coverage take less care,

because they bear less of the cost of any accident or other damage or loss, this would result in the same

relationship.

policies with higher deductibles than higher-risk drivers. This suggests that adverse

selection may not be occurring, or, if it is occurring, then scores may not reduce it.

B. Other Possible Effects of Credit-Based Insurance Scores

Innovations in risk prediction techniques like credit-based insurance scores may

affect the availability of insurance and some of the costs associated with selling

insurance. First, some consumers may have a broader range of options to choose from

when purchasing insurance. Because credit-based insurance scores predict risk more

accurately for consumers, insurance companies now may be willing to offer coverage to

some higher-risk consumers. In addition, credit-based insurance scores may make the

process of underwriting and rating quicker and cheaper, and competition between

insurance companies may cause cost savings from these process improvements to be

passed on to consumers in the form of lower premiums.

Insurance companies and industry representatives stated that the use of credit-

based insurance scores gives firms greater confidence in their ability to predict the risk

that consumers pose. That is, if firms have more confidence in their risk estimates, they

may be able to offer insurance to customers for whom they would otherwise not be able

to determine an appropriate premium. The American Insurance Association, for

example, has stated “(m)ore precise pricing enables insurers to accept greater risk by

ensuring that both good risks and more marginal risks are properly priced to reflect the

exposure they represent.”

Several firms, including The Hartford and MetLife Home and

Auto, have stated that the use of credit-based risk scores enabled them to offer policies to

AIA Comment, supra note 66, at 4.

higher-risk consumers than they had previously.

This could lead to higher-risk

consumers having more choices as they shop for insurance. No data, however, were

submitted or obtained to assess the extent to which credit-based insurance scores actually

have expanded insurance choices for higher-risk consumers.

In addition, several insurance companies and score developers emphasized that

the use of scores can save costs. Specifically, they asserted that the use of scores

facilitates automation, speeds up policy underwriting and rating, and otherwise reduces

the costs of underwriting and rating.

No data was submitted or obtained to allow the

FTC to develop reliable estimates of cost-savings associated with credit-based insurance

scores. Assuming that there are such savings, the FTC would anticipate that competition

in the market for automobile insurance would result in these savings being passed on to

consumers in the form of lower prices.

Further, banning the use of factors that are known to be correlated with risk could

have negative effects on insurance markets. If firms cannot adjust prices based on the

risk associated with a characteristic, they will have an incentive to refuse to offer policies

to people with the characteristic.

If the law prohibits firms from refusing to sell policies

to people with that characteristic, they will still have an incentive to try to avoid insuring

them. This could cause firms to expend resources on finding ways to avoid higher-risk

consumers, reducing the availability of insurance to higher-risk consumers and making

otherwise profitable distribution channels untenable.

Meeting between FTC staff and The Hartford (July 14, 2004); Meeting between FTC staff and MetLife

Home and Auto (July 12, 2004); Meeting between FTC staff and USAA, (July 14, 2004). See also AIA

Comment, supra note 66, at 7-8; NAMIC Comment, supra note 20, at 6-7.

AIA Comment, supra note 66, at 12; Fair Isaac Comment, supra note 18, at 15; State Farm Comment,

supra note 66, at 4.

Id. at 10.

A simple example illustrates the possible impact of banning the use of a

characteristic in making the decision whether to offer insurance to consumers. Assume

that geographic location is correlated with risk on insurance policies

and that firms are

allowed to refuse to sell insurance based on geography but are not allowed to charge

different prices based on geography. This would give insurance companies an incentive

to refuse to sell policies to people living in riskier areas.

If firms could not outright

refuse to sell policies based on geography, and could not charge different prices based on

geography, they would have an incentive to use other means to avoid insuring those who

live in more risky geographic areas, for example, not establishing offices, working with

independent agents, or advertising in these locations.

It is not clear, however, whether banning the use of credit-based insurance scores

would lead to distortions of the insurance market like those associated with banning the

use of geography. An insurance company does not see a consumer’s score until he or she

applies for insurance coverage. It therefore would be difficult for insurance companies to

directly avoid selling insurance to consumers with low scores. There may be, however,

different marketing approaches, such as alternative types of advertising, which bring in

consumers with different average scores. If firms cannot use scores to underwrite or rate,

they would have an incentive to market only through channels that bring in consumers

with higher scores. This could reduce the availability of information about insurance

options, particularly to consumers with lower insurance scores. No data was submitted or

obtained, however, to permit the FTC to determine whether restrictions on the use of

scores actually would have this type of effect.

Id.

Firms might specialize geographically, with firms with higher premiums offering policies everywhere but

mainly getting customers from high risk areas, while lower-cost firms refuse to write in high-risk areas.

Banning credit-based insurance scores may also give firms greater incentives to

invest in developing other risk-prediction tools. If the use of scores is banned, firms may

have an incentive to spend more on developing new risk variables to capture some of the

same risk prediction benefits of scores.

This could be seen as an unnecessary societal

cost, given that scoring technology has already been developed and scores are a fairly

low-cost risk prediction technique.

C. Effects on Residual Markets for Automobile Insurance

The introduction and growth in the use of credit-based insurance scores has taken

place during a time when one particular measure of the functioning of the market, the

share of consumers buying insurance through state-run “residual markets,” indicated the

market was working well. All states run some type of program to allow consumers to

purchase automobile insurance when they are unable to find a private firm willing to sell

them policies voluntarily.

To avoid attracting consumers who could otherwise obtain

private insurance coverage, these state-run programs typically charge higher prices than

private insurance companies.

Figure 7 shows the share of automobile policies that were purchased through

state-run programs during the years 1996 – 2003, broken down by states that allow the

use of credit-based insurance scores, and those that do not.

It shows that a larger share

of consumers participated in these programs in the states that did not allow the use of

See Cheng-Sheng Peter Wu, Deloitte and Touche, What to do When You Cannot Use Credit (Sept. 2000)

(presentation at the CAS Special Interest Seminar), available at

http://www.casact.org/education/specsem/f2005/handouts/credit.ppt

See https://www.aipso.com/about.asp.

The states identified as not allowing the use of scores during the relevant period of time are California,

New Jersey, Massachusetts, and Hawaii. Because of limitations in the data, Texas and South Carolina are

not included in either group.

scores. However, this was true both before and after the introduction of scores, and

therefore this difference in levels of participation presumably reflects other differences

between states. Figure 7 shows that the state-run program share fell during the second

half of the 1990s, as score were being introduced, and then leveled off after 2000. The

pattern is nearly identical in states that allowed the use of scores and states that did not.

Therefore, Figure 7 is probably best interpreted as meaning that scores at least did not

interfere with the smooth functioning of automobile insurance markets.

VI. EFFECTS OF SCORES ON PROTECTED CLASSES OF CONSUMERS

FACTA requires that the FTC analyze the extent to which the use of credit-based

insurance scores affects the availability and affordability of insurance for members of

certain categories of consumers. The statute mandates that the Commission consider the

impact of these scores on categories of consumers based on race, ethnicity, national

origin, geography, income, religion, age, sex, and marital status. In particular, the agency

was instructed to assess whether scores act as a proxy for membership in these groups.

In fulfilling the statutory mandate, the FTC focused its analysis on the effect of

credit-based insurance scores on consumers in different racial, ethnic, national origin, and

income groups. The Commission did not focus its assessment on consumers in different

religious groups because we are not aware of any reliable data relating scores to religious

affiliation. In addition, the FTC also did not focus its analysis on consumers in different

geographic, age, sex, and marital status groups. In most locations in the United States,

insurance companies can and do use geography, age, sex, and marital status directly in

determining automobile insurance premiums.

While credit-based insurance scores may

vary based on these factors, the direct effect of using these factors to price insurance far

outweighs any indirect effects these factors may have through their impact on scores.

The FTC therefore did not try to measure any such indirect effects.

A. Credit-Based Insurance Scores and Racial, Ethnic, and Income Groups

1. Difference in Scores across Groups

The FTC analyzed whether there was a relationship between credit-based

insurance scores and race, ethnicity, national origin, and income. In undertaking this

analysis, the Commission first reviewed and considered prior research. In 1999, the

Virginia State Corporation Commission’s Bureau of Insurance issued a report assessing

the relationship at the ZIP code level between scores and race as well as between scores

and income.

The report stated that “nothing in (our) analysis leads the Bureau to the

conclusion that income or race alone is a reliable predictor of credit scores.”

Nevertheless, the absence of more detailed information about the results of this study

leaves unclear the relationship between scores and race and income.

The State of Missouri Department of Insurance released a study in 2004 that

relied on similar data.

The Missouri study used ZIP-code level data on scores and race,

income, and other demographic variables. The scores analyzed were credit-based

insurance scores that twelve large insurance companies used for automobile or

While the use of income to underwrite policies or set rates may not be expressly prohibited in some

locations, it appears to be generally regarded as an illegitimate variable for those purposes.

Report of the State Corp. Comm’n’s Bureau of Ins. to the Sen. Commerce and Labor Comm. of the Gen.

Assemb. of Va., Use of Credit Reports in Underwriting (1999) [hereinafter Virginia Study].

Brent Kabler, Ph.D., Insurance-Based Credit Scores: Impact on Minority and Low Income Populations

in Missouri (Jan. 2004) [hereinafter Missouri Study], available at

http://www.insurance.mo.gov/reports/credscore.pdf

homeowners policies. The Missouri study found that scores were correlated with the

racial, ethnic, and income characteristics of ZIP codes. Specifically, as the proportion of

racial and ethnic minorities or lower-income consumers in a ZIP code increased, scores

decreased.

These correlations remained after controlling for education, marital status,

and housing values.

Unlike prior researchers, the Texas Department of Insurance (TDI) in its 2004

study moved beyond aggregate data and obtained data about individuals to analyze the

relationship between scores and race, ethnicity, and income. The TDI used automobile

and homeowners policy data from six large insurance companies. The TDI obtained race

data for each consumer from the Texas Department of Public Safety and ethnicity data

for each consumer from a Hispanic surname match. The TDI used median income for the

ZIP code in which consumers lived, because individual income information was not

available.

The TDI’s analysis of this data showed that African Americans and Hispanics

tended to have lower scores than Asians and whites.

100

It revealed mixed results for

income. For some insurance companies, consumers in higher-income areas had higher

scores, while this was not the case for other insurance companies. It is not clear whether

these different results for income reflect differences in the credit scoring models that the

insurance companies used, or in the mix of customers at each firm.

An attempt was also made in the Missouri study to use the ZIP code level data to draw inferences about

individual-level differences in credit scores by race and income. The results of this analysis were more

speculative, but did demonstrate that it would be very unlikely that the differences found at the ZIP code

level could have been found if there were no differences at the individual level.

100

The TDI characterized the scores in this way: “In general, Blacks have an average credit score that is

roughly 10% to 35% worse than the credit scores for Whites. Hispanics have an average credit score that is

roughly 5% to 25% worse than those for Whites. Asians have average credit scores that are about the same

or slightly worse than those for Whites.” 2004 Texas Report, supra note 41, at 13.

After reviewing the prior research, the FTC analyzed the information in its own

automobile insurance database to assess the relationship between scores and race,

ethnicity, national origin, and income. Figure 8 shows how non-Hispanic whites,

African Americans, Hispanics, and Asians are distributed across the range of scores.

The horizontal axis shows score deciles, and the vertical axis shows the share of each

group that falls in each decile. The deciles are defined using the overall distribution of

scores. If a group had the same distribution of scores as the overall sample, then 10% of

that group’s population therefore would fall in each of the ten deciles.

Figure 8 shows that non-Hispanic whites and Asians are fairly evenly distributed

across the score range, resulting in a roughly flat line near 10%. In contrast, African

Americans and Hispanics are strongly over-represented in the lowest deciles and under-

represented in the highest deciles. For example, 26% of African Americans are in the

group with the lowest 10% of credit-based insurance scores, while only 3% are in the

highest 10% of scores. Similarly, 19% of Hispanics are in the group with the lowest 10%

of scores, and 5% are in the highest 10% of scores.

Another way of measuring these differences is to look at where the median

person

101

for each racial or ethnic group falls in the overall distribution of scores. If

scores were distributed identically across racial and ethnic groups, the median score for

each group would equal the overall median – the 50

percentile. The FTC’s data show

that the median scores for non-Hispanic whites and Asians are quite similar to that of the

overall sample, with the median score for non-Hispanic whites and Asians falling in the

and 52

percentile, respectively. In contrast, the median scores for the African

101

One-half of the group will have a score lower than the median person and one-half will have a score

higher than the median person.

Americans and Hispanics are much lower, with the median scores of African Americans

and Hispanics falling in the 23

and 32

percentile, respectively. So, more than one-half

of all African Americans have credit scores in the lowest quarter of the overall score

distribution, and one-half of all Hispanics have credit scores in the lowest third of the

overall score distribution.

Figure 9 presents an alternative way of viewing these differences. It shows the

racial and ethnic makeup of each decile in the score distribution, which varies

considerably across the range of scores. Because non-Hispanic whites make up such a

large share of the populations, they are a majority in every score decile. But, as Figure 8

shows, African Americans and Hispanics are heavily over-represented in the lower score

deciles.

In addition to race and ethnicity, the FTC examined the relationship between

scores and national origin. To assess this relationship, the Commission compared scores

for foreign-born consumers with those of consumers born in the United States. The

scores for consumers born outside the United States were slightly lower than those of

consumers born in the United States, with the median score of the foreign-born

consumers falling in the 44

percentile of all scores.

The FTC also compared the scores for recent immigrants and other consumers.

102

Recent immigrants have scores that are slightly lower than other immigrants and lower

than consumers overall, with the median score for recent immigrants falling in the 39

percentile of all scores. We found that recent immigrants whose information is included

102

The FTC database does not contain information on the actual date of anyone’s arrival in the United

States. For this reason, recent immigrants were defined as people who first applied for a Social Security

card during the previous ten years, and who were 30 years old or older at the time of the sample. These

restrictions were an attempt to limit “recent immigrants” to people who arrived in the United States as

adults.

in the FTC database were much more likely to be Hispanic or Asian than consumers born

in the United States. This makes it complicated to evaluate and describe the relationship

between scores and race or ethnicity apart from the effect of national origin. Because

race and ethnicity are associated with much larger differences in scores than national

origin, the Commission focused its further analysis on race and ethnicity.

Finally, the FTC study evaluated the relationship between scores and income.

The Commission did not have access to information about the income of the particular

consumers in its database. The FTC instead used the median income of the United States

Census tract in which consumers live to divide them into low-to-moderate income,

middle income, and high-income neighborhood groups.

103

Figure 10 shows the share of

people in each income category in each decile of the distribution of scores. Low-to-

moderate income consumers are somewhat over-represented in the lower score deciles,

with 15% of these individuals in the lowest 10% of scores, and only 8% in the highest

10% of scores. Middle-income consumers are essentially evenly distributed across the

distribution of scores. High-income consumers are under-represented in the lowest 10%

of the score distribution, but otherwise fairly evenly distributed. Figure 11 shows the

income breakdown of each score decile. Again, it shows that there is some relationship

between neighborhood income and score.

The results for the FTC’s database show that as income increases, scores tend to

increase. These results, however, are much weaker than the results for race and ethnicity.

103

This approach follows methods used to analyze income in FRB studies of mortgage markets. The

groups were: (1) Low-to-moderate income: Tract median < 80% of MSA median income; (2) Middle

income: Tract median >= 80% of MSA median income and <%120 of MSA median income; and (3) High

income: Tract median >= %120 of MSA median income. As discussed in Appendix F, we have also done

much of the analysis using absolute median income, instead of income relative to the MSA, and the results

are not qualitatively different.

This may be because the relationship between score and income actually is weaker, or it

may simply be the result of only having data on income at the neighborhood level.

2. Possible Reasons for Differences in Scores across Groups

As discussed above, the FTC’s analysis shows a relationship between credit-based

insurance scores and race, ethnicity, and, to a lesser extent, income. The Commission

examined other information in its sample to determine what factors could account for

differences in scores among racial and ethnic groups. The FTC’s database contains some

information on factors that could explain some of the differences in scores among racial

and ethnic groups. Specifically, it includes information on the median income of the

neighborhood in which each consumer lives, and consumers who live in lower-income

neighborhoods tend to have lower scores. It also contains information from which the

age of the consumers whose score is in the database can be inferred,

104

and older

consumers tend to have higher scores. Finally, the FTC’s database contains information

about the gender of the consumers whose score is included (the “first named insured” on

the policy), and men in the FTC database tend to have higher scores than women,

although the difference in average score between men and women in the FTC database

cannot be generalized to the overall population.

105

104

For single-driver households, we know the age of the person for whom we have a credit score. For

multi-driver households, we need to make an assumption about whose age we have. We do this in several

ways. From Social Security Administration (“SSA”) data, we know the gender of the person whose credit

score we have. If there is only one driver in a household with that same gender, we assume that person is

the person for whom we have a credit score. If there are multiple people whose gender matches the SSA

data, we take the oldest, on the assumption that that person is most likely to be the first named insured.

105

We have a score for only one person covered by each policy. From examining our data, it is apparent

that in households with male and female adults (e.g., married couples), it is most often the male driver who

is the first named insured, and therefore the person for whom we have a score. About 75% of multi-driver

policies have a male first named insured, while the split for single-driver policies is 50/50. So, it appears

that the men for whom we have scores are much more likely to be married than the women for whom we

(continued)

Table 3 presents median neighborhood income, age, and gender for racial and

ethnic groups for consumers whose information is in the FTC database. It shows that

African Americans and Hispanics live in neighborhoods with lower median incomes than

non-Hispanic whites and Asians. It reveals that Hispanics and Asians are younger than

non-Hispanic whites and African Americans. It further shows that the African-American

customers in this sample are much more likely to be female than are customers in other

racial and ethnic groups.

106

All of these differences are consistent with African

Americans and Hispanics having lower credit scores.

Figure 12 shows the distribution of scores by race and ethnicity after controlling

for neighborhood income, age, and gender of the person scored. It shows that large

differences remain in the distributions of scores across racial and ethnic groups, and that

these differences are only slightly smaller than they were prior to controlling for these

factors.

107

In particular, prior to controlling for these factors, the median score for

African Americans and Hispanics was in the 23

and 32

percentiles, respectively.

After using these controls, the median score for African Americans and Hispanics rose to

the 27

and 37

percentiles, respectively. In short, consideration of neighborhood

income, age, and gender explains only a small part of the difference in credit-based

insurance scores between racial and ethnic groups. It is not clear what explains the rest

of the difference.

have scores. The differences in score by gender in the FTC database, therefore, cannot be interpreted as the

difference in scores that would be observed between all men and all women, because they also reflect

differences in credit score by marital status and household size.

106

Recall that age and gender, like score, are for the customer who was the “first named insured” on a

policy.

107

It is our understanding that the Federal Reserve Board is undertaking a similar analysis using a richer set

of data about each individual.

3. Impact of Differences in Scores on Premiums Paid

a. Effect on Those for Whom Scores Were Available

The FTC assessed the implications of the differences in credit-based insurance

scores for the premiums that members of different racial and ethnic groups would be

predicted to pay. As discussed above, the FTC database can be used to predict

differences in claims risk with and without the use of scores. These differences, in turn,

can be used to estimate the effects of scores on expected insurance premiums for racial,

ethnic, and income groups.

Figure 13 shows the results of the FTC’s analysis. These are graphs that show the

share of each group with different size changes in their predicted risk between models

where scores were not used and models where scores were used. Comparing across

groups clearly shows that a much larger share of African American and Hispanics had

increases in their predicted risk than did non-Hispanic whites and Asians. When scores

are used, the predicted risk decreased for 62% of non-Hispanic whites and 66% of

Asians. On the other hand, the predicted risk increased for 64% of African Americans

and 53% of Hispanics. These results flow from the fact that, as discussed above, the

scores for African Americans and Hispanics are lower on average than the scores of non-

Hispanic whites and Asians.

Table 4 shows the magnitude of changes in predicted risk for racial and ethnic

groups as a result of the use of scores. The average predicted risk increased by 10% for

African Americans and 4.2% for Hispanics, and dropped by 1.6% for non-Hispanic

whites and 4.9% for Asians.

108

b. Effect on Those for Whom Scores Were Not Available

The FTC also sought to determine whether the likelihood that a credit-based

insurance score could not be generated for a consumer varied across racial and ethnic

groups, and what impact any such differences would be expected to have on the

premiums paid by consumers. A score may not be available for a consumer for one of

two reasons: either it cannot be located for a consumer (a “no-hit”), or a consumer may

have a credit history file, but it may not contain information sufficient to calculate a

credit-based insurance score (a “thin file”).

The FTC database does not contain Social Security Administration race and

ethnicity data for most customers who were “no hits” or “thin files.”

109

The FTC

therefore used United States Census data to determine whether there are differences in

the proportions of racial and ethnic groups that do not have a credit score. Based on

block-level data, the Commission estimates that credit reports could not be located for

9.7% of African Americans, 9.2% of Hispanics, 7.8% of non-Hispanic whites, and 6.4%

of Asians. Similarly, 2.4% of Hispanics, 2.1% of African Americans, 1.8% of non-

108

The relatively large decrease in predicted risk for Asians relative to non-Hispanic whites was surprising,

given how similar the score distributions are for these two groups. In addition, the increase in predicted

risk for Hispanics was only half that of African Americans, even though Hispanics have average scores

closer to African American than to the overall population. Further examination of the results of the models

showed that the inclusion of scores affected the impact of other variables on predicted risk. This, in turn,

affected the predicted risk of Asians and Hispanics. In particular, the impact that short tenure with a firm

and low liability limits had on predicted risk shrank considerably when scores were included in the models.

Asians and Hispanics have low average tenure and low average liability limits, so when the impact of those

characteristics on predicted risk decreased, so did the average predicted risk of Asians and Hispanics.

109

The process of obtaining SSA race and ethnicity data relied on obtaining Social Security Numbers or

dates of birth from credit reports; thus we did not receive SSA information for people whose credit reports

could not be located, or who had very little information in their reports. Similarly, we do not have SSA

national origin information for these people, and therefore cannot analyze the impact on immigrants of a

lack of a credit-based insurance score.

Hispanic whites, and 1.8% of Asians had credit reports, but with too little information

available to calculate a score.

110

Note that because these results are based on geographic

data, they may not exactly reflect actual differences between racial and ethnic groups.

The FTC’s assessment indicates that consumers for whom scores were not

available appeared slightly riskier when scores were considered than when they were not.

The Commission compared the results from risk models without scores with results from

risk models with scores that also included categories for “no hit” and “thin file” in

making this determination. No-hit consumers were 1.06 times riskier in a model that

included controls for scores compared to a model that did not. Thin-file consumers were

1.02 times riskier in a model that included controls for scores compared to a model that

did not.

Given the relatively small differences across groups in the share of people who

were “no hits” or “thin files,” and the relatively small impact of not having a score on

predicted risk (as opposed to the large impacts of using scores on the predicted risk of

people in the lowest score deciles, for example), this is unlikely to be an important source

of differences in premiums across racial and ethnic groups. Again, this analysis is limited

by the lack of individual-level data on race and ethnicity for people for whom we do not

have credit scores.

110

The block data were used by assuming each person had a likelihood of being a member of each racial or

ethnic group that was proportional to the share of the population of each group in that person’s block. This

is implemented similarly to how imputed race/ethnicity information for SSA data are used. See Appendix

C for a discussion of that process.

B. Scores as a Proxy for Race and Ethnicity

Section 215 of FACTA mandates that the FTC create a statistical model of

insurance claims that includes credit-based insurance scores, standard non-credit risk

variables, and controls for protected classes under the Equal Credit Opportunity Act.

111

We understand this to require the agency to analyze whether credit-based insurance

scores act as a “proxy” for membership in these classes. As discussed above, we focused

our analysis on effects on different groups defined by race, ethnicity, and income.

Understanding how a proxy functions is critical to the FTC’s analysis. Insurance

companies build statistical models that relate a variety of characteristics of customers

(e.g., age or driving history) to risk. Firms then use these models to predict the average

claims that customers with those characteristics will generate, and these predictions of

risk play a central role in determining the premiums that firms charge.

The risk models that companies build do not include information about race,

ethnicity, or income. If there are large differences in average risk based on race,

ethnicity, or income, then models may attribute some of those differences in risk to other

variables included in the model that differ across these groups. The included variable

thus may act in whole or in part as a statistical “proxy” for the excluded variables of race,

ethnicity, or income.

112

The FTC sought to determine whether credit-based insurance scores act as a

proxy for race, ethnicity, and income in insurance decisions. To determine whether there

is such an effect, and, if so, its magnitude, the Commission conducted three related

111

FACTA § 215(a)(2) (2006); 15 U.S.C. §1681 note (2006).

112

The econometric term for this effect is “omitted variable bias.” The omission of a predictive variable

(such as race, ethnicity, or income) causes the estimated effect of a variable that is correlated with the

omitted variables, such as score, to be “biased” away from the true effect. In this scenario, the direction of

the bias would be to overstate the relationship between score and claims.

analyses. First, the Commission analyzed whether scores predict risk within racial,

ethnic, and income groups. If scores do not predict risk within any group defined by

race, ethnicity, and income, then the sole reason that scores predict risk in the general

population would be because they act as a proxy for membership in different groups.

Second, the Commission analyzed whether average risk differed substantially by

race, ethnicity, and income. If there were no substantial differences in the average risk

across racial and ethnic groups, then there would be no underlying difference for which

scores could act as a proxy. If there are substantial differences in risk across groups,

scores may in part act as a proxy, even if scores also predict risk within groups (and are

therefore not solely acting as a proxy for membership in different groups).

Third, the FTC created models that included controls for race, ethnicity, and

income, along with credit-based insurance scores and the full range of other predictive

variables. The Commission quantified the proxy effect of scores by measuring the

impact of including these additional controls on the estimated relationship between scores

and claims. To provide a basis for comparison, the FTC also conducted this analysis for

several other variables that are predictive of risk.

1. Do Scores Act Solely as a Proxy for Race, Ethnicity, or Income?

Whether credit-based insurance scores predict risk within racial, ethnic, and

income groups provides critical insight into whether scores are a proxy for membership

in these groups. If scores did not predict claims within racial, ethnic, and income groups,

the relationship between scores and claims must come from scores acting as a proxy for

race, ethnicity, and income. On the other hand, if scores do predict risk within groups,

then they do not serve solely as a proxy if used to assess risk for all consumers.

Therefore, the FTC analyzed whether scores predict risk within race, ethnicity, and

income groups.

The results of the FTC’s analysis are presented in Figure 14 for each racial and

ethnic group for each type of automobile insurance coverage. If credit-based insurance

scores predict the amount that insurance companies paid out in claims within each group,

there should be a downward slope on each graph.

113

In other words, as scores increase for

members of each group, the amount paid out on claims should be decreasing.

Although the relatively small sample size for the minority groups in the FTC

database (which is a particular problem for bodily injury coverage, which has relatively

few claims) leads to results that sometimes vary substantially from decile to decile, the

overall pattern observed is that the amount paid out decreases as credit-based insurance

scores increase for each group for each type of coverage.

114

With the exception of

collision coverage, very few of the decile and coverage combinations have estimated risk

for a given racial or ethnic group that is statistically significantly different from that of

the overall sample.

115

Because they show that scores predict risk within groups, these

results show that credit-based insurance scores do not predict risk solely by acting as a

113

These were estimated by including interaction terms between the race/ethnicity variable and the scores

variables. The coefficients on non-race/ethnicity non-score variables are therefore forced to be the same

across groups. Entirely separate models cannot be estimated for many race/ethnicity/coverage

combinations, because the small sample size of the minority groups often leads to the non-convergence of

the estimation procedure.

114

One cell that jumps out as being out of line with that pattern is the ninth decile for African Americans

for comprehensive coverage. Further investigation showed that this result was affected by an outlier; a

single individual with a very large claim, very low earned car years, and a very large nationally-

representative weight had a large impact on the estimated risk for this decile. There are also few African

Americans in the ninth decile. The difference between the estimated risk for African Americans in the

ninth decile and the overall sample in the ninth decile was not statistically significant. When this outlier

was dropped the risk estimate for this decile was similar to the surrounding deciles. The treatment of

outliers is discussed in Appendix F.

115

Statistical significance was determined using a bootstrap procedure with 500 replications. The bootstrap

procedure is discussed in Appendix D.

proxy for membership in racial and ethnic groups.

116

The FTC conducted the same analysis based on neighborhood income. These

results are shown in Figure 15. These graphs show a consistent negative relationship

between amount paid and credit-based insurance score for all neighborhood income

groups. In other words, as scores increased, claims decreased for all income groups.

In short, because scores do predict risk within racial, ethnic, and income groups,

they do not act solely as a proxy for these characteristics.

2. Differences in Average Risk by Race, Ethnicity, and Income

Even though scores do not act solely as a proxy for race, ethnicity, and income,

there may still be some proxy effect. For such a partial proxy effect to occur, there must

be differences in average risk among racial, ethnic, or income groups, i.e., scores can

only have a proxy effect if there is an underlying relationship for which scores can serve

as a proxy. To determine whether such differences exist, the FTC created models that

evaluated the relative amount paid on claims by race, ethnicity, and neighborhood income

for the four main types of automobile insurance coverage. These models included other

risk variables, but not scores. The results of this analysis are shown in Table 5. For

purposes of comparisons in these tables, the FTC assigned a relative value of 1 to the

amount of claims that would be expected to be paid to non-Hispanic white consumers and

to consumers living in high income neighborhoods.

Column (a) shows that Asians and Hispanics had a higher amount of claims paid

under property damage liability coverage than did African Americans and non-Hispanic

116

We also did the same analysis for “foreign born” and “recent immigrants.” The results were similar, and

scores are correlated with risk for those groups.

whites, although the difference was not statistically significant for Hispanics. It also

shows that there was very little relationship between the amount of property damage

liability claims and whether a consumer lives in a neighborhood with a low, middle, or

high income. While Asians did have more claims under property damage liability

coverage, as discussed above, our analysis showed that they had scores that were similar

to the scores of the overall distribution. Therefore, scores cannot act as proxy for being

Asian, so it is unlikely that scores could act as a proxy for race or ethnicity in a model of

property damage liability claims.

Columns (b) through (d) of Table 5 present results concerning the amount paid

out for bodily injury, collision, and comprehensive coverage, respectively. After

controlling for other risk factors, insurance companies paid out 48% more to African

Americans than non-Hispanic whites for bodily injury, 43% more for collision, and 63%

more for comprehensive coverage. Similarly, they paid out 25% more to Hispanics than

non-Hispanic whites for bodily injury, 33% more for collision, and 45% more for

comprehensive coverage. Insurance companies paid out 30% more to Asians than non-

Hispanic whites for collision coverage. These differences were all statistically

significant. The differences for bodily injury liability and comprehensive coverages

between Asians and non-Hispanic whites were relatively small and not statistically

significant. The large differences in average risk on comprehensive coverage for

Hispanics and African Americans should be treated with some caution, as the geographic

risk variable in the FTC database is not a very effective control for geographic variation

in risk on comprehensive coverage.

117

Table 5 shows that the differences among neighborhood income groups were

much smaller than those among racial and ethnic groups. The one substantial difference

in risk was that customers in low-income neighborhoods pose a 16% higher risk for

comprehensive coverage. Again, this may in part be due to the lack of an effective

geographic risk measure for comprehensive coverage.

118

These results show that there were substantial differences in the average risk of

consumers in different racial and ethnic groups for all four major automobile insurance

coverages.

119

For property damage liability coverage, Asians were the only group with

117

The geographic risk measure in the FTC database is based on property damage liability claims, which

result from accidents. The estimated effect of the geographic risk measure is much smaller in the

comprehensive coverage risk models than in the models for the other coverages, suggesting that it is a poor

control for geographic variation in comprehensive coverage risk. According to the Bureau of Justice

Statistics, African Americans and Hispanics are much more likely to be victims of automobile theft (a risk

covered by comprehensive coverage) than non-Hispanic whites. See Bureau of Justice Statistics file

cv0516.csv, available at www.ojp.usdoj.gov/bjs/pub/sheets/cvus/2005/cv0516.csv

; Bureau of Justice

Statistics file cv0517.csv, available at www.ojp.usdoj.gov/bjs/pub/sheets/cvus/2005/cv0517.csv. In the

absence of a good measure of the geographic variation in comprehensive coverage risk, race, ethnicity, and

neighborhood income are likely picking up some of that variation in risk (e.g., they may be acting as a

proxy for other characteristics of neighborhoods that affect comprehensive coverage risk). Additional

support for this hypothesis was found by estimating separate risk and severity models that included race,

ethnicity, and income controls. In those models, race, ethnicity, and income affected only frequency in the

property damage liability, bodily injury liability, and collision coverage models. In the comprehensive

coverage model, race, ethnicity, and income were strongly related to claim severity. This is consistent with

those variables being related to the likelihood of theft claims.

118

Id.

119

We found similar patterns when we used loss ratios as the measure of relative risk, instead of the direct

results of the risk models. The loss ratio is the ratio of payments companies made on claims divided by

premiums customers paid in. Using loss ratios, therefore, shows whether customers in different racial and

ethnic groups generated greater or lesser total payouts on claims, on average, than predicted by the

companies, as reflected in the premiums the customers were charged. Loss ratios were fairly similar across

groups for property damage liability coverage, with Hispanics and Asians generating somewhat more

claims relative to premiums than African Americans and non-Hispanic whites. For bodily injury liability

coverage, collision coverage, and comprehensive coverage, African Americans and Hispanics generated

higher claims relative to premiums than did non-Hispanic whites. The same was true for Asians for

collision coverage, although Asians had a substantially smaller loss ratio for comprehensive coverage than

did any other group. For example, the loss ratios of African Americans and Hispanics for collision

coverage were 83.9% and 85.6%, respectively, for Asians 78.2%, and for non-Hispanic whites the loss ratio

was 63.3%. Unlike in our risk models, the coverage with the largest differences across groups was bodily

injury liability coverage, as opposed to comprehensive coverage. This again suggests that part of the

reason we find such large differences in risk across groups for comprehensive coverage in our models is the

lack of a geographic risk measure that is specific to risk on comprehensive coverage. For the four

(continued)

significantly higher risk. For the other three coverages, Hispanics and African Americans

had substantially higher average payouts on claims than did non-Hispanic whites. Given

that Hispanics and African Americans have much lower credit-based insurance scores, on

average, than do non-Hispanic whites, there is the potential that scores could gain

additional predictive power by acting as a proxy for race and ethnicity in models of

claims under bodily injury, collision, and comprehensive coverages.

