Following is PART II of the testimony presented by Lawrence S. Powell, PhD before the United States House of Representatives Financial Services Committee Oversight & Investigations Subcommittee on May 21, 2008, by:
Lawrence S. Powell, Ph.D.
Research Fellow - The Independent Institute (http://www.independent.org/), and
Whitbeck-Beyer Chair of Insurance & Financial Services
University of Arkansas-Little Rock
***************************************
Section 2: Predictive Accuracy of Insurance Scores
The correlation between driving outcomes and credit information appears in academic literature as early as 1949 (Tillman and Hobbs, 1949). Over time, evidence of the empirical relation between automobile insurance losses and insurance scores has developed to address not only the simple correlation between insurance costs and insurance scores, but also the additional predictive power and accuracy insurance scores contribute to insurance pricing models containing traditional pricing variables.
In this section, I review methods and results from several studies investigating the relation between insurance scores and insurance losses. The findings consistently and conclusively demonstrate that insurance scores are highly correlated with losses. The studies also show that insurance scores supply information about insurance losses not contained in other underwriting and rating variables.
More than a dozen studies related to insurance scoring have appeared in the public domain in the last decade. To improve the exposition of information, I present evidence from various studies in order of increasing complexity. This does not exactly match the temporal order in which studies were released. Furthermore, many of studies produce very similar evidence and reach nearly identical conclusions. I make an effort to report from the most recent and clear studies.
The most basic result is the simple correlation between insurance scores and losses. A study conducted by the Texas Department of Insurance in 2004 (TDI, 2004), firmly establishes the simple correlation between insurance scores and losses. Using data representing approximately 2 million insurance policies, the authors group exposure units by deciles of credit scores and graph the coinciding average loss frequency and loss amount.
Figures 1 and 2 appear in TDI (2004) as Charts 7 and 9, respectively. Figure 1 shows that average loss per vehicle declines steadily across deciles of credit score. Those with the lowest scores average approximately $360 per vehicle, while those with the highest scores average approximately $175 per vehicle. Similarly, Figure 2 shows number of claims per 1,000 exposures decreasing from approximately 110 for those with the lowest scores to just over 60 for those with the highest scores. These results are qualitatively similar across all of the companies reporting automobile insurance data for the study.
Several other studies reach similar conclusions using data from nationally representative samples (EPIC, 2003 and FTC, 2007), rather than the single state sample used by TDI.
Critics of TDI (2004), including the Texas Department of Insurance itself, point out that simple correlation between a rating variable and losses is neither necessary nor sufficient to establish its validity as a predictor of losses. This is true because no variable can produce a more accurate prediction of losses alone than when combined with other accurate predictors of losses. Therefore, in addition to simple linear correlation between predictors and losses, one must also consider the interactions among a group of predictor variables. To do so requires multivariate analysis.
Multivariate analysis, as the name implies, involves analysis of two or more predictor variables at the same time. EPIC (2003), FCT (2007) and a second study by the Texas Department of Insurance (TDI, 2005) employ multivariate analysis to determine if insurance scores are risk related. I summarize the analysis and primary findings of these studies below.
TDI (2005) examines a large database of personal automobile and homeowners insurance policies in Texas. The authors performed multivariate analysis considering the interaction of insurance scores and several other common predictors of insurance losses. They find that the strong correlation between insurance scores and losses persists even when controlling for other underwriting factors. TDI (2005) concluded that, "credit scoring provides insurers with additional predictive information, distinct from other rating variables, which an insurer can use to better classify and rate risks based on differences in claim experience." The authors also find that "use [of insurance scoring] is justified actuarially and it adds value to the insurance transaction."
EPIC (2003) examines a nationally representative sample of insurance scores, underwriting data, and policy outcomes (losses). The study produces four primary findings: First, insurance scores are correlated with risk of loss, even after controlling for relationships with other variables. The correlation is due primarily to loss frequency rather than loss severity. Second, insurance scores are correlated with some other common risk factors; however, even after controlling for other factors, insurance scores significantly increase the accuracy of the risk assessment process. Third, insurance scores are very powerful predictors of loss relative to other common risk factors. Finally, results from the study apply generally to all states and regions.
