Genetic Prediction and Adverse Selection
In 1994 I published Genetic Testing: An Economic and Contractarian Analysis which discussed how genetic testing could undermine insurance markets. I also proposed a solution, genetic insurance, which would in essence insure people for changes in their health and life insurance premiums due to the revelation of genetic data. Later John Cochrane would independently create Time Consistent Health Insurance a generalized form of the same idea that would allow people to have long term health insurance without being tied to a single firm.
The Human Genome Project completed in 2003 but, somewhat surprisingly, insurance markets didn’t break down, even though genetic information became more common. We know from twin studies that genetic heritability is very large but it turned out that the effect from each gene variant is very small. Thus, only a few diseases can be predicted well using single-gene mutations. Since each SNP has only a small effect on disease, to predict how genes influence disease we would need data on hundreds of thousands, even millions of people, and millions of their SNPs across the genome and their diseases. Until recently, that has been cost-prohibitive and as a result the available genetic information lacked much predictive power.
In an impressive new paper, however, Azevedo, Beauchamp and Linnér (ABL) show that data from Genome-Wide Association Studies can be used to create polygenic risk indexes (PGIs) which can predict individual disease risk from the aggregate effects of many genetic variants. The data is prodigious:
We analyze data from the UK Biobank (UKB) (Bycroft et al., 2018; Sudlow et al., 2015). The UKB contains genotypic and rich health-related data for over 500,000 individuals from across the United Kingdom who were between 40 and 69 years old at recruitment (between 2006 and 2010). UKB data is linked to the UK’s National Health Service (NHS), which maintains detailed records of health events across the lifespan and with which 98% of the UK population is registered (Sudlow et al., 2015). In addition, all UKB participants took part in a baseline assessment, in which they provided rich environmental, family history, health, lifestyle, physical, and sociodemographic data, as well as blood, saliva, and urine samples.
The UKB contains genome-wide array data for ∼800,000 genetic variants for ∼488,000 participants.
So for each of these individuals ABL construct risk indexes and they ask how significant is this new information for buying insurance in the Critical Illness Insurance market:
Critical illness insurance (CII) pays out a lump sum in the event that the insured person gets diagnosed with any of the medical conditions listed on the policy (Brackenridge et al., 2006). The lump sum can be used as the policyholder wishes. The policy pays out once and is thereafter terminated.
… Major CII markets include Canada, the United Kingdom, Japan, Australia, India, China, and Germany. It is estimated that 20% of British workers were covered by a CII policy in 2009 (Gatzert and Maegebier, 2015). The global CII market has been valued at over $100 billion in 2021 and was projected to grow to over $350 billion by 2031 (Allied Market Research, 2022).
The answer, as you might have guessed by now, is very significant. Even though current PGIs explain only a fraction of total genetic risk, they are already predictive enough so that it would make sense for individuals with high measured risk to purchase insurance, while those with low-risk would opt out—leading to adverse selection that threatens the financial sustainability of the insurance market.
Today, the 500,000 people in the UK’s Biobank don’t know their PGIs but in principle they could and in the future they will. Indeed, as GWAS sample sizes increase, PGI betas will become more accurate and they will be applied to a greater fraction of an individual’s genome so individual PGIs will become increasingly predictive, exacerbating selection problems in insurance markets.
If my paper was a distant early warning, Azevedo, Beauchamp, and Linnér provide an early—and urgent—warning. Without reform, insurance markets risk unraveling. The authors explore potential solutions, including genetic insurance, community rating, subsidies, and risk adjustment. However, the effectiveness of these measures remains uncertain, and knee-jerk policies, such as banning insurers from using genetic information, could lead to the collapse of insurance altogether.