3. Controlling for Race, Ethnicity, and Income to Test for a Proxy Effect

a. Existence of a Proxy Effect

The FTC created models that evaluated the relative amount paid on claims by

score decile with and without controls for membership in racial, ethnic, and income

groups for the four main types of automobile insurance coverage. Table 6 shows the

results.

120

For purposes of comparisons on this Table, the FTC assigned the relative value

of 1 to: (1) the amount of claims that would be expected to be paid to consumers in the

highest 10% of credit-based insurance scores; (2) non-Hispanic white consumers; and (3)

consumers living in high-income Census tracts. For each coverage, the first column

shows the predicted relative amount of claims for credit-based insurance score deciles for

a model that does not include controls for race, ethnicity and income. The second

column for each coverage shows the results from models that include scores and controls

for the prohibited factors.

Comparing the two columns for property damage liability coverage (columns (a)

and (b)) reveals that there was very little difference in the impact of credit-based

coverages combined, the loss ratios of the four groups were: for non-Hispanic whites, 62%; for African

Americans, 80%; for Hispanics, 81%; and, for Asians, 67%.

120

Again, the models used here are Tweedie GLMs. Modeling details are given in Appendix D.

insurance scores on predicted risk based on whether the model included controls for

membership in a protected class. The only statistically significant difference was that the

estimated relative risk for the lowest score decile was larger when protected class

controls were included in the model.

121

This is opposite of the change that would occur if

scores were acting as a proxy. This lack of a proxy effect is not surprising, given that the

only statistically significant difference in risk by racial or ethnic group for this coverage

was that Asians had higher average risk. As pointed out above, because Asians have

similar scores, on average, as the population as a whole, scores cannot act as a proxy for

being Asian. The lack of any proxy effect for property damage liability coverage is made

very clear in Figure 16, which shows the estimated relationship between claims risk and

credit-based insurance scores from Table 6.

Table 6 shows that the results were somewhat different for bodily injury liability,

collision, and comprehensive coverage. These are the coverages for which African

Americans and Hispanics had substantially higher average total payouts on claims than

did non-Hispanic whites. The FTC’s analysis revealed that including these controls did

reduce somewhat the effect of scores on predicted risk for these three coverages. The

results show, however, that scores do continue to predict claims strongly if controls for

race, ethnicity, and income are included in the risk models, which means that scores do

not predict risk primarily by acting as a proxy for these characteristics. In addition to

Table 6, the results are presented in Figure 16, which shows the estimated relationship

between scores and risk, with and without controls for race, ethnicity, and income.

Controls for race, ethnicity, and income decreased the impact of scores on predicted risk

121

A 95% confidence interval for the difference between the score decile parameter estimates from the two

models was computed using a bootstrap procedure with 500 replications. Details of the bootstrap

procedure are provided in Appendix C.

for these coverages most for the lowest credit-based insurance score deciles (where

African Americans and Hispanics are disproportionately located), and these decreases

were statistically significant. In short, the FTC’s analysis indicates that credit-based

insurance scores appear to have some proxy effect for three of the four coverages studied,

but that this is not the primary source of their relationship with claims risk. In the next

section, we address the magnitude of the proxy effect.

b. Magnitude of a Proxy Effect

The FTC also sought to determine the magnitude of any proxy effect from the use

of credit-based insurance scores. Controlling for race and ethnicity had the largest impact

on the predicted effect of scores on risk for comprehensive coverage. See columns (g)

and (h) of Table 6. Without these controls, consumers in the lowest 10% of scores were

estimated to pose 1.95 times more risk than consumers in the highest 10%. With the

controls, consumers in the lowest 10% of scores were estimated to pose 1.74 times more

risk than consumers in the highest 10%. As discussed above, this result should be treated

with caution, because it could be affected by the lack of a good measure of the

geographic variation in comprehensive coverage risk.

Controlling for race and ethnicity had a smaller effect on the predicted impact of

scores on risk for bodily injury liability and collision coverage. For bodily injury liability

coverage, without these controls, consumers who are in the lowest 10% of credit-based

insurance scores were estimated to pose 2.20 times more risk than consumers in the

highest 10% of scores, while with controls they were estimated to pose only 2.10 times

more risk. See columns (c) and (d) of Table 6. For collision coverage, without controls,

consumers who are in the lowest 10% of credit-based insurance scores were estimated to

pose 2.03 times more risk than consumers in the highest 10% of scores, while with

controls they posed only 1.93 times more risk. See columns (e) and (f) of Table 6.

It may be difficult to interpret the magnitudes of the proxy effects by examining

changes in the predicted effects of scores on claims risk. An alternative way to measure

the magnitude of the proxy effect is to examine how it affects the impact of scores on the

predicted risk of different race and ethnicity groups. The information presented in Table

7 compares the impact of scores on predicted risk for different groups, with and without

race, ethnicity, and income controls. The first column in Table 7 shows that if scores

were used, then on average the predicted risk of African Americans increased by 10%

and Hispanics increased by 4.2%, while the predicted risk of non-Hispanic whites

dropped by 1.6% and Asians dropped 4.9%.

122

The second column shows the effects of

scores on the average predicted risk of the different groups using the impact of scores on

predicted risk that comes from models that include controls for race, ethnicity, and

income. When these score effects were used, the average predicted risk of African

Americans increased by 8.9% and Hispanics by 3.5%, while the predicted risk of non-

Hispanic whites decreased by 1.4% and Asians by 4.8%.

123

The change in the impact of

scores on predicted risk when race, ethnicity, and income controls were included was

statistically significant for all racial and ethnic groups. However, given that the use of

these controls when determining the effects of scores resulted in relatively small

decreases in the effect of scores on predicted risk for African Americans (10% versus

122

These are the same results that were presented in Table 4.

123

The effects of other variables are held constant between the two models. This was done by using the

estimated risk effects of non-credit risk variables from the models without race, ethnicity, and income

controls, and the estimated risk effects of the score deciles from the models with the controls. The

estimated risk effects of the race, ethnicity, and income controls were not used to predict risk. This hybrid

risk estimate produced an overall average predicted claims payout that was lower than the actual average

amount of claims payouts, so every individual’s predicted risk was then inflated by the ratio of actual

average claims over predicted average claims.

8.9%) and Hispanics (4.2% versus 3.5%), it is apparent that most of the effect of using

scores on these groups is not because scores act as a proxy for race, ethnicity, and

income.

To provide a basis for comparison in evaluating the importance of these proxy

effects, the FTC conducted the same analysis for several other standard risk variables.

This could only be done for a small set of the risk variables in the FTC database.

124

Variables that could be used were tenure (number of years the customer has been with the

company), the model year of the car, and the vehicle identification number (“VIN”),

which the FTC used to obtain information on vehicle characteristics like body type and

safety systems.

125

In addition, there are two risk variables in the FTC database that did

not come from the company policy-level database. These are the geographic risk

measure and the CLUE prior-claims data.

Table 8 shows the results of applying the FTC’s proxy-effect analysis to these

variables. The proxy-effect analysis was applied to these other variables in the same way

it was applied to scores.

126

These other variables have much smaller effects on the

average predicted risk of different racial and ethnic groups than do scores.

127

For three of

124

Most of the standard risk variables that came from the companies’ data had large numbers of missing

values, which reflects the fact that some companies did not collect or store information on some of the

variables. This means that evaluating these variables is complicated by the fact that when a group of

policies has “missing” as the value of a given variable, that may mean that most of the policies came from

the same company. When this is true, the effects of individual variables on risk may be confounded with

differences across companies in the average risk of their customers.

125

The VINs in the FTC database were truncated, so individual cars cannot be identified. While VIN is

missing for a substantial number of cars, this is mainly for cars in earlier model years. Newer model years

have relatively small numbers of missing values, roughly 12%, suggesting that the missing values are

unlikely to be driven primarily by differences across companies in reporting VINs.

126

The first column of Table 8 shows the difference in predicted risk between a model that does not include

the variable being tested and a model that does. The second column shows the difference between a model

that does not include the variable being tested and a model that does include the variable, where the impact

of the variable comes from a model that includes controls for race, ethnicity, and income.

127

There are several reasons that could explain why the impacts of these variables on the predicted risk of

different groups are not as large. It may be because the differences in these variables across groups are not

(continued)

the four variables, adding race, ethnicity, and income controls reduced the magnitude of

the impact that the variables had on the change in predicted risk for different groups.

Adding the geographic risk measure increased average predicted risk 5.4% for African

Americans, 3.3% for Hispanics, and 4.4% for Asians.

128

When controls were included

for race, ethnicity, and income, the impact of the geographic risk measure decreased to

4.7% for African Americans, 2.2% for Hispanics, and 3.6% for Asians. The effect of

tenure on predicted risk for different groups was also reduced by adding race, ethnicity,

and income controls, from 0.4% to 0.1% for African Americans, from 2.4% to 1.9% for

Hispanics, and from 2.1% to 1.7% for Asians. Finally, including race, ethnicity, and

income controls reduced the impact of prior claims on predicted risk from 2.4% to 2.2%

for African Americans, from 0.3% to 0.2% for Hispanics, and from 1.5% to 1.4% for

Asians. While these effects are small in absolute value, they are of a similar proportion

to the effects that these controls have on scores’ impact on the predicted risk of different

racial and ethnic groups. Thus, like scores, these other risk variables also gain some

predictive power from acting as proxies for race, ethnicity, or income.

In summary, the FTC’s analysis shows that credit-based insurance scores do

predict risk within different racial, ethnic, and income groups. Thus, they do not act

solely as a proxy for membership in these groups. Scores, however, do gain a small

amount of additional predictive power because of a proxy effect. Controlling for race

as large as the differences for scores, because the impacts of these variables on predicted risk are not as

large as the impact of scores, and/or because the impact that the inclusion of the variable has on the risk

associated with other variables is not as large as the impact that scores have.

128

Note that this is not a geographic risk measure used by any company, but rather a variable created for

the purpose of this study. In addition, the geographic risk measure is not a very effective control for risk on

comprehensive coverage. A better geographic risk control for comprehensive risk would likely have a

larger impact on the average predicted risk of African Americans and Hispanics, for comprehensive

coverage, and thus overall, given the large risk differences between African Americans and Hispanics

versus non-Hispanic whites for that coverage.

and ethnicity in estimating the relationship between scores and risk causes a small

reduction in the extent to which scores increase the expected risk of African Americans

and Hispanics. Finally, this small proxy effect is not limited to scores, but was found for

three of four other risk variables studied.

VII. ALTERNATIVE SCORING MODELS

FACTA directed the Commission to determine whether credit-based insurance

scoring models could be developed that would reduce the differences in scores for

consumers in protected classes relative to other consumers, yet continue to be effective

predictors of risk.

129

Because race and ethnicity account for the largest differences in

credit-based insurance scores among groups of consumers in the FTC database, the

agency focused on constructing an effective model that decreased differences among

racial and ethnic groups. To the extent practicable, the Commission also sought to build

an effective model that decreased differences among income groups.

As discussed above, credit-based insurance scores are calculated using models

that assign values to credit history variables to calculate numerical scores. To develop a

model that effectively predicts risk while reducing differences between racial and ethnic

groups, the FTC first created a baseline scoring model using the information in its

database. The Commission chose variables for its baseline model with regard only to

their power to generate a score that predicts risk as accurately as possible. The FTC then

used a number of different techniques to try to construct alternative scoring models that

129

FACTA § 215(a)(3) (2006); 15 U.S.C. § 1681 note (2006).

were as predictive as the FTC baseline model, yet had smaller differences in scores

among racial and ethnic groups.

The FTC was not able to develop a credit-based insurance scoring model that met

the dual objectives of maintaining predictive power and decreasing the differences in

scores between racial, ethnic, and income groups. This does not necessarily mean that a

model could not be constructed that meets these objectives. It does strongly suggest,

however, that there is no readily available scoring model or score development

methodology that would do so.

A. The FTC Baseline Model

Developing a baseline model to use for comparisons is the first step in

determining whether a model can produce scores that continue to predict risk but have

smaller differences by race and ethnicity.

130

The FTC used claims information in its

database, the non-credit risk variables in the database, and credit history variables that

were appended to the insurance policy data to build the model.

131

The FTC database

includes 180 credit history variables for each consumer in the development sample. This

is a set of variables that ChoicePoint developed over time for its own score-building, and

they are intended to capture all relevant information in a credit report.

132

130

Using either the ChoicePoint or FICO model as the base model would not be a useful test. Even a very

simplistic model developed with the FTC database is likely to do better at predicting claims in the FTC

database than either of those scores, because it is predicting “within sample.” That is, the model is

predicting the very claims that were used to develop it.

131

The development sample was limited to consumers for whom there is race, ethnicity, and income

information in the FTC database. This demographic information was used only to develop the alternative

scoring models, not the baseline model. Appendix G contains a description of the methodologies used to

produce this credit-based insurance score, as well as the other scoring models discussed in this section.

132

No ChoicePoint model uses all 180 variables, and many of these variables are not used in any model.

The Commission selected variables for its baseline model that would produce

credit-based insurance scores that were effective in predicting total dollars paid out on

claims per year,

133

after controlling for other non-credit risk factors, such as age and

driving history. This model was constructed without giving any consideration to race,

ethnicity, or income. Insurance companies and other private firms that develop scoring

models likewise build their models in a “race blind” fashion.

The variables that the FTC determined produced scores that were most predictive

of the claims of the consumers in its development sample are presented in Table 9.

134

shows the fifteen variables chosen and the scoring factor assigned to each of them.

135

calculate a score for a consumer, the factors for his or her values of each variable are

multiplied together.

136

The first five variables that enter the model each represent different aspects of a

credit report: (1) Delinquencies: presence of derogatory information on the file; (2)

Credit utilization: number of accounts with balance greater than 75% of the credit limit

or all-time high credit balance; (3) Age of accounts: average age of bank revolving

(credit card) accounts; (4) Inquiries; and (5) Type of Credit: presence of an open auto

finance account in the credit report.

137

The variables that entered the model later are all

133

The models are intended to be predictive of claims for all major types of coverage. For this reason,

claims were summed across coverages into a single measure of losses. Claims under first-party medical

coverage’s, “Med Pay” and personal injury protection, are also included in the “total losses” variable.

134

The variable descriptions are proprietary and confidential information of ChoicePoint. Some variable

descriptions have been made public previously. For other variables, we include only a general description

of the type of variable.

135

The Table shows the variables in the order in which they were chosen by the score-building

methodology. Variables chosen earlier are generally those that provide greater predictive power to the

scoring model.

136

The resulting score is the inverse of the relative predicted risk for the consumer. The inverse is used so

that higher scores are associated with lower predicted risk.

137

An auto finance account is an account with a lender associated with a car company, like GMAC or Ford

Motor Credit.

variations on these categories, with the exception of a variable that measures what share

of credit card accounts on the report are currently reported as “open.” The category with

the greatest impact on scores is delinquencies, which makes up six of the fifteen

variables.

138

The scores the FTC baseline model produces did predict risk. Figure 17 shows

the relationship between total claims paid out and the FTC credit-based insurance score

for the four major types of automobile coverage. Each graph shows three lines: (1) the

average total amount paid on claims by score decile in the development sample; (2) the

estimated relationship between scores and claims in the development sample from

models controlling for other risk factors; and, (3) the average total amount paid on claims

by score decile in CLUE data for the period June 2001 to December 2001, for people

who were not in the development sample (an “out of sample” check).

139

If the model

generated scores that effectively predicted risk, then the lines on the graph should slope

downward to the right. The FTC baseline model produced results consistent with this

expected pattern. For example, for bodily injury liability coverage, consumers in the

lowest 10% of scores were more than three and a half times riskier than consumers in the

highest 10% of scores. Even for property damage liability claims, which have the

weakest relationship with the FTC score, consumers in the lowest 10% of scores of the

138

In looking over the model, it is important to keep in mind that a piece of information in a credit report

can be represented in multiple ways and affect multiple variables. This means that care must be taken

when interpreting some of the results. For example, the score factors for variable C show that

delinquencies on a particular kind of account actually lead to a better score, which seems very strange in

isolation. But, it simply means that, in these data, a delinquency on that type of account is less indicative of

risk than delinquencies on other kinds of accounts, since there is another variable in the model that is a

broad measure of delinquencies and has a large negative impact on score.

139

The development sample consists only of the sub-sample of the FTC database for which we obtained

SSA race and ethnicity data. The development sub-sample includes everyone who had a claim in the

company data, so there was no way to use the company data to look at claims outside of the development

sample. Instead, we use CLUE data on claims for a different time period. We were able to use data on

roughly 800,000 policies for these checks.

development sample were more than twice as risky as consumers in the highest 10% of

scores. Consequently, the FTC baseline model is an effective predictor of risk. Figure 17

also shows that the FTC baseline model predicts risk for people outside the development

sample. This result is important in that it shows the FTC baseline model scores do not

simply predict the claims that were used to develop the model.

