2. Insurance, actuarial fairness and wider considerations regarding genomics

Insurance protects against losses associated with unpredictable events. While an event may be probable (such as some form of prolonged ill health) or certain (death), its timing and consequences are likely to be unpredictable. Faced with uncertainty about the timing, scope and extent of these events, individuals derive value from pooling risks with others in the population. Within a risk pool, the majority of individuals who do not make claims contribute to meet the cost of the minority who do make claims. Our discussion of insurance in this paper refers throughout to risk-rated insurance, as opposed to, for example, community-rated insurance which is a method of determining insurance premiums based on the overall risk profile of a community or group rather than on individual characteristics.

The premium and terms of risk-rated insurance contracts reflect the risk that a prospective customer may experience an event that gives rise to a claim, as well as the costs borne by the insurer in providing cover. Higher assessed risk generally results in higher premiums to be paid by customers, and/or more restrictive contract terms. The converse will generally be true for lower assessed risk. From an actuarial perspective, a fair insurance contract is one that accurately prices risk. Systematic mispricing of insurance by a single provider, in the sense of overcharging or undercharging certain groups given the risks and therefore the costs associated with each group, will result in a competitive disadvantage and will not be sustainable.

If genetic factors, broadly defined, influence the risk of insurable events, then their use in insurance underwriting will contribute to actuarial fairness in the pricing of risk, and the efficient operation of the insurance market as a whole. Reliance on actuarial fairness ‘expresses the moral judgment that fair underwriting practices must reflect the division of people according to actuarially accurate determination of their risks’ (Daniels, Reference Daniels and Rothstein2004). Wider considerations in relation to the use of genetics in insurance beyond actuarial concerns may involve access to insurance by different groups, the cost and quality of insurance, and privacy issues. These wider considerations motivate the existence of the Code, as well as other international examples that treat genetic information differently to other rating factors used in insurance underwriting.

These international examples include the Genetic Information Nondiscrimination Act (GINA) in the United States (Bélisle-Pipon et al., Reference Bélisle-Pipon, Vayena, Green and Glenn Cohen2019), which prohibits the use of genetic information in determining the offer of health insurance, but not necessarily other forms of insurance including life insurance. The Genetic Non-Discrimination Act (Bombard and Heim-Myers, Reference Bombard and Heim-Myers2018; Bélisle-Pipon et al., Reference Bélisle-Pipon, Vayena, Green and Glenn Cohen2019; Supreme Court of Canada, 2020) in Canada prohibits requesting disclosure of the results of genetic tests or being forced to take such tests in order to obtain access to goods and services including insurance. The Australian life insurance industry introduced a partial, self-regulated ban on the use of genetic results in 2019, and debates continue on whether this moratorium is fit for purpose (Tiller et al., Reference Tiller, Bakshi, Dowling, Keogh, McInerney-Leo, Barlow-Stewart, Boughtwood, Gleeson, Delatycki and Winship2024; Tiller and Lacaze, Reference Tiller and Lacaze2023).

As we assess emerging challenges posed by genomics to insurance, and specifically in the regulatory context of the United Kingdom, we consider challenges both to the process of underwriting (‘how much risk is attributable to a particular person?’), to these wider considerations that may give rise to departures from actuarial fairness in offers of insurance, and finally to the operation of the insurance market itself. We examine these challenges under three primary themes, as follows.

