I recently applied for some life insurance and was a bit surprised to see that my application was scored.
What I mean is, they gave me a physical, took blood and urine specimens, and then for each measurement (including various measurements of my blood and urine) gave a numerical score. Negative apparently means healthy and positive means unhealthy. For example, my blood pressure earned a score of zero, which apparently means that it’s not particularly healthy or unhealthy.
All of the numbers were added up, along with a fudge factor which supposedly corrected for my age and sex, resulting in a final total. Presumably if that number meets their threshold, you qualify for the insurance.
Is this a new thing? I think so. As far as I know, in the past, life insurance companies had “underwriting guidelines” For example to qualify for the preferred rates, your blood pressure had to meet a certain cutoff. It seems that with scoring, things are potentially more flexible.
Is this a better approach to life insurance underwriting? In theory it seems to make sense but something doesn’t seem quite right about it. If you have a lousy heart, can you really balance it out with a great liver? Also, something doesn’t seem right about just adding up all the numbers. Doesn’t it ignore the possibility that some of the factors interact in subtle ways?
Any actuaries here want to shed some light on this?
In theory - you are correct - simply adding things up isn’t as good as taking into account interactions. Sometimes the interactions don’t add as much as you think. There is also some advantage to having easy an easy to understand algorithm. Linear regression is basically taking each variable - multiplying it by a number and add them all up. In many cases it works well. Other methods that can work better (such as neural networks, random forests, GBM) do so at the cost of interpretation - also you can’t tell someone if their blood pressure went down 10 points they would save X dollars.
I’m guessing their method is probably a little bit of educated guessing as well.
Also keep in mind some of this has to do with pleasing upper management. I heard a guy who runs this stuff for Allstate (granted car not health) explain that they had trouble getting buyin from the powers that be on methods that used a sampling of data for training purposes.
Some methods get VERY computationally expensive the larger the data sets are that are used for “training” the data. You can get around this by using a random sample. As long as the sample is big enough (and truly random) - this is fine. But in this case - they couldn’t get the powers that be to accept it - as they wanted something that used “ALL” their customers.
Lots of times - simpler is better for many reasons - just trying to give an example of one of them.