To establish a baseline for evaluating the results of other models, the FTC also

measured the extent to which its model resulted in differences in scores among racial and

ethnic groups. Figure 18 shows how the four groups were distributed across the range of

FTC baseline-model scores. The horizontal axis shows score deciles, and the vertical

axis shows the share of each group that fell in each decile. The deciles were defined

using the overall distribution of scores, so if a given group had the same distribution of

scores as the overall sample, 10% of that group’s population would fall in each decile.

Figure 18 shows that the FTC baseline model produced lower scores for African

Americans and Hispanics than for non-Hispanic whites and Asians.

140

Table 9 also shows the breakdown of the different racial and ethnic groups across

the variables used in the FTC’s baseline model. The variables that show large differences

across racial and ethnic group are those relating to payment history (e.g., delinquencies)

and public records, and the variable for the share of accounts with high balances relative

to credit limits. The inclusion of these variables in the FTC baseline model explains why

African Americans and Hispanics had lower scores than non-Hispanic whites and Asians.

140

Note that these differences across racial and ethnic groups for the FTC baseline model are very similar

to those for the ChoicePoint scores discussed above, with the only substantial difference being that Asians

were less well represented in the higher score categories for the FTC baseline model than for ChoicePoint

scores.

B. Alternative Scoring Models

1. “Race Neutral” Scoring Models

The FTC credit-based insurance scoring model described in the previous section

provides a baseline for evaluating alternative models. To construct a model that was

“neutral” with respect to race, ethnicity, and income, the FTC created a model in which it

controlled directly for these factors.

141

“Neutral” in this context means that while the

scores produced by the model still may vary across groups, the variables used in the

scoring model should not derive predictive power from a relationship with race, ethnicity,

or income. Controls mitigate the impact of credit history variables that differ widely

among different racial, ethnic, or income groups, if those variables derive a substantial

portion of their power to predict losses from those differences. If controls are used for

race, ethnicity, and income, these variables become less predictive of risk. With this loss

in predictive power, these variables are either not selected for a scoring model at all, or, if

selected, they are not given as much weight.

Table 10 shows the scoring model that was produced if controls for race,

ethnicity, and income were used in the model building process. Most significantly, the

variables selected in this model that controls for race (a race “neutral” model) are

extremely similar to those in the FTC baseline model (a race “blind” model).

Specifically, only two of the fifteen variables are different between these two models, and

these two particular variables have a relatively weak effect on predicting risk. Despite

controlling for race, ethnicity, and income, a very similar set of credit history variables

141

Several authors have proposed this approach. See Elaine Fortowsky and Michael Lacour-Little, Credit

Scoring and Disparate Impact (Dec. 31, 2001), available at

http://fic.wharton.upenn.edu/fic/lacourpaper.pdf; Stephen L. Ross and John Yinger, T

HE COLOR OF CREDIT:

MORTGAGE DISCRIMINATION, RESEARCH METHODOLOGY, AND FAIR LENDING ENFORCEMENT (2002).

thus were found to be most predictive of claims. Even though some of these variables

have large differences across racial and ethnic groups, the variables were chosen not

because they vary by race, ethnicity, or income.

The Commission tried another approach to developing a race “neutral” model to

compare to the FTC baseline model. We constructed a credit-based insurance scoring

model using a development sample that included only non-Hispanic whites. Because

there were no other racial or ethnic groups in the sample used to construct such a model,

the predictive power of the variables selected cannot be attributed to any relationship

with race or ethnicity.

Table 11 shows the variables selected when a model was built using only non-

Hispanic whites as the development sample. Upon first examination, the variables

selected for this model appear quite different from the variables in the FTC baseline

model. Eight of the fifteen variables are different, including the variable with the second

greatest impact. However, there is an important similarity between the variables in these

two models. The same types of variables were found to be the most important:

delinquencies, inquiries, measures of high debt burden, age of the credit report, and type

of credit.

Both race-neutral models that the FTC developed predict risk within the

development sample about as well as the FTC baseline model. Figure 19 compares the

results for each of these models for each of the four types of automobile insurance

coverage.

142

These graphs show that the FTC baseline model (a race blind model)

produced very similar results for each type of coverage as models that controlled for race

142

Although only non-Hispanic whites were used to develop the “non-Hispanic whites” model, the results

shown here are for the complete development sample.

and ethnicity or that were developed using only non-Hispanic whites (race neutral

models). Given the similarity between the types of variables selected for use in these

models, it is not surprising that these scores have comparable power in predicting risk.

Just as their risk prediction is comparable to that of the FTC baseline model, the

race neutral models also display large differences in scores among racial and ethnic

groups. Figure 20 shows the distribution of scores for the different racial and ethnic

groups for the two race neutral models and the FTC baseline model. To facilitate

comparisons, each graph shows the results for all three models for a single racial or

ethnic group. For all groups except Asians, the distribution of people across deciles was

nearly identical for the three scoring models. For Asians, the FTC baseline model and

the model developed using controls for race, ethnicity, and income gave very similar

results. The model built using only non-Hispanic whites, however, produced a

distribution of scores for Asians that was more skewed towards lower scores.

In short, these comparisons show that, although the race neutral models that the

FTC built accurately predict risk, they do not decrease the differences in credit-based

insurance scores among racial and ethnic groups.

2. Model Discounting Variables with Large Differences by Race and

Ethnicity

In addition to developing race neutral models as possible alternatives, the FTC

also constructed alternative models that tried more directly to avoid selecting variables

with large differences among racial and ethnic groups. In building such models, the FTC

measured not just how well a given variable predicted claims, but how well it predicted

race and ethnicity. The FTC then chose the variables that contributed the most to

predicting risk and the least to predicting race and ethnicity.

Table 12 shows one of the models developed using this approach. It is very

different from the models described in the previous two sections. Most significantly,

there are no variables that relate directly to delinquencies, which Tables 9 – 11 showed

varied a great deal among racial and ethnic groups. Most variables selected relate to the

number and type of accounts that a consumer has. In addition, the discounted model

includes variables that relate to the age of the credit account and total indebtedness.

Figure 21 shows that the discounted model is much less predictive of risk than the

FTC baseline model for each of the four types of automobile insurance coverage. The

discounted model does produce credit-based insurance scores that predict risk. However,

each of these graphs shows that the relationship between the credit score and risk is much

weaker (flatter) for the discounted model than for the FTC baseline model. This shows

that this process of avoiding variables with large differences between groups resulted in a

model that is substantially less effective as a predictor of risk than the FTC baseline

model.

Figure 22 compares scores for each racial and ethnic group based on the results

obtained from the discounted model and the FTC baseline model. The model that assigns

consumers in a racial or ethnic group most closely to 10% in each decile (i.e., a flat line

at 10% on the vertical axis) shows the least differences based on race and ethnicity. Each

of these graphs shows that the discounted model resulted in scores with smaller

differences between members of racial and ethnic minority groups than did the FTC

baseline model. These differences were most substantial for African Americans. While

they were still slightly over-represented in the lower score categories, the scores from the

discounted model showed 14% of African Americans are in the bottom 10% of scores.

The scores from the FTC baseline model, in contrast, showed 27% of African Americans

in the bottom 10% of scores. Although the discounted model did substantially reduce the

differences in scores among members of racial and ethnic groups, as discussed above, it

also provides far less effective risk prediction.

In summary, the FTC’s inability to build a model that produces scores that

continues to predict risk accurately at the same time as narrowing the differences in

scores among racial and ethnic minority groups are by no means definitive. Perhaps

someone could develop a model that meets both of these objectives. The FTC’s inability

to build to such a model, however, strongly suggests that there is no readily available

approach for doing so.

VIII. CONCLUSION

The FTC’s analysis demonstrates that credit-based insurance scores are effective

predictors of risk under automobile insurance policies. Using scores is likely to make the

price of insurance conform more closely to the risk of loss that consumers pose, resulting,

on average, in higher-risk consumers paying higher premiums and lower-risk consumers

paying lower premiums. It has not been clearly established why scores are predictive of

risk.

Credit-based insurance scores may benefit consumers overall. Scores may permit

insurance companies to evaluate risk with greater accuracy, which may make them more

willing to offer insurance to higher-risk consumers. Scores also may make the process of

granting and pricing insurance quicker and cheaper, cost savings that may be passed on to

consumers in the form of lower premiums. However, little hard data was submitted or

available to the FTC to quantify the magnitude of these potential benefits to consumers.

Credit-based insurance scores are distributed differently among racial and ethnic

groups. The FTC’s analysis revealed that the use of scores for consumers whose

information was included in the FTC’s database caused the average predicted risk for

African Americans and Hispanics to increase by 10% and 4.2%, respectively, while it

caused the average predicted risk for non-Hispanic whites and Asians to decrease by

1.6% and 4.9%, respectively. These changes in predicted risk are likely to have an effect

on the insurance premiums that these groups on average pay.

Credit-based insurance scores predict risk within racial, ethnic, and income

groups. Scores have only a small effect as a “proxy” for membership in racial and ethnic

groups in estimating of insurance risk, remaining strong predictors of risk when controls

for race, ethnicity and income are included in risk models. The FTC’s analysis revealed

that the use of scores for consumers whose information was included in the FTC’s

database caused the average predicted risk for African Americans and Hispanics to

increase by 10% and 4.2%, respectively. The Commission’s analysis also showed that

using the effects of scores on predicted risk that come from models that include controls

for race, ethnicity, and income caused scores to increase the average predicted risk for

African Americans and Hispanics by 8.9% and 3.5%, respectively. The difference

between these two predictions for these two groups (1.1% and 0.7%, respectively) shows

that a relatively small portion of the impact of scores on these groups comes from scores

acting as a proxy for race, ethnicity, and income.

Finally, the FTC was not able to develop an alternative credit-based insurance

scoring model that would continue to predict risk effectively, yet decrease the differences

in scores on average among racial and ethnic groups. This does not mean that a model

could not be constructed that meets both of these objectives. It does strongly suggest,

however, that there is no readily available scoring model that would do so.

TABLES

Performance on Credit Obligations

Late payments/Delinquencies (-)

Collections (generally non-medical) (-)

Public records (judgments or bankruptcies) (-)

Credit-Seeking Behavior

Inquiries (generally non-insurance, non-medical) (-)

New accounts (-)

Use of Credit

Ratio of outstanding balances to available credit (-)

Length of Credit History

Age of oldest account (+)

Average age of all accounts (+)

Types of Credit Used

Department store trade lines (-)

Oil Company trade lines (-)

Travel and Entertainment trade lines (-)

Share of trade lines that are major bank credit cards or mortgages (+)

Note: (-) indicates that high values typically lead to a riskier score, and the converse for (+).

TABLE 1.

Typical Information Used in Credit-Based

Insurance Scoring Models

Score

Decile

verage Number

of Claims Per

Year of Coverage

(per hundred)

verage Cost per

Claim

Average Total

Paid on Claims

Per Year of

Coverage

[(a) x (b)]

(a) (b) (c)

Property Damage Liability Coverage

1 5.65 $2,100 $119

2 4.86 2,119 103

3 4.51 2,105 95

4 4.21 2,078 88

5 4.09 1,982 81

6 3.85 2,028 78

7 3.55 2,006 71

8 3.34 1,994 67

9 3.40 2,062 70

10 3.17 1,981 63

Overall 4.06 $2,053 $83

Bodily Injury Liablility Coverage

1 1.79 $8,560 $153

2 1.59 10,002 159

3 1.39 7,798 109

4 1.39 7,993 111

5 1.19 7,940 95

6 1.01 8,892 89

7 0.91 8,538 78

8 0.89 8,760 78

9 0.85 9,127 78

10 0.77 8,372 64

Overall 1.18 $8,609 $101

(continued…)

TABLE 2.

Claim Frequency, Claim Severity, and Average Total Amount Paid on Claims

Score

Decile

verage Number

of Claims Per

Year of Coverage

(per hundred)

verage Cost per

Claim

Average Total

Paid on Claims

Per Year of

Coverage

[(a) x (b)]

(a) (b) (c)

Collision Coverage

1 11.80 $2,364 $279

2 9.53 2,201 210

3 8.57 2,174 186

4 8.09 2,060 167

5 7.45 2,014 150

6 6.86 2,057 141

7 6.47 2,006 130

8 6.18 1,965 122

9 6.11 2,003 122

10 5.38 2,004 108

Overall 7.64 $2,112 $161

Comprehensive Coverage

1 11.50 $1,032 $119

2 9.69 879 85

3 9.06 828 75

4 9.06 773 70

5 8.34 773 64

6 8.07 752 61

7 7.46 774 58

8 7.42 718 53

9 7.03 722 51

10 6.95 688 48

Overall 8.44 $807 $68

Source: Analysis of FTC Automobile Insurance Policy Database

TABLE 2.

Claim Frequency, Claim Severity, and Average Total Amount Paid on Claims

(Continued)

Note: All numbers on this table represent actual means (i.e. , not derived from any risk modelling procedure).

Median Tract

Income

Median

Age Percent Male

(a) (b) (c)

African Americans $34,876 46 48%

Hispanics $38,475 42 60%

Asians $50,953 42 72%

Non-Hispanic Whites $44,356 48 68%

Source: Analysis of FTC Automobile Insurance Policy Database

TABLE 3.

Median Income and Age, and Gender Make-Up,

by Race and Ethnicity

Note: Age and gender are measured at the individual level. See section VI.A.2 of the report for a

discussion of how the age of the individual was determined. Neighborhood income is the median for the

Census tract where the individual lives. See Appendix C for details on the data sources and the

construction of the database.

Share With a

Decrease

Share With

an Increase

Percent Change in

Mean Predicted Risk

(a) (b) (c)

frican Americans

36% 64% 10.0%

Hispanics 47% 53% 4.2%

Asians 66% 34% - 4.9%

Non-Hispanic Whites

62% 38% - 1.6%

Overall 59% 41% 0.0%

Source: Analysis of FTC Automobile Insurance Policy Database

Note: Predicted change in the amount paid on claims was estimated by comparing individuals' predicted total claims from

risk models that include ChoicePoint Attract Standard Auto credit-based insurance scores with risk models that do not

include scores. (By construction, the average of all changes for the sample is zero.) Both of these models include a

standard set of risk variables as controls, and were run separately for property damage liability, bodily injury liability,

collision, and comprehensive coverage. In the final step we sum the predicted dollar risks for all four types of insurance

coverage with and without the use of credit-based insurance scores. See section VI.A.3 of the report for additional details

on this analysis. Modeling details and a description of the variables included in the models are provided in Appendix D.

TABLE 4.

Change in Predicted Amount Paid on Claims from Using Credit-Based

Insurance Scores, by Race and Ethnicity

Race and Ethnicity

African Americans 1.01 1.48 * 1.43 * 1.63 *

Hispanics 1.11 1.25 * 1.33 * 1.45 *

Asians 1.17 * 1.11 1.30 * 0.96

Non-Hispanic Whites 1.00 1.00 1.00 1.00

Neighborhood Income

Low 0.97 1.01 1.05 1.16 *

Middle 0.95 * 1.02 0.99 1.06 *

High 1.00 1.00 1.00 1.00

Notes:

Source: Analysis of FTC Automobile Insurance Policy Database

TABLE 5.

Estimated Relative Amount Paid on Claims,

by Race, Ethnicity, and Neighborhood Income

Asterisks indicate statistically significantly different from base category at 5% level.

1) For each variable – i.e. race and ethnicity, and neighborhood income – estimated amount paid on claims per year of

coverage is measured relative to a base category. For race and ethnicity, the base category is non-Hispanic whites;

and, for neighborhood income the base category is “high income” neighborhood.

2) Estimated relative amounts paid out on claims per year of coverage for each race, ethnicity and neighborhood income

category in each column are derived from Tweedie GLMs (Generalized Linear Models); which here include a set of

standard risk variables as controls, but not score. Since our GLM models are multiplicative, the relativities shown on this

table are equivalent to the exponentiated regression coefficients of the indicator variables for these categories. Modeling

details and a description of the variables included in the models are provided in Appendix D.

Comprehensive

Coverage

(d)

Property Damage

Liability Coverage

(a)

Bodily Injury

Liability Coverage

(b)

Collision

Coverage

(c)

Score Decile

1 1.70 * 1.73 * 2.20 * 2.10 *

2 1.52 * 1.53 * 2.14 * 2.07 *

3 1.43 * 1.44 * 1.75 * 1.72 *

4 1.35 * 1.35 * 1.66 * 1.65 *

5 1.24 * 1.24 * 1.37 * 1.36 *

6 1.23 * 1.23 * 1.26 * 1.26 *

7 1.13 * 1.12 * 1.15 1.14

8 1.07 1.07 1.13 1.13

9 1.12 * 1.12 * 1.21 1.21

10 1.00 1.00 1.00 1.00

Race and Ethnicity

African Americans - 0.93 - 1.29 *

Hispanics - 1.06 - 1.15

Asians - 1.20 * - 1.15

Non-Hispanic Whites - 1.00 - 1.00

Neighborhood

Income

Low - 0.96 - 0.98

Middle - 0.94 * - 1.00

High - 1.00 - 1.00

(continued. . .)

Asterisks indicate statistically significantly different from base category at 5% level.