The first relates to the fact that, as genetic testing routinely encompasses ever greater portions of the genome, the distinction in the Code between diagnostic and predictive genetic testing becomes blurred. A ‘typical’ person has around 100,000 rare variants in their genome (Auton et al., Reference Auton, Abecasis, Altshuler, Durbin, Abecasis, Bentley, Chakravarti, Clark, Donnelly, Eichler, Flicek, Gabriel, Gibbs, Green, Hurles, Knoppers, Korbel, Lander, Lee, Lehrach, Mardis, Marth, McVean, Nickerson, Schmidt, Sherry, Wang, Wilson, Gibbs, Boerwinkle, Doddapaneni, Han, Korchina, Kovar, Lee, Muzny, Reid, Zhu, Wang, Chang, Feng, Fang, Guo, Jian, Jiang, Jin, Lan, Li, Li, Li, Liu, Liu, Lu, Ma, Tang, Wang, Wang, Wu, Wu, Xu, Yin, Zhang, Zhang, Zhao, Zhao, Zheng, Lander, Altshuler, Gabriel, Gupta, Gharani, Toji, Gerry, Resch, Flicek, Barker, Clarke, Gil, Hunt, Kelman, Kulesha, Leinonen, McLaren, Radhakrishnan, Roa, Smirnov, Smith, Streeter, Thormann, Toneva, Vaughan, Zheng-Bradley, Bentley, Grocock, Humphray, James, Kingsbury, Lehrach, Sudbrak, Albrecht, Amstislavskiy, Borodina, Lienhard, Mertes, Sultan, Timmermann, Yaspo, Mardis, Wilson, Fulton, Fulton, Sherry, Ananiev, Belaia, Beloslyudtsev, Bouk, Chen, Church, Cohen, Cook, Garner, Hefferon, Kimelman, Liu, Lopez, Meric, O'Sullivan, Ostapchuk, Phan, Ponomarov, Schneider, Shekhtman, Sirotkin, Slotta, Zhang, McVean, Durbin, Balasubramaniam, Burton, Danecek, Keane, Kolb-Kokocinski, McCarthy, Stalker, Quail, Schmidt, Davies, Gollub, Webster, Wong, Zhan, Auton, Campbell, Kong, Marcketta, Gibbs, Yu, Antunes, Bainbridge, Muzny, Sabo, Huang, Wang, Coin, Fang, Guo, Jin, Li, Li, Li, Li, Lin, Liu, Luo, Shao, Xie, Ye, Yu, Zhang, Zheng, Zhu, Alkan, Dal, Kahveci, Marth, Garrison, Kural, Lee, Leong, Stromberg, Ward, Wu, Zhang, Daly, DePristo, Handsaker, Altshuler, Banks, Bhatia, del Angel, Gabriel, Genovese, Gupta, Li, Kashin, Lander, McCarroll, Nemesh, Poplin, Yoon, Lihm, Makarov, Clark, Gottipati, Keinan, Rodriguez-Flores, Korbel, Rausch, Fritz, Stütz, Flicek, Beal, Clarke, Datta, Herrero, McLaren, Ritchie, Smith, Zerbino, Zheng-Bradley, Sabeti, Shlyakhter, Schaffner, Vitti, Cooper, Ball, Stenson, Bentley, Barnes, Bauer, Keira Cheetham, Cox, Eberle, Humphray, Kahn, Murray, Peden, Shaw, Kenny, Batzer, Konkel, Walker, MacArthur, Lek, Sudbrak, Amstislavskiy, Herwig, Mardis, Ding, Koboldt, Larson, Ye and Gravel2015) – some of these may help diagnose a condition already known about, others may predict disease (with varying degrees of accuracy) and yet others be entirely uncertain as to what their effects may be. This blurs the distinction alluded to in the Code. This blurring may also increase the challenges of underwriting (given uncertainty associated with the interpretation of results), with wider considerations relating to the equitable processing of this information, and with the wider operation of the market (given the resources necessary to process increasing volumes of genetic data).

The second area relates to pharmacogenetics, which is the study of how genetic variation can affect an individual's response to medicines. However, predictions of response at an individual level are imperfect, the medicine in question may never be required, and identification of the risk of serious adverse events could increase the cost of future healthcare and therefore potentially the costs of insurance (if no effective alternative treatments are available) or reduce these costs (if treatment is more effective when informed by pharmacogenetics). Again, pharmacogenetics seems likely to blur the diagnostic/predictive distinction at the centre of the Code, and will likely have implications for underwriting and for equitable access to insurance.

The third area relates to predicting costs, mortality and related phenotypes from genotype, especially in the context of increasing availability and affordability of composite indices of disease liability such as ‘polygenic scores’ that measure a component of risk for common disease. These tests may be neither clearly diagnostic nor highly predictive, and the contribution of a polygenic score to absolute risk may be very small. Nonetheless, polygenic scores could add some value to underwriting beyond conventional factors, which, absent other considerations, would result in better pricing of risk. However, it is these other considerations that merit a wider debate on appropriate uses of this type of information. There may also be market-wide impacts under differential access to this information.

We explore each of these topics in more detail in the following sections.

