A post on Sam Friedman’s blog at National Underwriter has me thinking. (Danger, Will Robinson! Danger!). I posted a response at NU, but I’m going to extend my thoughts here.
Sam writes:
One of the most important and interesting developments at the recent NAIC meeting came in under the radar and took most insurers by surprise–a controversial call to force homeowners carriers to collect and disclose data on the race, gender and income bracket of their prospects and clients. In responding, insurers are damned if they do and damned if they don’t support the proposal. [...]
Obviously, Mr. Squires is eager for insurers to collect and provide access to such data to see once and for all if insurers are actively engaging in redlining, or are passively doing so–via disparate impact (in other words, even if their intent is non-discriminatory, the end result comes out that way, unfairly disadvantaging one racial, gender or economic group).
Just as obviously, insurers don’t want to collect such data because, for one, it might very well expose them to lawsuits over their underwriting patterns–however neutrally they are applied–and it might annoy or provoke those asked to list their race, income, etc., on a homeowners insurance application.
Mr. Squires argues effectively that he is not reinventing the wheel here. Essentially, he says he is simply calling for the same data required of home mortgage lenders under the federal Home Mortgage Disclosure Act. He told the NAIC that HMDA and other fair-lending laws helped improve access to credit for low-income and minority markets, and suggested the same might be said of insurance down the road.
The short answer I posted in the responses section still applies — if the goal is to collect the same demographic information that is already captured by mortgage brokers, then analysts ought to simply collect both insurance and mortgage demographic data, and merge the two data sets, avoiding pestering consumers unnecessarily.
However, with some additional thought (dusting off older thoughts, actually), I can offer a few points that ought to be addressed in any proposal to analyze industry data for evidence of “redlining” or other forms of unfair discrimination:
- The most important concern that must be addressed in any study proposal is exactly what it is that is going to be measured. Terms like “unfair discrimination” or “disparate impact” are philosophical or legal constructs. Sadly, philosophers and lawyers rarely speak the same language as actuaries and statisticians.
Until consensus can be reached between the industry and consumer advocates as to how such accusations can be measured, there will be dueling concerns that will complicate, if not derail, such efforts. The industry will not tolerate the assembly of a big mega-database which can be used in an uncontrolled, poorly-defined fishing expedition for data to embarrass or harass contributing carriers. Consumer advocates will not be satisfied in a study that could be viewed as too-limited, and too-controlled by the industry.
Thus, if analysis is going to be done, the very first step MUST be reaching consensus in defining metrics for our favorite legal terminology.
- Personally, I’m somewhat pessimistic about such consensus being possible. In an environment of political correctness, there seems to be extreme sensitivity and diminished awareness in the neutrality of certain terms.
“Discrimination” is not inherently bad; in fact it is risk classification that permits insurers to charge customers rates which approximate their exposure to risk…something that seems fair to many folks who think about the subject. After all, why should a childless, middle-aged couple have to pay more for their car insurance because your sixteen year-old boy is getting behind the wheel? And I’m sure that folks in North Dakota would just love the idea of chipping in a little extra into their homeowners insurance bills to subsidize the risk of a property-owner in Key West.
When testing for unfair discrimination, it is likely that correlations will be seen between certain rating variables or underwriting factors and certain demographic attributes. Some of that was seen when the FTC looked at allegations that credit scoring is a back door to racial discrimination for personal auto insurance. They did observe a correlation between score and race or income (a point that Birny and Bob like to mention), but they noted that credit retained its predictive power even within those groups of concern.
My own view is that to test for “disparate impact” of a particular variable, one should compile a representative sample of industry experience and demographic data. An actuary can then design a theoretically optimal rating algorithm with and without the variable of concern. If groups of interest do not see, in aggregate, significantly higher premiums “with” the variable than “without”, there is no disparate impact.
However, as long as advocates, politicians, and the media are prone to make too big a deal over correlation, the industry’s going to be reluctant to participate. And, if simple correlation is to be avoided…hell, we might as well have a Florida-like state-run insurance operation for all lines of personal P&C insurance, since a pure, socialized program would be the logical result of such a prohibition.
- The next major consideration is the collection of data. As an actuary who spends a fair amount of time on fishing expeditions in piles of data, there are certain data elements I resist collecting. I don’t need to know folks’ names, social security numbers, birthdates, etc. in the actual guts of my analysis. Sure, having that information comes in handy when trying to tie back to reality, or when trying to merge with outside information….but once all the data is collected and I’m ready to actually get to work on analysis, I don’t want to know personal details. It’s usually of little predictive value, and there are way too many downsides to risking even the perception of violating privacy, or otherwise misbehaving in inappropriate ways.
Similarly, insurers in general are leery of risking even the perception that they misuse certain data. The industry doesn’t want to know about race or income because if that information is known, we will be accused of abusing that knowledge even if we don’t.
Any study must keep insurers blind to such information. I don’t want to know about my customer’s skin color, or their AGI’s. I don’t want to ask, even if that data is fed to a consultant’s database, and never remains in-house.
The analysts who actually do the work in such a study need to acquire that information from other sources. Ideally, it should be data that has already been collected, and which can be merged with industry data for third-party analysis.
Here’s an idea for you — if unfair discrimination is such a concern, perhaps the necessary legislative or regulatory magic could be done to permit a third party to merge census data with insurer data.
- When you come down to it, I suspect that the biggest source of resistance from the industry is fear of what might happen if “something” were discovered.
Personally, I would welcome a study testing whether parts of personal lines rating and underwriting fail the “disparate impact” standard (assuming a reasonable interpretation of the standard). I doubt that any problems would be found…and on the off chance there are, it would be good to solve them.
I am not alone in having such thoughts. However, I am in the minority. Decision makers in much of the industry are less certain of the outcome (realistically, how could we know, since we don’t collect the data to test, to begin with), and we all know that if “something” were found, there would be a wave of civil rights lawsuits that could decimate the industry.
To do a study like this right…to get enough insurers willing to participate…safe harbor needs to be granted. Don’t excuse willful acts of unfair discrimination. But if industry conventions mean that “disparate impact” has been accidentally breached, let’s identify the problem and fix it going forward.
The relationship between the industry and its critics being what it is, I’m not particularly optimistic about any of those points being resolved. Thus any move towards studying whether insurers are inadvertently engaging in unfair discrimination are likely to be slow, contentious, and limited in their ability to satisfactorily answer questions for the major participants in the debate.
But if folks really wanted to answer the question, rather than just beat up on one another…consider the points above, and a productive answer should emerge.