TABLE 6.

Estimated Relative Amount Paid on Claims, by Score Decile, Race, Ethnicity,

and Neighborhood Income

Property Damage Liability

Coverage

Bodily Injury Liability

Coverage

(a) (b) (c) (d)

Coefficients in dashed boxes are statistically significantly different across models (within a given coverage

type) at the 5% level.

Score Decile

1 2.03 * 1.93 * 1.95 * 1.74 *

2 1.65 * 1.59 * 1.43 * 1.33 *

3 1.52 * 1.48 * 1.33 * 1.26 *

4 1.39 * 1.36 * 1.28 * 1.23 *

5 1.27 * 1.25 * 1.19 * 1.16 *

6 1.26 * 1.25 * 1.15 * 1.12 *

7 1.16 * 1.15 * 1.12 * 1.10 *

8 1.09 1.08 1.05 1.04

9 1.12 * 1.12 * 1.01 0.99

10 1.00 1.00 1.00 1.00

Race and Ethnicity

African Americans - 1.26 * - 1.46 *

Hispanics - 1.24 * - 1.36 *

Asians - 1.33 * - 0.97

Non-Hispanic Whites - 1.00 - 1.00

Neighborhood

Income

Low - 1.01 - 1.13 *

Middle - 0.97 - 1.04

High - 1.00 - 1.00

Notes:

Source: Analysis of FTC Automobile Insurance Policy Database

Asterisks indicate statistically significantly different from base category at 5% level.

TABLE 6.

Estimated Relative Amount Paid on Claims, by Score Decile, Race, Ethnicity,

and Neighborhood Income (Continued)

Coefficients in dashed boxes are statistically significantly different across models (within a given coverage

type) at the 5% level.

1) For each variable – score, race and ethnicity, and neighborhood income – estimated amount paid on

claims per year of coverage is measured relative to a base category. For scores, the base category is the

(highest) decile of scores; for race and ethnicity, the base category is non-Hispanic whites; and, for

neighborhood income the base category is “high income” neighborhood.

2) Estimated relative amounts paid out on claims per year of coverage for each race, ethnicity and

neighborhood income category in each column are derived from Tweedie GLMs (Generalized Linear

Models); which here include a set of standard risk variables as controls, as well as score deciles. Since

our GLM models are multiplicative, the relativities shown on this table are equivalent to the exponentiated

regression coefficients of the indicator variables for these categories. Modeling details and a description of

the variables included in the models are provided in Appendix D.

(e) (f) (g) (h)

Collision Coverage Comprehensive Coverage

Average Score Effect From

Model Without Race,

Ethnicity, and Income

Controls

Average Score Effect from

Model With Race, Ethnicity,

and Income Controls

(a) (b)

African Americans 10.0% 8.9%

Hispanics 4.2% 3.5%

Asians - 4.9% -4.8%

Non-Hispanic Whites - 1.6% -1.4%

Source: Analysis of FTC Automobile Insurance Policy Database

TABLE 7.

Change in Predicted Amount Paid on Claims from Using Credit-

Based Insurance Scores Without and With Controls for Race,

Ethnicity, and Income, by Race and Ethnicity

Notes:

Column (b): Results in this column are calculated by combining the estimated risk effects of the score

deciles from models with controls for race, ethnicity, and income with the estimated risk effects of non-

credit risk variables from the models used in column (a), which do not include these additional controls.

The estimated risk effects of race, ethnicity, and income were not used to predict risk. This hybrid risk

estimate produced an overall average predicted claims payout that was lower than the actual sample

average amount of claims payouts, so every individual’s predicted risk was then inflated by the ratio of

actual average claims over predicted average claims.

Column (a): Results in this column come from the same analysis that was used to create Table 4.

Predicted change in the amount paid on claims was estimated by comparing individual predicted risk

from risk models that include ChoicePoint Attract Standard Auto credit-based insurance scores with risk

models that do not include scores. All models include a standard set of risk variables as controls, and

were run separately for property damage liability, bodily injury liability, collision, and comprehensive

coverage (in the final step we sum the predicted dollar risks for all four types of insurance coverage); the

same is true for column (b). This procedure is described in section VI.A.3 of the report. Modeling details

and a description of the variables included in the models are provided in Appendix D.

Numbers for all race and ethnicity groups are statistically significantly different across the models

in columns (a) and (b) at the 5% level.

Average Effect of Variable

Without Race, Ethnicity, and

Income Controls

Average Effect of Variable

With Race, Ethnicity, and

Income Controls

(a) (b)

Geographic Risk

African Americans 5.4% 4.7%

Hispanics 3.3% 2.2%

Asians 4.4% 3.6%

Non-Hispanic Whites -1.3% -1.0%

Tenure

African Americans 0.4% 0.1%

Hispanics 2.4% 1.9%

Asians 2.1% 1.7%

Non-Hispanic Whites -0.5% -0.4%

Prior Claims

African Americans 2.4% 2.2%

Hispanics 0.3% 0.2%

Asians 1.5% 1.4%

Non-Hispanic Whites -0.3% -0.3%

(continued…)

TABLE 8.

Change in Predicted Amount Paid on Claims from Using Other Risk Variables,

Without and With Controls for Race, Ethnicity, and Income, by Race and

Ethnicity

Average Effect of Variable

Without Race, Ethnicity, and

Income Controls

Average Effect of Variable

With Race, Ethnicity, and

Income Controls

(a) (b)

Model Year & Other Car Attributes

African Americans -1.0% -1.2%

Hispanics 0.5% 0.5%

Asians 2.8% 2.6%

Non-Hispanic Whites 0.0% 0.0%

Notes:

Source: Analysis of FTC Automobile Insurance Policy Database

Column (a): Results in this column come from an analysis similar to that used to create Table 4 for score. Predicted change in

the amount paid on claims was estimated by comparing individual predicted risk from risk models that included the particular

variable being analyzed here with risk models that did not include the variable. All models include the standard set of risk

controls (including score), and were run separately for property damage liability, bodily injury liability, collision, and

comprehensive coverage (in the final step we sum the predicted dollar risks for all four types of insurance coverage); the same

is true for column (b). This procedure is described in section VI.A.3 of the report. Modeling details and a description of the

variables included in the models are provided in Appendix D.

Column (b): Results in this column are calculated by combining the estimated risk effects of the variable being analyzed from

models with controls for race, ethnicity, and income with the estimated risk effects of all other risk variables from the models

used in column (a), which do not include these additional controls. The estimated risk effects of race, ethnicity, and income

were not used to predict risk. This h

brid risk estimate produced an overall avera

e predicted claims pa

out that was lower tha

the actual sample average amount of claims payouts, so every individual’s predicted risk was then inflated by the ratio of actual

average claims over predicted average claims.

TABLE 8.

Change in Predicted Amount Paid on Claims from Using Other Risk Variables,

Without and With Controls for Race, Ethnicity, and Income, by Race and

Ethnicity (Continued)

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.14

84.5% 56.0% 69.9% 83.0%

1 or more 1.00

15.5% 44.0% 30.1% 17.0%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.25

43.2% 20.3% 28.9% 43.0%

1 - 2 1.16

24.9% 21.8% 24.6% 24.4%

2 - 3 1.09

13.1% 17.3% 15.9% 14.3%

3 - 6 1.04

14.0% 27.6% 23.3% 13.2%

6 or More 1.00

4.8% 13.0% 7.4% 5.1%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.16

3.2% 5.6% 5.5% 2.7%

0 - 24 0.67

3.6% 9.2% 10.0% 6.3%

24 - 51 0.80

10.4% 18.6% 18.6% 16.1%

51 - 64 0.83

9.4% 10.3% 12.0% 11.5%

64 - 99 0.84

34.8% 27.8% 31.3% 36.8%

99 - 205 0.87

36.2% 26.3% 21.5% 25.4%

205 or More 1.00

2.4% 2.1% 1.2% 1.3%

(continued. . .)

TABLE 9.

Baseline Credit-Based Insurance Scoring Model Developed by the FTC

1) Variable A: Presence of Certain Delinquencies or Adverse Public Records on the Credit File

2) Number of Accounts with Balance Greater than 75% of High Credit (Credit Limit)

3) Average Number of Months Bank Revolving Accounts Have Been Open

Share in each category, by race or ethnicity

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade lines 1.30

34.1% 21.6% 15.9% 19.8%

0 1.31

16.6% 13.9% 14.4% 16.2%

1 - 2 1.29

22.0% 22.8% 20.0% 21.3%

2 - 4 1.20

17.8% 23.2% 25.9% 23.6%

4 - 7 1.13

7.3% 12.6% 16.0% 12.7%

7 or more 1.00

2.4% 5.9% 7.8% 6.4%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.13

90.0% 84.1% 88.0% 85.3%

0 or more 1.00

10.0% 15.9% 12.0% 14.7%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.36

77.1% 47.4% 61.4% 75.8%

1 - 9 1.13

22.2% 50.3% 37.4% 23.6%

10 or more 1.00

0.7% 2.2% 1.2% 0.7%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.99

25.1% 23.8% 20.0% 25.1%

0 0.78

71.8% 67.0% 72.5% 71.4%

1 or more 1.00

3.1% 9.2% 7.5% 3.5%

(continued. . .)

6) Number of Accounts 30 Days Late or Worse in the Last 12 Months

4) Variable B: Relates to the Number of Inquiries on the File

Share in each category, by race or ethnicity

5) Number of Open Auto Finance Accounts

7) Variable C: Presence of Delinquencies on a Particular Kind of Account

TABLE 9.

Baseline Credit-Based Insurance Scoring Model Developed by the FTC (Continued)

Share in each category, by race or ethnicity

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.00

25.1% 23.8% 20.0% 25.1%

1 or more 1.17

72.3% 71.2% 74.4% 71.7%

6 or more 1.00

2.5% 5.0% 5.6% 3.2%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.75

5.4% 9.8% 8.9% 4.8%

0 - .135 0.89

2.8% 4.3% 3.5% 2.8%

> .135 1.00

91.7% 85.8% 87.7% 92.3%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.16

83.9% 54.1% 68.9% 82.3%

1 or more 1.00

16.1% 45.9% 31.1% 17.7%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - 6 0.82

31.7% 37.2% 38.3% 37.4%

6 - 9 0.87

15.1% 18.0% 18.0% 15.6%

9 - 20 0.89

26.3% 26.7% 25.0% 26.0%

20 or more 1.00

26.9% 18.1% 18.7% 21.0%

(continued. . .)

Share in each category, by race or ethnicity

10) Variable D: Presence of a Particular Kind of Delinquency on the Account

TABLE 9.

Baseline Credit-Based Insurance Scoring Model Developed by the FTC (Continued)

11) Age of Youngest Account (Months)

Share in each category, by race or ethnicity

8) Number of Department Store Accounts

9) Share of all Bank Revolving Accounts that are Open

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.89

0.2% 0.4% 0.5% 0.0%

0 - 3 0.82

3.7% 5.6% 4.6% 3.7%

3 or more 1.00

96.2% 94.1% 94.9% 96.3%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - .02 1.17

94.6% 80.7% 88.4% 93.8%

.02 - .14 1.20

2.7% 9.4% 6.0% 2.8%

> .14 1.00

2.8% 9.9% 5.6% 3.5%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.62

3.0% 5.0% 5.2% 2.4%

0 0.90

90.1% 75.3% 81.8% 89.9%

1 or more 1.00

6.9% 19.8% 12.9% 7.7%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.92

91.6% 93.7% 88.7% 91.2%

0 or more 1.00

8.4% 6.3% 11.3% 8.8%

Notes:

1) Variables in italics have not been described publicly, and ChoicePoint considers the descriptions of those variables to be

proprietary information.

2) This scoring model was developed to use credit history information to predict the relative risk posed by individuals, where

risk is defined as expected total dollars that would be paid out on claims in a year. To calculate a score for a given individual

with this model, the appropriate factors for each of the 15 variables are multiplied together. The resulting product is the

inverse of the estimated relative riskiness of the individual, based on the individual’s credit history. See Appendix E for a

detailed discussion of the score-building process.

Share in each category, by race or ethnicity

13) Variable F: A Ratio Relating to Delinquencies

14) Number of Bank Revolving Accounts Ever Bad Debt

15) Number of Open Oil Accounts

Share in each category, by race or ethnicity

12) Variable E: Relates to the Number of Accounts in the Credit File

TABLE 9.

Baseline Credit-Based Insurance Scoring Model Developed by the FTC (Continued)

Share in each category, by race or ethnicity

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.14 84.5% 56.0% 69.9% 83.0%

1 or more 1.00 15.5% 44.0% 30.1% 17.0%

2) Number of Accounts with Balance Greater than 75% of High Credit (Credit Limit)

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.26 43.2% 20.3% 28.9% 43.0%

1 - 2 1.17 24.9% 21.8% 24.6% 24.4%

2 - 3 1.11 13.1% 17.3% 15.9% 14.3%

3 - 6 1.05 14.0% 27.6% 23.3% 13.2%

6 or More 1.00 4.8% 13.0% 7.4% 5.1%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.13 3.2% 5.6% 5.5% 2.7%

0 - 24 0.67 3.6% 9.2% 10.0% 6.3%

24 - 51 0.80 10.4% 18.6% 18.6% 16.1%

51 - 64 0.82 9.4% 10.3% 12.0% 11.5%

64 - 99 0.84 34.8% 27.8% 31.3% 36.8%

99 - 205 0.87 36.2% 26.3% 21.5% 25.4%

205 or More 1.00 2.4% 2.1% 1.2% 1.3%

4) Number of Open Auto Finance Accounts

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.12 90.0% 84.1% 88.0% 85.3%

0 or more 1.00 10.0% 15.9% 12.0% 14.7%

(continued. . .)

Share in each category, by race or ethnicity

TABLE 10.

Credit-Based Insurance Scoring Model Developed by the FTC by Including Controls for

Race, Ethnicity, and Neighborhood Income in the Score-Building Process

Share in each category, by race or ethnicity

1) Variable A: Presence of Certain Delinquencies or Adverse Public Records on the Credit File

3) Average Number of Months Bank Revolving Accounts Have Been Open

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade lines 1.30 34.1% 21.6% 15.9% 19.8%

0 1.31 16.6% 13.9% 14.4% 16.2%

1 - 2 1.28 22.0% 22.8% 20.0% 21.3%

2 - 4 1.20 17.8% 23.2% 25.9% 23.6%

4 - 7 1.13 7.3% 12.6% 16.0% 12.7%

7 or more 1.00 2.4% 5.9% 7.8% 6.4%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.35 77.1% 47.4% 61.4% 75.8%

1 - 9 1.14 22.2% 50.3% 37.4% 23.6%

10 or more 1.00 0.7% 2.2% 1.2% 0.7%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.97 25.1% 23.8% 20.0% 25.1%

0 0.78 71.8% 67.0% 72.5% 71.4%

1 or more 1.00 3.1% 9.2% 7.5% 3.5%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.76 5.4% 9.8% 8.9% 4.8%

0 - .135 0.89 2.8% 4.3% 3.5% 2.8%

> .135 1.00 91.7% 85.8% 87.7% 92.3%

(continued. . .)

Share in each category, by race or ethnicity

5) Variable B: Relates to the Number of Inquiries on the File

6) Number of Accounts 30 Days Late or Worse in the Last 12 Months

TABLE 10.

Credit-Based Insurance Scoring Model Developed by the FTC by Including Controls for

Race, Ethnicity, and Neighborhood Income in the Score-Building Process (Continued)

Share in each category, by race or ethnicity

8) Share of all Bank Revolving Accounts that are Open

7) Variable C: Presence of Delinquencies on a Particular Kind of Account

Share in each category, by race or ethnicity

9) Number of Department Store Accounts

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.00 25.1% 23.8% 20.0% 25.1%

1 or more 1.15 72.3% 71.2% 74.4% 71.7%

6 or more 1.00 2.5% 5.0% 5.6% 3.2%

10) Age of Youngest Account (Months)

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - 6 0.81 31.7% 37.2% 38.3% 37.4%

6 - 9 0.87 15.1% 18.0% 18.0% 15.6%

9 - 20 0.89 26.3% 26.7% 25.0% 26.0%

20 or more 1.00 26.9% 18.1% 18.7% 21.0%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - 2 0.89 5.0% 9.9% 6.6% 4.1%

2 or more 1.00 95.0% 90.1% 93.4% 95.9%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.16 83.9% 54.1% 68.9% 82.3%

1 or more 1.00 16.1% 45.9% 31.1% 17.7%

(continued. . .)

Share in each category, by race or ethnicity

TABLE 10.

Credit-Based Insurance Scoring Model Developed by the FTC by Including Controls for

Race, Ethnicity, and Neighborhood Income in the Score-Building Process (Continued)

Share in each category, by race or ethnicity

11) Variable G: Relates to the Number of Accounts in the Credit File

12) Variable D: Presence of a Particular Kind of Delinquency on the Account

Share in each category, by race or ethnicity

13) Number of Open Personal Finance Accounts

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.90 82.0% 66.4% 73.6% 82.9%

0 - 2 0.97 14.9% 24.6% 21.5% 14.2%

2 or more 1.00 3.1% 9.0% 4.9% 2.9%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - .02 1.17 94.6% 80.7% 88.4% 93.8%

.02 - .14 1.20 2.7% 9.4% 6.0% 2.8%

> .14 1.00 2.8% 9.9% 5.6% 3.5%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.62 3.0% 5.0% 5.2% 2.4%

0 0.90 90.1% 75.3% 81.8% 89.9%

1 or more 1.00 6.9% 19.8% 12.9% 7.7%

Notes:

1) Variables in italics have not been described publicly, and ChoicePoint considers the descriptions of those variables to be

proprietary information.