3. The blurring of the distinction between diagnostic and predictive tests

The cost of genetic testing continues to fall, and the volume of data that such testing produces continues to increase (Horton and Lucassen, Reference Horton and Lucassen2019). Whilst in the past only certain variants were analysed based on a clinical suspicion of their presence, ‘genome first’ approaches facilitated by technological advances identify many more variants that may have implications for the individual concerned, and interpreting their significance can be very challenging. To place the scale of variation in context, there are on average 4–5 million differences between the reference human genome and any typical human genome (Auton et al., Reference Auton, Abecasis, Altshuler, Durbin, Abecasis, Bentley, Chakravarti, Clark, Donnelly, Eichler, Flicek, Gabriel, Gibbs, Green, Hurles, Knoppers, Korbel, Lander, Lee, Lehrach, Mardis, Marth, McVean, Nickerson, Schmidt, Sherry, Wang, Wilson, Gibbs, Boerwinkle, Doddapaneni, Han, Korchina, Kovar, Lee, Muzny, Reid, Zhu, Wang, Chang, Feng, Fang, Guo, Jian, Jiang, Jin, Lan, Li, Li, Li, Liu, Liu, Lu, Ma, Tang, Wang, Wang, Wu, Wu, Xu, Yin, Zhang, Zhang, Zhao, Zhao, Zheng, Lander, Altshuler, Gabriel, Gupta, Gharani, Toji, Gerry, Resch, Flicek, Barker, Clarke, Gil, Hunt, Kelman, Kulesha, Leinonen, McLaren, Radhakrishnan, Roa, Smirnov, Smith, Streeter, Thormann, Toneva, Vaughan, Zheng-Bradley, Bentley, Grocock, Humphray, James, Kingsbury, Lehrach, Sudbrak, Albrecht, Amstislavskiy, Borodina, Lienhard, Mertes, Sultan, Timmermann, Yaspo, Mardis, Wilson, Fulton, Fulton, Sherry, Ananiev, Belaia, Beloslyudtsev, Bouk, Chen, Church, Cohen, Cook, Garner, Hefferon, Kimelman, Liu, Lopez, Meric, O'Sullivan, Ostapchuk, Phan, Ponomarov, Schneider, Shekhtman, Sirotkin, Slotta, Zhang, McVean, Durbin, Balasubramaniam, Burton, Danecek, Keane, Kolb-Kokocinski, McCarthy, Stalker, Quail, Schmidt, Davies, Gollub, Webster, Wong, Zhan, Auton, Campbell, Kong, Marcketta, Gibbs, Yu, Antunes, Bainbridge, Muzny, Sabo, Huang, Wang, Coin, Fang, Guo, Jin, Li, Li, Li, Li, Lin, Liu, Luo, Shao, Xie, Ye, Yu, Zhang, Zheng, Zhu, Alkan, Dal, Kahveci, Marth, Garrison, Kural, Lee, Leong, Stromberg, Ward, Wu, Zhang, Daly, DePristo, Handsaker, Altshuler, Banks, Bhatia, del Angel, Gabriel, Genovese, Gupta, Li, Kashin, Lander, McCarroll, Nemesh, Poplin, Yoon, Lihm, Makarov, Clark, Gottipati, Keinan, Rodriguez-Flores, Korbel, Rausch, Fritz, Stütz, Flicek, Beal, Clarke, Datta, Herrero, McLaren, Ritchie, Smith, Zerbino, Zheng-Bradley, Sabeti, Shlyakhter, Schaffner, Vitti, Cooper, Ball, Stenson, Bentley, Barnes, Bauer, Keira Cheetham, Cox, Eberle, Humphray, Kahn, Murray, Peden, Shaw, Kenny, Batzer, Konkel, Walker, MacArthur, Lek, Sudbrak, Amstislavskiy, Herwig, Mardis, Ding, Koboldt, Larson, Ye and Gravel2015) – and many of these differences will have minor or unknown medical impact.

Indeed, large numbers of variants that have historically been considered to be pathogenic (i.e. associated with specific health outcomes) have in fact turned out to be common in individuals who do not show the associated phenotype, suggesting that either their original classification was wrong, or that their impact on health is more subtle or context-dependent than previously appreciated.

Beaumont and Wright (Reference Beaumont and Wright2022) illustrated the challenge of interpreting people's genomic data, showing that while large gene panels may maximise diagnostic yield, they are also likely to identify several variants that look hypothetically concerning though are probably benign; most people have at least one rare variant in the coding regions of the genome in panels containing over five hundred disease genes. Even for ‘well-understood’ pathogenic genetic variants, context matters: Jackson et al. (Reference Jackson, Weedon, Harrison, Wood, Ruth, Tyrrell and Wright2022) found that people with cancer-predisposing genetic variants were at significantly less elevated risk of cancer in the absence of a family history. These issues are likely to be amplified by initiatives to undertake whole genome sequencing of all newborn children within a population.