FTC (2007) also examines a large, nationally representative database to determine the relation between insurance scores and losses. The study finds that "even when noncredit 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." More specifically, they find 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.
These recent studies envelop a spectrum of backgrounds and data sources. Private groups and government agencies conduct them. They represent a single state and national samples. They employ different measures and methodologies. Nonetheless, they all reach the same general conclusion: that insurance scores are highly predictive of losses, even when controlling for other factors. As noted at the outset, insurers are unique in the U.S. economy as they do not know the ultimate cost of their product when they sell it so having a tool to more effectively predict losses helps insurers more fairly, for all consumers' benefit, price their products.
(The views expressed in this article/commentary are solely those of the author and do not necessarily represent the views of MyNewMarkets.com, the Insurance Journal or Wells Publishing.)
© 2010 by Wells Publishing, Inc. All Rights Reserved. Privacy Policy | Terms and Conditions | Advertise with Us | Submit Articles | Contact
Dr. Powell says that simultaneously looking at multiple variables lends to a more accurate predictive model than looking at a single variable in isolation. I don't disagree with that assessment assuming that there is causality rather than just casuality in the variables involved.
Consider the following, for which I couldn't find the name of the first person to raise this example to give them credit. Assume that a person is tracking drowning deaths. The researcher compares weather temperatures with increases in drowning deaths. This researcher being the "good" statistician that he is decides to add statewide popsicle consumption to his analysis. Now he has two variables to look at. So when he sees popsicle consumption go up he predicts a dramatic increase in drowning deaths. When the increase in drowning deaths doesn't materialize the statistician is left wondering. He was so sure of his analysis.
What he failed to realize is that there was a major manufacturer of popsicles that had a huge give away to increase popsicle consumption in the state and that caused popsicle consumption to go up.
The mistake he made was relating popsicle consumption to drowning deaths when causality was not present. The relationship was only a casual one.
Those promoting credit scoring are making this same fundamental error. There must be a causal link between poor credit score and higher frequency and severity of losses. If a person has declared bankruptcy due to huge medical bills because of cancer of a family member, that does NOT increase the likelihood that their house will burn down!!!
KEY POINT: There must be causality between a variable and a measured event to be predictive.
Thomas E. Nelson, CIC, CRM, ALCM
some do not; it is not a question of
"right" or "wrong" in their decision. It's use is a business decision to use specific rate determination tool in their business model. Some may agree or disagree that the use of credit scoring accurately determines loss frequency or severity; but it is undeniable that that credit scoring is a good indicator. The unfortunate reality is that those who are affected the most adversely by its use, are those who are least able to pay the increased cost of coverage.
Scoring is only a "head game", UNLESS a Large Number of actual claims are studied and a "direct" relationship is discovered between "status and severity".
Insurance has functioned well using this method for over 100 years sooooo we must assume it needs changing, Oh I know ! lets create a new "formula"!
I rank "scoring" right up there with "Cash flow underwriting"..............TROUBLE!
Thanks
MAA
The cancer scenario does, indeed, increase the moral hazard, as would know any beginning underwriter. People who are careless about their credit, whether it's forgetting to mail checks or simply putting off bill paying would be careless in other matters.
As anyone who has served in the USMC knows, inattention to small details (making sure that your uniform is clean, your locker organized, and etc.) will be reflected in all matters, large and small. A careless person is careless in all matters.
The popsicle analogy is specious, because the use of other variables in Mr. Lowell's study were varibles that reasonably are related to causes of insured losses. No one, other than Mr. Nelson, has ever contended that increased popsicle consumption leads to more drowings.