2) This scoring model was developed to use credit history information to predict the relative risk posed by individuals, where

risk is defined as expected total dollars that would be paid out on claims in a year. To calculate a score for a given individual

with this model, the appropriate factors for each of the 15 variables are multiplied together. The resulting product is the

inverse of the estimated relative riskiness of the individual, based on the individual’s credit history. This scoring model was

developed by including controls for race, ethnicity, and neighborhood income during the process of selecting variables for the

scoring model, and when estimating the final factors that are applied to the credit history variables. See Appendix E for a

detailed discussion of the score-building process.

Share in each category, by race or ethnicity

15) Number of Bank Revolving Accounts Ever Bad Debt

Share in each category, by race or ethnicity

14) Variable F: A Ratio Relating to Delinquencies

TABLE 10.

Credit-Based Insurance Scoring Model Developed by the FTC by Including Controls for

Race, Ethnicity, and Neighborhood Income in the Score-Building Process (Continued)

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.23 84.5% 56.0% 69.9% 83.0%

1 or more 1.00 15.5% 44.0% 30.1% 17.0%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade lines 1.25 34.1% 21.6% 15.9% 19.8%

0 - 2 1.25 38.5% 36.7% 34.3% 37.5%

2 or more 1.14 21.5% 29.5% 32.0% 29.3%

5 or more 1.00 6.0% 12.3% 17.8% 13.4%

3) Total Average Debt Burden

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

Invalid past due amount 0.89 0.6% 0.8% 0.9% 0.7%

0 - .19 1.20 41.7% 18.4% 26.2% 44.0%

.19 - .46 1.13 25.5% 22.5% 25.5% 24.7%

.46 - .81 1.06 24.4% 38.8% 33.8% 23.7%

> .81 1.00 7.7% 19.4% 13.6% 6.8%

4) Age of Youngest Account (Months)

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - 6 0.84 31.7% 37.2% 38.3% 37.4%

6 - 14 0.90 30.3% 35.5% 34.4% 31.1%

14 or more 1.00 38.0% 27.3% 27.3% 31.4%

(continued. . .)

Share in each category, by race or ethnicity

TABLE 11.

Credit-Based Insurance Scoring Model Developed by the FTC Using a Sample of Only

Non-Hispanic White Insurance Customers

Share in each category, by race or ethnicity

1) Variable A: Presence of Certain Delinquencies or Adverse Public Records on the Credit File

2) Variable B: Relates to the Number of Inquiries on the File

5) Number of Accounts 30 Days Late in the Last 24 Months

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.15 83.9% 65.3% 73.9% 83.7%

1 or more 1.00 16.1% 34.7% 26.1% 16.3%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.82 5.4% 9.8% 8.9% 4.8%

0 or more 1.00 94.6% 90.2% 91.1% 95.2%

7) Number of Open Auto Finance Accounts

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.10 90.0% 84.1% 88.0% 85.3%

0 or more 1.00 10.0% 15.9% 12.0% 14.7%

8) Average Number of Months Account have been Open

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - 32 0.68 3.8% 6.4% 9.8% 9.1%

32 - 75 0.90 30.5% 42.5% 45.2% 40.5%

75 - 118 0.95 41.7% 34.8% 32.7% 37.2%

118 or more 1.00 24.0% 16.4% 12.3% 13.2%

9) Number of Open Accounts

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - 12 1.10 81.3% 76.0% 76.8% 75.4%

12 or more 1.00 18.7% 24.0% 23.2% 24.6%

(continued. . .)

Share in each category, by race or ethnicity

TABLE 11.

Credit-Based Insurance Scoring Model Developed by the FTC Using a Sample of Only

Non-Hispanic White Insurance Customers (Continued)

6) Share of all Bank Revolving Accounts that are Open

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.28 98.8% 95.4% 97.7% 98.7%

1 or more 1.00 1.2% 4.6% 2.3% 1.3%

11) Ratio of Open Personal Financial Accounts to Total Open Accounts

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.90 82.0% 66.4% 73.6% 82.9%

0 or more 1.00 18.0% 33.6% 26.4% 17.1%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.16 83.9% 54.1% 68.9% 82.3%

1 or more 1.00 16.1% 45.9% 31.1% 17.7%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.90 25.1% 23.8% 20.0% 25.1%

0 0.81 71.8% 67.0% 72.5% 71.4%

1 or more 1.00 3.1% 9.2% 7.5% 3.5%

(continued. . .)

Share in each category, by race or ethnicity

10) Variable H: Presence of a Particular Kind of Delinquency on the Account

TABLE 11.

Credit-Based Insurance Scoring Model Developed by the FTC Using a Sample of Only

Non-Hispanic White Insurance Customers (Continued)

Share in each category, by race or ethnicity

12) Variable D: Presence of a Particular Kind of Delinquency on the Account

13) Variable C: Presence of Delinquencies on a Particular Kind of Account

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

Disputed 1.41 0.2% 0.5% 0.5% 0.0%

0 - 2 0.85 2.2% 5.0% 3.3% 2.4%

2 or more 1.00 97.6% 94.6% 96.2% 97.5%

15) Number of Bank Installment Accounts Ever Bad Debt

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.37 44.0% 46.8% 48.0% 49.9%

0 1.37 54.7% 49.5% 49.6% 48.7%

1 or more 1.00 1.3% 3.7% 2.4% 1.3%

Notes:

1) Variables in italics have not been described publicly, and ChoicePoint considers the descriptions of those variables to be

proprietary information.

2) This scoring model was developed to use credit history information to predict the relative risk posed by individuals, where

risk is defined as expected total dollars that would be paid out on claims in a year. To calculate a score for a given individual

with this model, the appropriate factors for each of the 15 variables are multiplied together. The resulting product is the

inverse of the estimated relative riskiness of the individual, based on the individual’s credit history. This scoring model was

developed using a development sample of only non-Hispanic white insurance customers. See Appendix E for a detailed

discussion of the score-building process.

Share in each category, by race or ethnicity

14) Variable I: Relates to the Number of Accounts in the Credit File

TABLE 11.

Credit-Based Insurance Scoring Model Developed by the FTC Using a Sample of Only

Non-Hispanic White Insurance Customers (Continued)

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.22 5.6% 10.0% 8.9% 4.9%

$0 - $1,000 1.34 36.0% 28.6% 34.4% 38.1%

$1000 - $3,000 1.25 20.1% 18.0% 17.9% 20.8%

$3,000 - $14,000 1.14 27.4% 31.9% 29.5% 25.5%

$14,000 or more 1.00 10.9% 11.5% 9.2% 10.7%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.72 0.2% 0.4% 0.5% 0.0%

0 - 3 0.80 3.7% 5.6% 4.6% 3.7%

3 or more 1.00 96.2% 94.1% 94.9% 96.3%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - .14 0.83 2.1% 2.6% 2.3% 1.6%

.14 - .27 0.89 8.3% 9.1% 8.1% 8.7%

.27 or more 1.00 89.6% 88.4% 89.6% 89.7%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.26 90.0% 84.1% 88.0% 85.3%

0 or more 1.00 10.0% 15.9% 12.0% 14.7%

(continued. . .)

TABLE 12.

Credit-Based Insurance Scoring Model Developed by the FTC by Discounting

Variables with Large Differences Across Racial and Ethnic Groups

Share in each category, by race or ethnicity

1) Variable J: Indebtedness on Accounts of a Particular Type

2) Variable E: Relates to the Number of Accounts in the Credit File

3) Share of all Accounts that are Open

4) Number of Open Auto Finance Accounts

Share in each category, by race or ethnicity

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.13 68.5% 68.5% 67.9% 69.4%

0 or more 1.00 31.5% 31.5% 32.1% 30.6%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.85 91.6% 93.7% 88.7% 91.2%

0 or more 1.00 8.4% 6.3% 11.3% 8.8%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.00 91.6% 93.7% 88.7% 91.2%

0 - .0741 0.86 4.6% 4.3% 6.2% 5.3%

.0741 or more 1.00 3.8% 2.1% 5.2% 3.5%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 1.15 79.3% 75.6% 74.7% 74.5%

1 or more 1.00 20.7% 24.4% 25.3% 25.5%

(continued. . .)

TABLE 12.

Credit-Based Insurance Scoring Model Developed by the FTC by Discounting

Variables with Large Differences Across Racial and Ethnic Groups (Continued)

Share in each category, by race or ethnicity

8) Number of Accounts Opened in the Last 3 Months

7) Ratio of Open Oil Accounts to Total Open Accounts

Share in each category, by race or ethnicity

5) Number of Open Bank Installment Accounts

6) Number of Open Oil Accounts

Share in each category, by race or ethnicity

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.06 36.1% 37.2% 34.0% 31.7%

1 - 5 1.06 51.5% 49.0% 52.3% 56.0%

5 or more 1.00 12.4% 13.7% 13.6% 12.3%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

0 - 2 0.88 98.3% 97.8% 98.5% 98.5%

2 or more 1.00 1.7% 2.2% 1.5% 1.5%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.77 5.4% 9.8% 8.9% 4.8%

0 - 6 0.96 77.8% 75.2% 74.6% 68.3%

6 or more 1.00 16.8% 15.0% 16.6% 26.9%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.99 32.0% 31.7% 28.1% 33.2%

0 - .36 0.93 58.9% 59.0% 61.6% 60.0%

.36 or more 1.00 9.1% 9.3% 10.3% 6.8%

(continued. . .)

TABLE 12.

Credit-Based Insurance Scoring Model Developed by the FTC by Discounting

Variables with Large Differences Across Racial and Ethnic Groups (Continued)

12) Ratio of Open Department Store Accounts to Total Open Accounts

Share in each category, by race or ethnicity

10) Age of Last Activity

9) Number of Credit Union Accounts

11) Variable K: Number of Accounts of a Particular Type

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 1.00 68.5% 68.5% 67.9% 69.4%

0 - .2917 1.10 27.5% 28.3% 28.6% 27.2%

.2917 or more 1.00 4.0% 3.2% 3.5% 3.4%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

$0 - $3,000 0.91 4.7% 6.8% 6.7% 4.6%

$3,000 or more 1.00 95.3% 93.2% 93.3% 95.4%

Category Factor

Non-Hispanic

Whites

African

Americans

Hispanics Asians

No trade line of this type 0.97 49.8% 48.9% 46.3% 45.1%

0 - .0789 0.95 7.4% 8.9% 8.6% 8.6%

.0789 or more 1.00 42.8% 42.2% 45.1% 46.3%

Notes:

1) Variables in italics have not been described publicly, and ChoicePoint considers the descriptions of those variables to be

proprietary information.

2) This scoring model was developed to use credit history information to predict the relative risk posed by individuals, where

risk is defined as expected total dollars that would be paid out on claims in a year. To calculate a score for a given individual

with this model, the appropriate factors for each of the 15 variables are multiplied together. The resulting product is the

inverse of the estimated relative riskiness of the individual, based on the individual’s credit history. This scoring model was

developed by discounting the predictive power of variables that had large differences across racial and ethnic groups, so

that those variables would be less likely to be chosen by the score-building procedure. See Appendix E for a detailed

discussion of the score-building process.

13) Ratio of Open Bank Installment Accounts to Total Open Accounts

Share in each category, by race or ethnicity

14) Variable L: Based on Total Available Credit

TABLE 12.

Credit-Based Insurance Scoring Model Developed by the FTC by Discounting

Variables with Large Differences Across Racial and Ethnic Groups (Continued)

15) Ratio of Open Credit Union Accounts to Total Open Accounts

Share in each category, by race or ethnicity

FIGURES

FIGURE 1.

Estimated Average Amount Paid Out on Claims,

Relative to Highest Score Decile

Without Controlling for Other Risk Variables

After Controlling for Other Risk Variables

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Bodily Injury

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 2.

Frequency and Average Size (Severity) of Claims,

Relative to Highest Score Decile

Frequency of Claims

Average Size of Claims (Severity)

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Frequency

or Severity

Bodily Injury

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Frequency

or Severity

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Frequency

or Severity

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Frequency

or Severity

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 3.

"CLUE" Claims Data:

Average Amount Paid Out on Claims,

Relative to Highest Score Decile

Company Submitted Data (July 2000 - June 2001)

Clue July 2000 - June 2001

Clue July 2001 - December 2001

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Bodily Injury

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relativ e

Claims

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 4.

By Model Year of Car:

Estimated Average Amount Paid Out on Claims,

Relative to Highest Score Decile

(Property Damage Liability Coverage)

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relative Claims

1992 and Prior

1993-1996

1997 and up

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 5.

Change in Predicted Amount Paid on Claims

from Using Scores

0.0

3.0

6.0

9.0

12.0

15.0

-30% -20% -10% 0% 10% 20% 30% 40% 50% 60%

Percent Change

% Share of

Sample

Percent with

Increase: 41%

Percent with

Decrease: 59%

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 6.

The Ratio of Uninsured Motorist Claims to

Liability Coverage Claims

(1996-2003)

States Allowing the Use of Credit-Based Insurance Scores

States Not Allowing the Use of Credit-Based Insurance Scores

See notes on Figures at the end of this section.

Source: Analysis of data from several National Association of Insurance Commissioners Database Reports.

Property Damage

Liability Coverage

0.1

0.2

0.3

0.4

1996 1997 1998 1999 2000 2001 2002 2003

Year

Ratio

Bodily Injury

Liability Coverage

0.1

0.2

0.3

0.4

1996 1997 1998 1999 2000 2001 2002 2003

Year

Ratio

FIGURE 7.

Share of Cars Insured through States' "Residual Market" Insurance Programs

(1996-2003)

States Allowing the Use of Credit-Based Insurance Scores

States Not Allowing the Use of Credit-Based Insurance Scores

1996 1997 1998 1999 2000 2001 2002 2003

Ye a r

Percent

See notes on Figures at the end of this section.

Source: Analysis of data from several National Association of Insurance Commissioners Database Reports.

FIGURE 8.

Distribution of Scores,

by Race and Ethnicity

10%

15%

20%

25%

30%

Lowest

Scores

Highest

ScoresDeciles of ChoicePoint Credit-Based Insurance Scores

Pe r c e n t

African Americans

Hispanics

Asians

Non-Hispanic Whites

Equal Distribution Line

FIGURE 9.

Distribution of Race and Ethnicity, by Score Decile

20%

40%

60%

80%

100%

Lowest

Scores

Highest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Percent

Non-Hispanic Whites African Americans Hispanics Asians

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 10.

Distribution of Scores,

by Neighborhood Income

10%

15%

20%

25%

30%

Low est

Scores

Highest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Percent

Low-to-Moderate Income

Middle

High

Equal Distribution Line

FIGURE 11.

Distribution of Neighborhood Income, by Score Decile

20%

40%

60%

80%

100%

Lowest

Scores

Highest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Pe rce n t

Low-to-Moderate Income Middle Income High Income

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 12.

Distribution of Scores by Race and Ethnicity,

After Controlling for Age, Gender, and Neighborhood Income

With Controls

Without Controls

Equal Distribution Line

Non-Hispanic Whites

10%

15%

20%

25%

30%

Low est

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Pe r c e n t

African Americans

10%

15%

20%

25%

30%

Low est

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Pe r c e n t

His p ani cs

10%

15%

20%

25%

30%

Low est

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Pe r c e n t

Asians

10%

15%

20%

25%

30%

Low est

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Pe r c e n t

Highest

Scores

Highest

Scores

Highest

Scores

Highest

Scores

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 13.

By Race and Ethnicity:

Change in Predicted Amount Paid on Claims from Using Scores

Non-Hispanic Whites

-30%-20%-10% 0% 10% 20% 30% 40% 50% 60%

Percent Change

% Share

of Group

Percent with

Decrease: 62%

Percent with

Increase: 38%

African Americans

-30% -20% -10% 0% 10% 20% 30% 40% 50% 60%

Percent Change

% Share

of Group

Percent with

Decrease: 36%

Per cent w ith

Increase: 64%

Hisp an ics

-30% -20% -10% 0% 10% 20% 30% 40% 50% 60%

Percent Change

% Share

of Group

Percent with

Decrease: 47%

Percent w ith

Increase: 53%

Asians

-30% -20% -10% 0% 10% 20% 30% 40% 50% 60%

Percent Change

% Share

of Group

Per c ent w ith

Decrease: 66%

Per cent w ith

Increase: 34%

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 14.

By Race and Ethnicity:

Estimated Average Amount Paid Out on Claims,

Relative to Non-Hispanic Whites in Highest Score Decile

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relat iv e

Claims

Bodily Injury

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relativ e

Claims

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relat iv e

Claims

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relat iv e

Claims

Non-Hispanic Whites

African Americans

Hispanics

Asians

Highest

Scores

Highest

Scores

Highest

Scores

Highest

Scores

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 15.