The predictive value of a specific variant identified via genetic testing in the absence of phenotype and positive family history may therefore be low (Horton et al., Reference Horton, Crawford, Freeman, Fenwick, Wright and Lucassen2019; Horton and Lucassen, Reference Horton and Lucassen2019; Horton and Lucassen, Reference Horton and Lucassen2022). As genetic tests become broader, a distinction has emerged between using genetic results for diagnoses (in tandem with other clinical information) and the use of genetic data for other purposes (such as prediction of disease risk) outside of a clear familial or phenotypic context. For example, in the former case, a high degree of confidence might be expected in reaching an overall diagnostic assessment for a particular individual. In the latter, inferences about disease risk are likely to be less meaningful at the level of the individual.

This changing distinction between diagnostic and predictive results may also be influenced in some contexts by prognostic information. For example, in some cases an underlying genetic cause for a clinical diagnosis may change the prognosis associated with a particular condition. For example, knowledge that an individual has congenital long QT syndrome (which is associated with irregular or abnormal heart beats) may change the prognosis associated with the risk of future cardiac arrests (Arthur et al., Reference Arthur, Ahmad and Postema Pieter2022).

There is also increasing interest in (Sud et al., Reference Sud, Horton, Hingorani, Tzoulaki, Turnbull, Houlston and Lucassen2023) aggregated summaries (typically referred to as polygenic risk scores) analysing many points of variation in the genome to estimate liability to disease incidence, disease progression or related outcomes. Does knowledge that an individual has high polygenic risk for a particular condition constitute a genetic ‘result’? Without additional information, knowledge that an individual is in the top decile or even the top percentile of a polygenic distribution for incident disease may indicate only a marginal increase in lifetime risk, and furthermore could miss most cases that occur in people in other centiles.

For example, women in the top 5 per cent of polygenic risk for ovarian cancer have a lifetime risk of 2.1 per cent for developing this condition, compared to a population average risk of 1.6 per cent (Sud et al., Reference Sud, Turnbull and Houlston2021). The overall distribution of the polygenic risk at the population level may be informative for disease aetiology, even if knowledge of an individual's polygenic risk in itself does not contribute much, if anything, to knowledge of ‘which particular individuals will succumb’ (Davey Smith, Reference Smith2011) to the condition of interest. Below, we explore other potential consequences of polygenic scores for insurance in the context of increasing knowledge of the association between genotype and healthcare costs.

These discussions illustrate that the distinction between predictive and diagnostic codes in the Code, based primarily on the presence or absence of symptoms, may be increasingly difficult to defend since this binary distinction may not be especially relevant to the insurance decision. Of note in this regard is that the UK Government opened a consultation in the second half of 2023 on whether (amongst other issues for which evidence was requested) the Code should widen the considerations used to characterise whether a predictive test should be disclosed. The consultation was structured around four questions: (1) How useful is the genetic test for characterising the risk of developing a condition? (2) How many people take the test? (3) What is the impact of the condition in terms of the length and quality of life of people who develop it? (4) What is the potential for reducing the risk of developing the condition and managing its effects if it develops?

We note that these questions highlight a recognition that topics such as penetrance need consideration (since the ‘usefulness’ of a genetic test in characterising disease risk is a function of penetrance) and that this moves away from the symptomatic/asymptomatic distinction that is central to the current Code. For example, highly penetrant risk-increasing BRCA variants are predictive of future risk and may have material consequences for future healthcare costs (of particular interest to a health insurer) and for mortality (of particular interest to a life insurer).

4. Pharmacogenetics

The complexities in defining which results merit intervention, and when, also arise in the context of pharmacogenetics. Response to medicines varies between individuals, in part, because of genetic variation. If a pharmacogenetic test determines that an individual is less likely to respond to a certain medication, it may be recommended to select an alternative treatment (if available). However, this alternative medicine may, on average in the population, be less effective than the ‘first-line’ therapy. Although this individual is receiving the most suitable treatment for them, it remains an inferior therapeutic strategy compared to what an individual without the pharmacogenetic variant would receive. Alongside the health consequences for the individual concerned, this has implications for potential future treatment costs. Pharmacogenetic testing may predict treatment options to some extent and confer an increased or decreased chance of response to therapy, but such findings are not usually thought of as diagnostic, again highlighting that the diagnostic/predictive classification may be less useful in this setting.

