[ExI] dna to search

[Anders Sandberg]

William Flynn Wallace <foozler83 at gmail.com> , 9/11/2014 12:53 AM:
She was unfazed by the genetics issues: genes are weaker indicators of future health than actual behaviour,  (from Anders)

Well, how about Huntingdon's chorea and breast cancer? (for two of many) Clear genetic causes, and no hope for treatment for chorea.  Who would write any kind of insurance for a person who is certain to develop H's chorea?) 
Actually, if you sign insurance you have to state if there is Huntingdon in your family already. However, it is a bit of an outlier: dominant, incurable and expensive. 
Genetic causes do not mean it is automatically worth changing the premiums, apparently. I have seen data on Alzheimers that shows that known genetic predictors simply are not reliable enough to matter insurance-wise. Breast cancer is multifactorial, and the BRCA version is the only one where early detection merits a premium change (or early intervention). 
The things that likely would matter for insurance are probability changes in high-probability illnesses (cardiovascular, cancers) or the presence of rare, but expensive conditions that become significantly more likely. Looking at my own 23andMe data, it suggests that maybe my premiums ought to go up a bit because of a higher risk of cardiovascular conditions (perhaps balanced by a lower risk of type II diabetes) - except that this is eminently moveable using lifestyle interventions, so knowing about my exercise and diet would tell my insurer more than knowing my genes. And when we get down to the plethora of cancer-gene interactions, most of these cancers are rare enough that even a fairly big change in probability doesn't affect the expected health costs above the noise level.

====Simple model: Imagine that a condition X will have a cost C it it occurs, and has a base probability P0. The actual probability P=P0(1+aL+bG), where L is lifestyle and G is genetic factors (0 means no effect) and a,b small constants. The expected cost of X is C P0 (1+aL+bG) if we assume independence of L and G. However, the total expected cost is the sum across all conditions: E[C] = sum_i C_i P0_i (1+a_i L_i + b_i G_i). Here we are again assuming independence, which is problematic: if you die of X, you cannot die of Y, but I have not had breakfast yet, so I will handwave this. The P0s are skew distributed: there are loads of rare illnesses, and a few common ones. I would guess that they roughly follow a power-law: let's set P0_i = i^-alpha, where alpha>1 is a parameter denoting how common rare illnesses are. I think, based on the fact that hospitals are not treating just a single dominant disorder, that alpha is likely somewhere around 2.5
So, assume you figure out that you have increased risk of condition i. Then your expected costs go up by C_i P0_i b_i. If i is randomly distributed as i^-2, then the expected i is around 3, and P0=3^-2.5. So the change in expectation is  0.064*C_i b_i. This tells us that if the general noise level Std[C] is much larger than this, it is likely not worth checking. Now, the Std[C] for this example depends on the distributions of all the different factors which I definitely do not have the mettle to guess, but I would guess it is pretty big since P0 has infinite variance (ah, those delightful power-laws!) Even if all P0s were equal, if we assume b's tend to be relatively small, the sum is dominated by the C_iP_0 terms and the variance becomes due to the variance in treatment costs - which I think I remember is another heavy-tailed distribution. So unless C_i or b_i is *unusually* high - like in Huntingdon - or you have an effect on a high P0_i condition - then the insurer will not care much.
And if it can be offset by a monitorable change in L_i, so much better. In a sense lifestyle changes are like (usually) low-cost treatments: you can move that term into the C term.
Mom always said I should become an actuary. 

Anders Sandberg, Future of Humanity Institute Philosophy Faculty of Oxford University
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20141109/e40a8a47/attachment.html>