Consider the example of a person who goes to the doctor for a particular condition. The doctor takes the list of symptoms down, does some blood tests and other inexpensive medical testing to determine the cause of the condition and then runs a credit score on the person. The medical tests turn out not to be determinative, as is often the case, so the doctor looks at the credit report supplied by a medical credit scoring company who provides a "medical score" The doctor uses the "medical score" to determine whether more testing should be done. The medical credit scoring company has captured data that correlates liver failure with poorer credit scores. Their data indicates that liver failure, which ultimately results in death, is unlikely in this person's case because his credit score is over 580. Hence the doctor decides that the abnormalities seen in the blood tests are likely transients from an infection the patient had two months earlier rather than taking the time to ask more questions and run more tests to see whether the blood abnormalities are really precursors of impending liver failure. Notice how even an inverse relationship based upon a high credit score is a potential problem.
While to my knowledge no "medical credit scoring" is presently done, if you knew a doctor treating you was making medical decisions on that basis, would you want to continue with that doctor? Personally I would be very uncomfortable with someone using credit scoring to make real life decisions affecting me based upon what is a "casual" not at "causal" relationship between credit scoring and life altering decisions. It works the same with "insurance scoring".
All it would take is for credit scoring companies to look at the correlations between credit scores and early deaths and supply such data to health insurers or life insurers to have your access to health insurance or life insurance denied or up-priced, not based upon your actual medical conditions but based rather upon correlations that exist between the way people pay their bills and the likelihood of early death or the likelihood of future medical costs. How would you feel if you had your health insurance declined or the premium significantly increased because you were late on a few credit card payments or some telephone company "slammed" you and you had a credit dispute over it?
Perhaps some actual examples would help of colossal foul-ups that occurred because "smart" people assumed simple correlation (i.e. where there was no actual causation proved) was a good enough basis for real life decisions.
Do you remember the tomato e coli scare of June 2008? Growers, packers, shippers, wholesalers, grocery retailers and restaurants lost hundreds of millions of dollars on that one. This occurred because researchers were looking at correlations to find what caused so many people to come down with e coli poisoning. What did they have in common? They ate fresh salsa. They concluded, without any sample of a tomato with e coli contamination, that it must be the tomatoes because salsa has fresh tomatoes. All the tomatoes were pulled. People still got e coli poisoning. They next concluded it must be the cilantro because fresh salsa also has fresh cilantro. It wasn't the fresh cilantro either. It finally turned out to be the fresh peppers.
Maybe you remember that coffee was bad for you. Then it was good for you. Then it was bad for you. Then it was good for some and bad for others. Then it was good for some conditions and bad for other conditions.
Then there was glucosamine and chondroitin. First it was good for you and then it didn't do anything.
Then there was St. John's wort. First it was good for you then it wasn't.
The list goes on and on. Among the things that affected the studies people relied upon for their "expert opinions" promoting this or that causative relationship, were flawed findings of direct causation (we studied this and this percentage of the time we found this condition followed that event or condition), researcher bias (research was paid for by someone benefiting somehow from the research), inadequate number of samples tested (we looked at 10,000 people in the Detroit area who took this drug, Notice the geographical constraint on data.) or no other factor could have caused this result. (For that read we couldn't think of anything else to test that would fall within our budget.)
Those that are promoters of credit scoring or "insurance scoring" don't even know why a correlation exists between lower credit scores and higher frequency of claims. (The purveyors of credit scoring information do get paid handsomely for the data they supply insurers. That can be a source of bias.) Denying access to insurance or raising the cost of auto or homeowners insurance based upon such correlation is the logical equivalent of the example of the doctor making medical decisions based upon medical credit scoring or raising life and health insurance premiums or denying access to life or health insurance to persons who have lower credit scores.
Think how easy it would be for credit scoring companies to correlate credit scores to public information on ages and causes of death. They could then offer such to health and life insurers who could use the information in the same way property and casualty insurers do. Don't think they wouldn't use that information that way. If they did use credit information that way you can be sure it would result in people dying because of lack of access to health insurance or in families left destitute because of lack of access to life insurance.