For example, CYP2C19 is an important drug metabolising enzyme (Gaedigk et al., Reference Gaedigk, Sangkuhl, Whirl-Carrillo, Klein and Steven Leeder2017). Genetic variation in the CYP2C19 gene is associated with diminished tolerance, treatment failure, and adverse reactions for many medicines (Botton et al., Reference Botton, Whirl-Carrillo, Del Tredici, Sangkuhl, Cavallari, Agúndez, Duconge, Michael Lee, Woodahl, Claudio-Campos, Daly, Klein, Pratt, Scnott and Gaedigk2021). For instance, CYP2C19 catalyses the activation of clopidogrel, a widely prescribed anti-platelet drug. Individuals with two CYP2C19 loss-of-function alleles (‘poor metabolisers’) will respond less well to clopidogrel compared to the rest of the population (Scott et al., Reference Scott, Sangkuhl, Stein, Hulot, Mega, Roden, Klein, Sabatine, Johnson and Shuldiner2013).

A genetic result showing that an individual carries a loss of function variant in CYP2C19 means that their CYP2C19 enzyme will have reduced activity. As such, CYP2C19 genotyping could be considered diagnostic in nature. However, the negative clinical impact of being a CYP2C19 poor metaboliser is only experienced in certain contexts, such as when the individual is prescribed clopidogrel. In that regard, one could consider the test to be more akin to a predictive test. This blurring between the diagnostic and predictive creates challenges when attempting to consider the insurance implications of a given pharmacogenetic test result.

5. Predicting costs, mortality and related phenotypes from genotype

Genetic rating factors that influence the propensity to incur healthcare costs and that influence mortality will be relevant to the actuarial pricing of health and life insurance products. We consider whether these kinds of rating factors might be feasible, and if so the types of wider considerations that merit scrutiny around their use.

Recent evidence has quantified the heritability of future healthcare costs. Heritability refers the proportion of variance in a phenotype that is attributable to genetic variance in a given population. Lakhani et al. (Reference Lakhani, Tierney, Manrai, Yang, Visscher and Patel2019) used an American health insurance dataset to examine the heritability of monthly healthcare cost amongst 56,396 twin pairs. They estimated that the heritability of average monthly cost was 0.29, meaning that 29 per cent of the variance in average monthly healthcare cost between individuals in that study population was attributable to genetic factors. de Zeeuw et al. (Reference de Zeeuw, Voort, Schoonhoven, Nivard, Emery, Hottenga, Willemsen, Dykstra, Zarrabi, Kartopawiro and Boomsma2021) studied 16,726 participants in the Netherlands Twin Register and estimated similar heritabilities between 0.29 and 0.38. Although debates continue about the extent to which twin studies might over-estimate heritability (Young, Reference Young2019), these estimates for healthcare cost heritability suggest measures of genetic liability to incur healthcare costs are potentially relevant to insurance underwriting.

Mendelian randomisation analyses – the use of common genetic variation indicating liability to particular phenotypes in causal instrumental variable analyses – has demonstrated that genotypes associated with a variety of diseases (Dixon et al., Reference Dixon, Harrison, Hollingworth, Davies and Smith2022a), traits (Dixon et al., Reference Dixon, Hollingworth, Harrison, Davies and Smith2020; Hazewinkel et al., Reference Hazewinkel, Richmond, Wade and Dixon2022; Lee et al., Reference Lee, Jukarainen, Dixon, Davies, Smith, Natarajan and Ganna2022) and behaviours (Dixon et al., Reference Dixon, Sallis, Munafo, Smith and Howe2022b) are also associated with healthcare costs and with closely related outcomes such as rates of inpatient hospital admission (Hazewinkel et al., Reference Hazewinkel, Richmond, Wade and Dixon2022). The use of cost phenotypes will be more consequential in health systems that rely on private healthcare insurance, and are unlikely to have a significant impact in tax-payer funded, universal and free-at-delivery health system such as the NHS in the UK.

However, these types of consideration may be important for life insurance in many countries. In relation to mortality risk, Karlsson Linnér and Koellinger (Reference Karlsson Linnér and Koellinger2022) found that a polygenic score could detect a substantially shorter median lifespan in the top decile of total genetic liability independent of other factors used in conventional insurance underwriting. There are emerging examples (such as Insurance Newsnet, 2023) of life insurers using polygenic risk scores for health conditions (rather than mortality risk per se) as a means of encouraging behavioural change such as improved adherence to medication in light of personal knowledge of a polygenic risk.