By Neighborhood Income:

Estimated Average Amount Paid Out on Claims,

Relative to People in Highest Score Decile in High Income Areas

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relat iv e

Claims

Bodily Injury

Liability Coverage

0.7

1.0

1.3

1.6

1.9

2.2

2.5

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Score

Relat iv e

Claims

Low-to-Moderate Income

Middle Income

High Income

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 16.

Estimated Average Amount Paid Out on Claims,

Relative to Highest Score Decile, with and without Controls for Race, Ethnicity,

and Neighborhood Income

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relat iv e

Claims

Bodily Injury

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highest

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relat iv e

Claims

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relat iv e

Claims

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

Lowest

Scores

Highes

Scores

Deciles of ChoicePoint

Credit-Based Insurance Scores

Relat iv e

Claims

Without Race, Ethnicity and Neighborhood Income Controls

With Race, Ethnicity and Neighborhood Income Controls

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 17.

FTC Baseline Model -

Estimated Average Amount Paid Out on Claims,

Relative to Highest Score Decile

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Low est

Scores

Highes

Scores

Deciles of FTC

Credit-Based Insurance Scores

Relat iv e

Claims

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Low est

Scores

Highes

Scores

Deciles of FTC

Credit-Based Insurance Scores

Relat iv e

Claims

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Low est

Scores

Highes

Scores

Deciles of FTC

Credit-Based Insurance Scores

Relat iv e

Claims

Bodily Injury

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

3.6

4.0

Lowest

Scores

Highes

Scores

Deciles of FTC

Credit-Based Insurance Scores

Relative

Claims

Within Sample with Controls

Within Sample without Controls

Out of Sample without Controls

Note that the vertical scale on these graphs is different than for previous graphs of relative claims and score deciles.

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 18.

Distribution of FTC Baseline Model Credit-Based Insurance Scores,

by Race and Ethnicity

10%

15%

20%

25%

30%

Lowest

Scores

Highest

Scores

Deciles of FTC Credit-Based

Insurance Scores

Percent

African Americans

Hispanics

Asians

Non-Hispanic Whites

Equal Distribution Line

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 19.

FTC Score Models

Built Controlling for Race, Ethnicity, and Neighborhood Income:

Estimated Average Amount Paid Out on Claims,

Relative to Highest Score Decile

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Lowest

Scores

Highes

Scores

Deciles of FTC Credit-Based

Insurance Scores

Relat iv e

Claims

Bodily Injury

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

3.6

4.0

Low est

Scores

Highes

Scores

Deciles of FTC Credit-Based

Insurance Scores

Relat iv e

Claims

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Lowest

Scores

Highes

Scores

Deciles of FTC Credit-Based

Insurance Scores

Relat iv e

Claims

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Lowest

Scores

Highes

Scores

Deciles of FTC Credit-Based

Insurance Scores

Relat iv e

Claims

FTC Baseline Model

Model Built with Race, Ethnicity, and Income Controls

Model Built with Non-Hispanic Whites Only

Note that the vertical scale on these graphs is different than for some previous graphs of relative claims and score deciles.

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 20.

Distribution of FTC Credit-Based Insurance Scores,

by Race and Ethnicity (A)

Non-Hispanic Whites

10%

15%

20%

25%

30%

Low est

Scores

Deciles of FTC Credit-Based

Insurance Scores

Pe rc e n t

African Americans

10%

15%

20%

25%

30%

Low est

Scores

Deciles of FTC Credit-Based

Insurance Scores

Pe r c e n t

His p ani cs

10%

15%

20%

25%

30%

Low est

Scores

Deciles of FTC Credit-Based

Insurance Scores

Pe r c e n t

Asians

10%

15%

20%

25%

30%

Low est

Scores

Deciles of FTC Credit-Based

Insurance Scores

Pe r c e n t

FTC Baseline Model

Model Built with Race, Ethnicity, and Income Controls

Model Built with Non-Hispanic Whites Only

Equal Distribution Line

Highest

Scores

Highest

Scores

Highest

Scores

Highest

Scores

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 21.

An Additional FTC Credit-Based Insurance Scoring Model:

The "Discounted Predictiveness" Model

Estimated Average Amount Paid Out on Claims,

Relative to Highest Score Decile

Property Damage

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Lowest

Scores

Highes

Scores

Deciles of FTC Credit-Based

Insurance Scores

Relat iv e

Claims

Bodily Injury

Liability Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

3.6

4.0

Lowest

Scores

Highes

Scores

Deciles of FTC Credit-Based

Insurance Scores

Relat iv e

Claims

Collision

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Lowest

Scores

Highes

Scores

Deciles of FTC Credit-Based

Insurance Scores

Relat iv e

Claims

Comprehensive

Coverage

0.8

1.2

1.6

2.0

2.4

2.8

3.2

Lowest

Scores

Highes

Scores

Deciles of FTC Credit-Based

Insurance Scores

Relat iv e

Claims

FTC Baseline Model

“Discounted Predictiveness” Model

Note that the vertical scale on these graphs is different than for some previous graphs of relative claims and score deciles.

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

FIGURE 22.

Distribution of FTC Credit-Based Insurance Scores,

by Race and Ethnicity (B)

Non-Hispanic Whites

10%

15%

20%

25%

30%

Low est

Scores

Deciles of FTC Credit-Based

Insurance Scores

Pe r c e n t

African Americans

10%

15%

20%

25%

30%

Low est

Scores

Deciles of FTC Credit-Based

Insurance Scores

Pe r c e n t

His p ani cs

10%

15%

20%

25%

30%

Low est

Scores

Deciles of FTC Credit-Based

Insurance Scores

Pe r c e n t

Asians

10%

15%

20%

25%

30%

Low est

Scores

Deciles of FTC Credit-Base

Insurance Scores

Pe r c e n t

FTC Baseline Model

“Discounted Predictiveness” Model

Equal Distribution Line

Highest

Scores

Highest

Scores

Highest

Scores

Highest

Scores

See notes on Figures at the end of this section.

Source: Analysis of FTC Automobile Insurance Policy Database

Notes on Figures

Figure 1:

The lines labeled “without controlling for other variables” show the actual average

amount paid out on claims per year of coverage for each score decile, relative to the

highest score decile. These are derived from the information in Table 2. For example,

the relativity for the lowest decile on the PD graph has a value of 1.89. This number is

calculated from column (c) on Table 2; by taking the average total paid on PD claims per

year of coverage for the 1st decile ($118.73) and dividing it by the respective value for the

10th decile ($62.70).

The lines labeled “after controlling for other variables” show the predicted amount paid

out on claims per year of coverage for each score decile, relative to the highest score

decile, from Tweedie GLMs (Generalized Linear Models) of claims risk that included

score and a set of standard risk variables as controls. Since our GLM models are

multiplicative, the relativities shown by these lines are equivalent to the exponentiated

coefficients of the score decile indicator variables. Modeling details and a description of

the variables in the models are provided in Appendix D.

Figure 2:

The lines labeled “frequency of claims” show the predicted number of claims per year of

coverage for each score decile, relative to the highest score decile, from Poisson GLM

models (“Poisson Regressions”) that included score and a set of standard risk variables as

controls. Since our GLM models are multiplicative, the relativities shown by these lines

are equivalent to the exponentiated coefficients of the score decile indicator variables.

Modeling details and a description of the variables in the models are provided in

Appendix D.

The lines labeled “average size of claims” show the predicted average size of claims for

each score decile, relative to the highest score decile, from Gamma GLM models that

included score and a set of standard risk variables as controls. Since our GLM model is

multiplicative, the relativities shown by these lines are equivalent to the exponentiated

coefficients of the score decile indicator variables. Modeling details and a description of

the variables in the models are provided in Appendix D.

Figure 3:

“CLUE” stands for Comprehensive Loss Underwriting Exchange. This informational/

database exchange service is run by ChoicePoint, which collects data on claims from

most major automobile insurance firms in the United States. These data allow firms to

determine whether a potential new customer has filed a claim under a previous policy

with another firm, and use that information in underwriting and rating.

Each line on this graph shows the average total amount paid out on claims per year of

coverage for each score decile, relative to the highest decile. These results do not include

controls for other risk variables because reliable non-credit risk variables are not

available for the CLUE claims data. For this figure we use the full sample of 1.4 million

policies, as opposed to the set of policies within the sub-sample of 400,000 normally

used. This is because the latter would have proved a very limited sub-sample for the

CLUE analysis for the half a year period moving forward, i.e., for July 2001 to December

2001. See Appendix C for a description of the company-provided claims data and the

CLUE database and claims data.

Figure 4:

Each line shows the predicted amount paid out on claims per year of coverage for each

score decile, relative to the highest score decile, for each of three ranges of car model

years from a Tweedie GLM risk model of claims that included score and a set of standard

risk variables as controls. The different lines for the three groups of model years were

estimated by interacting three model year range indicator variables with the score decile

indicator variables. Modeling details and a description of the variables included in the

models are provided in Appendix D.

Figure 5:

Predicted change in premium was estimated by comparing individuals' predicted total

claims from risk models that included ChoicePoint Attract Standard Auto credit-based

insurance score decile indicator variables with risk models that did not include scores.

(By construction, the average of all changes is zero.) Both of these models were run

separately for property damage liability, bodily injury liability, collision, and

comprehensive coverage. In the final step we summed the predicted dollar risks for all

four types of insurance coverage with and without the use of credit-based insurance

scores. See section V.A. of the report for additional details on this analysis. Modeling

details and a description of the variables included in the models are provided in Appendix

Figure 6:

Analysis based on data from several National Association of Insurance Commissioners

Database Reports. (e.g., National Association of Insurance Commissioners, “Auto

Insurance Database Report 2003/2004” (2006)) The states included in the category

“states not allowing the use of credit-based insurance scores” are California, New Jersey,

Massachusetts, and Hawaii. The category "states allowing the use of credit-based

insurance scores" includes all other states, except South Carolina and Texas (for which

complete information was not provided in the NAIC reports).

Credit-based insurance scores for use in automobile insurance were first commercially

available in 1995, and were widely adopted by insurance companies (in states that

allowed their use) during the late 1990s.

Figure 7:

The “residual market” consists of state-sponsored programs to sell insurance to drivers

who are unable to purchase insurance in the normal “voluntary” market. Analysis based

on data from several National Association of Insurance Commissioners Database

Reports. (e.g., National Association of Insurance Commissioners, “Auto Insurance

Database Report 2003/2004” (2006)) The states included in the category “states not

allowing the use of credit-based insurance scores” are California, New Jersey,

Massachusetts, and Hawaii. The category "states allowing the use of credit-based

insurance scores" includes all other states, except South Carolina and Texas (for which

information was not provided in the NAIC report).

Credit-based insurance scores for use in automobile insurance were first commercially

available in 1995, and were widely adopted by insurance companies (in states that

allowed their use) during the late 1990s.

Figure 8:

Each line shows the share of each racial and ethnic group that is in each of the ten deciles

of the ChoicePoint Attract Standard Auto credit-based insurance score. If each racial and

ethnic group had the same distribution of scores, 10% of each group would be in each

decile.

Figure 9:

[No Notes]

Figure 10:

Each line shows the share of each neighborhood income group that is in each of the ten

deciles of the ChoicePoint Attract Standard Auto credit-based insurance score. If each

neighborhood income group had the same distribution of scores, 10% of each group

would be in each decile.

Figure 11:

[No Notes]

Figure 12:

Each line shows the share of each racial and ethnic group that is in each of the ten deciles

of the ChoicePoint Attract Standard Auto credit-based insurance score after controlling

for age, gender, and neighborhood income. This was calculated based on the residuals

from an Ordinary Least Squares regression of ChoicePoint Attract Standard Auto credit-

based insurance scores on age, gender, and neighborhood income. If each racial and

ethnic group had the same distribution of scores, after controlling for age, gender, and

neighborhood income, 10% of each group would be in each decile.

Figure 13:

Predicted change in premium was estimated by comparing individuals' predicted total

claims from risk models that included ChoicePoint Attract Standard Auto credit-based

insurance scores with risk models that did not include scores. By construction, the

average of all changes for the entire sample is zero as in Figure 5, but the changes by race

or ethnic group are not. See note for Figure 5 above or section V.A. of the report for

additional details on this analysis. Modeling details and a description of the variables

included in the models are provided in Appendix D.

Figure 14:

Each line shows the predicted amount paid out on claims per year of coverage for each

score decile, relative to non-Hispanic whites in the highest score decile, from a Tweedie

GLM risk model of claims that included score and a set of standard risk variables as

controls. These values were generated by interacting the race and ethnicity indicator

variables with the score decile indicator variables. The score decile cut-points used are

the same across all race and ethnicity groups (these are the same deciles used for all

previous Figures). Thus, given the race and ethnicity distributions across score deciles

observed in Figure 8, there are relatively few African Americans and Hispanics in each of

the higher score deciles intervals (i.e., fewer than 10% of their group). Modeling details

and a description of the variables included in the models are provided in Appendix D.

The differences in the estimates of the amount paid out in claims in higher score deciles

versus the bottom score decile, within each race group, are generally statistically

significant (at the 5% level), except for Asians (where they are only significant for

comprehensive coverage). We also estimated the slope for each race and ethnicity group

using a continuous score (as opposed to deciles), and found a statistically significant

downward sloping relationship between score and the amount paid out in claims within

each group, with the exception of bodily injury and property damage for Asians.

Property damage for Asians did have a downward slope but was significant only at the

10% level. Note that Asians are the smallest race or ethnic group in our sample.

Figure 15:

Each line shows the predicted amount paid out on claims per year of coverage for each

score decile, relative to the residents of high-income neighborhoods in the highest score

decile, from a Tweedie GLM risk model of claims that included score and a set of

standard risk variables as controls. These values were generated by interacting the

neighborhood income category indicator variables with the score decile indicator

variables. Modeling details and a description of the variables included in the models are

provided in Appendix D.

Figure 16:

Each line shows the predicted amount paid out on claims per year of coverage for each

score decile, relative to the highest score decile, from a Tweedie GLM risk model of

claims that included score and a set of standard risk variables as controls. Since our

GLM model is multiplicative, the relativities shown by this line are the exponentiated

coefficients of the score decile indicator variables. The lines labeled “with race,

ethnicity, and neighborhood controls” come from a model that also included indicator

variables for race, ethnicity, and Census tract median income category. Modeling details

and a description of the variables included in the models are provided in Appendix D.

Figure 17:

The line labeled “Within Sample” shows the predicted amount paid out on claims per

year of coverage for each score decile relative to the highest score decile, of the FTC

baseline model, from Tweedie GLM risk models of claims that included score and a set

of standard risk variables as controls. Modeling details and a description of the variables

included in the models are provided in Appendix D. Details on the score building

process are provided in Appendix E.

The line labeled “Within Sample without Controls” shows the average total amount paid

out on claims per year of coverage for each score decile relative to the highest decile, of

the FTC baseline model, without controlling for any other risk variables. (This line is

shown for comparison with the “Out of Sample” values below, for which we do not have

controls.)

The “Out of Sample” line is based on CLUE claims data and shows the average total

amount paid out on claims per year of coverage for each score decile relative to the

highest decile, of the FTC baseline model, without controlling for any other risk variables

(since reliable non-credit risk variables are not available in CLUE). This “Out of

Sample” line is for the period July 2001 to December 2001, and uses CLUE claims data

only for individuals who were not in the score development sample.

The development sample consisted only of the sub-sample of the FTC database for which

we obtained SSA race and ethnicity data, which includes everyone who had a claim in the

company data, so there is no way to use the company data to look at claims outside of the

development sample. Therefore, we use CLUE data on claims for a different time period

and for a different set of people instead (we were able to use data on roughly 800,000

policies for this from the original 1.4 million dataset). See Appendix C for a description

of the CLUE database and claims data. Details on the score building process are

provided in Appendix E.

(Note that the vertical scale on the graphs in this Figure rises higher than it does for

previous graphs of relative claims and score deciles in Figures 1-4 and Figures 14-16)

Figure 18:

Each line shows the share of each racial and ethnic group that is in each of the ten deciles

of the scores produced by the FTC’s baseline credit-based insurance scoring model. If

each racial and ethnic group had the same distribution of scores, 10% of each group

would be in each decile.

Figure 19:

Each line shows the predicted amount paid out on claims per year of coverage for each

score decile, relative to the highest score decile, from Tweedie GLM risk models of

claims that included score and a set of standard risk variables as controls. Since our

GLM models are multiplicative, the relativities shown by these lines are equivalent to the

exponentiated coefficients of the score decile indicator variables. The lines labeled

“baseline model” use scores from the FTC baseline scoring model. The lines labeled

“race, ethnicity, and income controls model” use scores from a model built by controlling

for those variables during the score building process. The lines labeled “Non-Hispanic

whites model” come from a scoring model built using a development sample made up

exclusively of non-Hispanic white insurance customers. Modeling details and a

description of the variables included in the models are provided in Appendix D. Details

on the score building process are provided in Appendix E.

