# The Behavioral Economics of 23AndMe Results

Over the holidays, I received a 23andMe genetic testing kit as a gift. In this post, I’ll take a look at my results through the lens of prospect theory, which aims to quantify our perception of risk. 23andMe results estimate your lifetime likelihood of various medical conditions, making them a great dataset for testing out these concepts in behavioral economics.

## Prospect Theory: A Quick Example

Suppose I offer you a bet. I flip a fair coin once. Heads, you gain \$1000; tails, you lose \$ 900. Do you take the bet? Probability dictates that you should, since you would expect to come out $(1000 - 900) / 2 = 50$ dollars ahead.

If you’re like most people, though, you have a powerful aversion to losing \$900. This aversion is powerful enough that you’ll decline the bet. This only makes sense if losing \$ 900 has a greater negative value than the positive value of gaining \$1000. In other words, your perception of value is non-linear. This perception underpins many real-world bets that make little sense from an expected utility standpoint: • Insurance: the buyer takes a small certain loss to avoid the risk of an uncertain and potentially large loss; • Security: the property owner takes a small certain loss to reduce the risk of an uncertain and potentially large loss; • Lotteries: the ticket holder takes a small certain loss to try for an extremely unlikely but massive gain. Prospect theory creates a mathematical framework for understanding our perceptions of value and risk. It’s an awesome paper with highly approachable mathematics. Definitely recommended reading for anyone interested in economics, game theory, and the like. ## My Disease Risk Among other things, 23andMe can estimate your lifetime risk of various diseases. This information is divided into three categories according to whether your risk is highly elevated, highly decreased, or roughly typical: Within each category, the diseases are ordered by decreasing confidence rating, then by decreasing absolute risk. The different confidence levels are as follows: 1. Preliminary Research: fewer than 100 people studied 2. Preliminary Research: fewer than 750 people studied 3. Preliminary Research: a single study with 750+ participants 4. Established Research: multiple studies with 750+ participants The 23andMe dashboard doesn’t show estimated risk for lower-confidence findings, but I can fetch that information through their API. Armed with my raw risk data, I can now play around with a pair of alternative disease rankings. As an anchoring point, here’s my disease ranking by absolute risk: The alternative ranking metric code can be found here. ### Relative-Risk Sorting An obvious ranking metric is relative risk, which is already provided in the disease listing. ### Perceived Relative-Risk Sorting In prospect theory, a probability$ p $has perceived weight$ w(p) $. In Advances in Prospect Theory: Cumulative Representation of Uncertainty, Tversky and Kahneman fit$ w $to observed results for subjects evaluating bets similar to the one listed above: The corresponding equations are for positive prospects and for negative prospects. If my risk is$ p_0 $and the general risk is$ p $, I can define my perceived relative risk as if$ p_0 < p \$ (I consider decreased risk as a positive prospect!) and

otherwise. The resulting rankings are pretty close to ordinary relative risk, but not identical:

The difference is due to distortion of small probabilities.

## Further Ideas

So far, I’ve only used the probability weighting functions from prospect theory. I could also assign values to each disease:

• Financial: how much does this disease cost to treat on average?
• Reduced Life Expectancy: how many years of potential life do these diseases claim on average?
• Lifestyle: how much impact would this have on my highly active lifestyle? On others who might have to support me?

This is definitely morbid, but it’s also potentially worthwhile. After all, insurance companies make very detailed estimates of our risk. They will certainly incorporate genetic data into their models as soon as it’s legal to do so. If we want to make more informed decisions in situations involving risk, from medical insurance to lottery tickets, it helps to understand how we value different outcomes.

## Appendix: How To Use The 23andMe API

The 23andMe API is actually quite easy to use, and their documentation is excellent. Nevertheless, I’ll list the flow I went through to get my genetic disease risk data. If you’re unfamiliar with OAuth 2.0, the Google API docs include this primer complete with cute stick figure diagrams.

First, I log into the API console and create an app. This gives me the client_id and client_secret parameters I need to initiate the flow. Looking at the 23andMe API reference, I see that I need permissions under the analyses scope, so I request that by visiting

in the browser. I’m redirected to

which gives me the auth code I need to request a token:

Finally, I use my shiny new access_token to get my genetic risk data: