Sunday, 2 December 2012

Right and Wrong

That could be a tricky subject, but the post on insurance and sex discrimination, and particularly some of the replies, led me to ponder how one can rationally decide if this change is right or wrong. I.e. questioning myself (as one should always do).

In any case in making a decision one has to consider the underlying objectives, and how you justify picking those objectives even. Often you can reduce a problem to "obvious" right and wrong objectives.

The insurance one is a tad tricky. Naturally, as a scientist, I considered the factors like statistics, and how "wrong" it is inherently to remove source data (gender) from the equation. It is, of course, right to question if the statistics on gender do in fact create a reliable predictor of future risk for insurers. I am sure they ask themselves this on every factor, including gender, and I assume the answer is yes. But essentially what we are saying is that there is a correlation between gender and risk. As any statistician knows you have to be careful with correlations. If A correlates with B, it could mean A causes B, it could mean B causes A, it could mean C is involved somehow and that neither A causes B or B causes A. It can also be a fluke, but usually there would be some cause. A recent xkcd is a good example. It shows correlations by showing maps of where certain things are common in the country. They line up - they correlate. But do that simply because they are just population maps.

One nice counter example of the use of statistics and discrimination was gievn to me. It was basically "Statistically black people are involved in more crime - therefore employers should be able to use that statistic to decide not to employ black people".  A big problem with that is that it can cause a loop. If you don't employ one group of people it is likely they will be more involved in crime, bolstering your original statistic, regardless of the original cause. Before long you have a self justifying policy.

However, when it comes to insurance, a business that is wholly about statistics, it makes sense that any factor which statistically predicts future risk is worth using. It is also very hard to see how this could create a feedback loop. Indeed, it is easy to see how high prices on insurance lead to less driving or driving slower and cheaper to insure cars, so reducing risk in future. A negative feedback that damps the effect.

So, from a purely mathematical point of view, it is obvious that insurers should be able to use any and all factors they can in deciding risk. Any policy has to result in a stable solution which is morally acceptable given the naturally selfish nature of people and companies. It is, obviously, in their commercial interests to get it right, and so to find factors which fail to be good predictors and remove them. So allowing insurers to use any factors they can, including gender, leads to more accurate risk prediction.

Now, is that "right"? Well, to tell that one has to consider whether insurance with more or less accurate risk prediction is "right". Turns out that is a lot harder to come to any conclusion on. One common method is to consider the extreme cases, and see if they seem right or wrong - that can help align your moral compass on such things.

The problem is that it comes down to the age old difference between communism and capitalism!

Lets assume you want more accurate predictions. The premise being that it is "unfair" for careful driver to pay high premiums to help pay for people that have claims. The extreme of this is an insurance company with access to a time machine. That changes insurance. It becomes a pre-payment installment plan to pay for what will happen. "Certainly sir, you can have a 27 year term life insurance policy, but I'm afraid the quote for 28 year term is rather more expensive". Arguably paying for your own claims is "fair", for some definitions of "fair".

At the other extreme you have insurance with no prediction of risk, just a cost to cover. That means everyone paying the same amount for insurance, even people that do not drive (after all pedestrians and passengers benefit too). That is "fair" as everyone pays the same. In some ways this is what we aim for in the NHS.

I can't decide which is better. Obviously, as extremes, they are both a problem. They would both cause feedback loops affecting behaviour. But without knowing which direction is better you cannot say if a change from one status-quo to another is right or wrong.

The change to remove gender is a change in the direction of the "everyone pays the same" end of the scale. If that is, in fact, "right", surely we need more changes towards that end of the scale.

Fun, isn't it?


  1. I don't think everyone pays the same makes sense - it just means they can't adjust for the risk and most people pay more to make up for the riskiest drivers paying less.

    What we should be doing is making the risk analysis *more* granular not less - I'm in favour of the in-car monitoring.. it introduces the idea of having a personal risk profile - become a better driver and your premiums drop.

  2. I think the current/previous situation is laughable - and there for the simple reason below:

    Insurance quote for woman + spouse (male) on a car is/was less than man + spouse (female) on a car.

    Same couple - pays less if the insurance for one/both cars is in the female members name.

    I strongly suspect that accident risk is actually better measured as a risk per mile driven, and that in current societal norms men generally drive more than women.

    It is however _easier_ to quote on gender than predicting and insuring against annual mileage. - Perhaps the more accurate statistical indicator is more costly to administer?