(Note that the vertical scale on the graphs in this Figure rises higher than it does for

previous graphs of relative claims and score deciles in Figures 1-4 and Figures 14-16)

Figure 20:

Each line shows the share of each racial and ethnic group that is in each of the ten deciles

of three FTC credit-based insurance scoring models. The lines labeled “baseline model”

use scores from the FTC baseline scoring model. The lines labeled “race, ethnicity, and

income controls model” use scores from a model built by controlling for those variables

during the score building process. The lines labeled “Non-Hispanic whites model” come

from a scoring model built using a development sample made up exclusively of non-

Hispanic white insurance customers. If each racial and ethnic group had the same

distribution of scores, 10% of each group would be in each decile. Details on the score

building process are provided in Appendix E.

Figure 21:

Each line shows the predicted relative amount paid out on claims per year of coverage for

each score decile, relative to the highest score decile, from Tweedie GLM risk models of

claims that included score and a set of standard risk variables as controls. Since our

GLM models are multiplicative, the relativities shown by these lines are equivalent to the

exponentiated coefficients of the score decile indicator variables. The lines labeled

“baseline model” use scores from the FTC baseline scoring model. The lines labeled

“discounted predictiveness model” use scores from a model built by discounting the

power of a variable to predict risk based on how different the variable was across racial

and ethnic groups. Modeling details and a description of the variables included in the

models are provided in Appendix D. Details on the score building process are provided

in Appendix E.

(Note that the vertical scale on the graphs in this Figure rises higher than it does for

previous graphs of relative claims and score deciles in Figures 1-4 and Figures 14-16)

Figure 22:

Each line shows the share of each racial and ethnic group that is in each of the ten deciles

of two FTC credit-based insurance scoring models. The lines labeled “baseline model”

use scores from the FTC baseline scoring model. The lines labeled “discounted

predictiveness model” use scores from a model built by discounting the power of a

variable to predict risk based on how different the variable was across racial and ethnic

groups. If each racial and ethnic group had the same distribution of scores, 10% of each

group would be in each decile. Details on the score building process are provided in

Appendix E.

APPENDIX A

TEXT OF SECTION 215 OF THE FACT ACT

SEC. 215. STUDY OF EFFECTS OF CREDIT SCORES AND CREDIT-

BASEDINSURANCE SCORES ON AVAILABILITY AND AFFORDABILITYOF

FINANCIAL PRODUCTS.

(a) STUDY REQUIRED.—The Commission and the Board, in consultation with the

Office of Fair Housing and Equal Opportunity of the Department of Housing and Urban

Development, shall conduct a study of—

(1) the effects of the use of credit scores and credit-based insurance scores on the

availability and affordability of financial products and services, including credit

cards, mortgages, auto loans, and property and casualty insurance;

(2) the statistical relationship, utilizing a multivariate analysis that controls for

prohibited factors under the Equal Credit Opportunity Act and other known risk

factors, between credit scores and credit-based insurance scores and the

quantifiable risks and actual losses experienced by businesses;

(3) the extent to which, if any, the use of credit scoring models, credit scores, and

credit-based insurance scores impact on the availability and affordability of credit

and insurance to the extent information is currently available or is available

through proxies, by geography, income, ethnicity, race, color, religion, national

origin, age, sex, marital status, and creed, including the extent to which the

consideration or lack of consideration of certain factors by credit scoring systems

could result in negative or differential treatment of protected classes under the

Equal Credit Opportunity Act, and the extent to which, if any, the use of

underwriting systems relying on these models could achieve comparable results

through the use of factors with less negative impact; and

(4) the extent to which credit scoring systems are used by businesses, the factors

considered by such systems, and the effects of variables which are not considered

by such systems.

(b) PUBLIC PARTICIPATION.—The Commission shall seek public input about the

prescribed methodology and research design of the study described in subsection (a),

including from relevant Federal regulators, State insurance regulators, community, civil

rights, consumer, and housing groups.

(1) IN GENERAL.—Before the end of the 24-month period beginning on the date

of enactment of this Act, the Commission shall submit a detailed report on the

study conducted pursuant to subsection (a) to the Committee on Financial

Services of the House of Representatives and the Committee on Banking,

Housing, and Urban Affairs of the Senate.

(2) CONTENTS OF REPORT.—The report submitted under paragraph (1) shall

include the findings and conclusions of the Commission, recommendations to

address specific areas of concerns addressed in the study, and recommendations

for legislative or administrative action that the Commission may determine to be

necessary to ensure that credit and credit-based insurance scores are used

appropriately and fairly to avoid negative effects.

APPENDIX B

REQUESTS FOR PUBLIC COMMENT

[Billing Code 6750-01-P]

FEDERAL TRADE COMMISSION

RIN 3084- [AA94]

Public Comment on Methodology and Research Design for Conducting a Study of

the Effects of Credit Scores and Credit-Based Insurance Scores on Availability and

Affordability of Financial Products

AGENCY: Federal Trade Commission

ACTION: Notice and request for public comment.

SUMMARY: The Fair and Accurate Credit Transactions Act of 2003 (“FACT Act” or

“Act”) requires the Federal Trade Commission (“FTC” or “Commission”) and the

Federal Reserve Board (“Board”) to conduct a study on the effects of credit scores and

credit-based insurance scores on the availability and affordability of financial products.

These products include credit cards, mortgages, auto loans, and property and casualty

insurance. The Act requires the FTC to seek public input about “the prescribed

methodology and research design of the study.” As part of its efforts to fulfill its

obligations under the Act, the FTC seeks public comment on how the FTC and the Board

should conduct the study.

DATES: Comments must be received by August 16, 2004.

ADDRESSES: Public comments are invited, and may be filed with the Commission in

either paper or electronic form. Comments should refer to “FACT Act Scores Study,

Matter No. P044804,” to facilitate their organization. A comment filed in paper form

should include this reference both in the text and on the envelope, and should be mailed

or delivered to: Federal Trade Commission/Office of the Secretary, Room H-159 (Annex

N), 600 Pennsylvania Avenue, N.W., Washington, D.C. 20580. The FTC urges that any

comment filed in paper form be sent by courier or overnight service, if possible, because

U.S. postal mail in the Washington area and at the Commission is subject to delay due to

heightened security precautions.

Comments that do not contain any nonpublic information may be filed in

electronic form (in ASCII format, WordPerfect, or Microsoft Word) as a part of or as an

attachment to email messages directed to: [email protected]. If a comment

contains nonpublic information, it must be filed in paper (rather than electronic) form,

and the first page of the document must be clearly labeled “Confidential.”

143

The FTC Act and other laws the Commission administers permit the collection of

public comments to consider and use in this proceeding as appropriate. All timely and

responsive public comments, whether filed in paper or electronic form, will be considered

by the Commission, and will be available to the public on the FTC Web site, to the extent

practicable, at www.ftc.gov. As a matter of discretion, the FTC makes every effort to

remove home contact information for individuals from the public comments it receives

before placing those comments on the FTC Web site. More information, including

routine uses permitted by the Privacy Act, may be found in the FTC’s privacy policy, at

http://www.ftc.gov/ftc/privacy.htm.

FOR FURTHER INFORMATION CONTACT: Jesse Leary, Deputy Assistant

Director, (202) 326-3480, Division of Consumer Protection, Bureau of Economics,

Federal Trade Commission, 600 Pennsylvania Avenue, N.W., Washington, DC 20580.

143

Commission Rule 4.2(d), 16 CFR 4.2(d). The comment must also be accompanied by an explicit

request for confidential treatment, including the factual and legal basis for the request, and must identify

the specific portions of the comment to be withheld from the public record. The request will be granted or

denied by the Commission’s General Counsel, consistent with applicable law and the public interest. See

Commission Rule 4.9(c), 16 CFR 4.9(c).

SUPPLEMENTARY INFORMATION:

I. Background

The FACT Act was signed into law on December 4, 2003. Fair and Accurate

Credit Transactions Act of 2003, Pub. L. No. 108-159 (2003). In general, the Act amends

the Fair Credit Reporting Act (“FCRA”) to enhance the accuracy of consumer reports and

to allow consumers to exercise greater control regarding the type and amount of

marketing solicitations they receive. To promote increasingly efficient national credit

markets, the FACT Act also establishes uniform national standards in key areas of

regulation regarding consumer report information. The Act contains a number of

provisions intended to combat consumer fraud and related crimes, including identity

theft, and to assist its victims. Finally, the Act requires a number of studies be conducted

on credit reporting and related issues.

Section 215 of the FACT Act requires the FTC and the Board, in consultation

with the Office of Fair Housing and Equal Opportunity of the Department of Housing and

Urban Development, to conduct a study on the effects of credit scores and credit-based

insurance scores on the availability and affordability of financial products. These

products include mortgages, auto loans, credit cards, and property and casualty insurance.

Section 215 further requires the FTC and the Board to study: 1) “the statistical

relationship, utilizing a multivariate analysis that controls for prohibited factors under the

Equal Credit Opportunity Act and other known risk factors, between credit scores and

credit-based insurance scores and the quantifiable risks and actual losses;” and 2) “the

extent to which, if any, the use of credit scoring models, credit scores, and credit-based

insurance scores impact on the availability and affordability of credit to the extent

information is currently available or is available through proxies, by geography, income,

ethnicity, race, color, religion, national origin, age, sex, marital status, and creed,

including the extent to which the consideration or lack of consideration of certain factors

by credit scoring systems could result in negative or differential treatment of the

protected classes, under the Equal Credit Opportunity Act, and the extent to which, if

any, the use of underwriting systems relying on these models could achieve comparable

results through the use of factors with less negative impact.”

The study is due December 4, 2005.

II. Request for Comments

The Act requires the FTC to seek public input about “the prescribed methodology

and research design of the study.” As part of its efforts to fulfill its obligations under the

Act, the FTC seeks public comment on how the FTC and the Board should conduct the

study. Public comment is requested on all aspects of the study. In addition, the FTC seeks

comment on the following questions:

1. How should the effects of credit scores and credit based insurance scores on the

price and availability of mortgages, auto loans, credit cards, other credit products, and

property and casualty insurance be studied? What is a reasonable methodology for

measuring the price and availability of mortgages, auto loans, credit cards, other credit.

Products, and property and casualty insurance, and the impact of credit scores and credit

based insurance scores on those prices and availability?

2. An effect can often only be measured relative to a counterfactual (that is,

relative to some hypothetical alternative situation). To determine the effects of credit

scores on the price and availability of credit products, what is a reasonable counterfactual

to the current use of credit scores? To determine the effects of credit-based insurance

scores on the price and availability of property and casualty insurance, what is a

reasonable counterfactual to the current use of credit-based insurance scores?

3. Paragraph (a)(2) of Section 215 requires a study of “the statistical relationship,

utilizing a multivariate analysis that controls for prohibited factors under the

(ECOA) and other known risk factors, between credit scores and credit-based insurance

scores and the quantifiable risks and actual losses experienced by businesses.” (The

ECOA “prohibited factors” are race, color, religion, national origin, sex or marital status,

and age.) What is an appropriate multivariate technique for studying this relationship?

What data would be required to undertake such an analysis? What data are available to

undertake such an analysis?

4. What is an appropriate methodology to determine whether the use of credit

scores or credit based insurance scores results in “negative or differential treatment” of

ECOA-protected classes?

5. What is an appropriate methodology to determine whether the use of specific

factors in credit scores or credit based insurance scores results in “negative or differential

treatment” of ECOA protected classes?

6. What is an appropriate methodology to determine whether there are factors that

are not considered by credit scores or credit based insurance scores that result in

“negative or differential treatment” of ECOA protected classes?

7. In order to address paragraphs (a)(2) and (a)(3) of Section 215, data are needed

on the geography, income, ethnicity, race, color, religion, national origin, age, sex,

marital status, or creed of borrowers, potential borrowers, insurance customers, or

potential insurance customers. Are these data available, and if so, where?

8. If the data discussed in question 7 are not available, what proxies are available

for the geography, income, ethnicity, race, color, religion, national origin, age, sex,

marital status, or creed of borrowers, potential borrowers, insurance customers, or

potential insurance customers?

9. If there are proxies for the geography, income, ethnicity, race, color, religion,

national origin, age, sex, marital status, or creed of borrowers, potential borrowers,

insurance customers, or potential insurance customers, what type of analysis would allow

inferences to be drawn using the proxies instead of actual data on individual

characteristics? What limitations are there to the inferences that can be drawn using

proxies in place of data on individual characteristics?

10. One potential proxy for individual characteristics may be Census data about

the location where a borrower or insurance customer resides. What type of analysis

would allow inferences to be drawn using data about the characteristics of the location

where a borrower or insurance customer resides instead of data on individual

characteristics? What limitations are there to the inferences that can be drawn using data

about the characteristics of the location where a borrower or insurance customer resides

in place of data on individual characteristics?

Authority: Sec. 112(b), Pub. L. 108-159, 117 Stat. 1956 (15 U.S.C. 1681c-1).

By direction of the Commission.

Donald S. Clark

Secretary

[Billing Code 6750-01-P]

FEDERAL TRADE COMMISSION

RIN [3084-AA94]

Public Comment on Data, Studies, or Other Evidence Related to the Effects of

Credit Scores and Credit-Based Insurance Scores on the Availability and

Affordability of Financial Products

AGENCY: Federal Trade Commission

ACTION: Notice and request for public comment.

SUMMARY: The Fair and Accurate Credit Transactions Act of 2003 (“FACT Act” or

“Act”) requires the Federal Trade Commission (“FTC” or “Commission”) and the

Federal Reserve Board (“Board”) to conduct a study on the effects of credit scores and

credit-based insurance scores on the availability and affordability of financial products.

These products include credit cards, mortgages, auto loans, and property and casualty

insurance. As part of its efforts to fulfill its obligations under the Act, the FTC seeks

public comment on any evidence the FTC and the Board should consider in conducting

the study.

DATES: Comments must be received by April 25, 2005.

ADDRESSES: Public comments are invited, and may be filed with the Commission in

either paper or electronic form. Comments filed in paper form should refer to “FACT

Act Scores Study” both in the text and on the envelope, to facilitate their organization,

and should be mailed or delivered to: Federal Trade Commission/Office of the Secretary,

Room H-159 (Annex Z), 600 Pennsylvania Avenue, N.W., Washington, D.C. 20580.

The FTC requests that any comment filed in paper form be sent by courier or overnight

service, if possible, because U.S. postal mail in the Washington area and at the

Commission is subject to delay due to heightened security precautions. Comments may

be filed in electronic form by clicking on the following:

https://secure.commentworks.com/FTCCreditScoreStudy/ and following the instructions

on the web-based form. If a comment contains confidential information, it must be filed

in paper (rather than electronic) form, and the first page of the document must be clearly

labeled “Confidential.”

144

To ensure that the Commission considers an electronic comment, you must file it

on the web-based form at https://secure.commentworks.com/FTCCreditScoreStudy/.

You also may visit http://www.regulations.gov to read this Notice, and may file an

electronic comment through that website. The Commission will consider all comments

that regulations.gov forwards to it.

The FTC Act and other laws the Commission administers permit the collection of

public comments to consider and use in this proceeding as appropriate. All timely and

responsive public comments, whether filed in paper or electronic form, will be considered

by the Commission, and will be available to the public on the FTC Web site, to the extent

practicable, at www.ftc.gov. As a matter of discretion, the FTC makes every effort to

remove home contact information for individuals from the public comments it receives

before placing those comments on the FTC Web site. More Information, including

routine uses permitted by the Privacy Act, may be found in the FTC’s privacy policy, at

http://www.ftc.gov/ftc/privacy.htm.

FOR FURTHER INFORMATION CONTACT:

144

Commission Rule 4.2(d), 16 CFR 4.2(d). The comment must also be accompanied by an explicit

request for confidential treatment, including the factual and legal basis for the request, and must identify

the specific portions of the comment to be withheld from the public record. The request will be granted or

denied by the Commission’s General Counsel, consistent with applicable law and the public interest. See

Commission Rule 4.9(c), 16 CFR 4.9(c).

Jesse Leary, Deputy Assistant Director, (202) 326-3480, Division of Consumer

Protection, Bureau of Economics, Federal Trade Commission, 600 Pennsylvania Avenue,

N.W., Washington, DC 20580.

SUPPLEMENTARY INFORMATION:

I. Background

The FACT Act was signed into law on December 4, 2003. Fair and Accurate

Credit Transactions Act of 2003, Pub. L. No. 108-159 (2003). In general, the Act amends

the Fair Credit Reporting Act (“FCRA”) to enhance the accuracy of consumer reports and

to allow consumers to exercise greater control regarding the type and amount of

marketing solicitations they receive. The Act contains a number of provisions intended to

combat consumer fraud and related crimes, including identity theft, and to assist its

victims. Finally, the Act requires that a number of studies be conducted on credit

reporting and related issues.

Section 215 of the FACT Act requires the FTC and the Board, in consultation

with the Office of Fair Housing and Equal Opportunity of the Department of Housing and

Urban Development, to conduct a study on the effects of credit scores and credit based

insurance scores on the availability and affordability of financial products. These

products include mortgages, auto loans, credit cards, and property and casualty insurance.

Section 215 further requires the FTC and the Board to study: 1) “the statistical

relationship, utilizing a multivariate analysis that controls for prohibited factors under the

Equal Credit Opportunity Act and other known risk factors, between credit scores and

credit-based insurance scores and the quantifiable risks and actual losses;” and 2) “the

extent to which, if any, the use of credit scoring models, credit scores, and credit-based