Sheila's Fails? The statistics of biological risk
Why the ECJ insurance judgment might not be the right road after all
Comment Yesterday’s ruling by the European Courts may have stirred the general public to a wide-ranging and not altogether informed debate on the issues of gender discrimination. Less obvious, but in the long run more serious, is the fundamental challenge it poses to the way in which two pillars of the establishment – the financial services and the equalities industries – use and abuse statistical concepts and statistical thinking.
On the surface, this was a victory for equality, a defeat for the insurers: in the long run, this result may have equally uncomfortable implications for both parties.
The central issue is the use of probabilities and probability data. Insurance, for all its outward respectability, is a form of gambling. Its modern origins lie in the coffee houses of London and Amsterdam, as merchants figured out that it was likely to be more profitable in the long run if they clubbed together to bet on whether their ships would return home safely.
For a relatively modest outlay, they could lay off the risk of individual ships falling prey to extreme weather. Informal clubs became formal syndicates: those prepared to take on the risk became less directly associated with the risk itself – and the insurance industry was born.
Similar developments followed the Great Fire in 1666, with the modern practice of gambling against household risk swiftly taking on a veneer of respectability entirely lacking in its Roman ancestor ("Pay your fire risk premium, or I bet this lighted torch might just fall on that pile of bedding over there!").
And that, for all the window-dressing, is pretty much all insurance is: a bet against the possibility of something untoward happening. Like bookies, insurers can spread risk around (through the use of re-insurance), and they'll take bets on some things bookies won't. Life insurance, for instance, is no more than a sophisticated gamble based on the most likely date of an individual’s death.
Despite that, the real income in insurance is not in the premiums – which do little more than cover the underlying risk – but in the investment income that accrues from investing billions of pounds of premiums in the months before that money has to be paid out in claims.
Insurers make much of "targeted risk", but that is not an issue that they obviously need to bother with. As explained above, it is a pooling of risk for social purposes. This is why, across the world, most states have now some form of regulation of the process of “red-lining” – the exclusion of individuals from insurance on the basis of some high-risk group characteristic they possess.
If it were possible to predict, Minority Report style, exactly who would be a crime victim, who would suffer a particular accident in the next 12 months, the insurance industry would cease to exist. It only works because we don’t know exactly who will be victim of some unfortunate circumstance that it makes sense to calculate the average probability of such an outcome, as well as the average cost of each incident, based on past occurrence, and provision accordingly.
So why are there different premiums for men and women? Why Sheila’s Wheels? The answer lies in competition and compromise.
First, as insurers identify new data sources that in turn allow them to quantify risk more precisely, they are able to fit premiums much more closely to the characteristics of specific groups. Admiral owed much of its early growth in the motor insurance market to exceedingly clever under-writing that allowed it to underwrite risks in parts of London that other insurers found difficult or impossible.
Second, sharp premiums are attractive to the punter – so if you wish to attract premium income for investment purposes, you need products that are competitively priced.
All of which, of course, flies absolutely in the face of ideas of insurance as a social good: and none of which explains why across a series of press reports yesterday, industry insiders were suggesting that removal of gender from the risk equation would inevitably lead to an upward pressure on prices, with premiums stabilising at a level that is, in aggregate, above the current aggregate.
The Reg asked the Association of British Insurers why this might be the case. It does not comment on pricing matters - nor, though, would the UK’s largest insurers. This leaves that question as something of a puzzle: the aggregate risk remains the same. All that has happened is that one predictor has been removed from the pot: the total cost to customers ought not to change.
If insurance is regarded as social good, the ruling by the European Court makes more sense. Whether it is viewed as right depends on the extent to which individuals believe society should equalise risk sharing across categories. Young males are riskier on average behind a steering wheel: older women do have longer life expectancies.
Those are statistical facts, and the debate boils down to whether we wish, as a society, to use them – to use gender - as a factor in apportioning risk and social benefit. The European Court of Justice was clear that we shouldn’t – and in that sense was wholly consistent with the broad thrust of equalities legislation as it has evolved over the last few decades.
Just the first worm from a pretty big can
There are many ways in which men and women are observably statistically different – from work absenteeism rates, to job commitment – and it would make as much, or as little sense to pre-vet individuals for employment based on their gender as it does to use gender to measure insurance risk.
Where the court may have scored something of an own goal is in its attempt to relegate statistical evidence to a special, less valid category than other forms of evidence.
If we go back to the original view of Advocate General Juliane Kokott last September, she is of the opinion that "the exception in question [insurance] does not relate to any clear biological differences between insured persons. On the contrary, it concerns cases in which different insurance risks can at most be associated statistically with gender."
This has been distilled, since, into the view that "statistical" differences are not the same as "biological" ones. That is a peculiar view – and one that is also at odds with the way in which equality law works.
Direct discrimination is, quite simply, discrimination on the grounds of a particular "protected characteristic".
"No women, blacks or gays" would be direct discrimination. Indirect discrimination involves the application of a condition that, although applied equally, tends to hit one group disproportionately by comparison with another. "No one under six foot" is indirect discrimination because it tends to affect women more than men. It is a statistical fact that women tend to be shorter than men, so such a condition, unless required for clear operational reasons, would be unlawful.
So the law permits statistical facts. In fact, a reading of judgments in this area would suggest an active encouragement of statistics used in this way. As one English court declared not that long ago: where recruitment outcomes, in terms of relative frequency, can be shown to be statistically disproportionate, it is likely that a discriminatory policy is being applied, even if unintended. In such case, an employer would be guilty of discrimination and would need to change their recruitment practices – or face penalties.
What did the Advocate General mean? In her earlier opinion, she appears to draw a distinction between biological factors, such as the costs associated with pregnancy, and statistical factors that do not represent any clear biological differences. It’s a hard distinction to maintain – and certainly one that is not otherwise held to in law.
If this raises issues for the European Court, it also opens an entire can of worms for the insurance industry.
When it comes to forecasting future outcomes, the statistician’s task boils down to identifying the degree of variance at play in possible outcomes, and apportioning that variance to underlying factors. Random variability is excluded.
What’s left tends to be mostly due to three or four main factors, distributed in geometric fashion: analysis of most human behaviours often gives rise to a series of explanatory components, with around 50 per cent of the variance taken up by the first, 25 per cent by the second, and so on.
If, as seems likely, gender is a high-ranking component of human variability, the ability of the insurance industry to predict outcomes has just been significantly reduced.
What then of other factors? Make and model of car, for instance?
Here comes a problem. Insofar as make and model are factors independent of gender, they can still be used as risk predictors. But where they overlap – where the risk due to type of car driven links directly to gender – then this, too, has just been removed from the equation. Dual driver policies are cheaper as inherently less risky – but if they tend to be linked to one gender more than the other, they could also now be at risk.
Ultimately, that is the real issue for insurers. They can’t simply exclude inner city areas from insurance, because in today’s UK that would almost certainly result in a degree of indirect racial discrimination.
From December 2012, they cannot use gender explicitly as a factor in setting premiums: in the long run, though, as they pore over their detailed charts and risk calculations, the picture may be far worse: because many of the other factors they might instinctively put in place of gender could well correlate with gender. So they could soon be outlawed too.
Whatever happens, this is not the last we will hear of this case. Implicit in the interaction of statistical fact and statistical discrimination is a view of society that calls for a much greater evening out of difference than we are used to at present. If that is what we want, this ruling is a positive step forward: if not, it is a door opening into a world that some will find increasingly difficult to bear. ®