    1. I think that Chris has picked out the answer to a lot of this in one sentance

      [quote]I strongly suspect that accident risk is actually better measured as a risk per mile driven, and that in current societal norms men generally drive more than women.[/quote]

      I'm certain that there are stats out there, but until women drive as far as men on average, the difference in accident rates when not adjusted per mile will be higher for men. My understanding was that this was part of the problem, the insurance companies are welcome to 'discriminate' based on sex, provided that they can prove that the link is more than casual. If they can show that an average woman driver, in exactly the same car, doing the same miles, at the same time, in the same circumstances as an average male driver would be safer, then they can use that to 'discriminate' on, and offer women a lower premium, if they can't, then they can't use it in their pricing model. My understanding is that even with all of their data, the insurance companies can't prove that. Does this mean that if women shop around and match their insurance closer to their circumstances, that womens insurance will always be higher? No. Because if they do drive less, then they can declare that they drive less miles, and then get cheaper insurance cover.

      The difficulty (as ever) is in making that link to accidents per mile driven.

      To an extent, the best way to solve this would be via an upgraded version of the current in-car tracking boxes, making them even more personalised. A fully pay-as-you-drive system if you like, were you pay for a number of hours or miles of insurance at a predetermined personal rate. You could even put in a taxi like monitor so you can see what you're spending. you could go as far as to dynamically vary the rate could be variable, depending on things like time driven so far, road type driven on, time of day etc. This could be handled by ensuring that each driver had a nfc enabled keyfob, so that the car could tell who was driving it. This would allow the analysis of the car data to be attached to an individual driver, rather than an individual car, to give a far more personalised insurance premium than before.

      This does of course bring civil liberties, and data protection issues into the mix, which is a whole different kettle of fish.

  3. The problem the law faces is distinguishing between correlation and causation - it is simpler to ban discrimination on certain touchy areas (gender, race etc) than it is to work out whether there is a causative effect from a particular bit of discrimination. You've had a race-based example of possible causation as a comment on a previous post; however, you've not considered the possibility that insurance for women is cheaper because there is widespread discrimination against women.

    For example, we noted that quotes for me were cheaper for a 170 PS Skoda Superb than they were for my wife; the reverse was true with a 60 PS Skoda Citigo. Now, there are lots of confounding factors in here - like the difference in no claims discount between us, driving history, power outputs of both cars etc, but it is at least possible that the difference on average between men and women comes down to women finding it disproportionately cheaper to drive smaller, lower power-to-weight ratio cars than their male counterparts, and therefore, when choosing a car, choosing one that is safer for an inexperienced driver (because the insurance is much cheaper that way). This acts as a feedback loop - men are more likely to choose high-risk cars, because the reward for being sensible is lower for men than for women; this makes it more likely that men will be in high-cost accidents, increasing insurance for men, making the difference between a safe but "dull" car and a "fun" but risky car less significant to men than it is to women (or, looked at differently, making the reward for choosing a safe but dull car higher for women).

    How do you prove that there isn't such a feedback loop going on? How do you prove that women's "cheaper" car insurance isn't simply a result of women getting cheap insurance on "dull" cars while being excluded (relatively speaking) from bigger, more powerful vehicles?

  4. I see you've carried on the train of thought from one post to another, perhaps my last comment should have been on this page? :)

    I guess TonyHoyle's comment has a point about in car GPS (although I'm not a fan of the implications). Bit I would modify his statement to say you can make risk analysis more granualar *as long as it is something you can have some impact on* i.e. some way of control.

    To go back to my last posts example of medical insurance. Reading your DNA as not insuring because of some heredity condition should be illegal to even measure. However reducing or increasing premiums because of your weight condition or the fact you partake in dangerous sports _should_ be taken into consideration.

    So to my mind, insurance (for the person taking it out) is to nullify or reduce the impact of a particular risk condition.

    If I am an overweight, smoking male and I wanted to nullify or reduce the impact of a heart attack on my life, I could either reduce my BMI and give up smoking or just pay more medical insurance to cover a heart attack condition to do this.

    However, if my DNA says I will have a heart attack at age 55 -no matter what- then this shouldn't be measured in the first place for insurance purposes (until we get to the point it can be fixed as easily as it can be measured, and even then the insurance shouldn't measure it until it has been covered!)

    Now I've just successfully argued against discrimination between male/female (you can't control this) which I was arguing for in your last post. What fun! :) I'm going back to work :p