Do Credit‐Based Insurance Scores Proxy for Income in Predicting Auto Claim Risk?

• Author: Daniel Schwarcz,Joshua C. Teitelbaum,Darcy Steeg Morris
• DOI: http://doi.org/10.1111/jels.12151
• Published date: 01 June 2017
• Date: 01 June 2017

Do Credit-Based Insurance Scores Proxy

for Income in Predicting Auto Claim Risk?

Darcy Steeg Morris, Daniel Schwarcz, and Joshua C. Teitelbaum*

Property and casualty insurers often use credit-based insurance scores in their underwriting

and rating processes. The practice is controversial---many consumer groups oppose it, and

most states regulate it, in part out of concern that insurance scores proxy for policyholder

income in predicting claim risk. We offer new evidence on this issue in the context of auto

insurance. Prior studies on the subject suffer from the limitation that they rely solely on

aggregate measures of income, such as the median income in a policyholder’s census tract

or zip code. We analyze a panel of households that purchased auto and home policies from

a U.S. insurance company. Because we observe the households’ home policies as well as

their auto policies, we are able to employ two measures of income: the median income in a

household’s census tract, an aggregate measure, and the insured value of the household’s

dwelling, a policyholder-level measure. Using these measures, we find that insurance scores

do not proxy for income in a standard actuarial model of auto claim risk.

I. Introduction

Insurance companies are in the business of classifying policyholders into risk-based cate-

gories. Insurers build actuarial models to relate policyholder characteristics to claim

risk. They then use these models to predict the rates at which policyholders with differ-

ent characteristics will generate claims. These claim rate predictions, along with predic-

tions about claim amounts, play a central role in determining whether insurers offer

coverage to prospective policyholders and, if so, the premiums that they set for this

coverage.

*Address correspondence to Joshua C. Teitelbaum, Georgetown University Law Center, 600 New Jersey Ave. NW,

Washington, DC 20001; email: jct48@law.georgetown.edu. Morris is Research Mathematical Statistician, Center

for Statistical Research and Methodology, U.S. Census Bureau; Schwarcz is Julius E. Davis Professor of Law,

University of Minnesota Law School; Teitelbaum is Professor of Law, Georgetown University Law Center.

Teitelbaum acknowledges financial support from National Science Foundation grant SES-1031136. For helpful

comments and feedback, we thank the editor, three anonymous referees, and participants at the following con-

ferences and seminars: the American Risk and Insurance Association Annual Meeting, the American Law and

Economics Association Annual Meeting, the European Association of Law and Economics Annual Conference,

the Annual Meeting of the Section on Insurance Law of the Association of American Law Schools, the Law and

Economics Seminar at Boston University School of Law, the Faculty Workshop at George Washington University

Law School, the Law and Economics Colloquium at George Mason University School of Law, and the Law and

Economics Workshop at the University of Virginia School of Law. The views expressed are those of the authors

and not necessarily those of the U.S. Census Bureau.

397

Journal of Empirical Legal Studies

Volume 14, Issue 2, 397–423, June 2017

In the United States, federal and state laws limit the scope of insurers’ risk classifi-

cation schemes. These laws often restrict insurers’ capacity to discriminate among policy-

holders on the basis of characteristics, such as race and gender, that are viewed as

potentially suspect classifications in a wide variety of settings (Brilmayer et al. 1983).

They also may target discrimination on the basis of other policyholder characteristics,

such as income or wealth, that are potentially suspect in the specific context of insur-

ance, though not always in other settings. Finally, insurance anti-discrimination laws may

regulate insurers’ use of characteristics, such as occupation and zip code, that are not

independently suspect, but that may correlate with one or more suspect classifications.

This final form of regulation is often quite controversial, resulting in widely varying legal

regimes across states (Avraham et al. 2014a, 2014b).

Credit-based insurance scores (“insurance scores”) are perhaps the most impor-

tant example in the category of policyholder characteristics that are regulated because

they potentially correlate with suspect classifications. Many property and casualty insur-

ers use insurance scores in their actuarial models for their automobile and homeowners

coverage lines. The widespread use of insurance scores in these lines of coverage stems

from a simple fact: they are predictive of claim risk (Miller & Smith 2003; Golden et al.

2016). At the same time, however, insurance scores may be correlated with one or more

suspect classifications, including, most importantly, race and income. For this reason,

most states regulate insurers’ use of insurance scores in auto and home insurance, and

a few states ban their use altogether (Avraham et al. 2014a, 2014b).

Any correlation between insurance scores, on the one hand, and race or income,

on the other, is potentially troubling from a policy standpoint for two reasons that are

not always clearly distinguished. The first is that insurance scoring may have a disparate

impact on racial minorities and low-income households, causing members of these

groups to pay higher premiums on average. Whether this fact, by itself, warrants legal

intervention is highly controversial and context dependent (O’Neill 2007). Generally,

however, critics have had little success arguing that insurers’ classification schemes

should be limited solely because they have a disparate impact on certain groups. This is

particularly true outside of the domain of race and certain types of property insurance,

where enhanced federal scrutiny applies due to the Fair Housing Act. But even under

the Fair Housing Act, insurance practices that have a disparate impact on protected

groups are generally permissible if no less discriminatory alternative is available.

1

The second, and more potent, reason for regulating the use of insurance scores is

that they may proxy for race or income. A correlation between insurance scores and a

suspect classifier such as race or income is a necessary, but not a sufficient, condition

for insurance scores to operate as a proxy for the suspect classifier. In addition to such

a correlation, this “proxy variable” argument presumes that the predictive power of

1

Implementation of the Fair Housing Act’s Discriminatory Effects Standard, 78 Fed. Reg. 11,460 (Feb. 15, 2013),

http://portal.hud.gov/hudportal/documents/huddoc?id5discriminatoryeffectrule.pdf. Recently, the Supreme

Court held that disparate impact claims are indeed cognizable under the Fair Housing Act, giving new vitality to

this issue. See Texas Dep’t of Hous. & Cmty. Affairs v. Inclusive Communities Project, Inc., 135 S. Ct. 2507

(2015).

398 Morris et al.