6. Towards deliberative processes

The foregoing considered issues that were specific to the regulatory regime of the UK (specifically the Code), as well as more general issues that may arise both in the UK and elsewhere. A fundamental issue underlying both the specific and general issues to which these emerging developments in genomics give rise for insurance relates to whether and how genetic information merits distinct treatment as a risk factor for risk-rated insurance. The complexities include the uneven way in which genetic information may be revealed (and potentially disclosed) across types of individual, diseases, risk levels and at points in the life course.

A central task that for any new regulatory response that involves wider considerations than the actuarial pricing of genetic risk will therefore be to identify individuals and groups on whom these costs fall and on whom they ought to fall. We do not take a position on which normative perspectives are necessarily appropriate, but instead outline the parameters in which such a debate may be held.

For example, normative considerations might suggest people at risk of some types of condition ought not to lose access to insurance or to face higher premiums because of their genetic risk. A per se rejection of risk-based pricing for these individuals shifts their costs onto those with lower (genetic) risk, with the effect that a cross-subsidy is created from lower to higher-risk individuals. This cross-subsidy may be very modest in scale, although the scale of this impact remains to be determined and overall consequence of rejecting risk-based pricing is unlikely to ‘net off’ to zero costs. For example, depending on the nature of the risks and proportions of people in each risk category, it is possible that the reduction in costs for high risks is not as great as the increase in costs for the low risks. In this scenario, the aggregate costs of insurance become higher for society albeit the extent of increase remains to be established. Moreover, any potential gains from actuarial pricing of genetic risk (such as improved medication adherence incentivised by life insurance policies that request information on polygenic risk) would not be realised.

The risks faced by individuals are not abolished simply because some groups do not face actuarially fair risk-based pricing for their insurance products; instead, there will necessarily be impacts on the price and availability of insurance. There may be second-order effects on the dynamics of competition amongst insurance. For example, adverse selection is still possible even if those with the highest genetic risk do not face premiums based on their actuarial risk. Higher prices amongst the lower risk groups may reduce their use of insurance, resulting in future avoidable health impairments and economic detriment. On the other hand, appropriate regulatory protections could prevent individuals being penalised from undertaking genetic testing where indicated (Filipova-Neumann and Hoy, Reference Filipova-Neumann and Hoy2014).

The analytic challenges therefore involve characterising and modelling the trade-offs involved (Ossa and Towse, Reference Ossa and Towse2004; Wilson, Reference Wilson2006). These trade-offs include the quantitative impact of genetic information on the terms and price of insurance, increasing knowledge of genetic contributions to disease and mortality, the predictive capacity of polygenic and other genetic risk scores, the demand for insurance and the behavioural responses of consumers. If genetic data cannot be used in risk based-pricing, then variables correlated with these data could potentially be used (if available) and this may frustrate to some degree restrictions on the use of genetic data (Pope and Sydnor, Reference Pope and Sydnor2011; Aseervatham et al., Reference Aseervatham, Lex and Spindler2016). The feasibility of finding robust proxies for genetic data may be more limited than for more traditional risk factors such as gender (Oxera, 2010) but this is likely to vary by context.

The normative issues require, we suggest, a deliberative approach to identify what deviations from actuarially fair pricing are appropriate. The final product of a deliberative process is ‘guidance shaped by judgements’ (Culyer, Reference Culyer2006). What judgements matter or ought to matter in this context? A fundamental consideration is to establish which conditions and which genetic risks (and their consequences) society assesses should be borne by the individual and which should be shared more widely.

A starting point is perhaps to recognise the value of insurance, and the need to ensure its sustainable provision. An overarching approach could be to minimise the overall cost of insurance (potentially subject to specific exemptions) and the costs of which are either managed through cross-subsidy imposed on lower risk groups or met through some other mechanism. An alternative starting point could be to ensure that those facing the greatest possible health burden (however defined) from genetic conditions retain access on reasonable terms.

A multitude of other models could be proposed, each of which will embody their own trade-offs and give rise to different cost profiles and health outcomes. This will also give rise to a host of ethical issues. Should the treatment of genetic information be different from other fundamental personal characteristics that influence insurance premiums, such as sex and age? Is discrimination by insurance companies on the basis of genotype normatively the same as discrimination according to sex, the use of which as a rating factor the European Union prohibited in 2012 (Commission, 2012)? Can and should there be separate ethical frameworks for different types of insurance?